diff --git a/data/fairseq/fairseq/criterions/__init__.py b/data/fairseq/fairseq/criterions/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..ecd65d34adec4222ac8781106560ebc5dc2622f5 --- /dev/null +++ b/data/fairseq/fairseq/criterions/__init__.py @@ -0,0 +1,36 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. +"""isort:skip_file""" + +import importlib +import os + +from fairseq import registry +from fairseq.criterions.fairseq_criterion import ( # noqa + FairseqCriterion, + LegacyFairseqCriterion, +) +from omegaconf import DictConfig + + +( + build_criterion_, + register_criterion, + CRITERION_REGISTRY, + CRITERION_DATACLASS_REGISTRY, +) = registry.setup_registry( + "--criterion", base_class=FairseqCriterion, default="cross_entropy" +) + + +def build_criterion(cfg: DictConfig, task, from_checkpoint=False): + return build_criterion_(cfg, task, from_checkpoint=from_checkpoint) + + +# automatically import any Python files in the criterions/ directory +for file in sorted(os.listdir(os.path.dirname(__file__))): + if file.endswith(".py") and not file.startswith("_"): + file_name = file[: file.find(".py")] + importlib.import_module("fairseq.criterions." + file_name) diff --git a/data/fairseq/fairseq/criterions/__pycache__/__init__.cpython-38.pyc b/data/fairseq/fairseq/criterions/__pycache__/__init__.cpython-38.pyc new file mode 100644 index 0000000000000000000000000000000000000000..f8eab2a8068d2e223b15722d205ede56ba107460 Binary files /dev/null and b/data/fairseq/fairseq/criterions/__pycache__/__init__.cpython-38.pyc differ diff --git a/data/fairseq/fairseq/criterions/__pycache__/adaptive_loss.cpython-38.pyc b/data/fairseq/fairseq/criterions/__pycache__/adaptive_loss.cpython-38.pyc new file mode 100644 index 0000000000000000000000000000000000000000..821d7121da54b4c33f4fe0bc501c2bd19a3fbf3e Binary files /dev/null and b/data/fairseq/fairseq/criterions/__pycache__/adaptive_loss.cpython-38.pyc differ diff --git a/data/fairseq/fairseq/criterions/__pycache__/composite_loss.cpython-38.pyc b/data/fairseq/fairseq/criterions/__pycache__/composite_loss.cpython-38.pyc new file mode 100644 index 0000000000000000000000000000000000000000..53150593359b396fb26d5c4eb1cb5a57858f48ac Binary files /dev/null and b/data/fairseq/fairseq/criterions/__pycache__/composite_loss.cpython-38.pyc differ diff --git a/data/fairseq/fairseq/criterions/__pycache__/cross_entropy.cpython-38.pyc b/data/fairseq/fairseq/criterions/__pycache__/cross_entropy.cpython-38.pyc new file mode 100644 index 0000000000000000000000000000000000000000..8196179df410addac829fe62f01d66db1743627f Binary files /dev/null and b/data/fairseq/fairseq/criterions/__pycache__/cross_entropy.cpython-38.pyc differ diff --git a/data/fairseq/fairseq/criterions/__pycache__/ctc.cpython-38.pyc b/data/fairseq/fairseq/criterions/__pycache__/ctc.cpython-38.pyc new file mode 100644 index 0000000000000000000000000000000000000000..678225c14e22c626017cb37b3c86e01d85f0e71c Binary files /dev/null and b/data/fairseq/fairseq/criterions/__pycache__/ctc.cpython-38.pyc differ diff --git a/data/fairseq/fairseq/criterions/__pycache__/fairseq_criterion.cpython-38.pyc b/data/fairseq/fairseq/criterions/__pycache__/fairseq_criterion.cpython-38.pyc new file mode 100644 index 0000000000000000000000000000000000000000..8254330ea3fd31c1d1d3913ca36e8acd52053184 Binary files /dev/null and b/data/fairseq/fairseq/criterions/__pycache__/fairseq_criterion.cpython-38.pyc differ diff --git a/data/fairseq/fairseq/criterions/__pycache__/fastspeech2_loss.cpython-38.pyc b/data/fairseq/fairseq/criterions/__pycache__/fastspeech2_loss.cpython-38.pyc new file mode 100644 index 0000000000000000000000000000000000000000..b3c11aa1810f672364fb307162c4233f519cb38e Binary files /dev/null and b/data/fairseq/fairseq/criterions/__pycache__/fastspeech2_loss.cpython-38.pyc differ diff --git a/data/fairseq/fairseq/criterions/__pycache__/hubert_criterion.cpython-38.pyc b/data/fairseq/fairseq/criterions/__pycache__/hubert_criterion.cpython-38.pyc new file mode 100644 index 0000000000000000000000000000000000000000..e9d3523507a358bcd5a0a6ec791f1a743efe525b Binary files /dev/null and b/data/fairseq/fairseq/criterions/__pycache__/hubert_criterion.cpython-38.pyc differ diff --git a/data/fairseq/fairseq/criterions/__pycache__/label_smoothed_cross_entropy.cpython-38.pyc b/data/fairseq/fairseq/criterions/__pycache__/label_smoothed_cross_entropy.cpython-38.pyc new file mode 100644 index 0000000000000000000000000000000000000000..ba369a983e2b4b6ccaa35c315da5c8a6c61f6a31 Binary files /dev/null and b/data/fairseq/fairseq/criterions/__pycache__/label_smoothed_cross_entropy.cpython-38.pyc differ diff --git a/data/fairseq/fairseq/criterions/__pycache__/label_smoothed_cross_entropy_latency_augmented.cpython-38.pyc b/data/fairseq/fairseq/criterions/__pycache__/label_smoothed_cross_entropy_latency_augmented.cpython-38.pyc new file mode 100644 index 0000000000000000000000000000000000000000..c8dc9660457f0d8c9227a377f33bffe06419fc0b Binary files /dev/null and b/data/fairseq/fairseq/criterions/__pycache__/label_smoothed_cross_entropy_latency_augmented.cpython-38.pyc differ diff --git a/data/fairseq/fairseq/criterions/__pycache__/label_smoothed_cross_entropy_with_alignment.cpython-38.pyc b/data/fairseq/fairseq/criterions/__pycache__/label_smoothed_cross_entropy_with_alignment.cpython-38.pyc new file mode 100644 index 0000000000000000000000000000000000000000..2d992862c59726a62b7942e3a62ed2227cb6b0b5 Binary files /dev/null and b/data/fairseq/fairseq/criterions/__pycache__/label_smoothed_cross_entropy_with_alignment.cpython-38.pyc differ diff --git a/data/fairseq/fairseq/criterions/__pycache__/label_smoothed_cross_entropy_with_ctc.cpython-38.pyc b/data/fairseq/fairseq/criterions/__pycache__/label_smoothed_cross_entropy_with_ctc.cpython-38.pyc new file mode 100644 index 0000000000000000000000000000000000000000..675187c9d50eab34e6a7ddefb42e50669bec1f4a Binary files /dev/null and b/data/fairseq/fairseq/criterions/__pycache__/label_smoothed_cross_entropy_with_ctc.cpython-38.pyc differ diff --git a/data/fairseq/fairseq/criterions/__pycache__/label_smoothed_cross_entropy_with_rdrop.cpython-38.pyc b/data/fairseq/fairseq/criterions/__pycache__/label_smoothed_cross_entropy_with_rdrop.cpython-38.pyc new file mode 100644 index 0000000000000000000000000000000000000000..1fec63db287c7f6dc65116a368fbcf2c495cfbe9 Binary files /dev/null and b/data/fairseq/fairseq/criterions/__pycache__/label_smoothed_cross_entropy_with_rdrop.cpython-38.pyc differ diff --git a/data/fairseq/fairseq/criterions/__pycache__/legacy_masked_lm.cpython-38.pyc b/data/fairseq/fairseq/criterions/__pycache__/legacy_masked_lm.cpython-38.pyc new file mode 100644 index 0000000000000000000000000000000000000000..10987e1e3c7c50ff2bfa84657ed25c6487b9cf89 Binary files /dev/null and b/data/fairseq/fairseq/criterions/__pycache__/legacy_masked_lm.cpython-38.pyc differ diff --git a/data/fairseq/fairseq/criterions/__pycache__/masked_lm.cpython-38.pyc b/data/fairseq/fairseq/criterions/__pycache__/masked_lm.cpython-38.pyc new file mode 100644 index 0000000000000000000000000000000000000000..1050f3660231467741909bc0951bccf917672576 Binary files /dev/null and b/data/fairseq/fairseq/criterions/__pycache__/masked_lm.cpython-38.pyc differ diff --git a/data/fairseq/fairseq/criterions/__pycache__/model_criterion.cpython-38.pyc b/data/fairseq/fairseq/criterions/__pycache__/model_criterion.cpython-38.pyc new file mode 100644 index 0000000000000000000000000000000000000000..3911c45956886a2b30268ca79db01e4768f3f246 Binary files /dev/null and b/data/fairseq/fairseq/criterions/__pycache__/model_criterion.cpython-38.pyc differ diff --git a/data/fairseq/fairseq/criterions/__pycache__/nat_loss.cpython-38.pyc b/data/fairseq/fairseq/criterions/__pycache__/nat_loss.cpython-38.pyc new file mode 100644 index 0000000000000000000000000000000000000000..13ce9055bb248d44d686e75825706b98c10f4111 Binary files /dev/null and b/data/fairseq/fairseq/criterions/__pycache__/nat_loss.cpython-38.pyc differ diff --git a/data/fairseq/fairseq/criterions/__pycache__/sentence_prediction.cpython-38.pyc b/data/fairseq/fairseq/criterions/__pycache__/sentence_prediction.cpython-38.pyc new file mode 100644 index 0000000000000000000000000000000000000000..e69b20808006c54c9c2ef8331413bf9e95f4f666 Binary files /dev/null and b/data/fairseq/fairseq/criterions/__pycache__/sentence_prediction.cpython-38.pyc differ diff --git a/data/fairseq/fairseq/criterions/__pycache__/sentence_prediction_adapters.cpython-38.pyc b/data/fairseq/fairseq/criterions/__pycache__/sentence_prediction_adapters.cpython-38.pyc new file mode 100644 index 0000000000000000000000000000000000000000..3c670c35617e7f3a9ede70ace285be2a075f1494 Binary files /dev/null and b/data/fairseq/fairseq/criterions/__pycache__/sentence_prediction_adapters.cpython-38.pyc differ diff --git a/data/fairseq/fairseq/criterions/__pycache__/sentence_ranking.cpython-38.pyc b/data/fairseq/fairseq/criterions/__pycache__/sentence_ranking.cpython-38.pyc new file mode 100644 index 0000000000000000000000000000000000000000..b47036ba19d56fe769975d0d1aebe535b8d1a12f Binary files /dev/null and b/data/fairseq/fairseq/criterions/__pycache__/sentence_ranking.cpython-38.pyc differ diff --git a/data/fairseq/fairseq/criterions/__pycache__/speech_dlm_criterion.cpython-38.pyc b/data/fairseq/fairseq/criterions/__pycache__/speech_dlm_criterion.cpython-38.pyc new file mode 100644 index 0000000000000000000000000000000000000000..704658ad8421882e2a9312177defd1ea1322291c Binary files /dev/null and b/data/fairseq/fairseq/criterions/__pycache__/speech_dlm_criterion.cpython-38.pyc differ diff --git a/data/fairseq/fairseq/criterions/__pycache__/speech_to_speech_criterion.cpython-38.pyc b/data/fairseq/fairseq/criterions/__pycache__/speech_to_speech_criterion.cpython-38.pyc new file mode 100644 index 0000000000000000000000000000000000000000..019f7d8c4bddeaf23805f04eb904db7945d0baf5 Binary files /dev/null and b/data/fairseq/fairseq/criterions/__pycache__/speech_to_speech_criterion.cpython-38.pyc differ diff --git a/data/fairseq/fairseq/criterions/__pycache__/speech_ulm_criterion.cpython-38.pyc b/data/fairseq/fairseq/criterions/__pycache__/speech_ulm_criterion.cpython-38.pyc new file mode 100644 index 0000000000000000000000000000000000000000..771e944552ccfad03ea950e7c30919bc0588f244 Binary files /dev/null and b/data/fairseq/fairseq/criterions/__pycache__/speech_ulm_criterion.cpython-38.pyc differ diff --git a/data/fairseq/fairseq/criterions/__pycache__/tacotron2_loss.cpython-38.pyc b/data/fairseq/fairseq/criterions/__pycache__/tacotron2_loss.cpython-38.pyc new file mode 100644 index 0000000000000000000000000000000000000000..dfe16f622a1864fa7652fafe1f29b9e127392518 Binary files /dev/null and b/data/fairseq/fairseq/criterions/__pycache__/tacotron2_loss.cpython-38.pyc differ diff --git a/data/fairseq/fairseq/criterions/__pycache__/wav2vec_criterion.cpython-38.pyc b/data/fairseq/fairseq/criterions/__pycache__/wav2vec_criterion.cpython-38.pyc new file mode 100644 index 0000000000000000000000000000000000000000..b4ff5f900b1511d717483a264f396e78db3fd37e Binary files /dev/null and b/data/fairseq/fairseq/criterions/__pycache__/wav2vec_criterion.cpython-38.pyc differ diff --git a/data/fairseq/fairseq/criterions/adaptive_loss.py b/data/fairseq/fairseq/criterions/adaptive_loss.py new file mode 100644 index 0000000000000000000000000000000000000000..fc1ac8540461ab402e6ba6b4b40afe363774fffc --- /dev/null +++ b/data/fairseq/fairseq/criterions/adaptive_loss.py @@ -0,0 +1,124 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import math +from dataclasses import dataclass + +import torch.nn.functional as F +from fairseq import utils +from fairseq.logging import metrics +from fairseq.criterions import FairseqCriterion, register_criterion +from fairseq.dataclass import FairseqDataclass +from fairseq.dataclass.constants import DDP_BACKEND_CHOICES +from omegaconf import II + + +@dataclass +class AdaptiveLossConfig(FairseqDataclass): + sentence_avg: bool = II("optimization.sentence_avg") + ddp_backend: DDP_BACKEND_CHOICES = II("distributed_training.ddp_backend") + + +@register_criterion("adaptive_loss", dataclass=AdaptiveLossConfig) +class AdaptiveLoss(FairseqCriterion): + """This is an implementation of the loss function accompanying the adaptive softmax approximation for + graphical processing units (GPU), described in the paper "Efficient softmax approximation for GPUs" + (http://arxiv.org/abs/1609.04309).""" + + def __init__(self, task, sentence_avg): + super().__init__(task) + self.sentence_avg = sentence_avg + + @classmethod + def build_criterion(cls, cfg: AdaptiveLossConfig, task): + if cfg.ddp_backend in {"c10d", "pytorch_ddp"}: + raise Exception( + "AdaptiveLoss is not compatible with the PyTorch " + "version of DistributedDataParallel. Please use " + "`--ddp-backend=legacy_ddp` instead." + ) + return cls(task, cfg.sentence_avg) + + def forward(self, model, sample, reduce=True): + """Compute the loss for the given sample. + + Returns a tuple with three elements: + 1) the loss + 2) the sample size, which is used as the denominator for the gradient + 3) logging outputs to display while training + """ + + assert ( + hasattr(model.decoder, "adaptive_softmax") + and model.decoder.adaptive_softmax is not None + ) + adaptive_softmax = model.decoder.adaptive_softmax + + net_output = model(**sample["net_input"]) + orig_target = model.get_targets(sample, net_output) + + nsentences = orig_target.size(0) + orig_target = orig_target.view(-1) + + bsz = orig_target.size(0) + + logits, target = adaptive_softmax(net_output[0], orig_target) + assert len(target) == len(logits) + + loss = net_output[0].new(1 if reduce else bsz).zero_() + + for i in range(len(target)): + if target[i] is not None: + assert target[i].min() >= 0 and target[i].max() <= logits[i].size(1) + loss += F.cross_entropy( + logits[i], + target[i], + ignore_index=self.padding_idx, + reduction="sum" if reduce else "none", + ) + + orig = utils.strip_pad(orig_target, self.padding_idx) + ntokens = orig.numel() + sample_size = sample["target"].size(0) if self.sentence_avg else ntokens + logging_output = { + "loss": loss.data, + "ntokens": ntokens, + "nsentences": nsentences, + "sample_size": sample_size, + } + return loss, sample_size, logging_output + + @staticmethod + def reduce_metrics(logging_outputs) -> None: + """Aggregate logging outputs from data parallel training.""" + loss_sum = utils.item(sum(log.get("loss", 0) for log in logging_outputs)) + ntokens = utils.item(sum(log.get("ntokens", 0) for log in logging_outputs)) + sample_size = utils.item( + sum(log.get("sample_size", 0) for log in logging_outputs) + ) + + metrics.log_scalar( + "loss", loss_sum / sample_size / math.log(2), sample_size, round=3 + ) + if sample_size != ntokens: + metrics.log_scalar( + "nll_loss", loss_sum / ntokens / math.log(2), ntokens, round=3 + ) + metrics.log_derived( + "ppl", lambda meters: utils.get_perplexity(meters["nll_loss"].avg) + ) + else: + metrics.log_derived( + "ppl", lambda meters: utils.get_perplexity(meters["loss"].avg) + ) + + @staticmethod + def logging_outputs_can_be_summed() -> bool: + """ + Whether the logging outputs returned by `forward` can be summed + across workers prior to calling `reduce_metrics`. Setting this + to True will improves distributed training speed. + """ + return True diff --git a/data/fairseq/fairseq/criterions/composite_loss.py b/data/fairseq/fairseq/criterions/composite_loss.py new file mode 100644 index 0000000000000000000000000000000000000000..98e835fa6e4c0bcad062df9c519701bf795c98be --- /dev/null +++ b/data/fairseq/fairseq/criterions/composite_loss.py @@ -0,0 +1,100 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from fairseq import utils +from fairseq.criterions import LegacyFairseqCriterion, register_criterion +from torch import nn + + +@register_criterion("composite_loss") +class CompositeLoss(LegacyFairseqCriterion): + """This is a composite loss that, given a list of model outputs and a list of targets, + computes an average of losses for each output-target pair""" + + def __init__(self, args, task): + super().__init__(args, task) + self.underlying_criterion = args.underlying_criterion + + @staticmethod + def add_args(parser): + """Add criterion-specific arguments to the parser.""" + # fmt: off + parser.add_argument('--underlying-criterion', type=str, metavar='VAL', required=True, + help='underlying criterion to use for the composite loss') + # fmt: on + + @staticmethod + def build_underlying_criterion(args, task): + saved_criterion = args.criterion + args.criterion = args.underlying_criterion + assert saved_criterion != args.underlying_criterion + underlying_criterion = task.build_criterion(args) + args.criterion = saved_criterion + return underlying_criterion + + @classmethod + def build_criterion(cls, args, task): + underlying_criterion = CompositeLoss.build_underlying_criterion(args, task) + + class FakeModel(nn.Module): + def __init__(self, model, net_out, target): + super().__init__() + self.model = model + self.net_out = net_out + self.target = target + + def forward(self, **unused): + return self.net_out + + def get_normalized_probs(self, net_output, log_probs, sample=None): + return self.model.get_normalized_probs( + net_output, log_probs, sample=sample + ) + + def get_targets(self, *unused): + return self.target + + @property + def decoder(self): + return self.model.decoder + + class _CompositeLoss(LegacyFairseqCriterion): + def __init__(self, args, task, underlying_criterion): + super().__init__(args, task) + self.underlying_criterion = underlying_criterion + + def forward(self, model, sample, reduce=True): + net_outputs = model(**sample["net_input"]) + targets = sample["target"] + + bsz = targets[0].size(0) + loss = net_outputs[0][0].new(1 if reduce else bsz).float().zero_() + + sample_size = 0 + logging_output = {} + for o, t in zip(net_outputs[0], targets): + m = FakeModel(model, (o, net_outputs[1]), t) + sample["target"] = t + l, ss, logging_output = self.underlying_criterion(m, sample, reduce) + loss += l + sample_size += ss + + loss.div_(len(targets)) + sample_size /= len(targets) + + logging_output["loss"] = utils.item(loss.data) if reduce else loss.data + return loss, sample_size, logging_output + + @staticmethod + def aggregate_logging_outputs(logging_outputs): + return underlying_criterion.__class__.aggregate_logging_outputs( + logging_outputs + ) + + @staticmethod + def reduce_metrics(logging_outputs) -> None: + underlying_criterion.__class__.reduce_metrics(logging_outputs) + + return _CompositeLoss(args, task, underlying_criterion) diff --git a/data/fairseq/fairseq/criterions/cross_entropy.py b/data/fairseq/fairseq/criterions/cross_entropy.py new file mode 100644 index 0000000000000000000000000000000000000000..24d6bcd6128f9bfb7667adcdb3e7f001cc57a523 --- /dev/null +++ b/data/fairseq/fairseq/criterions/cross_entropy.py @@ -0,0 +1,91 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import math +from dataclasses import dataclass + +import torch.nn.functional as F +from fairseq import utils +from fairseq.logging import metrics +from fairseq.criterions import FairseqCriterion, register_criterion +from fairseq.dataclass import FairseqDataclass +from omegaconf import II + + +@dataclass +class CrossEntropyCriterionConfig(FairseqDataclass): + sentence_avg: bool = II("optimization.sentence_avg") + + +@register_criterion("cross_entropy", dataclass=CrossEntropyCriterionConfig) +class CrossEntropyCriterion(FairseqCriterion): + def __init__(self, task, sentence_avg): + super().__init__(task) + self.sentence_avg = sentence_avg + + def forward(self, model, sample, reduce=True): + """Compute the loss for the given sample. + + Returns a tuple with three elements: + 1) the loss + 2) the sample size, which is used as the denominator for the gradient + 3) logging outputs to display while training + """ + net_output = model(**sample["net_input"]) + loss, _ = self.compute_loss(model, net_output, sample, reduce=reduce) + sample_size = ( + sample["target"].size(0) if self.sentence_avg else sample["ntokens"] + ) + logging_output = { + "loss": loss.data, + "ntokens": sample["ntokens"], + "nsentences": sample["target"].size(0), + "sample_size": sample_size, + } + return loss, sample_size, logging_output + + def compute_loss(self, model, net_output, sample, reduce=True): + lprobs = model.get_normalized_probs(net_output, log_probs=True) + lprobs = lprobs.view(-1, lprobs.size(-1)) + target = model.get_targets(sample, net_output).view(-1) + loss = F.nll_loss( + lprobs, + target, + ignore_index=self.padding_idx, + reduction="sum" if reduce else "none", + ) + return loss, loss + + @staticmethod + def reduce_metrics(logging_outputs) -> None: + """Aggregate logging outputs from data parallel training.""" + loss_sum = sum(log.get("loss", 0) for log in logging_outputs) + ntokens = sum(log.get("ntokens", 0) for log in logging_outputs) + sample_size = sum(log.get("sample_size", 0) for log in logging_outputs) + + # we divide by log(2) to convert the loss from base e to base 2 + metrics.log_scalar( + "loss", loss_sum / sample_size / math.log(2), sample_size, round=3 + ) + if sample_size != ntokens: + metrics.log_scalar( + "nll_loss", loss_sum / ntokens / math.log(2), ntokens, round=3 + ) + metrics.log_derived( + "ppl", lambda meters: utils.get_perplexity(meters["nll_loss"].avg) + ) + else: + metrics.log_derived( + "ppl", lambda meters: utils.get_perplexity(meters["loss"].avg) + ) + + @staticmethod + def logging_outputs_can_be_summed() -> bool: + """ + Whether the logging outputs returned by `forward` can be summed + across workers prior to calling `reduce_metrics`. Setting this + to True will improves distributed training speed. + """ + return True diff --git a/data/fairseq/fairseq/criterions/ctc.py b/data/fairseq/fairseq/criterions/ctc.py new file mode 100644 index 0000000000000000000000000000000000000000..368213cb2b05bcfcec3ae84aef68c82bd792492b --- /dev/null +++ b/data/fairseq/fairseq/criterions/ctc.py @@ -0,0 +1,325 @@ +# All rights reserved. +# +# This source code is licensed under the license found in the LICENSE file in +# the root directory of this source tree. An additional grant of patent rights +# can be found in the PATENTS file in the same directory. + +import math +from argparse import Namespace +from dataclasses import dataclass, field +from omegaconf import II +from typing import Optional + +import torch +import torch.nn.functional as F + +from fairseq import utils +from fairseq.logging import metrics +from fairseq.criterions import FairseqCriterion, register_criterion +from fairseq.dataclass import FairseqDataclass +from fairseq.data.data_utils import post_process +from fairseq.tasks import FairseqTask +from fairseq.logging.meters import safe_round + + +@dataclass +class CtcCriterionConfig(FairseqDataclass): + zero_infinity: bool = field( + default=False, + metadata={"help": "zero inf loss when source length <= target length"}, + ) + sentence_avg: bool = II("optimization.sentence_avg") + post_process: str = field( + default="letter", + metadata={ + "help": "how to post process predictions into words. can be letter, " + "wordpiece, BPE symbols, etc. " + "See fairseq.data.data_utils.post_process() for full list of options" + }, + ) + wer_kenlm_model: Optional[str] = field( + default=None, + metadata={ + "help": "if this is provided, use kenlm to compute wer (along with other wer_* args)" + }, + ) + wer_lexicon: Optional[str] = field( + default=None, + metadata={"help": "lexicon to use with wer_kenlm_model"}, + ) + wer_lm_weight: float = field( + default=2.0, + metadata={"help": "lm weight to use with wer_kenlm_model"}, + ) + wer_word_score: float = field( + default=-1.0, + metadata={"help": "lm word score to use with wer_kenlm_model"}, + ) + wer_sil_weight: float = field( + default=0, + metadata={"help": "lm word score to use with wer_kenlm_model"}, + ) + + wer_args: Optional[str] = field( + default=None, + metadata={ + "help": "DEPRECATED: tuple of (wer_kenlm_model, wer_lexicon, wer_lm_weight, wer_word_score)" + }, + ) + + +@register_criterion("ctc", dataclass=CtcCriterionConfig) +class CtcCriterion(FairseqCriterion): + def __init__( + self, cfg: CtcCriterionConfig, task: FairseqTask, rdrop_alpha: int = 0.0 + ): + super().__init__(task) + self.blank_idx = ( + task.target_dictionary.index(task.blank_symbol) + if hasattr(task, "blank_symbol") + else 0 + ) + self.pad_idx = task.target_dictionary.pad() + self.eos_idx = task.target_dictionary.eos() + self.post_process = cfg.post_process + + self.rdrop_alpha = rdrop_alpha + + if cfg.wer_args is not None: + ( + cfg.wer_kenlm_model, + cfg.wer_lexicon, + cfg.wer_lm_weight, + cfg.wer_word_score, + ) = eval(cfg.wer_args) + + if cfg.wer_kenlm_model is not None and cfg.wer_kenlm_model != "": + from examples.speech_recognition.w2l_decoder import W2lKenLMDecoder + + dec_args = Namespace() + dec_args.nbest = 1 + dec_args.criterion = "ctc" + dec_args.kenlm_model = cfg.wer_kenlm_model + dec_args.lexicon = cfg.wer_lexicon + dec_args.beam = 50 + dec_args.beam_size_token = min(50, len(task.target_dictionary)) + dec_args.beam_threshold = min(50, len(task.target_dictionary)) + dec_args.lm_weight = cfg.wer_lm_weight + dec_args.word_score = cfg.wer_word_score + dec_args.sil_weight = cfg.wer_sil_weight + dec_args.unk_weight = -math.inf + dec_args.sil_weight = 0 + + self.w2l_decoder = W2lKenLMDecoder(dec_args, task.target_dictionary) + else: + self.w2l_decoder = None + + self.zero_infinity = cfg.zero_infinity + self.sentence_avg = cfg.sentence_avg + + def forward(self, model, sample, reduce=True, **kwargs): + net_output = model(**sample["net_input"]) + lprobs = model.get_normalized_probs( + net_output, log_probs=True + ).contiguous() # (T, B, C) from the encoder + + # CTC loss is calculated over duplicated inputs + # sample is already duplicated for R-Drop + if self.rdrop_alpha > 0: + for k, v in sample.items(): + if k in ["target", "target_lengths"]: + sample[k] = torch.cat([v, v.clone()], dim=0) + elif k == "net_input": + if sample[k]["src_tokens"].size(1) != sample[k]["src_lengths"].size( + 0 + ): + # for decoder CTC loss + sample[k]["src_lengths"] = torch.cat( + [ + sample[k]["src_lengths"], + sample[k]["src_lengths"].clone(), + ], + dim=0, + ) + + if "src_lengths" in sample["net_input"]: + input_lengths = sample["net_input"]["src_lengths"] + else: + if net_output["padding_mask"] is not None: + non_padding_mask = ~net_output["padding_mask"] + input_lengths = non_padding_mask.long().sum(-1) + else: + input_lengths = lprobs.new_full( + (lprobs.size(1),), lprobs.size(0), dtype=torch.long + ) + + pad_mask = (sample["target"] != self.pad_idx) & ( + sample["target"] != self.eos_idx + ) + targets_flat = sample["target"].masked_select(pad_mask) + if "target_lengths" in sample: + target_lengths = sample["target_lengths"] + else: + target_lengths = pad_mask.sum(-1) + + with torch.backends.cudnn.flags(enabled=False): + loss = F.ctc_loss( + lprobs, + targets_flat, + input_lengths, + target_lengths, + blank=self.blank_idx, + reduction="sum", + zero_infinity=self.zero_infinity, + ) + + ntokens = ( + sample["ntokens"] if "ntokens" in sample else target_lengths.sum().item() + ) + + sample_size = sample["target"].size(0) if self.sentence_avg else ntokens + logging_output = { + "loss": utils.item(loss.data), # * sample['ntokens'], + "ntokens": ntokens, + "nsentences": sample["id"].numel(), + "sample_size": sample_size, + } + + if not model.training: + import editdistance + + with torch.no_grad(): + lprobs_t = lprobs.transpose(0, 1).float().contiguous().cpu() + + c_err = 0 + c_len = 0 + w_errs = 0 + w_len = 0 + wv_errs = 0 + for lp, t, inp_l in zip( + lprobs_t, + sample["target_label"] + if "target_label" in sample + else sample["target"], + input_lengths, + ): + lp = lp[:inp_l].unsqueeze(0) + + decoded = None + if self.w2l_decoder is not None: + decoded = self.w2l_decoder.decode(lp) + if len(decoded) < 1: + decoded = None + else: + decoded = decoded[0] + if len(decoded) < 1: + decoded = None + else: + decoded = decoded[0] + + p = (t != self.task.target_dictionary.pad()) & ( + t != self.task.target_dictionary.eos() + ) + targ = t[p] + targ_units = self.task.target_dictionary.string(targ) + targ_units_arr = targ.tolist() + + toks = lp.argmax(dim=-1).unique_consecutive() + pred_units_arr = toks[toks != self.blank_idx].tolist() + + c_err += editdistance.eval(pred_units_arr, targ_units_arr) + c_len += len(targ_units_arr) + + targ_words = post_process(targ_units, self.post_process).split() + + pred_units = self.task.target_dictionary.string(pred_units_arr) + pred_words_raw = post_process(pred_units, self.post_process).split() + + if decoded is not None and "words" in decoded: + pred_words = decoded["words"] + w_errs += editdistance.eval(pred_words, targ_words) + wv_errs += editdistance.eval(pred_words_raw, targ_words) + else: + dist = editdistance.eval(pred_words_raw, targ_words) + w_errs += dist + wv_errs += dist + + w_len += len(targ_words) + + logging_output["wv_errors"] = wv_errs + logging_output["w_errors"] = w_errs + logging_output["w_total"] = w_len + logging_output["c_errors"] = c_err + logging_output["c_total"] = c_len + + return loss, sample_size, logging_output + + @staticmethod + def reduce_metrics(logging_outputs) -> None: + """Aggregate logging outputs from data parallel training.""" + + loss_sum = utils.item(sum(log.get("loss", 0) for log in logging_outputs)) + ntokens = utils.item(sum(log.get("ntokens", 0) for log in logging_outputs)) + nsentences = utils.item( + sum(log.get("nsentences", 0) for log in logging_outputs) + ) + sample_size = utils.item( + sum(log.get("sample_size", 0) for log in logging_outputs) + ) + + metrics.log_scalar( + "loss", loss_sum / sample_size / math.log(2), sample_size, round=3 + ) + metrics.log_scalar("ntokens", ntokens) + metrics.log_scalar("nsentences", nsentences) + if sample_size != ntokens: + metrics.log_scalar( + "nll_loss", loss_sum / ntokens / math.log(2), ntokens, round=3 + ) + + c_errors = sum(log.get("c_errors", 0) for log in logging_outputs) + metrics.log_scalar("_c_errors", c_errors) + c_total = sum(log.get("c_total", 0) for log in logging_outputs) + metrics.log_scalar("_c_total", c_total) + w_errors = sum(log.get("w_errors", 0) for log in logging_outputs) + metrics.log_scalar("_w_errors", w_errors) + wv_errors = sum(log.get("wv_errors", 0) for log in logging_outputs) + metrics.log_scalar("_wv_errors", wv_errors) + w_total = sum(log.get("w_total", 0) for log in logging_outputs) + metrics.log_scalar("_w_total", w_total) + + if c_total > 0: + metrics.log_derived( + "uer", + lambda meters: safe_round( + meters["_c_errors"].sum * 100.0 / meters["_c_total"].sum, 3 + ) + if meters["_c_total"].sum > 0 + else float("nan"), + ) + if w_total > 0: + metrics.log_derived( + "wer", + lambda meters: safe_round( + meters["_w_errors"].sum * 100.0 / meters["_w_total"].sum, 3 + ) + if meters["_w_total"].sum > 0 + else float("nan"), + ) + metrics.log_derived( + "raw_wer", + lambda meters: safe_round( + meters["_wv_errors"].sum * 100.0 / meters["_w_total"].sum, 3 + ) + if meters["_w_total"].sum > 0 + else float("nan"), + ) + + @staticmethod + def logging_outputs_can_be_summed() -> bool: + """ + Whether the logging outputs returned by `forward` can be summed + across workers prior to calling `reduce_metrics`. Setting this + to True will improves distributed training speed. + """ + return True diff --git a/data/fairseq/fairseq/criterions/fairseq_criterion.py b/data/fairseq/fairseq/criterions/fairseq_criterion.py new file mode 100644 index 0000000000000000000000000000000000000000..0b1e64a8e36c55be90d7e3f854effd99ed5bcc44 --- /dev/null +++ b/data/fairseq/fairseq/criterions/fairseq_criterion.py @@ -0,0 +1,121 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import inspect +from typing import Any, Dict, List + +from fairseq import utils +from fairseq.logging import metrics +from fairseq.dataclass import FairseqDataclass +from fairseq.dataclass.utils import gen_parser_from_dataclass +from torch.nn.modules.loss import _Loss + + +class FairseqCriterion(_Loss): + def __init__(self, task): + super().__init__() + self.task = task + if hasattr(task, "target_dictionary"): + tgt_dict = task.target_dictionary + self.padding_idx = tgt_dict.pad() if tgt_dict is not None else -100 + + @classmethod + def add_args(cls, parser): + """Add criterion-specific arguments to the parser.""" + dc = getattr(cls, "__dataclass", None) + if dc is not None: + gen_parser_from_dataclass(parser, dc()) + + @classmethod + def build_criterion(cls, cfg: FairseqDataclass, task): + """Construct a criterion from command-line args.""" + # arguments in the __init__. + init_args = {} + for p in inspect.signature(cls).parameters.values(): + if ( + p.kind == p.POSITIONAL_ONLY + or p.kind == p.VAR_POSITIONAL + or p.kind == p.VAR_KEYWORD + ): + # we haven't implemented inference for these argument types, + # but PRs welcome :) + raise NotImplementedError("{} not supported".format(p.kind)) + + assert p.kind in {p.POSITIONAL_OR_KEYWORD, p.KEYWORD_ONLY} + + if p.name == "task": + init_args["task"] = task + elif p.name == "cfg": + init_args["cfg"] = cfg + elif hasattr(cfg, p.name): + init_args[p.name] = getattr(cfg, p.name) + elif p.default != p.empty: + pass # we'll use the default value + else: + raise NotImplementedError( + "Unable to infer Criterion arguments, please implement " + "{}.build_criterion".format(cls.__name__) + ) + return cls(**init_args) + + def forward(self, model, sample, reduce=True): + """Compute the loss for the given sample. + + Returns a tuple with three elements: + 1) the loss + 2) the sample size, which is used as the denominator for the gradient + 3) logging outputs to display while training + """ + raise NotImplementedError + + @staticmethod + def aggregate_logging_outputs( + logging_outputs: List[Dict[str, Any]] + ) -> Dict[str, Any]: + """Aggregate logging outputs from data parallel training.""" + utils.deprecation_warning( + "The aggregate_logging_outputs API is deprecated. " + "Please use the reduce_metrics API instead." + ) + raise NotImplementedError + + @classmethod + def reduce_metrics(cls, logging_outputs: List[Dict[str, Any]]) -> None: + """Aggregate logging outputs from data parallel training.""" + utils.deprecation_warning( + "Criterions should implement the reduce_metrics API. " + "Falling back to deprecated aggregate_logging_outputs API." + ) + agg_logging_outputs = cls.aggregate_logging_outputs(logging_outputs) + for k, v in agg_logging_outputs.items(): + if k in {"nsentences", "ntokens", "sample_size"}: + continue + metrics.log_scalar(k, v) + + @staticmethod + def logging_outputs_can_be_summed() -> bool: + """ + Whether the logging outputs returned by `forward` can be summed + across workers prior to calling `reduce_metrics`. Setting this + to True will improves distributed training speed. + """ + return False + + +class LegacyFairseqCriterion(FairseqCriterion): + def __init__(self, args, task): + super().__init__(task=task) + self.args = args + + utils.deprecation_warning( + "Criterions should take explicit arguments instead of an " + "argparse.Namespace object, please update your criterion by " + "extending FairseqCriterion instead of LegacyFairseqCriterion." + ) + + @classmethod + def build_criterion(cls, args, task): + """Construct a criterion from command-line args.""" + return cls(args, task) diff --git a/data/fairseq/fairseq/criterions/fastspeech2_loss.py b/data/fairseq/fairseq/criterions/fastspeech2_loss.py new file mode 100644 index 0000000000000000000000000000000000000000..ab7cd08e3bd9c8d5c4be017095034b18362d77e0 --- /dev/null +++ b/data/fairseq/fairseq/criterions/fastspeech2_loss.py @@ -0,0 +1,137 @@ +# Copyright (c) 2017-present, Facebook, Inc. +# All rights reserved. +# +# This source code is licensed under the license found in the LICENSE file in +# the root directory of this source tree. An additional grant of patent rights +# can be found in the PATENTS file in the same directory. + +from typing import List, Dict, Any +from dataclasses import dataclass, field + +import torch +import torch.nn.functional as F + +from fairseq import utils +from fairseq.logging import metrics +from fairseq.criterions import FairseqCriterion, register_criterion +from fairseq.dataclass import FairseqDataclass +from fairseq.data.data_utils import lengths_to_mask +from fairseq.models.fairseq_model import FairseqEncoderModel + + +@dataclass +class FastSpeech2CriterionConfig(FairseqDataclass): + ctc_weight: float = field(default=0.0, metadata={"help": "weight for CTC loss"}) + + +@register_criterion("fastspeech2", dataclass=FastSpeech2CriterionConfig) +class FastSpeech2Loss(FairseqCriterion): + def __init__(self, task, ctc_weight): + super().__init__(task) + self.ctc_weight = ctc_weight + + def forward(self, model: FairseqEncoderModel, sample, reduction="mean"): + src_tokens = sample["net_input"]["src_tokens"] + src_lens = sample["net_input"]["src_lengths"] + tgt_lens = sample["target_lengths"] + _feat_out, _feat_out_post, _, log_dur_out, pitch_out, energy_out = model( + src_tokens=src_tokens, + src_lengths=src_lens, + prev_output_tokens=sample["net_input"]["prev_output_tokens"], + incremental_state=None, + target_lengths=tgt_lens, + speaker=sample["speaker"], + durations=sample["durations"], + pitches=sample["pitches"], + energies=sample["energies"], + ) + + src_mask = lengths_to_mask(sample["net_input"]["src_lengths"]) + tgt_mask = lengths_to_mask(sample["target_lengths"]) + + pitches, energies = sample["pitches"], sample["energies"] + pitch_out, pitches = pitch_out[src_mask], pitches[src_mask] + energy_out, energies = energy_out[src_mask], energies[src_mask] + + feat_out, feat = _feat_out[tgt_mask], sample["target"][tgt_mask] + l1_loss = F.l1_loss(feat_out, feat, reduction=reduction) + if _feat_out_post is not None: + l1_loss += F.l1_loss(_feat_out_post[tgt_mask], feat, reduction=reduction) + + pitch_loss = F.mse_loss(pitch_out, pitches, reduction=reduction) + energy_loss = F.mse_loss(energy_out, energies, reduction=reduction) + + log_dur_out = log_dur_out[src_mask] + dur = sample["durations"].float() + dur = dur.half() if log_dur_out.type().endswith(".HalfTensor") else dur + log_dur = torch.log(dur + 1)[src_mask] + dur_loss = F.mse_loss(log_dur_out, log_dur, reduction=reduction) + + ctc_loss = torch.tensor(0.0).type_as(l1_loss) + if self.ctc_weight > 0.0: + lprobs = model.get_normalized_probs((_feat_out,), log_probs=True) + lprobs = lprobs.transpose(0, 1) # T x B x C + src_mask = lengths_to_mask(src_lens) + src_tokens_flat = src_tokens.masked_select(src_mask) + ctc_loss = ( + F.ctc_loss( + lprobs, + src_tokens_flat, + tgt_lens, + src_lens, + reduction=reduction, + zero_infinity=True, + ) + * self.ctc_weight + ) + + loss = l1_loss + dur_loss + pitch_loss + energy_loss + ctc_loss + + sample_size = sample["nsentences"] + logging_output = { + "loss": utils.item(loss.data), + "ntokens": sample["ntokens"], + "nsentences": sample["nsentences"], + "sample_size": sample_size, + "l1_loss": utils.item(l1_loss.data), + "dur_loss": utils.item(dur_loss.data), + "pitch_loss": utils.item(pitch_loss.data), + "energy_loss": utils.item(energy_loss.data), + "ctc_loss": utils.item(ctc_loss.data), + } + return loss, sample_size, logging_output + + @classmethod + def reduce_metrics(cls, logging_outputs: List[Dict[str, Any]]) -> None: + ns = [log.get("sample_size", 0) for log in logging_outputs] + ntot = sum(ns) + ws = [n / (ntot + 1e-8) for n in ns] + for key in [ + "loss", + "l1_loss", + "dur_loss", + "pitch_loss", + "energy_loss", + "ctc_loss", + ]: + vals = [log.get(key, 0) for log in logging_outputs] + val = sum(val * w for val, w in zip(vals, ws)) + metrics.log_scalar(key, val, ntot, round=3) + metrics.log_scalar("sample_size", ntot, len(logging_outputs)) + + # inference metrics + if "targ_frames" not in logging_outputs[0]: + return + n = sum(log.get("targ_frames", 0) for log in logging_outputs) + for key, new_key in [ + ("mcd_loss", "mcd_loss"), + ("pred_frames", "pred_ratio"), + ("nins", "ins_rate"), + ("ndel", "del_rate"), + ]: + val = sum(log.get(key, 0) for log in logging_outputs) + metrics.log_scalar(new_key, val / n, n, round=3) + + @staticmethod + def logging_outputs_can_be_summed() -> bool: + return False diff --git a/data/fairseq/fairseq/criterions/hubert_criterion.py b/data/fairseq/fairseq/criterions/hubert_criterion.py new file mode 100644 index 0000000000000000000000000000000000000000..262874b582aa4765981fd9cc958c7221596d681e --- /dev/null +++ b/data/fairseq/fairseq/criterions/hubert_criterion.py @@ -0,0 +1,195 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import math +import re +from dataclasses import dataclass, field +from typing import List, Optional + +import torch +import torch.nn.functional as F +from fairseq import utils +from fairseq.logging import metrics +from fairseq.criterions import FairseqCriterion, register_criterion +from fairseq.dataclass import FairseqDataclass + + +@dataclass +class HubertCriterionConfig(FairseqDataclass): + pred_masked_weight: float = field( + default=1.0, + metadata={"help": "weight for predictive loss for masked frames"}, + ) + pred_nomask_weight: float = field( + default=0.0, + metadata={"help": "weight for predictive loss for unmasked frames"}, + ) + loss_weights: Optional[List[float]] = field( + default=None, + metadata={"help": "weights for additional loss terms (not first one)"}, + ) + log_keys: List[str] = field( + default_factory=lambda: [], + metadata={"help": "output keys to log"}, + ) + + +@register_criterion("hubert", dataclass=HubertCriterionConfig) +class HubertCriterion(FairseqCriterion): + def __init__( + self, + task, + pred_masked_weight, + pred_nomask_weight, + loss_weights=None, + log_keys=None, + ): + super().__init__(task) + self.pred_masked_weight = pred_masked_weight + self.pred_nomask_weight = pred_nomask_weight + self.loss_weights = loss_weights + self.log_keys = [] if log_keys is None else log_keys + + def forward(self, model, sample, reduce=True, log_pred=False): + """Compute the loss for the given sample. + Returns a tuple with three elements: + 1) the loss + 2) the sample size, which is used as the denominator for the gradient + 3) logging outputs to display while training + """ + net_output = model(target_list=sample["target_list"], **sample["net_input"]) + loss = 0.0 + sample_size = 0 + logging_output = {} + reduction = "sum" if reduce else "none" + + loss_m_list = [] + logp_m_list = model.get_logits(net_output, True) + targ_m_list = model.get_targets(net_output, True) + assert self.pred_masked_weight == 0 or len(logp_m_list) > 0 + for i, (logp_m, targ_m) in enumerate(zip(logp_m_list, targ_m_list)): + loss_m = F.cross_entropy(logp_m, targ_m, reduction=reduction) + loss_m_list.append(loss_m) + logging_output[f"loss_m_{i}"] = loss_m.detach().item() + if self.pred_masked_weight > 0: + loss += self.pred_masked_weight * sum(loss_m_list) + sample_size += targ_m_list[0].numel() + + loss_u_list = [] + logp_u_list = model.get_logits(net_output, False) + targ_u_list = model.get_targets(net_output, False) + assert self.pred_nomask_weight == 0 or len(logp_u_list) > 0 + for i, (logp_u, targ_u) in enumerate(zip(logp_u_list, targ_u_list)): + loss_u = F.cross_entropy(logp_u, targ_u, reduction=reduction) + loss_u_list.append(loss_u) + logging_output[f"loss_u_{i}"] = loss_u.detach().item() + if self.pred_nomask_weight > 0: + loss += self.pred_nomask_weight * sum(loss_u_list) + sample_size += targ_u_list[0].numel() + + if self.loss_weights is not None: + assert hasattr(model, "get_extra_losses") + extra_losses, names = model.get_extra_losses(net_output) + if torch.is_tensor(extra_losses): + extra_losses = [extra_losses] + names = [names] + if len(self.loss_weights) == 1 and len(extra_losses) != 1: + self.loss_weights = [self.loss_weights[0]] * len(extra_losses) + assert len(extra_losses) == len( + self.loss_weights + ), f"{len(extra_losses)}, {len(self.loss_weights)}" + for p, n, coef in zip(extra_losses, names, self.loss_weights): + if coef != 0 and p is not None: + p = coef * p.float() * sample_size + loss += p + logging_output[f"loss_{n}"] = p.item() + + logging_output = { + "loss": loss.item() if reduce else loss, + "ntokens": sample_size, + "nsentences": sample["id"].numel(), + "sample_size": sample_size, + **logging_output, + } + + for lk in self.log_keys: + if lk in net_output: + logging_output[lk] = float((net_output[lk])) + + def compute_correct(logits): + if logits.numel() == 0: + return 0, 0 + else: + assert logits.dim() > 1, logits.shape + max = logits.argmax(-1) == 0 + min = logits.argmin(-1) == 0 + both = max & min + corr = max.long().sum().item() - both.long().sum().item() + count = max.numel() + return corr, count + + with torch.no_grad(): + for i, logp_m in enumerate(logp_m_list): + corr_m, count_m = compute_correct(logp_m) + logging_output[f"correct_m_{i}"] = corr_m + logging_output[f"count_m_{i}"] = count_m + + for i, logp_u in enumerate(logp_u_list): + corr_u, count_u = compute_correct(logp_u) + logging_output[f"correct_u_{i}"] = corr_u + logging_output[f"count_u_{i}"] = count_u + + return loss, sample_size, logging_output + + @staticmethod + def reduce_metrics(logging_outputs) -> None: + """Aggregate logging outputs from data parallel training (copied from normal cross entropy).""" + loss_sum = sum(log.get("loss", 0) for log in logging_outputs) + ntokens = sum(log.get("ntokens", 0) for log in logging_outputs) + sample_size = sum(log.get("sample_size", 0) for log in logging_outputs) + + metrics.log_scalar( + "loss", loss_sum / sample_size / math.log(2), sample_size, round=3 + ) + if sample_size != ntokens: + metrics.log_scalar( + "nll_loss", loss_sum / ntokens / math.log(2), ntokens, round=3 + ) + metrics.log_derived( + "ppl", lambda meters: utils.get_perplexity(meters["nll_loss"].avg) + ) + else: + metrics.log_derived( + "ppl", lambda meters: utils.get_perplexity(meters["loss"].avg) + ) + + counts = {} + for lk in logging_outputs[0].keys(): + if lk.startswith("count_"): + val = sum(log[lk] for log in logging_outputs) + metrics.log_scalar(lk, val) + counts[lk] = val + + for lk in logging_outputs[0].keys(): + if lk.startswith("loss_"): + val = sum(log[lk] for log in logging_outputs) + metrics.log_scalar(lk, val / sample_size / math.log(2), round=3) + elif lk.startswith("correct_"): + val = sum(log[lk] for log in logging_outputs) + metrics.log_scalar(lk, val / counts[re.sub("correct", "count", lk)]) + + @staticmethod + def aggregate_logging_outputs(logging_outputs): + """Aggregate logging outputs from data parallel training.""" + raise NotImplementedError() + + @staticmethod + def logging_outputs_can_be_summed() -> bool: + """ + Whether the logging outputs returned by `forward` can be summed + across workers prior to calling `reduce_metrics`. Setting this + to True will improves distributed training speed. + """ + return False diff --git a/data/fairseq/fairseq/criterions/label_smoothed_cross_entropy.py b/data/fairseq/fairseq/criterions/label_smoothed_cross_entropy.py new file mode 100644 index 0000000000000000000000000000000000000000..325679bb1678928b9fe644293b39f00115300a15 --- /dev/null +++ b/data/fairseq/fairseq/criterions/label_smoothed_cross_entropy.py @@ -0,0 +1,168 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import math +from dataclasses import dataclass, field + +import torch +from fairseq import utils +from fairseq.logging import metrics +from fairseq.criterions import FairseqCriterion, register_criterion +from fairseq.dataclass import FairseqDataclass +from omegaconf import II + + +@dataclass +class LabelSmoothedCrossEntropyCriterionConfig(FairseqDataclass): + label_smoothing: float = field( + default=0.0, + metadata={"help": "epsilon for label smoothing, 0 means no label smoothing"}, + ) + report_accuracy: bool = field( + default=False, + metadata={"help": "report accuracy metric"}, + ) + ignore_prefix_size: int = field( + default=0, + metadata={"help": "Ignore first N tokens"}, + ) + sentence_avg: bool = II("optimization.sentence_avg") + + +def label_smoothed_nll_loss(lprobs, target, epsilon, ignore_index=None, reduce=True): + if target.dim() == lprobs.dim() - 1: + target = target.unsqueeze(-1) + nll_loss = -lprobs.gather(dim=-1, index=target) + smooth_loss = -lprobs.sum(dim=-1, keepdim=True) + if ignore_index is not None: + pad_mask = target.eq(ignore_index) + nll_loss.masked_fill_(pad_mask, 0.0) + smooth_loss.masked_fill_(pad_mask, 0.0) + else: + nll_loss = nll_loss.squeeze(-1) + smooth_loss = smooth_loss.squeeze(-1) + if reduce: + nll_loss = nll_loss.sum() + smooth_loss = smooth_loss.sum() + eps_i = epsilon / (lprobs.size(-1) - 1) + loss = (1.0 - epsilon - eps_i) * nll_loss + eps_i * smooth_loss + return loss, nll_loss + + +@register_criterion( + "label_smoothed_cross_entropy", dataclass=LabelSmoothedCrossEntropyCriterionConfig +) +class LabelSmoothedCrossEntropyCriterion(FairseqCriterion): + def __init__( + self, + task, + sentence_avg, + label_smoothing, + ignore_prefix_size=0, + report_accuracy=False, + ): + super().__init__(task) + self.sentence_avg = sentence_avg + self.eps = label_smoothing + self.ignore_prefix_size = ignore_prefix_size + self.report_accuracy = report_accuracy + + def forward(self, model, sample, reduce=True): + """Compute the loss for the given sample. + + Returns a tuple with three elements: + 1) the loss + 2) the sample size, which is used as the denominator for the gradient + 3) logging outputs to display while training + """ + net_output = model(**sample["net_input"]) + loss, nll_loss = self.compute_loss(model, net_output, sample, reduce=reduce) + sample_size = ( + sample["target"].size(0) if self.sentence_avg else sample["ntokens"] + ) + logging_output = { + "loss": loss.data, + "nll_loss": nll_loss.data, + "ntokens": sample["ntokens"], + "nsentences": sample["target"].size(0), + "sample_size": sample_size, + } + if self.report_accuracy: + n_correct, total = self.compute_accuracy(model, net_output, sample) + logging_output["n_correct"] = utils.item(n_correct.data) + logging_output["total"] = utils.item(total.data) + return loss, sample_size, logging_output + + def get_lprobs_and_target(self, model, net_output, sample): + lprobs = model.get_normalized_probs(net_output, log_probs=True) + target = model.get_targets(sample, net_output) + if self.ignore_prefix_size > 0: + # lprobs: B x T x C + lprobs = lprobs[:, self.ignore_prefix_size :, :].contiguous() + target = target[:, self.ignore_prefix_size :].contiguous() + return lprobs.view(-1, lprobs.size(-1)), target.view(-1) + + def compute_loss(self, model, net_output, sample, reduce=True): + lprobs, target = self.get_lprobs_and_target(model, net_output, sample) + loss, nll_loss = label_smoothed_nll_loss( + lprobs, + target, + self.eps, + ignore_index=self.padding_idx, + reduce=reduce, + ) + return loss, nll_loss + + def compute_accuracy(self, model, net_output, sample): + lprobs, target = self.get_lprobs_and_target(model, net_output, sample) + mask = target.ne(self.padding_idx) + n_correct = torch.sum( + lprobs.argmax(1).masked_select(mask).eq(target.masked_select(mask)) + ) + total = torch.sum(mask) + return n_correct, total + + @classmethod + def reduce_metrics(cls, logging_outputs) -> None: + """Aggregate logging outputs from data parallel training.""" + loss_sum = sum(log.get("loss", 0) for log in logging_outputs) + nll_loss_sum = sum(log.get("nll_loss", 0) for log in logging_outputs) + ntokens = sum(log.get("ntokens", 0) for log in logging_outputs) + sample_size = sum(log.get("sample_size", 0) for log in logging_outputs) + + metrics.log_scalar( + "loss", loss_sum / sample_size / math.log(2), sample_size, round=3 + ) + metrics.log_scalar( + "nll_loss", nll_loss_sum / ntokens / math.log(2), ntokens, round=3 + ) + metrics.log_derived( + "ppl", lambda meters: utils.get_perplexity(meters["nll_loss"].avg) + ) + + total = utils.item(sum(log.get("total", 0) for log in logging_outputs)) + if total > 0: + metrics.log_scalar("total", total) + n_correct = utils.item( + sum(log.get("n_correct", 0) for log in logging_outputs) + ) + metrics.log_scalar("n_correct", n_correct) + metrics.log_derived( + "accuracy", + lambda meters: round( + meters["n_correct"].sum * 100.0 / meters["total"].sum, 3 + ) + if meters["total"].sum > 0 + else float("nan"), + ) + + @staticmethod + def logging_outputs_can_be_summed() -> bool: + """ + Whether the logging outputs returned by `forward` can be summed + across workers prior to calling `reduce_metrics`. Setting this + to True will improves distributed training speed. + """ + return True diff --git a/data/fairseq/fairseq/criterions/label_smoothed_cross_entropy_latency_augmented.py b/data/fairseq/fairseq/criterions/label_smoothed_cross_entropy_latency_augmented.py new file mode 100644 index 0000000000000000000000000000000000000000..6eaedab9cf35dd6636e3463fdab1f0d4f9dda7e4 --- /dev/null +++ b/data/fairseq/fairseq/criterions/label_smoothed_cross_entropy_latency_augmented.py @@ -0,0 +1,221 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from dataclasses import dataclass, field +import torch +from fairseq import utils +from fairseq.logging import metrics +from fairseq.criterions import register_criterion +from fairseq.criterions.label_smoothed_cross_entropy import ( + LabelSmoothedCrossEntropyCriterion, + LabelSmoothedCrossEntropyCriterionConfig, +) + +try: + from simuleval.metrics.latency import ( + AverageLagging, + AverageProportion, + DifferentiableAverageLagging, + ) + + LATENCY_METRICS = { + "average_lagging": AverageLagging, + "average_proportion": AverageProportion, + "differentiable_average_lagging": DifferentiableAverageLagging, + } +except ImportError: + LATENCY_METRICS = None + + +@dataclass +class LabelSmoothedCrossEntropyCriterionLatencyAugmentConfig( + LabelSmoothedCrossEntropyCriterionConfig +): + latency_avg_weight: float = field( + default=0.0, + metadata={"help": "weight fot average latency loss."}, + ) + latency_var_weight: float = field( + default=0.0, + metadata={"help": "weight fot variance latency loss."}, + ) + latency_avg_type: str = field( + default="differentiable_average_lagging", + metadata={"help": "latency type for average loss"}, + ) + latency_var_type: str = field( + default="variance_delay", + metadata={"help": "latency typ for variance loss"}, + ) + latency_gather_method: str = field( + default="weighted_average", + metadata={"help": "method to gather latency loss for all heads"}, + ) + latency_update_after: int = field( + default=0, + metadata={"help": "Add latency loss after certain steps"}, + ) + + +@register_criterion( + "latency_augmented_label_smoothed_cross_entropy", + dataclass=LabelSmoothedCrossEntropyCriterionLatencyAugmentConfig, +) +class LatencyAugmentedLabelSmoothedCrossEntropyCriterion( + LabelSmoothedCrossEntropyCriterion +): + def __init__( + self, + task, + sentence_avg, + label_smoothing, + ignore_prefix_size, + report_accuracy, + latency_avg_weight, + latency_var_weight, + latency_avg_type, + latency_var_type, + latency_gather_method, + latency_update_after, + ): + super().__init__( + task, sentence_avg, label_smoothing, ignore_prefix_size, report_accuracy + ) + assert LATENCY_METRICS is not None, "Please make sure SimulEval is installed." + + self.latency_avg_weight = latency_avg_weight + self.latency_var_weight = latency_var_weight + self.latency_avg_type = latency_avg_type + self.latency_var_type = latency_var_type + self.latency_gather_method = latency_gather_method + self.latency_update_after = latency_update_after + + def forward(self, model, sample, reduce=True): + net_output = model(**sample["net_input"]) + # 1. Compute cross entropy loss + loss, nll_loss = self.compute_loss(model, net_output, sample, reduce=reduce) + + # 2. Compute cross latency loss + latency_loss, expected_latency, expected_delays_var = self.compute_latency_loss( + model, sample, net_output + ) + + if self.latency_update_after > 0: + num_updates = getattr(model.decoder, "num_updates", None) + assert ( + num_updates is not None + ), "model.decoder doesn't have attribute 'num_updates'" + if num_updates <= self.latency_update_after: + latency_loss = 0 + + loss += latency_loss + + sample_size = ( + sample["target"].size(0) if self.sentence_avg else sample["ntokens"] + ) + + logging_output = { + "loss": loss.data, + "nll_loss": nll_loss.data, + "ntokens": sample["ntokens"], + "nsentences": sample["target"].size(0), + "sample_size": sample_size, + "latency": expected_latency, + "delays_var": expected_delays_var, + "latency_loss": latency_loss, + } + + if self.report_accuracy: + n_correct, total = self.compute_accuracy(model, net_output, sample) + logging_output["n_correct"] = utils.item(n_correct.data) + logging_output["total"] = utils.item(total.data) + return loss, sample_size, logging_output + + def compute_latency_loss(self, model, sample, net_output): + assert ( + net_output[-1].encoder_padding_mask is None + or not net_output[-1].encoder_padding_mask[:, 0].any() + ), "Only right padding on source is supported." + # 1. Obtain the expected alignment + alpha_list = [item["alpha"] for item in net_output[1].attn_list] + num_layers = len(alpha_list) + bsz, num_heads, tgt_len, src_len = alpha_list[0].size() + + # bsz * num_layers * num_heads, tgt_len, src_len + alpha_all = torch.cat(alpha_list, dim=1).view(-1, tgt_len, src_len) + + # 2 compute expected delays + # bsz * num_heads * num_layers, tgt_len, src_len for MMA + steps = ( + torch.arange(1, 1 + src_len) + .unsqueeze(0) + .unsqueeze(1) + .expand_as(alpha_all) + .type_as(alpha_all) + ) + + expected_delays = torch.sum(steps * alpha_all, dim=-1) + + target_padding_mask = ( + model.get_targets(sample, net_output) + .eq(self.padding_idx) + .unsqueeze(1) + .expand(bsz, num_layers * num_heads, tgt_len) + .contiguous() + .view(-1, tgt_len) + ) + + src_lengths = ( + sample["net_input"]["src_lengths"] + .unsqueeze(1) + .expand(bsz, num_layers * num_heads) + .contiguous() + .view(-1) + ) + expected_latency = LATENCY_METRICS[self.latency_avg_type]( + expected_delays, src_lengths, None, target_padding_mask=target_padding_mask + ) + + # 2.1 average expected latency of heads + # bsz, num_layers * num_heads + expected_latency = expected_latency.view(bsz, -1) + if self.latency_gather_method == "average": + # bsz * tgt_len + expected_latency = expected_delays.mean(dim=1) + elif self.latency_gather_method == "weighted_average": + weights = torch.nn.functional.softmax(expected_latency, dim=1) + expected_latency = torch.sum(expected_latency * weights, dim=1) + elif self.latency_gather_method == "max": + expected_latency = expected_latency.max(dim=1)[0] + else: + raise NotImplementedError + + expected_latency = expected_latency.sum() + avg_loss = self.latency_avg_weight * expected_latency + + # 2.2 variance of expected delays + expected_delays_var = ( + expected_delays.view(bsz, -1, tgt_len).var(dim=1).mean(dim=1) + ) + expected_delays_var = expected_delays_var.sum() + var_loss = self.latency_avg_weight * expected_delays_var + + # 3. Final loss + latency_loss = avg_loss + var_loss + + return latency_loss, expected_latency, expected_delays_var + + @classmethod + def reduce_metrics(cls, logging_outputs) -> None: + super().reduce_metrics(logging_outputs) + latency = sum(log.get("latency", 0) for log in logging_outputs) + delays_var = sum(log.get("delays_var", 0) for log in logging_outputs) + latency_loss = sum(log.get("latency_loss", 0) for log in logging_outputs) + nsentences = sum(log.get("nsentences", 0) for log in logging_outputs) + metrics.log_scalar("latency", latency.float() / nsentences, nsentences, round=3) + metrics.log_scalar("delays_var", delays_var / nsentences, nsentences, round=3) + metrics.log_scalar( + "latency_loss", latency_loss / nsentences, nsentences, round=3 + ) diff --git a/data/fairseq/fairseq/criterions/label_smoothed_cross_entropy_with_alignment.py b/data/fairseq/fairseq/criterions/label_smoothed_cross_entropy_with_alignment.py new file mode 100644 index 0000000000000000000000000000000000000000..b55f65e5cc7e9e949208786a4974b4ba09b0de66 --- /dev/null +++ b/data/fairseq/fairseq/criterions/label_smoothed_cross_entropy_with_alignment.py @@ -0,0 +1,131 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import math + +from fairseq import utils +from fairseq.logging import metrics +from fairseq.criterions import register_criterion + +from .label_smoothed_cross_entropy import ( + LabelSmoothedCrossEntropyCriterion, + LabelSmoothedCrossEntropyCriterionConfig, +) + +from dataclasses import dataclass, field + + +@dataclass +class LabelSmoothedCrossEntropyCriterionWithAlignmentConfig( + LabelSmoothedCrossEntropyCriterionConfig +): + alignment_lambda: float = field( + default=0.05, metadata={"help": "weight for the alignment loss"} + ) + + +@register_criterion( + "label_smoothed_cross_entropy_with_alignment", + dataclass=LabelSmoothedCrossEntropyCriterionWithAlignmentConfig, +) +class LabelSmoothedCrossEntropyCriterionWithAlignment( + LabelSmoothedCrossEntropyCriterion +): + def __init__(self, task, sentence_avg, label_smoothing, alignment_lambda): + super().__init__(task, sentence_avg, label_smoothing) + self.alignment_lambda = alignment_lambda + + def forward(self, model, sample, reduce=True): + """Compute the loss for the given sample. + + Returns a tuple with three elements: + 1) the loss + 2) the sample size, which is used as the denominator for the gradient + 3) logging outputs to display while training + """ + net_output = model(**sample["net_input"]) + loss, nll_loss = self.compute_loss(model, net_output, sample, reduce=reduce) + sample_size = ( + sample["target"].size(0) if self.sentence_avg else sample["ntokens"] + ) + logging_output = { + "loss": utils.item(loss.data) if reduce else loss.data, + "nll_loss": utils.item(nll_loss.data) if reduce else nll_loss.data, + "ntokens": sample["ntokens"], + "nsentences": sample["target"].size(0), + "sample_size": sample_size, + } + + alignment_loss = None + + # Compute alignment loss only for training set and non dummy batches. + if "alignments" in sample and sample["alignments"] is not None: + alignment_loss = self.compute_alignment_loss(sample, net_output) + + if alignment_loss is not None: + logging_output["alignment_loss"] = utils.item(alignment_loss.data) + loss += self.alignment_lambda * alignment_loss + + return loss, sample_size, logging_output + + def compute_alignment_loss(self, sample, net_output): + attn_prob = net_output[1]["attn"][0] + bsz, tgt_sz, src_sz = attn_prob.shape + attn = attn_prob.view(bsz * tgt_sz, src_sz) + + align = sample["alignments"] + align_weights = sample["align_weights"].float() + + if len(align) > 0: + # Alignment loss computation. align (shape [:, 2]) contains the src-tgt index pairs corresponding to + # the alignments. align_weights (shape [:]) contains the 1 / frequency of a tgt index for normalizing. + loss = -( + (attn[align[:, 1][:, None], align[:, 0][:, None]]).log() + * align_weights[:, None] + ).sum() + else: + return None + + return loss + + @staticmethod + def reduce_metrics(logging_outputs) -> None: + """Aggregate logging outputs from data parallel training.""" + loss_sum = utils.item(sum(log.get("loss", 0) for log in logging_outputs)) + nll_loss_sum = utils.item( + sum(log.get("nll_loss", 0) for log in logging_outputs) + ) + alignment_loss_sum = utils.item( + sum(log.get("alignment_loss", 0) for log in logging_outputs) + ) + ntokens = utils.item(sum(log.get("ntokens", 0) for log in logging_outputs)) + sample_size = utils.item( + sum(log.get("sample_size", 0) for log in logging_outputs) + ) + + metrics.log_scalar( + "loss", loss_sum / sample_size / math.log(2), sample_size, round=3 + ) + metrics.log_scalar( + "nll_loss", nll_loss_sum / ntokens / math.log(2), ntokens, round=3 + ) + metrics.log_scalar( + "alignment_loss", + alignment_loss_sum / sample_size / math.log(2), + sample_size, + round=3, + ) + metrics.log_derived( + "ppl", lambda meters: utils.get_perplexity(meters["nll_loss"].avg) + ) + + @staticmethod + def logging_outputs_can_be_summed() -> bool: + """ + Whether the logging outputs returned by `forward` can be summed + across workers prior to calling `reduce_metrics`. Setting this + to True will improves distributed training speed. + """ + return True diff --git a/data/fairseq/fairseq/criterions/label_smoothed_cross_entropy_with_ctc.py b/data/fairseq/fairseq/criterions/label_smoothed_cross_entropy_with_ctc.py new file mode 100644 index 0000000000000000000000000000000000000000..f2e8cdf3bfe0caea99125c6f9607dff9495891cf --- /dev/null +++ b/data/fairseq/fairseq/criterions/label_smoothed_cross_entropy_with_ctc.py @@ -0,0 +1,97 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import math +from dataclasses import dataclass, field + +import torch +import torch.nn.functional as F + +from fairseq import utils +from fairseq.logging import metrics +from fairseq.criterions import register_criterion +from fairseq.criterions.label_smoothed_cross_entropy import ( + LabelSmoothedCrossEntropyCriterion, + LabelSmoothedCrossEntropyCriterionConfig, +) +from fairseq.data.data_utils import lengths_to_mask + + +@dataclass +class LabelSmoothedCrossEntropyWithCtcCriterionConfig( + LabelSmoothedCrossEntropyCriterionConfig +): + ctc_weight: float = field(default=1.0, metadata={"help": "weight for CTC loss"}) + + +@register_criterion( + "label_smoothed_cross_entropy_with_ctc", + dataclass=LabelSmoothedCrossEntropyWithCtcCriterionConfig, +) +class LabelSmoothedCrossEntropyWithCtcCriterion(LabelSmoothedCrossEntropyCriterion): + def __init__( + self, + task, + sentence_avg, + label_smoothing, + ignore_prefix_size, + report_accuracy, + ctc_weight, + ): + super().__init__( + task, sentence_avg, label_smoothing, ignore_prefix_size, report_accuracy + ) + self.ctc_weight = ctc_weight + + def forward(self, model, sample, reduce=True): + net_output = model(**sample["net_input"]) + loss, nll_loss = self.compute_loss(model, net_output, sample, reduce=reduce) + + ctc_loss = torch.tensor(0.0).type_as(loss) + if self.ctc_weight > 0.0: + ctc_lprobs, ctc_lens = model.get_ctc_output(net_output, sample) + ctc_tgt, ctc_tgt_lens = model.get_ctc_target(sample) + ctc_tgt_mask = lengths_to_mask(ctc_tgt_lens) + ctc_tgt_flat = ctc_tgt.masked_select(ctc_tgt_mask) + reduction = "sum" if reduce else "none" + ctc_loss = ( + F.ctc_loss( + ctc_lprobs, + ctc_tgt_flat, + ctc_lens, + ctc_tgt_lens, + reduction=reduction, + zero_infinity=True, + ) + * self.ctc_weight + ) + loss += ctc_loss + + sample_size = ( + sample["target"].size(0) if self.sentence_avg else sample["ntokens"] + ) + logging_output = { + "loss": utils.item(loss.data), + "nll_loss": utils.item(nll_loss.data), + "ctc_loss": utils.item(ctc_loss.data), + "ntokens": sample["ntokens"], + "nsentences": sample["target"].size(0), + "sample_size": sample_size, + } + if self.report_accuracy: + n_correct, total = self.compute_accuracy(model, net_output, sample) + logging_output["n_correct"] = utils.item(n_correct.data) + logging_output["total"] = utils.item(total.data) + return loss, sample_size, logging_output + + @classmethod + def reduce_metrics(cls, logging_outputs) -> None: + super().reduce_metrics(logging_outputs) + loss_sum = sum(log.get("ctc_loss", 0) for log in logging_outputs) + sample_size = sum(log.get("sample_size", 0) for log in logging_outputs) + + metrics.log_scalar( + "ctc_loss", loss_sum / sample_size / math.log(2), sample_size, round=3 + ) diff --git a/data/fairseq/fairseq/criterions/label_smoothed_cross_entropy_with_rdrop.py b/data/fairseq/fairseq/criterions/label_smoothed_cross_entropy_with_rdrop.py new file mode 100644 index 0000000000000000000000000000000000000000..47ee263a8de63261e4c8838ba44fe269553f5f3b --- /dev/null +++ b/data/fairseq/fairseq/criterions/label_smoothed_cross_entropy_with_rdrop.py @@ -0,0 +1,177 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import math +from dataclasses import dataclass, field + +import torch + +from fairseq import utils +from fairseq.logging import metrics +from fairseq.criterions import register_criterion +from fairseq.criterions.label_smoothed_cross_entropy import ( + LabelSmoothedCrossEntropyCriterion, + LabelSmoothedCrossEntropyCriterionConfig, + label_smoothed_nll_loss, +) + + +@dataclass +class RdropLabelSmoothedCrossEntropyCriterionConfig( + LabelSmoothedCrossEntropyCriterionConfig +): + rdrop_alpha: float = field( + default=0.0, + metadata={"help": "alpha for r-drop, 0 means no r-drop"}, + ) + + +@register_criterion( + "label_smoothed_cross_entropy_with_rdrop", + dataclass=RdropLabelSmoothedCrossEntropyCriterionConfig, +) +class RdropLabelSmoothedCrossEntropyCriterion(LabelSmoothedCrossEntropyCriterion): + def __init__( + self, + task, + sentence_avg, + label_smoothing, + ignore_prefix_size=0, + report_accuracy=False, + rdrop_alpha=0.0, + ): + super().__init__( + task, + sentence_avg, + label_smoothing, + ignore_prefix_size=ignore_prefix_size, + report_accuracy=report_accuracy, + ) + self.sentence_avg = sentence_avg + self.eps = label_smoothing + self.ignore_prefix_size = ignore_prefix_size + self.report_accuracy = report_accuracy + self.rdrop_alpha = rdrop_alpha + + def forward(self, model, sample, reduce=True, net_output=None): + """Compute the loss for the given sample. + + Returns a tuple with three elements: + 1) the loss + 2) the sample size, which is used as the denominator for the gradient + 3) logging outputs to display while training + """ + if net_output is None: + if self.rdrop_alpha > 0 and sample["net_input"]["src_tokens"].size( + 0 + ) == sample["target"].size(0): + sample = duplicate_input(sample) + net_output = model(**sample["net_input"]) + loss, nll_loss, rdrop_kl_loss = self.compute_loss( + model, net_output, sample, reduce=reduce + ) + sample_size = ( + sample["target"].size(0) if self.sentence_avg else sample["ntokens"] + ) + logging_output = { + "loss": loss.data, + "nll_loss": nll_loss.data, + "ntokens": sample["ntokens"], + "nsentences": sample["target"].size(0), + "sample_size": sample_size, + } + if self.report_accuracy: + n_correct, total = self.compute_accuracy(model, net_output, sample) + logging_output["n_correct"] = utils.item(n_correct.data) + logging_output["total"] = utils.item(total.data) + if self.rdrop_alpha > 0: + logging_output["rdrop_kl_loss"] = utils.item(rdrop_kl_loss.data) + return loss, sample_size, logging_output + + def get_lprobs_and_target(self, model, net_output, sample): + lprobs = model.get_normalized_probs(net_output, log_probs=True) + target = model.get_targets(sample, net_output) + if self.rdrop_alpha > 0 or target.size(0) != lprobs.size(0): + target = torch.cat([target, target.clone()], dim=0) + + if self.ignore_prefix_size > 0: + # lprobs: B x T x C + lprobs = lprobs[:, self.ignore_prefix_size :, :].contiguous() + target = target[:, self.ignore_prefix_size :].contiguous() + return lprobs.view(-1, lprobs.size(-1)), target.view(-1) + + def compute_loss(self, model, net_output, sample, reduce=True): + lprobs, target = self.get_lprobs_and_target(model, net_output, sample) + loss, nll_loss = label_smoothed_nll_loss( + lprobs, + target, + self.eps, + ignore_index=self.padding_idx, + reduce=reduce, + ) + + if self.rdrop_alpha > 0: + pad_mask = target[: target.size(0) // 2].unsqueeze(-1).eq(self.padding_idx) + rdrop_kl_loss = compute_kl_loss(model, net_output, pad_mask) + loss += self.rdrop_alpha * rdrop_kl_loss + else: + rdrop_kl_loss = loss.new_zeros(1) + return loss, nll_loss, rdrop_kl_loss + + @classmethod + def reduce_metrics(cls, logging_outputs) -> None: + """Aggregate logging outputs from data parallel training.""" + super().reduce_metrics(logging_outputs) + + sample_size = sum(log.get("sample_size", 0) for log in logging_outputs) + + rdrop_kl_loss = utils.item( + sum(log.get("rdrop_kl_loss", 0) for log in logging_outputs) + / sample_size + / math.log(2) + ) + if rdrop_kl_loss > 0: + metrics.log_scalar("rdrop_kl_loss", rdrop_kl_loss) + + +def duplicate_input(sample): + if "net_input" in sample.keys(): + sample_input = sample["net_input"] + else: + sample_input = sample + + for k, v in sample_input.items(): + if isinstance(v, torch.Tensor): + sample_input[k] = torch.cat([v, v.clone()], dim=0) + if "net_input" in sample.keys(): + sample["net_input"] = sample_input + else: + sample = sample_input + return sample + + +def compute_kl_loss(model, net_output, pad_mask=None, reduce=True): + net_prob = model.get_normalized_probs(net_output, log_probs=True) + net_prob_tec = model.get_normalized_probs(net_output, log_probs=False) + + net_prob = net_prob.view(-1, net_prob.size(-1)) + net_prob_tec = net_prob_tec.view(-1, net_prob_tec.size(-1)) + + p, q = torch.split(net_prob, net_prob.size(0) // 2, dim=0) + p_tec, q_tec = torch.split(net_prob_tec, net_prob_tec.size(0) // 2, dim=0) + + p_loss = torch.nn.functional.kl_div(p, q_tec, reduction="none") + q_loss = torch.nn.functional.kl_div(q, p_tec, reduction="none") + + if pad_mask is not None: + p_loss.masked_fill_(pad_mask, 0.0) + q_loss.masked_fill_(pad_mask, 0.0) + + if reduce: + p_loss = p_loss.sum() + q_loss = q_loss.sum() + + loss = (p_loss + q_loss) / 2 + return loss diff --git a/data/fairseq/fairseq/criterions/legacy_masked_lm.py b/data/fairseq/fairseq/criterions/legacy_masked_lm.py new file mode 100644 index 0000000000000000000000000000000000000000..5cf70df2ab97eef1ec454ddc8ccaf5a86cc3c153 --- /dev/null +++ b/data/fairseq/fairseq/criterions/legacy_masked_lm.py @@ -0,0 +1,178 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import math + +import torch +import torch.nn.functional as F +from fairseq import utils +from fairseq.logging import metrics +from fairseq.criterions import FairseqCriterion, register_criterion + + +def compute_cross_entropy_loss(logits, targets, ignore_index=-100): + """ + Function to compute the cross entropy loss. The default value of + ignore_index is the same as the default value for F.cross_entropy in + pytorch. + """ + assert logits.size(0) == targets.size( + -1 + ), "Logits and Targets tensor shapes don't match up" + + loss = F.nll_loss( + F.log_softmax(logits, -1, dtype=torch.float32), + targets, + reduction="sum", + ignore_index=ignore_index, + ) + return loss + + +@register_criterion("legacy_masked_lm_loss") +class LegacyMaskedLmLoss(FairseqCriterion): + """ + Implementation for the loss used in masked language model (MLM) training. + This optionally also computes the next sentence prediction (NSP) loss and + adds it to the overall loss based on the specified args. There are three + cases to consider: + 1) Generic MLM training without NSP loss. In this case sentence_targets + and sentence_logits are both None. + 2) BERT training without NSP loss. In this case sentence_targets is + not None but sentence_logits is None and we should not be computing + a sentence level loss. + 3) BERT training with NSP loss. In this case both sentence_targets and + sentence_logits are not None and we should be computing a sentence + level loss. The weight of the sentence level loss is specified as + an argument. + """ + + def __init__(self, task, masked_lm_only, nsp_loss_weight): + super().__init__(task) + self.masked_lm_only = masked_lm_only + self.nsp_loss_weight = nsp_loss_weight + + @staticmethod + def add_args(parser): + """Args for MaskedLM Loss""" + # Default for masked_lm_only is False so as to not break BERT training + parser.add_argument( + "--masked-lm-only", + default=False, + action="store_true", + help="compute MLM loss only", + ) + parser.add_argument( + "--nsp-loss-weight", + default=1.0, + type=float, + help="weight for next sentence prediction" " loss (default 1)", + ) + + def forward(self, model, sample, reduce=True): + """Compute the loss for the given sample. + Returns a tuple with three elements: + 1) the loss + 2) the sample size, which is used as the denominator for the gradient + 3) logging outputs to display while training + """ + lm_logits, output_metadata = model(**sample["net_input"]) + + # reshape lm_logits from (N,T,C) to (N*T,C) + lm_logits = lm_logits.view(-1, lm_logits.size(-1)) + lm_targets = sample["lm_target"].view(-1) + lm_loss = compute_cross_entropy_loss(lm_logits, lm_targets, self.padding_idx) + + # compute the number of tokens for which loss is computed. This is used + # to normalize the loss + ntokens = utils.strip_pad(lm_targets, self.padding_idx).numel() + loss = lm_loss / ntokens + nsentences = sample["nsentences"] + # nsentences = 0 + + # Compute sentence loss if masked_lm_only is False + sentence_loss = None + if not self.masked_lm_only: + sentence_logits = output_metadata["sentence_logits"] + sentence_targets = sample["sentence_target"].view(-1) + # This needs to be recomputed due to some differences between + # TokenBlock and BlockPair dataset. This can be resolved with a + # refactor of BERTModel which we will do in the future. + # TODO: Remove this after refactor of BERTModel + nsentences = sentence_targets.size(0) + + # Check for logits being none which can happen when remove_heads + # is set to true in the BERT model. Ideally we should set + # masked_lm_only to true in this case, but that requires some + # refactor in the BERT model. + if sentence_logits is not None: + sentence_loss = compute_cross_entropy_loss( + sentence_logits, sentence_targets + ) + + loss += self.nsp_loss_weight * (sentence_loss / nsentences) + + # NOTE: as we are summing up per token mlm loss and per sentence nsp loss + # we don't need to use sample_size as denominator for the gradient + # here sample_size is just used for logging + sample_size = 1 + logging_output = { + "loss": utils.item(loss.data) if reduce else loss.data, + "lm_loss": utils.item(lm_loss.data) if reduce else lm_loss.data, + # sentence loss is not always computed + "sentence_loss": ( + (utils.item(sentence_loss.data) if reduce else sentence_loss.data) + if sentence_loss is not None + else 0.0 + ), + "ntokens": ntokens, + "nsentences": nsentences, + "sample_size": sample_size, + } + return loss, sample_size, logging_output + + @staticmethod + def reduce_metrics(logging_outputs) -> None: + """Aggregate logging outputs from data parallel training.""" + lm_loss_sum = sum(log.get("lm_loss", 0) for log in logging_outputs) + sentence_loss_sum = sum(log.get("sentence_loss", 0) for log in logging_outputs) + ntokens = sum(log.get("ntokens", 0) for log in logging_outputs) + nsentences = sum(log.get("nsentences", 0) for log in logging_outputs) + sample_size = sum(log.get("sample_size", 0) for log in logging_outputs) + agg_loss = sum(log.get("loss", 0) for log in logging_outputs) + + metrics.log_scalar( + "loss", + agg_loss / sample_size / math.log(2) if sample_size > 0 else 0.0, + sample_size, + round=3, + ) + metrics.log_scalar( + "lm_loss", + lm_loss_sum / ntokens / math.log(2) if ntokens > 0 else 0.0, + ntokens, + round=3, + ) + metrics.log_scalar( + "sentence_loss", + sentence_loss_sum / nsentences / math.log(2) if nsentences > 0 else 0.0, + nsentences, + round=3, + ) + metrics.log_scalar( + "nll_loss", + lm_loss_sum / ntokens / math.log(2) if ntokens > 0 else 0.0, + ntokens, + round=3, + ) + + @staticmethod + def logging_outputs_can_be_summed() -> bool: + """ + Whether the logging outputs returned by `forward` can be summed + across workers prior to calling `reduce_metrics`. Setting this + to True will improves distributed training speed. + """ + return True diff --git a/data/fairseq/fairseq/criterions/masked_lm.py b/data/fairseq/fairseq/criterions/masked_lm.py new file mode 100644 index 0000000000000000000000000000000000000000..09ddd9f3e6d41c4771521bb187ad981223e09e95 --- /dev/null +++ b/data/fairseq/fairseq/criterions/masked_lm.py @@ -0,0 +1,99 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from dataclasses import dataclass +import math +from omegaconf import II + +import torch +from fairseq import modules, utils +from fairseq.logging import metrics +from fairseq.criterions import FairseqCriterion, register_criterion +from fairseq.dataclass import FairseqDataclass + + +@dataclass +class MaskedLmConfig(FairseqDataclass): + tpu: bool = II("common.tpu") + + +@register_criterion("masked_lm", dataclass=MaskedLmConfig) +class MaskedLmLoss(FairseqCriterion): + """ + Implementation for the loss used in masked language model (MLM) training. + """ + + def __init__(self, cfg: MaskedLmConfig, task): + super().__init__(task) + self.tpu = cfg.tpu + + def forward(self, model, sample, reduce=True): + """Compute the loss for the given sample. + + Returns a tuple with three elements: + 1) the loss + 2) the sample size, which is used as the denominator for the gradient + 3) logging outputs to display while training + """ + masked_tokens = sample["target"].ne(self.padding_idx) + sample_size = masked_tokens.int().sum() + + # Rare: when all tokens are masked, project all tokens. + # We use torch.where to avoid device-to-host transfers, + # except on CPU where torch.where is not well supported + # (see github.com/pytorch/pytorch/issues/26247). + if self.tpu: + masked_tokens = None # always project all tokens on TPU + elif masked_tokens.device == torch.device("cpu"): + if not masked_tokens.any(): + masked_tokens = None + else: + masked_tokens = torch.where( + masked_tokens.any(), + masked_tokens, + masked_tokens.new([True]), + ) + + logits = model(**sample["net_input"], masked_tokens=masked_tokens)[0] + targets = model.get_targets(sample, [logits]) + if masked_tokens is not None: + targets = targets[masked_tokens] + + loss = modules.cross_entropy( + logits.view(-1, logits.size(-1)), + targets.view(-1), + reduction="sum", + ignore_index=self.padding_idx, + ) + + logging_output = { + "loss": loss if self.tpu else loss.data, + "ntokens": sample["ntokens"], + "nsentences": sample["nsentences"], + "sample_size": sample_size, + } + return loss, sample_size, logging_output + + @staticmethod + def reduce_metrics(logging_outputs) -> None: + """Aggregate logging outputs from data parallel training.""" + loss_sum = sum(log.get("loss", 0) for log in logging_outputs) + sample_size = sum(log.get("sample_size", 0) for log in logging_outputs) + + metrics.log_scalar( + "loss", loss_sum / sample_size / math.log(2), sample_size, round=3 + ) + metrics.log_derived( + "ppl", lambda meters: utils.get_perplexity(meters["loss"].avg) + ) + + @staticmethod + def logging_outputs_can_be_summed() -> bool: + """ + Whether the logging outputs returned by `forward` can be summed + across workers prior to calling `reduce_metrics`. Setting this + to True will improves distributed training speed. + """ + return True diff --git a/data/fairseq/fairseq/criterions/model_criterion.py b/data/fairseq/fairseq/criterions/model_criterion.py new file mode 100644 index 0000000000000000000000000000000000000000..4c020ddbd2fde9a879d94fe63e745d3e6b9a627e --- /dev/null +++ b/data/fairseq/fairseq/criterions/model_criterion.py @@ -0,0 +1,177 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import logging +from dataclasses import dataclass, field +from typing import Dict, List + +import torch + +from fairseq import utils +from fairseq.logging import metrics +from fairseq.criterions import FairseqCriterion, register_criterion +from fairseq.dataclass import FairseqDataclass +from fairseq.logging.meters import safe_round + + +logger = logging.getLogger(__name__) + + +@dataclass +class ModelCriterionConfig(FairseqDataclass): + loss_weights: Dict[str, float] = field( + default_factory=dict, + metadata={"help": "weights for the loss terms"}, + ) + log_keys: List[str] = field( + default_factory=list, + metadata={"help": "additional output keys to log"}, + ) + can_sum: bool = True + + +@register_criterion("model", dataclass=ModelCriterionConfig) +class ModelCriterion(FairseqCriterion): + """ + This criterion relies on the model to supply losses. + The losses should be a dictionary of name -> scalar returned by + the model either by including it in the net_output dict or by + implementing a get_losses(net_output, sample) method. The final loss is + a scaled sum of all losses according to weights in loss_weights. + If no weights are provided, then all losses are scaled by 1.0. + + The losses will be automatically logged. Additional keys from + net_output dict can be logged via the log_keys parameter. + """ + + def __init__(self, task, loss_weights=None, log_keys=None, can_sum=True): + super().__init__(task) + self.loss_weights = loss_weights + self.log_keys = log_keys + self.can_sum = can_sum + + def forward(self, model, sample, reduce=True): + net_output = model(**sample["net_input"]) + + scaled_losses = {} + + if hasattr(model, "get_losses"): + losses = model.get_losses(net_output, sample) + elif isinstance(net_output, dict) and "losses" in net_output: + losses = net_output["losses"] + else: + raise Exception("Could not retrieve losses") + + for lk, p in losses.items(): + try: + coef = 1.0 if len(self.loss_weights) == 0 else self.loss_weights[lk] + except KeyError: + logger.error( + f"weight for loss {lk} is not in loss_weights ({self.loss_weights})" + ) + raise + if coef != 0 and p is not None: + scaled_losses[lk] = coef * p.float().sum() + + loss = sum(scaled_losses.values()) + + if "sample_size" in net_output: + sample_size = net_output["sample_size"] + else: + sample_size = loss.numel() + + if reduce and loss.numel() > 1: + loss = loss.sum() + + logging_output = { + "loss": loss.data, + "ntokens": sample_size, + "nsentences": sample["id"].numel(), + "sample_size": sample_size, + "_world_size": 1, + } + + for lk in self.log_keys: + if lk in net_output and net_output[lk] is not None: + if not torch.is_tensor(net_output[lk]) or net_output[lk].numel() == 1: + logging_output[lk] = float(net_output[lk]) + elif lk.startswith("_"): + logging_output[lk] = net_output[lk] + else: + for i, v in enumerate(net_output[lk]): + logging_output[f"{lk}_{i}"] = float(v) + + if len(scaled_losses) > 1: + for lk, l in scaled_losses.items(): + if l.numel() > 1: + l = l.sum() + logging_output[f"loss_{lk}"] = l.item() + + if "logs" in net_output: + for lgw in net_output["logs"]: + logging_output[lgw] = net_output["logs"][lgw] + + return loss, sample_size, logging_output + + @staticmethod + def reduce_metrics(logging_outputs) -> None: + """Aggregate logging outputs from data parallel training.""" + loss_sum = utils.item(sum(log.get("loss", 0) for log in logging_outputs)) + ntokens = utils.item(sum(log.get("ntokens", 0) for log in logging_outputs)) + nsentences = utils.item( + sum(log.get("nsentences", 0) for log in logging_outputs) + ) + sample_size = utils.item( + sum(log.get("sample_size", 0) for log in logging_outputs) + ) + + metrics.log_scalar("loss", loss_sum / sample_size, sample_size, round=3) + metrics.log_scalar("ntokens", ntokens) + metrics.log_scalar("nsentences", nsentences) + metrics.log_scalar("sample_size", sample_size) + + builtin_keys = { + "loss", + "ntokens", + "nsentences", + "sample_size", + "_world_size", + } + + world_size = utils.item( + sum(log.get("_world_size", 0) for log in logging_outputs) + ) + + for k in logging_outputs[0]: + if k not in builtin_keys and not k.startswith("_"): + val = sum(log.get(k, 0) for log in logging_outputs) + if k.startswith("loss_"): + metrics.log_scalar(k, val / sample_size, sample_size, round=3) + else: + metrics.log_scalar(k, val / world_size, round=3) + + correct = sum(log.get("correct", 0) for log in logging_outputs) + total = sum(log.get("count", 0) for log in logging_outputs) + + if total > 0: + metrics.log_scalar("_correct", correct) + metrics.log_scalar("_total", total) + + metrics.log_derived( + "accuracy", + lambda meters: safe_round( + meters["_correct"].sum / meters["_total"].sum, 5 + ) + if meters["_total"].sum > 0 + else float("nan"), + ) + + def logging_outputs_can_be_summed(self) -> bool: + """ + Whether the logging outputs returned by `forward` can be summed + across workers prior to calling `reduce_metrics`. Setting this + to True will improves distributed training speed. + """ + return self.can_sum diff --git a/data/fairseq/fairseq/criterions/nat_loss.py b/data/fairseq/fairseq/criterions/nat_loss.py new file mode 100644 index 0000000000000000000000000000000000000000..fc0bdaf8510d8feae20779aad49b53a4d84d37db --- /dev/null +++ b/data/fairseq/fairseq/criterions/nat_loss.py @@ -0,0 +1,181 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import math + +import torch +import torch.nn.functional as F +from fairseq import utils +from fairseq.logging import metrics +from fairseq.criterions import FairseqCriterion, register_criterion +from fairseq.dataclass import FairseqDataclass +from torch import Tensor + +from dataclasses import dataclass, field + + +@dataclass +class LabelSmoothedDualImitationCriterionConfig(FairseqDataclass): + label_smoothing: float = field( + default=0.0, + metadata={"help": "epsilon for label smoothing, 0 means no label smoothing"}, + ) + + +@register_criterion("nat_loss", dataclass=LabelSmoothedDualImitationCriterionConfig) +class LabelSmoothedDualImitationCriterion(FairseqCriterion): + def __init__(self, task, label_smoothing): + super().__init__(task) + self.label_smoothing = label_smoothing + + def _compute_loss( + self, outputs, targets, masks=None, label_smoothing=0.0, name="loss", factor=1.0 + ): + """ + outputs: batch x len x d_model + targets: batch x len + masks: batch x len + + policy_logprob: if there is some policy + depends on the likelihood score as rewards. + """ + + def mean_ds(x: Tensor, dim=None) -> Tensor: + return ( + x.float().mean().type_as(x) + if dim is None + else x.float().mean(dim).type_as(x) + ) + + if masks is not None: + outputs, targets = outputs[masks], targets[masks] + + if masks is not None and not masks.any(): + nll_loss = torch.tensor(0) + loss = nll_loss + else: + logits = F.log_softmax(outputs, dim=-1) + if targets.dim() == 1: + losses = F.nll_loss(logits, targets.to(logits.device), reduction="none") + + else: # soft-labels + losses = F.kl_div(logits, targets.to(logits.device), reduction="none") + losses = losses.sum(-1) + + nll_loss = mean_ds(losses) + if label_smoothing > 0: + loss = ( + nll_loss * (1 - label_smoothing) - mean_ds(logits) * label_smoothing + ) + else: + loss = nll_loss + + loss = loss * factor + return {"name": name, "loss": loss, "nll_loss": nll_loss, "factor": factor} + + def _custom_loss(self, loss, name="loss", factor=1.0): + return {"name": name, "loss": loss, "factor": factor} + + def forward(self, model, sample, reduce=True): + """Compute the loss for the given sample. + Returns a tuple with three elements: + 1) the loss + 2) the sample size, which is used as the denominator for the gradient + 3) logging outputs to display while training + """ + nsentences, ntokens = sample["nsentences"], sample["ntokens"] + + # B x T + src_tokens, src_lengths = ( + sample["net_input"]["src_tokens"], + sample["net_input"]["src_lengths"], + ) + tgt_tokens, prev_output_tokens = sample["target"], sample["prev_target"] + + outputs = model(src_tokens, src_lengths, prev_output_tokens, tgt_tokens) + losses, nll_loss = [], [] + + for obj in outputs: + if outputs[obj].get("loss", None) is None: + _losses = self._compute_loss( + outputs[obj].get("out"), + outputs[obj].get("tgt"), + outputs[obj].get("mask", None), + outputs[obj].get("ls", 0.0), + name=obj + "-loss", + factor=outputs[obj].get("factor", 1.0), + ) + else: + _losses = self._custom_loss( + outputs[obj].get("loss"), + name=obj + "-loss", + factor=outputs[obj].get("factor", 1.0), + ) + + losses += [_losses] + if outputs[obj].get("nll_loss", False): + nll_loss += [_losses.get("nll_loss", 0.0)] + + loss = sum(l["loss"] for l in losses) + nll_loss = sum(l for l in nll_loss) if len(nll_loss) > 0 else loss.new_tensor(0) + + # NOTE: + # we don't need to use sample_size as denominator for the gradient + # here sample_size is just used for logging + sample_size = 1 + logging_output = { + "loss": loss.data, + "nll_loss": nll_loss.data, + "ntokens": ntokens, + "nsentences": nsentences, + "sample_size": sample_size, + } + + for l in losses: + logging_output[l["name"]] = ( + utils.item(l["loss"].data / l["factor"]) + if reduce + else l[["loss"]].data / l["factor"] + ) + + return loss, sample_size, logging_output + + @staticmethod + def reduce_metrics(logging_outputs) -> None: + """Aggregate logging outputs from data parallel training.""" + sample_size = utils.item( + sum(log.get("sample_size", 0) for log in logging_outputs) + ) + loss = utils.item(sum(log.get("loss", 0) for log in logging_outputs)) + nll_loss = utils.item(sum(log.get("nll_loss", 0) for log in logging_outputs)) + + metrics.log_scalar( + "loss", loss / sample_size / math.log(2), sample_size, round=3 + ) + metrics.log_scalar( + "nll_loss", nll_loss / sample_size / math.log(2), sample_size, round=3 + ) + metrics.log_derived( + "ppl", lambda meters: utils.get_perplexity(meters["loss"].avg) + ) + + for key in logging_outputs[0]: + if key[-5:] == "-loss": + val = sum(log.get(key, 0) for log in logging_outputs) + metrics.log_scalar( + key[:-5], + val / sample_size / math.log(2) if sample_size > 0 else 0.0, + sample_size, + round=3, + ) + + @staticmethod + def logging_outputs_can_be_summed() -> bool: + """ + Whether the logging outputs returned by `forward` can be summed + across workers prior to calling `reduce_metrics`. Setting this + to True will improves distributed training speed. + """ + return True diff --git a/data/fairseq/fairseq/criterions/sentence_prediction.py b/data/fairseq/fairseq/criterions/sentence_prediction.py new file mode 100644 index 0000000000000000000000000000000000000000..298b80576814d785e8722ae4726698a0a928a20f --- /dev/null +++ b/data/fairseq/fairseq/criterions/sentence_prediction.py @@ -0,0 +1,288 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import math +from dataclasses import dataclass, field +from itertools import chain + +import numpy as np +import torch +import torch.nn.functional as F +from sklearn.metrics import f1_score +from sklearn.metrics import matthews_corrcoef as _matthews_corrcoef +from scipy.stats import pearsonr, spearmanr + +from fairseq.logging import metrics +from fairseq.criterions import FairseqCriterion, register_criterion +from fairseq.dataclass import FairseqDataclass +from fairseq.logging.meters import safe_round + + +def simple_accuracy(preds, labels): + return (preds == labels).mean() + + +def acc_and_f1(preds, labels): + acc = simple_accuracy(preds, labels) + f1 = f1_score(y_true=labels, y_pred=preds) + return { + "acc": acc, + "f1": f1, + "acc_and_f1": (acc + f1) / 2, + } + + +def pearson_and_spearman(preds, labels): + pearson_corr = pearsonr(preds, labels)[0] + spearman_corr = spearmanr(preds, labels)[0] + return { + "pearson": pearson_corr, + "spearmanr": spearman_corr, + "corr": (pearson_corr + spearman_corr) / 2, + } + + +def matthews_corrcoef(preds, labels): + # make it consistent with other metrics taking (preds, labels) as input + mcc = _matthews_corrcoef(labels, preds) + return mcc + + +@dataclass +class SentencePredictionConfig(FairseqDataclass): + classification_head_name: str = field( + default="sentence_classification_head", + metadata={"help": "name of the classification head to use"}, + ) + regression_target: bool = field( + default=False, + ) + report_mcc: bool = False + report_acc_and_f1: bool = False + report_pearson_and_spearman: bool = False + + +@register_criterion("sentence_prediction", dataclass=SentencePredictionConfig) +class SentencePredictionCriterion(FairseqCriterion): + def __init__(self, cfg: SentencePredictionConfig, task): + super().__init__(task) + self.classification_head_name = cfg.classification_head_name + self.regression_target = cfg.regression_target + self.keep_pred_and_targ = ( + cfg.report_mcc or cfg.report_acc_and_f1 or cfg.report_pearson_and_spearman + ) + self.report_mcc = cfg.report_mcc + self.report_acc_and_f1 = cfg.report_acc_and_f1 + self.report_pearson_and_spearman = cfg.report_pearson_and_spearman + self.label_dict = task.label_dictionary + + def forward(self, model, sample, reduce=True): + """Compute the loss for the given sample. + + Returns a tuple with three elements: + 1) the loss + 2) the sample size, which is used as the denominator for the gradient + 3) logging outputs to display while training + """ + assert ( + hasattr(model, "classification_heads") + and self.classification_head_name in model.classification_heads + ), "model must provide sentence classification head for --criterion=sentence_prediction" + + logits, _ = model( + **sample["net_input"], + features_only=True, + classification_head_name=self.classification_head_name, + ) + targets = model.get_targets(sample, [logits]).view(-1) + sample_size = targets.numel() + + if not self.regression_target: + lprobs = F.log_softmax(logits, dim=-1, dtype=torch.float32) + task_loss = F.nll_loss(lprobs, targets, reduction="sum") + else: + logits = logits.view(-1).float() + targets = targets.float() + task_loss = F.mse_loss(logits, targets, reduction="sum") + + logging_output = {} + loss = task_loss + # mha & ffn regularization update + if ( + hasattr(model, "args") + and hasattr(model.args, "mha_reg_scale_factor") + and model.args.mha_reg_scale_factor != 0.0 + ): + mha_reg_loss = model._get_adaptive_head_loss() + loss += mha_reg_loss + logging_output.update({"mha_reg_loss": mha_reg_loss}) + if ( + hasattr(model, "args") + and hasattr(model.args, "ffn_reg_scale_factor") + and model.args.ffn_reg_scale_factor != 0.0 + ): + ffn_reg_loss = model._get_adaptive_ffn_loss() + loss += ffn_reg_loss + logging_output.update({"ffn_reg_loss": ffn_reg_loss}) + + logging_output.update( + { + "loss": loss.data, + "ntokens": sample["ntokens"], + "nsentences": sample_size, + "sample_size": sample_size, + } + ) + if not self.regression_target: + preds = logits.argmax(dim=1) + logging_output["ncorrect"] = (preds == targets).sum() + if self.keep_pred_and_targ and not model.training: + if self.regression_target: + logging_output["pred"] = logits.detach().cpu().tolist() + logging_output["targ"] = targets.detach().cpu().tolist() + else: + # remove offset `self.label_dict.nspecial` from OffsetTokensDataset + preds = self.label_dict.string(preds + self.label_dict.nspecial).split() + targets = self.label_dict.string( + targets + self.label_dict.nspecial + ).split() + logging_output["pred"] = list(map(int, preds)) + logging_output["targ"] = list(map(int, targets)) + + if self.report_mcc: + logging_output["report_mcc"] = True + if self.report_acc_and_f1: + logging_output["report_acc_and_f1"] = True + if self.report_pearson_and_spearman: + logging_output["report_pearson_and_spearman"] = True + + return loss, sample_size, logging_output + + @staticmethod + def reduce_metrics(logging_outputs) -> None: + """Aggregate logging outputs from data parallel training.""" + loss_sum = sum(log.get("loss", 0) for log in logging_outputs) + ntokens = sum(log.get("ntokens", 0) for log in logging_outputs) + nsentences = sum(log.get("nsentences", 0) for log in logging_outputs) + sample_size = sum(log.get("sample_size", 0) for log in logging_outputs) + mha_reg_loss_sum = sum(log.get("mha_reg_loss", 0) for log in logging_outputs) + ffn_reg_loss_sum = sum(log.get("ffn_reg_loss", 0) for log in logging_outputs) + + metrics.log_scalar( + "loss", loss_sum / sample_size / math.log(2), sample_size, round=3 + ) + if mha_reg_loss_sum: + metrics.log_scalar( + "mha_reg_loss", + mha_reg_loss_sum / sample_size / math.log(2), + sample_size, + round=3, + ) + if ffn_reg_loss_sum: + metrics.log_scalar( + "ffn_reg_loss", + ffn_reg_loss_sum / sample_size / math.log(2), + sample_size, + round=3, + ) + if sample_size != ntokens: + metrics.log_scalar( + "nll_loss", loss_sum / ntokens / math.log(2), ntokens, round=3 + ) + + if len(logging_outputs) > 0 and "ncorrect" in logging_outputs[0]: + ncorrect = sum(log.get("ncorrect", 0) for log in logging_outputs) + metrics.log_scalar( + "accuracy", 100.0 * ncorrect / nsentences, nsentences, round=1 + ) + + # Metrics used by GLUE + pred = np.array( + list(chain.from_iterable(log.get("pred", []) for log in logging_outputs)) + ) + targ = np.array( + list(chain.from_iterable(log.get("targ", []) for log in logging_outputs)) + ) + if len(pred): + metrics.log_concat_tensor("pred", torch.from_numpy(pred), dim=0) + metrics.log_concat_tensor("targ", torch.from_numpy(targ), dim=0) + if any("report_mcc" in log for log in logging_outputs): + metrics.log_derived( + "mcc", + lambda meters: safe_round( + matthews_corrcoef( + meters["pred"].tensor.numpy(), + meters["targ"].tensor.numpy(), + ) + * 100, + 1, + ), + ) + if any("report_acc_and_f1" in log for log in logging_outputs): + metrics.log_derived( + "acc_and_f1", + lambda meters: safe_round( + acc_and_f1( + meters["pred"].tensor.numpy(), + meters["targ"].tensor.numpy(), + )["acc_and_f1"] + * 100, + 1, + ), + ) + metrics.log_derived( + "f1", + lambda meters: safe_round( + acc_and_f1( + meters["pred"].tensor.numpy(), + meters["targ"].tensor.numpy(), + )["f1"] + * 100, + 1, + ), + ) + if any("report_pearson_and_spearman" in log for log in logging_outputs): + metrics.log_derived( + "pearson_and_spearman", + lambda meters: safe_round( + pearson_and_spearman( + meters["pred"].tensor.numpy(), + meters["targ"].tensor.numpy(), + )["corr"] + * 100, + 1, + ), + ) + metrics.log_derived( + "pearson", + lambda meters: safe_round( + pearson_and_spearman( + meters["pred"].tensor.numpy(), + meters["targ"].tensor.numpy(), + )["pearson"] + * 100, + 1, + ), + ) + metrics.log_derived( + "spearman", + lambda meters: safe_round( + pearson_and_spearman( + meters["pred"].tensor.numpy(), + meters["targ"].tensor.numpy(), + )["spearmanr"] + * 100, + 1, + ), + ) + + @staticmethod + def logging_outputs_can_be_summed() -> bool: + """ + Whether the logging outputs returned by `forward` can be summed + across workers prior to calling `reduce_metrics`. Setting this + to True will improves distributed training speed. + """ + return True diff --git a/data/fairseq/fairseq/criterions/sentence_prediction_adapters.py b/data/fairseq/fairseq/criterions/sentence_prediction_adapters.py new file mode 100644 index 0000000000000000000000000000000000000000..8a873a45b3b121730c8a27d64facfa2922f8eb88 --- /dev/null +++ b/data/fairseq/fairseq/criterions/sentence_prediction_adapters.py @@ -0,0 +1,63 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import torch +import torch.nn.functional as F +from fairseq.criterions import register_criterion +from fairseq.criterions.sentence_prediction import ( + SentencePredictionCriterion, + SentencePredictionConfig, +) + + +@register_criterion("sentence_prediction_adapters", dataclass=SentencePredictionConfig) +class SentencePredictionCriterionAdapters(SentencePredictionCriterion): + def forward(self, model, sample, reduce=True): + """Compute the loss for the given sample. + + Returns a tuple with three elements: + 1) the loss + 2) the sample size, which is used as the denominator for the gradient + 3) logging outputs to display while training + """ + assert ( + hasattr(model, "classification_heads") + and self.classification_head_name in model.classification_heads + ), "model must provide sentence classification head for --criterion=sentence_prediction" + + if not hasattr(sample, "lang_id"): + # If no language ID is given, we fall back to English + lang_id = ["en_XX"] * sample["nsentences"] + else: + lang_id = sample["lang_id"] + + logits, _ = model( + **sample["net_input"], + features_only=True, + classification_head_name=self.classification_head_name, + lang_id=lang_id, + ) + targets = model.get_targets(sample, [logits]).view(-1) + sample_size = targets.numel() + + if not self.regression_target: + lprobs = F.log_softmax(logits, dim=-1, dtype=torch.float32) + loss = F.nll_loss(lprobs, targets, reduction="sum") + else: + logits = logits.view(-1).float() + targets = targets.float() + loss = F.mse_loss(logits, targets, reduction="sum") + + logging_output = { + "loss": loss.data, + "ntokens": sample["ntokens"], + "nsentences": sample_size, + "sample_size": sample_size, + } + if not self.regression_target: + preds = logits.argmax(dim=1) + logging_output["ncorrect"] = (preds == targets).sum() + + return loss, sample_size, logging_output diff --git a/data/fairseq/fairseq/criterions/sentence_ranking.py b/data/fairseq/fairseq/criterions/sentence_ranking.py new file mode 100644 index 0000000000000000000000000000000000000000..bfb9f058f9208a1eb1218715df0f6f2183085dd9 --- /dev/null +++ b/data/fairseq/fairseq/criterions/sentence_ranking.py @@ -0,0 +1,121 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import math + +import torch +import torch.nn.functional as F +from fairseq import utils +from fairseq.logging import metrics +from fairseq.criterions import FairseqCriterion, register_criterion + + +@register_criterion("sentence_ranking") +class SentenceRankingCriterion(FairseqCriterion): + def __init__(self, task, ranking_head_name, save_predictions, num_classes): + super().__init__(task) + self.ranking_head_name = ranking_head_name + if save_predictions is not None: + self.prediction_h = open(save_predictions, "w") + else: + self.prediction_h = None + self.num_classes = num_classes + + def __del__(self): + if self.prediction_h is not None: + self.prediction_h.close() + + @staticmethod + def add_args(parser): + # fmt: off + parser.add_argument('--save-predictions', metavar='FILE', + help='file to save predictions to') + parser.add_argument('--ranking-head-name', + default='sentence_classification_head', + help='name of the ranking head to use') + # fmt: on + + def forward(self, model, sample, reduce=True): + """Compute ranking loss for the given sample. + + Returns a tuple with three elements: + 1) the loss + 2) the sample size, which is used as the denominator for the gradient + 3) logging outputs to display while training + """ + assert ( + hasattr(model, "classification_heads") + and self.ranking_head_name in model.classification_heads + ), "model must provide sentence ranking head for --criterion=sentence_ranking" + + scores = [] + for idx in range(self.num_classes): + score, _ = model( + **sample["net_input{idx}".format(idx=idx + 1)], + classification_head_name=self.ranking_head_name, + ) + scores.append(score) + + logits = torch.cat(scores, dim=1) + sample_size = logits.size(0) + + if "target" in sample: + targets = model.get_targets(sample, [logits]).view(-1) + lprobs = F.log_softmax(logits, dim=-1, dtype=torch.float32) + loss = F.nll_loss(lprobs, targets, reduction="sum") + else: + targets = None + loss = torch.tensor(0.0, requires_grad=True) + + if self.prediction_h is not None: + preds = logits.argmax(dim=1) + for i, (id, pred) in enumerate(zip(sample["id"].tolist(), preds.tolist())): + if targets is not None: + label = targets[i].item() + print("{}\t{}\t{}".format(id, pred, label), file=self.prediction_h) + else: + print("{}\t{}".format(id, pred), file=self.prediction_h) + + logging_output = { + "loss": loss.data, + "ntokens": sample["ntokens"], + "nsentences": sample_size, + "sample_size": sample_size, + } + if targets is not None: + logging_output["ncorrect"] = (logits.argmax(dim=1) == targets).sum() + + return loss, sample_size, logging_output + + @staticmethod + def reduce_metrics(logging_outputs) -> None: + """Aggregate logging outputs from data parallel training.""" + loss_sum = sum(log.get("loss", 0) for log in logging_outputs) + ntokens = sum(log.get("ntokens", 0) for log in logging_outputs) + nsentences = sum(log.get("nsentences", 0) for log in logging_outputs) + sample_size = sum(log.get("sample_size", 0) for log in logging_outputs) + + metrics.log_scalar( + "loss", loss_sum / sample_size / math.log(2), sample_size, round=3 + ) + if sample_size != ntokens: + metrics.log_scalar( + "nll_loss", loss_sum / ntokens / math.log(2), ntokens, round=3 + ) + + if len(logging_outputs) > 0 and "ncorrect" in logging_outputs[0]: + ncorrect = sum(log.get("ncorrect", 0) for log in logging_outputs) + metrics.log_scalar( + "accuracy", 100.0 * ncorrect / nsentences, nsentences, round=1 + ) + + @staticmethod + def logging_outputs_can_be_summed() -> bool: + """ + Whether the logging outputs returned by `forward` can be summed + across workers prior to calling `reduce_metrics`. Setting this + to True will improves distributed training speed. + """ + return True diff --git a/data/fairseq/fairseq/criterions/speech_dlm_criterion.py b/data/fairseq/fairseq/criterions/speech_dlm_criterion.py new file mode 100644 index 0000000000000000000000000000000000000000..888818011408fd81e5cd3e3c9b074e5082702c79 --- /dev/null +++ b/data/fairseq/fairseq/criterions/speech_dlm_criterion.py @@ -0,0 +1,335 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import math +from dataclasses import dataclass, field +from typing import Optional + +import torch.nn.functional as F +from fairseq import metrics, utils +from fairseq.criterions import FairseqCriterion, register_criterion +from fairseq.dataclass import FairseqDataclass +from omegaconf import II + + +@dataclass +class SpeechDLMCriterionConfig(FairseqDataclass): + sentence_avg: bool = II("optimization.sentence_avg") + main_and_cross_weights: Optional[str] = field( + default="1,0", + metadata={ + "help": "Comma-separated list of weights of Main-channel vs Cross-channel Prediction Losses" + "(default: 1,0)" + }, + ) + general_unit_loss_weight: float = field( + default=0, + metadata={ + "help": "The weight of the General Prediction Loss (Next-step Unit Prediction Loss)" + "(default: 0)" + }, + ) + edge_unit_loss_weight: float = field( + default=1, + metadata={"help": "The weight of the Edge Unit Prediction Loss" "(default: 1)"}, + ) + duration_loss_weight: float = field( + default=1, + metadata={ + "help": "The weight of the Edge Unit Duration Prediction Loss" + "(default: 1)" + }, + ) + + +@register_criterion("speech_dlm_criterion", dataclass=SpeechDLMCriterionConfig) +class SpeechDLMCriterion(FairseqCriterion): + """Criteron for the SpeechDLM model as described in the paper: + https://arxiv.org/pdf/2203.16502.pdf + + There are 3 possible losses depending on the targets of the model: + - general_unit_loss : The next unit prediction loss, corresponding to + 'next' target + - edge_unit_loss : The edge unit prediction loss, corresponding to + 'edge' target + - duration_loss : The duration prediction loss, corresponding to + 'duration' target + """ + + def __init__( + self, + task, + sentence_avg, + main_and_cross_weights, + general_unit_loss_weight, + edge_unit_loss_weight, + duration_loss_weight, + ): + super().__init__(task) + self.sentence_avg = sentence_avg + + self.channels = task.channels + self.targets = task.targets + self.delayed_duration_target = task.delayed_duration_target + + self.main_channel_weight = float(main_and_cross_weights.split(",")[0]) + self.cross_channel_weight = float(main_and_cross_weights.split(",")[1]) + assert self.main_channel_weight >= 0 and self.cross_channel_weight >= 0 + + self.channel_weights = { + channel: weight + for channel, weight in zip(self.channels, task.channel_weights) + } + + self.target_weights = {} + for t in self.targets: + if t == "next": + self.target_weights[t] = general_unit_loss_weight + assert ( + general_unit_loss_weight > 0 + ), "Expect a positive --general-unit-loss-weight for next unit prediction" + elif t == "edge": + self.target_weights[t] = edge_unit_loss_weight + assert ( + edge_unit_loss_weight > 0 + ), "Expect a positive --edge-unit-loss-weight for edge unit prediction" + elif t == "duration": + self.target_weights[t] = duration_loss_weight + assert ( + duration_loss_weight > 0 + ), "Expect a positive --duration-loss-weight for duration prediction" + + def forward(self, model, sample, reduce=True): + """Compute the loss for the given sample. + + Returns a tuple with three elements: + 1) the loss + 2) the sample size, which is used as the denominator for the gradient + 3) logging outputs to display while training + """ + net_output = model(**sample["net_input"]) + loss_dict, stats_dict = self.compute_loss( + model, net_output, sample, reduce=reduce + ) + nsentences = sample["net_input"]["src_tokens"][self.channels[0]].size(0) + + logging_output = { + "nsentences": nsentences, + } + logging_output["nsentences"] = nsentences + + loss_all = {t: 0 for t in self.targets} + correct_all = {t: 0 for t in self.targets} + count_all = {t: 0 for t in self.targets} + ntokens_all = 0 + sample_size_all = 0 + for channel in loss_dict: + for pred_channel in loss_dict[channel]: + # Get ntokens & sample_size + ntokens = sample["net_input"]["src_tokens"][channel].numel() + sample_size = nsentences if self.sentence_avg else ntokens + prefix = "[{}-{}]".format(channel, pred_channel) + log_keys = { + "next": "general_token", + "edge": "edge_token", + "duration": "edge_duration", + } + + # Log & Update the sizes + logging_output["{}ntokens".format(prefix)] = ntokens + logging_output["{}sample_size".format(prefix)] = sample_size + ntokens_all += ntokens + sample_size_all += sample_size + + for t in self.targets: + log_key = log_keys[t] + loss = loss_dict[channel][pred_channel][t] + correct, count = stats_dict[channel][pred_channel][t] + + # Log the statistics + logging_output["{}{}_loss".format(prefix, log_key)] = loss.data + logging_output["{}{}_correct".format(prefix, log_key)] = correct + logging_output["{}{}_count".format(prefix, log_key)] = count + + # Scale the training loss by weights + target_loss = loss * self.channel_weights[channel] + if pred_channel == channel: + target_loss = target_loss * self.main_channel_weight + else: + target_loss = target_loss * self.cross_channel_weight + # Normalize the losses in the training by the number of edges + if t in ["edge", "duration"]: + target_loss = target_loss / count * sample_size + + # Update the statistics + loss_all[t] += target_loss + correct_all[t] += correct + count_all[t] += count + + # Logging the average statistics + logging_output["ntokens"] = ntokens_all + logging_output["sample_size"] = sample_size_all + for t in self.targets: + log_key = { + "next": "general_token", + "edge": "edge_token", + "duration": "edge_duration", + }[t] + logging_output["{}_loss".format(log_key)] = loss_all[t].data + logging_output["{}_correct".format(log_key)] = correct_all[t] + logging_output["{}_count".format(log_key)] = count_all[t] + + # Define the training loss + training_loss = 0 + for t in self.targets: + training_loss += loss_all[t] * self.target_weights[t] + logging_output["loss"] = training_loss.data + + return training_loss, sample_size_all, logging_output + + def compute_loss(self, model, net_output, sample, reduce=True): + # Get the model outputs and target + lprobs_dict = model.get_normalized_probs(net_output, log_probs=True) + target_dict = model.get_targets(sample, net_output) + + # Init the dictionaries + loss_dict, stats_dict = {}, {} + + for channel in lprobs_dict: + # Init the dictionaries + loss_dict[channel], stats_dict[channel] = {}, {} + + for pred_channel in lprobs_dict[channel]: + # Init the dictionaries + loss_dict[channel][pred_channel] = {} + stats_dict[channel][pred_channel] = {} + + # Get token & duration predictions + outputs = lprobs_dict[channel][pred_channel] + if not isinstance(outputs, dict): + token_lprobs = outputs + else: + token_lprobs = outputs["pred_token"] + dur_preds = outputs["pred_duration"] + dur_preds = dur_preds.view(-1) + token_lprobs = token_lprobs.view(-1, token_lprobs.size(-1)) + token_preds = token_lprobs.argmax(dim=-1) + + # Get edge indices + if "edge" in self.targets or "duration" in self.targets: + edge_indices = target_dict["edge_indices"][pred_channel] + + # Compute loss and statistics + for t in self.targets: + if t in ["next", "edge"]: + if t == "next": + target = target_dict["next"][pred_channel].view(-1) + lprobs = token_lprobs + preds = token_preds + elif t == "edge": + target = target_dict["edge"][pred_channel] + lprobs = token_lprobs[edge_indices] + preds = token_preds[edge_indices] + + loss = F.nll_loss( + lprobs, + target, + ignore_index=self.padding_idx, + reduction="sum" if reduce else "none", + ) + elif t == "duration": + target = target_dict["duration"][pred_channel] + if self.delayed_duration_target: + duration_indices = edge_indices + 1 + if duration_indices[-1] == len(dur_preds): + duration_indices = duration_indices[:-1] + target = target[:-1] + else: + duration_indices = edge_indices + preds = dur_preds[duration_indices] + + loss = F.l1_loss( + preds, + target, + reduction="sum" if reduce else "none", + ) + preds = preds.round() + + correct = (preds == target).sum().float().cpu().item() + count = float(target.size(0)) + + loss_dict[channel][pred_channel][t] = loss + stats_dict[channel][pred_channel][t] = (correct, count) + + return loss_dict, stats_dict + + @staticmethod + def reduce_metrics(logging_outputs) -> None: + """Aggregate logging outputs from data parallel training.""" + logging_keys = next(iter(logging_outputs)).keys() + channels = [item[:-7] for item in logging_keys if item.endswith("ntokens")] + target_prefixes = set( + [ + item[:-5].split("]")[-1] + for item in logging_keys + if item.endswith("_loss") + ] + ) + for channel_prefix in channels: + for target_prefix in target_prefixes: + prefix = "{}{}".format(channel_prefix, target_prefix) + count_sum = sum( + log.get("{}_count".format(prefix), 0) for log in logging_outputs + ) + correct_sum = sum( + log.get("{}_correct".format(prefix), 0) for log in logging_outputs + ) + loss_sum = sum( + log.get("{}_loss".format(prefix), 0) for log in logging_outputs + ) + + if "duration" not in target_prefix: + # we divide by log(2) to convert the loss from base e to base 2 + metrics.log_scalar( + "{}_loss".format(prefix), + loss_sum / count_sum / math.log(2), + count_sum, + round=3, + ) + metrics.log_derived( + "{}_ppl".format(prefix), + lambda meters, prefix=prefix: utils.get_perplexity( + meters["{}_loss".format(prefix)].avg + ), + ) + else: + # for duration we don't need to divide by log(2) + metrics.log_scalar( + "{}_loss".format(prefix), + loss_sum / count_sum, + count_sum, + round=3, + ) + + accuracy = 100 * correct_sum / count_sum + metrics.log_scalar("{}_pred_acc".format(prefix), accuracy, round=3) + + # Logging training loss + sample_size = sum(log.get("sample_size", 0) for log in logging_outputs) + loss_sum = sum(log.get("loss", 0) for log in logging_outputs) + + # we divide by log(2) to convert the loss from base e to base 2 + metrics.log_scalar( + "loss", loss_sum / sample_size / math.log(2), sample_size, round=3 + ) + + @staticmethod + def logging_outputs_can_be_summed() -> bool: + """ + Whether the logging outputs returned by `forward` can be summed + across workers prior to calling `reduce_metrics`. Setting this + to True will improves distributed training speed. + """ + return True diff --git a/data/fairseq/fairseq/criterions/speech_to_speech_criterion.py b/data/fairseq/fairseq/criterions/speech_to_speech_criterion.py new file mode 100644 index 0000000000000000000000000000000000000000..06a825214013bf9d7d39d683895d90166efbae3f --- /dev/null +++ b/data/fairseq/fairseq/criterions/speech_to_speech_criterion.py @@ -0,0 +1,517 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import logging +import math +from collections import OrderedDict + +import torch + +from fairseq import utils +from fairseq.logging import metrics +from fairseq.criterions import register_criterion +from fairseq.criterions.ctc import CtcCriterion +from fairseq.criterions.label_smoothed_cross_entropy_with_rdrop import ( + RdropLabelSmoothedCrossEntropyCriterion, + RdropLabelSmoothedCrossEntropyCriterionConfig, + duplicate_input, +) +from fairseq.criterions.tacotron2_loss import ( + Tacotron2Criterion, + Tacotron2CriterionConfig, +) + +logger = logging.getLogger(__name__) + + +class MultitaskCriterion: + def __init__(self, multitask_tasks, rdrop_alpha=0.0): + self.rdrop_alpha = rdrop_alpha + self.rdrop_alpha_mtl = rdrop_alpha + + self.multitask_criterion = OrderedDict() + self.multitask_loss_weight = OrderedDict() + for task_name, task_obj in multitask_tasks.items(): + if task_obj.args.get_loss_weight(0) == 0: + logger.info(f"Skip {task_name} loss criterion") + continue + + rdrop_alpha_task = task_obj.args.rdrop_alpha + if rdrop_alpha_task is None: + rdrop_alpha_task = rdrop_alpha + self.rdrop_alpha_mtl = rdrop_alpha_task + logger.info(f"rdrop_alpha is set to {rdrop_alpha_task} for {task_name}") + + if task_obj.args.decoder_type == "ctc": + self.multitask_criterion[task_name] = CtcCriterion( + task_obj.args.criterion_cfg, + task_obj, + rdrop_alpha=rdrop_alpha_task, + ) + else: + self.multitask_criterion[ + task_name + ] = RdropLabelSmoothedCrossEntropyCriterion( + task_obj, + task_obj.args.criterion_cfg.sentence_avg, + label_smoothing=task_obj.args.criterion_cfg.label_smoothing, + rdrop_alpha=rdrop_alpha_task, + ) + + def set_multitask_loss_weight(self, task_name, weight=0.0): + self.multitask_loss_weight[task_name] = weight + + def get_multitask_loss(self, model, sample, model_out): + logging_output = {} + loss = 0.0 + for task_name, task_criterion in self.multitask_criterion.items(): + layer_id = task_criterion.task.args.input_layer + if isinstance(task_criterion, CtcCriterion): + if task_criterion.task.args.input_from == "encoder": + if len(model_out["encoder_padding_mask"]) > 0: + non_padding_mask = ~model_out["encoder_padding_mask"][0] + input_lengths = non_padding_mask.long().sum(-1) + else: + out = model_out["encoder_states"][layer_id] + input_lengths = out.new_full( + (out.shape[1],), out.shape[0] + ).long() + + task_sample = { + "net_input": { + "src_tokens": model_out["encoder_states"][ + layer_id + ], # check batch idx + "src_lengths": input_lengths, + }, + "id": sample["id"], + } + else: + task_sample = { + "net_input": { + "src_tokens": model_out["inner_states"][layer_id], + "src_lengths": sample["target_lengths"], + }, + "id": sample["id"], + } + else: + task_sample = { + "net_input": { + "src_tokens": sample["multitask"][task_name]["net_input"][ + "prev_output_tokens" + ], + "encoder_out": { + "encoder_out": [model_out["encoder_states"][layer_id]], + "encoder_padding_mask": model_out["encoder_padding_mask"], + }, + } + } + + for key in ["target", "target_lengths", "ntokens"]: + task_sample[key] = sample["multitask"][task_name][key] + + if task_name == getattr(model, "mt_task_name", None): + decoder_out = model_out["mt_decoder_out"] + else: + decoder_out = None + task_loss, task_sample_size, task_logging_output = task_criterion( + model.multitask_decoders[task_name], task_sample, net_output=decoder_out + ) + + loss = loss + self.multitask_loss_weight[task_name] * task_loss + task_logging_output["loss_weight"] = self.multitask_loss_weight[task_name] + logging_output[task_name] = task_logging_output + return loss, logging_output + + @classmethod + def reduce_metrics(cls, logging_outputs) -> None: + for task_name in logging_outputs[0]["multitask"].keys(): + # different criterion may return different logging + # currently only reduce on loss, the most common one + # ideally the way that losses are reduced should also depend on the task type + loss_sum = sum( + log["multitask"][task_name].get("loss", 0) for log in logging_outputs + ) + sample_size = sum( + log["multitask"][task_name].get("sample_size", 0) + for log in logging_outputs + ) + + metrics.log_scalar( + f"multitask_{task_name}_loss", + loss_sum / sample_size / math.log(2), + sample_size, + round=3, + ) + + loss_weight = logging_outputs[0]["multitask"][task_name].get( + "loss_weight", 0 + ) + metrics.log_scalar( + f"multitask_{task_name}_loss_weight", + loss_weight, + weight=0, + priority=250, + ) + + +@register_criterion( + "speech_to_unit", dataclass=RdropLabelSmoothedCrossEntropyCriterionConfig +) +class SpeechToUnitMultitaskTaskCriterion( + RdropLabelSmoothedCrossEntropyCriterion, MultitaskCriterion +): + def __init__( + self, + task, + sentence_avg, + label_smoothing, + ignore_prefix_size=0, + report_accuracy=False, + rdrop_alpha=0.0, + ): + super().__init__( + task, + sentence_avg, + label_smoothing, + ignore_prefix_size, + report_accuracy, + rdrop_alpha, + ) + MultitaskCriterion.__init__(self, task.multitask_tasks, rdrop_alpha) + + def forward(self, model, sample, reduce=True): + net_input_concat = { + "src_tokens": sample["net_input"]["src_tokens"], + "src_lengths": sample["net_input"]["src_lengths"], + "prev_output_tokens": sample["net_input"]["prev_output_tokens"], + "tgt_speaker": sample["net_input"].get("tgt_speaker", None), + "return_all_hiddens": True, + } + + if self.rdrop_alpha > 0 or self.rdrop_alpha_mtl > 0: + net_input_concat = duplicate_input(net_input_concat) + + net_output, extra = model(**net_input_concat) + loss, nll_loss, rdrop_kl_loss = self.compute_loss( + model, [net_output], sample, reduce=reduce + ) + sample_size = ( + sample["target"].size(0) if self.sentence_avg else sample["ntokens"] + ) + logging_output = { + "loss": loss.data, + "nll_loss": nll_loss.data, + "ntokens": sample["ntokens"], + "nsentences": sample["target"].size(0), + "sample_size": sample_size, + } + if self.report_accuracy: + n_correct, total = self.compute_accuracy(model, [net_output], sample) + logging_output["n_correct"] = utils.item(n_correct.data) + logging_output["total"] = utils.item(total.data) + if self.rdrop_alpha > 0: + logging_output["rdrop_kl_loss"] = utils.item(rdrop_kl_loss.data) + + if len(self.multitask_criterion) == 0: + return loss, sample_size, logging_output + + # multitask + multitask_loss, multitask_log = self.get_multitask_loss(model, sample, extra) + loss += multitask_loss + logging_output["multitask"] = multitask_log + + return loss, sample_size, logging_output + + @classmethod + def reduce_metrics(cls, logging_outputs) -> None: + super().reduce_metrics(logging_outputs) + + # inference metrics + if "targ_frames" in logging_outputs[0]: + n = sum(log.get("norm_frames", 0) for log in logging_outputs) + for key, new_key in [ + ("mcd_loss", "mcd_loss"), + ("pred_frames", "pred_ratio"), + ("nins", "ins_rate"), + ("ndel", "del_rate"), + ]: + val = sum(log.get(key, 0) for log in logging_outputs) + metrics.log_scalar(new_key, val / n, n, round=3) + + if "multitask" not in logging_outputs[0]: + return + + MultitaskCriterion.reduce_metrics(logging_outputs) + + @staticmethod + def logging_outputs_can_be_summed() -> bool: + """ + Whether the logging outputs returned by `forward` can be summed + across workers prior to calling `reduce_metrics`. Setting this + to True will improves distributed training speed. + """ + return False + + +@register_criterion( + "speech_to_unit_2pass", dataclass=RdropLabelSmoothedCrossEntropyCriterionConfig +) +class SpeechToUnit2passMultitaskTaskCriterion(SpeechToUnitMultitaskTaskCriterion): + def __init__( + self, + task, + sentence_avg, + label_smoothing, + ignore_prefix_size=0, + report_accuracy=False, + rdrop_alpha=0.0, + ): + super().__init__( + task, + sentence_avg, + label_smoothing, + ignore_prefix_size, + report_accuracy, + rdrop_alpha, + ) + + def forward(self, model, sample, reduce=True): + net_input_concat = { + "src_tokens": sample["net_input"]["src_tokens"], + "src_lengths": sample["net_input"]["src_lengths"], + "prev_output_tokens": sample["net_input"]["prev_output_tokens"], + "prev_output_tokens_mt": sample["multitask"][model.mt_task_name][ + "net_input" + ]["prev_output_tokens"], + "tgt_speaker": sample["net_input"].get("tgt_speaker", None), + "return_all_hiddens": True, + } + if getattr(model, "asr_task_name", None) is not None: + net_input_concat["prev_output_tokens_asr"] = sample["multitask"][ + model.asr_task_name + ]["net_input"]["prev_output_tokens"] + + if self.rdrop_alpha > 0 or self.rdrop_alpha_mtl > 0: + net_input_concat = duplicate_input(net_input_concat) + + net_output, extra = model(**net_input_concat) + loss, nll_loss, rdrop_kl_loss = self.compute_loss( + model, [net_output], sample, reduce=reduce + ) + + sample_size = ( + sample["target"].size(0) if self.sentence_avg else sample["ntokens"] + ) + logging_output = { + "loss": loss.data, + "nll_loss": nll_loss.data, + "ntokens": sample["ntokens"], + "nsentences": sample["target"].size(0), + "sample_size": sample_size, + } + if self.report_accuracy: + n_correct, total = self.compute_accuracy(model, [net_output], sample) + logging_output["n_correct"] = utils.item(n_correct.data) + logging_output["total"] = utils.item(total.data) + if self.rdrop_alpha > 0: + logging_output["rdrop_kl_loss"] = utils.item(rdrop_kl_loss.data) + + if len(self.multitask_criterion) == 0: + return loss, sample_size, logging_output + + # multitask + multitask_loss, multitask_log = self.get_multitask_loss(model, sample, extra) + loss += multitask_loss + logging_output["multitask"] = multitask_log + + return loss, sample_size, logging_output + + +@register_criterion("speech_to_spectrogram", dataclass=Tacotron2CriterionConfig) +class SpeechToSpectrogramMultitaskTaskCriterion(Tacotron2Criterion, MultitaskCriterion): + def __init__( + self, + task, + sentence_avg, + use_guided_attention_loss, + guided_attention_loss_sigma, + bce_pos_weight, + ctc_weight, + ): + super().__init__( + task, + sentence_avg, + use_guided_attention_loss, + guided_attention_loss_sigma, + bce_pos_weight, + ctc_weight, + ) + MultitaskCriterion.__init__(self, task.multitask_tasks) + + def forward(self, model, sample, reduction="mean"): + bsz, max_len, _ = sample["target"].size() + feat_tgt = sample["target"] + feat_len = sample["target_lengths"].view(bsz, 1).expand(-1, max_len) + eos_tgt = torch.arange(max_len).to(sample["target"].device) + eos_tgt = eos_tgt.view(1, max_len).expand(bsz, -1) + eos_tgt = (eos_tgt == (feat_len - 1)).float() + + feat_out, eos_out, extra = model( + src_tokens=sample["net_input"]["src_tokens"], + src_lengths=sample["net_input"]["src_lengths"], + prev_output_tokens=sample["net_input"]["prev_output_tokens"], + tgt_speaker=sample["net_input"]["tgt_speaker"], + target_lengths=sample["target_lengths"], + return_all_hiddens=True, + ) + + l1_loss, mse_loss, eos_loss = self.compute_loss( + extra["feature_out"], + feat_out, + eos_out, + feat_tgt, + eos_tgt, + sample["target_lengths"], + reduction, + ) + attn_loss = torch.tensor(0.0).type_as(l1_loss) + if self.guided_attn is not None: + attn_loss = self.guided_attn( + extra["attn"], + sample["net_input"]["src_lengths"], + sample["target_lengths"], + reduction, + ) + loss = ( + l1_loss + mse_loss + eos_loss + attn_loss + ) # do not include ctc loss as there's no text target + + sample_size = sample["nsentences"] if self.sentence_avg else sample["ntokens"] + logging_output = { + "loss": utils.item(loss.data), + "ntokens": sample["ntokens"], + "nsentences": sample["nsentences"], + "sample_size": sample_size, + "l1_loss": utils.item(l1_loss.data), + "mse_loss": utils.item(mse_loss.data), + "eos_loss": utils.item(eos_loss.data), + "attn_loss": utils.item(attn_loss.data), + } + + if len(self.multitask_criterion) == 0: + return loss, sample_size, logging_output + + # multitask + multitask_loss, multitask_log = self.get_multitask_loss(model, sample, extra) + loss += multitask_loss + logging_output["multitask"] = multitask_log + return loss, sample_size, logging_output + + @classmethod + def reduce_metrics(cls, logging_outputs) -> None: + super().reduce_metrics(logging_outputs) + + # inference metrics + if "targ_frames" in logging_outputs[0]: + n = sum(log.get("norm_frames", 0) for log in logging_outputs) + for key, new_key in [ + ("mcd_loss", "mcd_loss"), + ("pred_frames", "pred_ratio"), + ("nins", "ins_rate"), + ("ndel", "del_rate"), + ]: + val = sum(log.get(key, 0) for log in logging_outputs) + metrics.log_scalar(new_key, val / n, n, round=3) + + if "multitask" not in logging_outputs[0]: + return + + MultitaskCriterion.reduce_metrics(logging_outputs) + + +@register_criterion("speech_to_spectrogram_2pass", dataclass=Tacotron2CriterionConfig) +class SpeechToSpectrogram2passMultitaskTaskCriterion( + SpeechToSpectrogramMultitaskTaskCriterion +): + def __init__( + self, + task, + sentence_avg, + use_guided_attention_loss, + guided_attention_loss_sigma, + bce_pos_weight, + ctc_weight, + ): + super().__init__( + task, + sentence_avg, + use_guided_attention_loss, + guided_attention_loss_sigma, + bce_pos_weight, + ctc_weight, + ) + + def forward(self, model, sample, reduction="mean"): + bsz, max_len, _ = sample["target"].size() + feat_tgt = sample["target"] + feat_len = sample["target_lengths"].view(bsz, 1).expand(-1, max_len) + eos_tgt = torch.arange(max_len).to(sample["target"].device) + eos_tgt = eos_tgt.view(1, max_len).expand(bsz, -1) + eos_tgt = (eos_tgt == (feat_len - 1)).float() + + feat_out, eos_out, extra = model( + src_tokens=sample["net_input"]["src_tokens"], + src_lengths=sample["net_input"]["src_lengths"], + prev_output_tokens=sample["net_input"]["prev_output_tokens"], + prev_output_tokens_mt=sample["multitask"][model.mt_task_name]["net_input"][ + "prev_output_tokens" + ], + tgt_speaker=sample["net_input"]["tgt_speaker"], + target_lengths=sample["target_lengths"], + return_all_hiddens=True, + ) + + l1_loss, mse_loss, eos_loss = self.compute_loss( + extra["feature_out"], + feat_out, + eos_out, + feat_tgt, + eos_tgt, + sample["target_lengths"], + reduction, + ) + attn_loss = torch.tensor(0.0).type_as(l1_loss) + if self.guided_attn is not None: + attn_loss = self.guided_attn( + extra["attn"], + sample["net_input"]["src_lengths"], + sample["target_lengths"], + reduction, + ) + loss = ( + l1_loss + mse_loss + eos_loss + attn_loss + ) # do not include ctc loss as there's no text target + + sample_size = sample["nsentences"] if self.sentence_avg else sample["ntokens"] + logging_output = { + "loss": utils.item(loss.data), + "ntokens": sample["ntokens"], + "nsentences": sample["nsentences"], + "sample_size": sample_size, + "l1_loss": utils.item(l1_loss.data), + "mse_loss": utils.item(mse_loss.data), + "eos_loss": utils.item(eos_loss.data), + "attn_loss": utils.item(attn_loss.data), + } + + if len(self.multitask_criterion) == 0: + return loss, sample_size, logging_output + + # multitask + multitask_loss, multitask_log = self.get_multitask_loss(model, sample, extra) + loss += multitask_loss + logging_output["multitask"] = multitask_log + return loss, sample_size, logging_output diff --git a/data/fairseq/fairseq/criterions/speech_ulm_criterion.py b/data/fairseq/fairseq/criterions/speech_ulm_criterion.py new file mode 100644 index 0000000000000000000000000000000000000000..eea74bae2641e285c56e86feb6e9866464c9673f --- /dev/null +++ b/data/fairseq/fairseq/criterions/speech_ulm_criterion.py @@ -0,0 +1,126 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import torch +from dataclasses import dataclass, field + +import torch.nn.functional as F +from fairseq.logging import metrics +from fairseq.tasks import FairseqTask +from fairseq.criterions import FairseqCriterion, register_criterion +from fairseq.dataclass import FairseqDataclass +from omegaconf import II + + +@dataclass +class SpeechUnitLmCriterionConfig(FairseqDataclass): + sentence_avg: bool = II("optimization.sentence_avg") + loss_weights: str = field( + default="1.;0.0;0.0", + metadata={ + "help": "Weights of the losses that correspond to token, duration, and F0 streams" + }, + ) + discrete_duration: bool = II("task.discrete_duration") + discrete_f0: bool = II("task.discrete_f0") + + +def mae_loss(pred, targ, mask, reduce=True): + if pred.ndim == 3: + pred = pred.squeeze(2) + else: + assert pred.ndim == 2 + loss = (pred.float() - targ.float()).abs() * (~mask).float() + loss = loss.sum() if reduce else loss.view(-1) + return loss + + +def nll_loss(pred, targ, mask, reduce=True): + lprob = F.log_softmax(pred, dim=-1) + loss = F.nll_loss(lprob.view(-1, lprob.size(-1)), targ.view(-1), reduction="none") + loss = loss * (~mask).float().view(-1) + loss = loss.sum() if reduce else loss.view(-1) + return loss + + +@register_criterion("speech_unit_lm_criterion", dataclass=SpeechUnitLmCriterionConfig) +class SpeechUnitLmCriterion(FairseqCriterion): + def __init__(self, cfg: SpeechUnitLmCriterionConfig, task: FairseqTask): + super().__init__(task) + self.sentence_avg = cfg.sentence_avg + self.weights = torch.tensor([float(w) for w in cfg.loss_weights.split(";")]) + assert self.weights.size(0) == 3 + assert (self.weights >= 0.0).all() + + self.dur_loss_fn = nll_loss if cfg.discrete_duration else mae_loss + self.f0_loss_fn = nll_loss if cfg.discrete_f0 else mae_loss + + def forward(self, model, sample, reduce=True): + """Compute the loss for the given sample. + + Returns a tuple with three elements: + 1) the loss + 2) the sample size, which is used as the denominator for the gradient + 3) logging outputs to display while training + """ + net_output = model(**sample["net_input"]) + + token_loss = nll_loss( + net_output["token"], sample["target"], sample["mask"], reduce + ) + dur_loss = self.dur_loss_fn( + net_output["duration"], + sample["dur_target"], + sample["dur_mask"], + reduce, + ) + f0_loss = self.f0_loss_fn( + net_output["f0"], + sample["f0_target"], + sample["f0_mask"], + reduce, + ) + loss = self.weights.to(token_loss.device) * torch.stack( + [token_loss, dur_loss, f0_loss], dim=-1 + ) + loss = loss.sum() if reduce else loss.sum(-1) + + sample_size = ( + sample["target"].size(0) if self.sentence_avg else sample["ntokens"] + ) + logging_output = { + "loss": loss.detach().sum().item(), + "token_loss": token_loss.detach().sum().item(), + "dur_loss": dur_loss.detach().sum().item(), + "f0_loss": f0_loss.detach().sum().item(), + "ntokens": sample["ntokens"], + "nsentences": sample["target"].size(0), + "sample_size": sample_size, + } + return loss, sample_size, logging_output + + @staticmethod + def reduce_metrics(logging_outputs) -> None: + """Aggregate logging outputs from data parallel training.""" + loss_sum = sum(log.get("loss", 0) for log in logging_outputs) + token_loss_sum = sum(log.get("token_loss", 0) for log in logging_outputs) + dur_loss_sum = sum(log.get("dur_loss", 0) for log in logging_outputs) + f0_loss_sum = sum(log.get("f0_loss", 0) for log in logging_outputs) + + sample_size = sum(log.get("sample_size", 0) for log in logging_outputs) + + metrics.log_scalar("loss", loss_sum / sample_size, sample_size, round=3) + + metrics.log_scalar( + "token_loss", token_loss_sum / sample_size, sample_size, round=3 + ) + + metrics.log_scalar("dur_loss", dur_loss_sum / sample_size, sample_size, round=3) + + metrics.log_scalar("f0_loss", f0_loss_sum / sample_size, sample_size, round=3) + + @staticmethod + def logging_outputs_can_be_summed() -> bool: + return True diff --git a/data/fairseq/fairseq/criterions/tacotron2_loss.py b/data/fairseq/fairseq/criterions/tacotron2_loss.py new file mode 100644 index 0000000000000000000000000000000000000000..4113fdc5489f1e0c787b60735086ae9d073c8e17 --- /dev/null +++ b/data/fairseq/fairseq/criterions/tacotron2_loss.py @@ -0,0 +1,227 @@ +# Copyright (c) 2017-present, Facebook, Inc. +# All rights reserved. +# +# This source code is licensed under the license found in the LICENSE file in +# the root directory of this source tree. An additional grant of patent rights +# can be found in the PATENTS file in the same directory. + +import logging +from dataclasses import dataclass, field +from functools import lru_cache +from typing import Any, Dict, List + +import torch +import torch.nn.functional as F +from omegaconf import II + +from fairseq import utils +from fairseq.logging import metrics +from fairseq.criterions import FairseqCriterion, register_criterion +from fairseq.data.data_utils import lengths_to_mask +from fairseq.dataclass import FairseqDataclass + +logger = logging.getLogger(__name__) + + +@dataclass +class Tacotron2CriterionConfig(FairseqDataclass): + bce_pos_weight: float = field( + default=1.0, + metadata={"help": "weight of positive examples for BCE loss"}, + ) + use_guided_attention_loss: bool = field( + default=False, + metadata={"help": "use guided attention loss"}, + ) + guided_attention_loss_sigma: float = field( + default=0.4, + metadata={"help": "weight of positive examples for BCE loss"}, + ) + ctc_weight: float = field(default=0.0, metadata={"help": "weight for CTC loss"}) + sentence_avg: bool = II("optimization.sentence_avg") + + +class GuidedAttentionLoss(torch.nn.Module): + """ + Efficiently Trainable Text-to-Speech System Based on Deep Convolutional + Networks with Guided Attention (https://arxiv.org/abs/1710.08969) + """ + + def __init__(self, sigma): + super().__init__() + self.sigma = sigma + + @staticmethod + @lru_cache(maxsize=8) + def _get_weight(s_len, t_len, sigma): + grid_x, grid_y = torch.meshgrid(torch.arange(t_len), torch.arange(s_len)) + grid_x = grid_x.to(s_len.device) + grid_y = grid_y.to(s_len.device) + w = (grid_y.float() / s_len - grid_x.float() / t_len) ** 2 + return 1.0 - torch.exp(-w / (2 * (sigma**2))) + + def _get_weights(self, src_lens, tgt_lens): + bsz, max_s_len, max_t_len = len(src_lens), max(src_lens), max(tgt_lens) + weights = torch.zeros((bsz, max_t_len, max_s_len)) + for i, (s_len, t_len) in enumerate(zip(src_lens, tgt_lens)): + weights[i, :t_len, :s_len] = self._get_weight(s_len, t_len, self.sigma) + return weights + + @staticmethod + def _get_masks(src_lens, tgt_lens): + in_masks = lengths_to_mask(src_lens) + out_masks = lengths_to_mask(tgt_lens) + return out_masks.unsqueeze(2) & in_masks.unsqueeze(1) + + def forward(self, attn, src_lens, tgt_lens, reduction="mean"): + weights = self._get_weights(src_lens, tgt_lens).to(attn.device) + masks = self._get_masks(src_lens, tgt_lens).to(attn.device) + loss = (weights * attn.transpose(1, 2)).masked_select(masks) + loss = torch.sum(loss) if reduction == "sum" else torch.mean(loss) + return loss + + +@register_criterion("tacotron2", dataclass=Tacotron2CriterionConfig) +class Tacotron2Criterion(FairseqCriterion): + def __init__( + self, + task, + sentence_avg, + use_guided_attention_loss, + guided_attention_loss_sigma, + bce_pos_weight, + ctc_weight, + ): + super().__init__(task) + self.sentence_avg = sentence_avg + self.bce_pos_weight = bce_pos_weight + + self.guided_attn = None + if use_guided_attention_loss: + self.guided_attn = GuidedAttentionLoss(guided_attention_loss_sigma) + self.ctc_weight = ctc_weight + + def forward(self, model, sample, reduction="mean"): + bsz, max_len, _ = sample["target"].size() + feat_tgt = sample["target"] + feat_len = sample["target_lengths"].view(bsz, 1).expand(-1, max_len) + eos_tgt = torch.arange(max_len).to(sample["target"].device) + eos_tgt = eos_tgt.view(1, max_len).expand(bsz, -1) + eos_tgt = (eos_tgt == (feat_len - 1)).float() + src_tokens = sample["net_input"]["src_tokens"] + src_lens = sample["net_input"]["src_lengths"] + tgt_lens = sample["target_lengths"] + + feat_out, eos_out, extra = model( + src_tokens=src_tokens, + src_lengths=src_lens, + prev_output_tokens=sample["net_input"]["prev_output_tokens"], + incremental_state=None, + target_lengths=tgt_lens, + speaker=sample["speaker"], + ) + + l1_loss, mse_loss, eos_loss = self.compute_loss( + extra["feature_out"], + feat_out, + eos_out, + feat_tgt, + eos_tgt, + tgt_lens, + reduction, + ) + attn_loss = torch.tensor(0.0).type_as(l1_loss) + if self.guided_attn is not None: + attn_loss = self.guided_attn(extra["attn"], src_lens, tgt_lens, reduction) + ctc_loss = torch.tensor(0.0).type_as(l1_loss) + if self.ctc_weight > 0.0: + net_output = (feat_out, eos_out, extra) + lprobs = model.get_normalized_probs(net_output, log_probs=True) + lprobs = lprobs.transpose(0, 1) # T x B x C + src_mask = lengths_to_mask(src_lens) + src_tokens_flat = src_tokens.masked_select(src_mask) + ctc_loss = ( + F.ctc_loss( + lprobs, + src_tokens_flat, + tgt_lens, + src_lens, + reduction=reduction, + zero_infinity=True, + ) + * self.ctc_weight + ) + loss = l1_loss + mse_loss + eos_loss + attn_loss + ctc_loss + + sample_size = sample["nsentences"] if self.sentence_avg else sample["ntokens"] + logging_output = { + "loss": utils.item(loss.data), + "ntokens": sample["ntokens"], + "nsentences": sample["nsentences"], + "sample_size": sample_size, + "l1_loss": utils.item(l1_loss.data), + "mse_loss": utils.item(mse_loss.data), + "eos_loss": utils.item(eos_loss.data), + "attn_loss": utils.item(attn_loss.data), + "ctc_loss": utils.item(ctc_loss.data), + } + return loss, sample_size, logging_output + + def compute_loss( + self, + feat_out, + feat_out_post, + eos_out, + feat_tgt, + eos_tgt, + tgt_lens, + reduction="mean", + ): + mask = lengths_to_mask(tgt_lens) + _eos_out = eos_out[mask].squeeze() + _eos_tgt = eos_tgt[mask] + _feat_tgt = feat_tgt[mask] + _feat_out = feat_out[mask] + _feat_out_post = feat_out_post[mask] + + l1_loss = F.l1_loss(_feat_out, _feat_tgt, reduction=reduction) + F.l1_loss( + _feat_out_post, _feat_tgt, reduction=reduction + ) + mse_loss = F.mse_loss(_feat_out, _feat_tgt, reduction=reduction) + F.mse_loss( + _feat_out_post, _feat_tgt, reduction=reduction + ) + eos_loss = F.binary_cross_entropy_with_logits( + _eos_out, + _eos_tgt, + pos_weight=torch.tensor(self.bce_pos_weight), + reduction=reduction, + ) + return l1_loss, mse_loss, eos_loss + + @classmethod + def reduce_metrics(cls, logging_outputs: List[Dict[str, Any]]) -> None: + ns = [log.get("sample_size", 0) for log in logging_outputs] + ntot = sum(ns) + ws = [n / (ntot + 1e-8) for n in ns] + for key in ["loss", "l1_loss", "mse_loss", "eos_loss", "attn_loss", "ctc_loss"]: + vals = [log.get(key, 0) for log in logging_outputs] + val = sum(val * w for val, w in zip(vals, ws)) + metrics.log_scalar(key, val, ntot, round=3) + metrics.log_scalar("sample_size", ntot, len(logging_outputs)) + + # inference metrics + if "targ_frames" not in logging_outputs[0]: + return + n = sum(log.get("targ_frames", 0) for log in logging_outputs) + for key, new_key in [ + ("mcd_loss", "mcd_loss"), + ("pred_frames", "pred_ratio"), + ("nins", "ins_rate"), + ("ndel", "del_rate"), + ]: + val = sum(log.get(key, 0) for log in logging_outputs) + metrics.log_scalar(new_key, val / n, n, round=3) + + @staticmethod + def logging_outputs_can_be_summed() -> bool: + return False diff --git a/data/fairseq/fairseq/criterions/wav2vec_criterion.py b/data/fairseq/fairseq/criterions/wav2vec_criterion.py new file mode 100644 index 0000000000000000000000000000000000000000..3975468487704e053fe7634257f443ee2c396616 --- /dev/null +++ b/data/fairseq/fairseq/criterions/wav2vec_criterion.py @@ -0,0 +1,231 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import math +from dataclasses import dataclass, field +from typing import List, Optional + +import torch +import torch.nn.functional as F +from fairseq import utils +from fairseq.logging import metrics +from fairseq.criterions import FairseqCriterion, register_criterion +from fairseq.dataclass import FairseqDataclass +from fairseq.logging.meters import safe_round +from fairseq.utils import is_xla_tensor + + +@dataclass +class Wav2VecCriterionConfig(FairseqDataclass): + infonce: bool = field( + default=False, + metadata={ + "help": "if set, uses cross entropy instead of binary cross entropy (i.e. InfoNCE loss)" + }, + ) + loss_weights: Optional[List[float]] = field( + default=None, + metadata={"help": "weights for additional loss terms (not first one)"}, + ) + log_keys: List[str] = field( + default_factory=lambda: [], + metadata={"help": "output keys to log"}, + ) + + +@register_criterion("wav2vec", dataclass=Wav2VecCriterionConfig) +class Wav2vecCriterion(FairseqCriterion): + def __init__(self, task, infonce=False, loss_weights=None, log_keys=None): + super().__init__(task) + self.infonce = infonce + self.loss_weights = loss_weights + self.log_keys = [] if log_keys is None else log_keys + + def forward(self, model, sample, reduce=True): + """Compute the loss for the given sample. + + Returns a tuple with three elements: + 1) the loss + 2) the sample size, which is used as the denominator for the gradient + 3) logging outputs to display while training + """ + net_output = model(**sample["net_input"]) + logits = model.get_logits(net_output).float() + target = model.get_targets(sample, net_output) + self.xla = is_xla_tensor(logits) + + # XXX: handle weights on xla. + weights = None + if hasattr(model, "get_target_weights") and not self.infonce: + weights = model.get_target_weights(target, net_output) + if torch.is_tensor(weights): + weights = weights.float() + + losses = [] + + reduction = "none" if ((not reduce) or self.xla) else "sum" + if self.infonce: + loss = F.cross_entropy(logits, target, reduction=reduction) + else: + loss = F.binary_cross_entropy_with_logits( + logits, target.float(), weights, reduction=reduction + ) + + if self.xla: + # tpu-comment: since dynamic shapes lead to recompilations on xla, + # we don't shrink tensors using mask_indices. + # Instead, we use mask indices to adjust loss. + mi = ( + sample["net_input"]["mask_indices"] + .transpose(0, 1) # logits are transposed in `model.get_logits` + .reshape(logits.size(0)) + ) + loss = (loss * mi).sum() if reduce else (loss * mi) + + if "sample_size" in sample: + sample_size = sample["sample_size"] + elif "mask_indices" in sample["net_input"]: + sample_size = sample["net_input"]["mask_indices"].sum() + else: + sample_size = target.numel() if self.infonce else target.long().sum().item() + losses.append(loss.detach().clone()) + + if self.loss_weights is not None: + assert hasattr(model, "get_extra_losses") + extra_losses = model.get_extra_losses(net_output) + if torch.is_tensor(extra_losses): + extra_losses = [extra_losses] + if len(self.loss_weights) == 1 and len(extra_losses) != 1: + self.loss_weights = [self.loss_weights[0]] * len(extra_losses) + assert len(extra_losses) == len( + self.loss_weights + ), f"{len(extra_losses)}, {len(self.loss_weights)}" + for p, coef in zip(extra_losses, self.loss_weights): + if coef != 0 and p is not None: + p = coef * p.float() * sample_size + loss += p + losses.append(p) + + logging_output = { + "loss": loss.item() if (reduce and not self.xla) else loss.detach(), + "ntokens": sample_size, + "nsentences": sample["id"].numel(), + "sample_size": sample_size, + } + + for lk in self.log_keys: + # Only store "logits" and "target" for computing MAP and MAUC + # during validation + if lk == "logits": + if not self.training: + logging_output["logits"] = logits.cpu().numpy() + elif lk == "target": + if not self.training: + # If the targets have been mixed with the predictions of + # teacher models, find the original targets + if hasattr(model, "get_original_targets"): + original_target = model.get_original_targets(sample, net_output) + else: + original_target = target + logging_output["target"] = original_target.cpu().numpy() + elif lk in net_output: + value = net_output[lk] + if not is_xla_tensor(value): + value = float(value) + logging_output[lk] = value + + if len(losses) > 1: + for i, l in enumerate(losses): + logging_output[f"loss_{i}"] = l.item() if not self.xla else l.detach() + + if self.infonce: + with torch.no_grad(): + if logits.numel() == 0: + corr = 0 + count = 0 + else: + assert logits.dim() > 1, logits.shape + max = logits.argmax(-1) == 0 + min = logits.argmin(-1) == 0 + if is_xla_tensor(logits): + max, min = max * mi, min * mi + both = max & min + corr = max.long().sum() - both.long().sum() + count = mi.sum() + else: + both = max & min + corr = max.long().sum().item() - both.long().sum().item() + count = float(max.numel()) + + logging_output["correct"] = corr + logging_output["count"] = count + + return loss, sample_size, logging_output + + @staticmethod + def reduce_metrics(logging_outputs) -> None: + """Aggregate logging outputs from data parallel training.""" + loss_sum = utils.item(sum(log.get("loss", 0) for log in logging_outputs)) + ntokens = utils.item(sum(log.get("ntokens", 0) for log in logging_outputs)) + nsentences = utils.item( + sum(log.get("nsentences", 0) for log in logging_outputs) + ) + sample_size = utils.item( + sum(log.get("sample_size", 0) for log in logging_outputs) + ) + + metrics.log_scalar( + "loss", loss_sum / (sample_size or 1) / math.log(2), sample_size, round=3 + ) + metrics.log_scalar("ntokens", ntokens) + metrics.log_scalar("nsentences", nsentences) + + correct = sum(log.get("correct", 0) for log in logging_outputs) + metrics.log_scalar("_correct", correct) + + total = sum(log.get("count", 0) for log in logging_outputs) + metrics.log_scalar("_total", total) + + if total > 0: + metrics.log_derived( + "accuracy", + lambda meters: safe_round( + meters["_correct"].sum / meters["_total"].sum, 5 + ) + if meters["_total"].sum > 0 + else float("nan"), + ) + + builtin_keys = { + "loss", + "ntokens", + "nsentences", + "sample_size", + "correct", + "count", + } + + for k in logging_outputs[0]: + if k not in builtin_keys: + val = sum(log.get(k, 0) for log in logging_outputs) + if k.startswith("loss"): + metrics.log_scalar( + k, val / (sample_size or 1) / math.log(2), sample_size, round=3 + ) + else: + metrics.log_scalar(k, val / len(logging_outputs), round=3) + + # FIXME: revert when gather based xla reduction is implemented + # @staticmethod + # def logging_outputs_can_be_summed() -> bool: + def logging_outputs_can_be_summed(self) -> bool: + """ + Whether the logging outputs returned by `forward` can be summed + across workers prior to calling `reduce_metrics`. Setting this + to True will improves distributed training speed. + """ + # XXX: Gather based reduction not implemented for xla yet. + # So we fall to sum based reduction for xla. + return self.xla diff --git a/data/fairseq/fairseq/data/encoders/__pycache__/__init__.cpython-38.pyc b/data/fairseq/fairseq/data/encoders/__pycache__/__init__.cpython-38.pyc new file mode 100644 index 0000000000000000000000000000000000000000..7de1d815a4fb7095aa69efd74f94f84d7212e08e Binary files /dev/null and b/data/fairseq/fairseq/data/encoders/__pycache__/__init__.cpython-38.pyc differ diff --git a/data/fairseq/fairseq/data/encoders/__pycache__/byte_bpe.cpython-38.pyc b/data/fairseq/fairseq/data/encoders/__pycache__/byte_bpe.cpython-38.pyc new file mode 100644 index 0000000000000000000000000000000000000000..46cb92de88083fe43a09c2efdc1da069fb05ea58 Binary files /dev/null and b/data/fairseq/fairseq/data/encoders/__pycache__/byte_bpe.cpython-38.pyc differ diff --git a/data/fairseq/fairseq/data/encoders/__pycache__/byte_utils.cpython-38.pyc b/data/fairseq/fairseq/data/encoders/__pycache__/byte_utils.cpython-38.pyc new file mode 100644 index 0000000000000000000000000000000000000000..0e8113a07ef2f15af750ea08d6310386f5ba4e2d Binary files /dev/null and b/data/fairseq/fairseq/data/encoders/__pycache__/byte_utils.cpython-38.pyc differ diff --git a/data/fairseq/fairseq/data/encoders/__pycache__/gpt2_bpe_utils.cpython-38.pyc b/data/fairseq/fairseq/data/encoders/__pycache__/gpt2_bpe_utils.cpython-38.pyc new file mode 100644 index 0000000000000000000000000000000000000000..0626789ef9474c303b7a10e97cec35f2f18666f7 Binary files /dev/null and b/data/fairseq/fairseq/data/encoders/__pycache__/gpt2_bpe_utils.cpython-38.pyc differ diff --git a/data/fairseq/fairseq/data/encoders/byte_bpe.py b/data/fairseq/fairseq/data/encoders/byte_bpe.py new file mode 100644 index 0000000000000000000000000000000000000000..31e3a0627827f19ca7f0b58da45e46d40a80c3bf --- /dev/null +++ b/data/fairseq/fairseq/data/encoders/byte_bpe.py @@ -0,0 +1,48 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + + +from dataclasses import dataclass, field + +from fairseq import file_utils +from fairseq.data.encoders import register_bpe +from fairseq.data.encoders.byte_utils import ( + SPACE, + SPACE_ESCAPE, + byte_encode, + smart_byte_decode, +) +from fairseq.dataclass import FairseqDataclass + + +@dataclass +class ByteBpeConfig(FairseqDataclass): + sentencepiece_model_path: str = field( + default="???", metadata={"help": "path to sentencepiece model"} + ) + + +@register_bpe("byte_bpe", dataclass=ByteBpeConfig) +class ByteBPE(object): + def __init__(self, cfg): + vocab = file_utils.cached_path(cfg.sentencepiece_model_path) + try: + import sentencepiece as spm + + self.sp = spm.SentencePieceProcessor() + self.sp.Load(vocab) + except ImportError: + raise ImportError( + "Please install sentencepiece with: pip install sentencepiece" + ) + + def encode(self, x: str) -> str: + byte_encoded = byte_encode(x) + return SPACE.join(self.sp.EncodeAsPieces(byte_encoded)) + + @staticmethod + def decode(x: str) -> str: + unescaped = x.replace(SPACE, "").replace(SPACE_ESCAPE, SPACE) + return smart_byte_decode(unescaped) diff --git a/data/fairseq/fairseq/data/encoders/byte_utils.py b/data/fairseq/fairseq/data/encoders/byte_utils.py new file mode 100644 index 0000000000000000000000000000000000000000..a305c080926c2d094b7e8ae48f5331da82025a75 --- /dev/null +++ b/data/fairseq/fairseq/data/encoders/byte_utils.py @@ -0,0 +1,51 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import re + + +WHITESPACE_NORMALIZER = re.compile(r"\s+") +SPACE = chr(32) +SPACE_ESCAPE = chr(9601) +# excluding non-breaking space (160) here +PRINTABLE_LATIN = set( + list(range(32, 126 + 1)) + list(range(161, 172 + 1)) + list(range(174, 255 + 1)) +) +BYTE_TO_BCHAR = { + b: chr(b) if b in PRINTABLE_LATIN else chr(256 + b) for b in range(256) +} +BCHAR_TO_BYTE = {bc: b for b, bc in BYTE_TO_BCHAR.items()} + + +def byte_encode(x: str) -> str: + normalized = WHITESPACE_NORMALIZER.sub(SPACE, x) + return "".join([BYTE_TO_BCHAR[b] for b in normalized.encode("utf-8")]) + + +def byte_decode(x: str) -> str: + try: + return bytes([BCHAR_TO_BYTE[bc] for bc in x]).decode("utf-8") + except ValueError: + return "" + + +def smart_byte_decode(x: str) -> str: + output = byte_decode(x) + if output == "": + # DP the best recovery (max valid chars) if it's broken + n_bytes = len(x) + f = [0 for _ in range(n_bytes + 1)] + pt = [0 for _ in range(n_bytes + 1)] + for i in range(1, n_bytes + 1): + f[i], pt[i] = f[i - 1], i - 1 + for j in range(1, min(4, i) + 1): + if f[i - j] + 1 > f[i] and len(byte_decode(x[i - j : i])) > 0: + f[i], pt[i] = f[i - j] + 1, i - j + cur_pt = n_bytes + while cur_pt > 0: + if f[cur_pt] == f[pt[cur_pt]] + 1: + output = byte_decode(x[pt[cur_pt] : cur_pt]) + output + cur_pt = pt[cur_pt] + return output diff --git a/data/fairseq/fairseq/data/encoders/bytes.py b/data/fairseq/fairseq/data/encoders/bytes.py new file mode 100644 index 0000000000000000000000000000000000000000..f88f8f6929f5b6bdb0db470be9ebedf8fe1f752d --- /dev/null +++ b/data/fairseq/fairseq/data/encoders/bytes.py @@ -0,0 +1,34 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + + +from fairseq.data.encoders import register_bpe +from fairseq.data.encoders.byte_utils import ( + SPACE, + SPACE_ESCAPE, + byte_encode, + smart_byte_decode, +) + + +@register_bpe("bytes") +class Bytes(object): + def __init__(self, *unused): + pass + + @staticmethod + def add_args(parser): + pass + + @staticmethod + def encode(x: str) -> str: + encoded = byte_encode(x) + escaped = encoded.replace(SPACE, SPACE_ESCAPE) + return SPACE.join(list(escaped)) + + @staticmethod + def decode(x: str) -> str: + unescaped = x.replace(SPACE, "").replace(SPACE_ESCAPE, SPACE) + return smart_byte_decode(unescaped) diff --git a/data/fairseq/fairseq/data/encoders/gpt2_bpe.py b/data/fairseq/fairseq/data/encoders/gpt2_bpe.py new file mode 100644 index 0000000000000000000000000000000000000000..e661426a73c7e735f7054bcb04281bf1649bb46c --- /dev/null +++ b/data/fairseq/fairseq/data/encoders/gpt2_bpe.py @@ -0,0 +1,45 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from dataclasses import dataclass, field + +from fairseq import file_utils +from fairseq.data.encoders import register_bpe +from fairseq.dataclass import FairseqDataclass + +from .gpt2_bpe_utils import get_encoder + + +DEFAULT_ENCODER_JSON = "https://dl.fbaipublicfiles.com/fairseq/gpt2_bpe/encoder.json" +DEFAULT_VOCAB_BPE = "https://dl.fbaipublicfiles.com/fairseq/gpt2_bpe/vocab.bpe" + + +@dataclass +class GPT2BPEConfig(FairseqDataclass): + gpt2_encoder_json: str = field( + default=DEFAULT_ENCODER_JSON, metadata={"help": "path to encoder.json"} + ) + gpt2_vocab_bpe: str = field( + default=DEFAULT_VOCAB_BPE, metadata={"help": "path to vocab.bpe"} + ) + + +@register_bpe("gpt2", dataclass=GPT2BPEConfig) +class GPT2BPE(object): + def __init__(self, cfg): + encoder_json = file_utils.cached_path(cfg.gpt2_encoder_json) + vocab_bpe = file_utils.cached_path(cfg.gpt2_vocab_bpe) + self.bpe = get_encoder(encoder_json, vocab_bpe) + + def encode(self, x: str) -> str: + return " ".join(map(str, self.bpe.encode(x))) + + def decode(self, x: str) -> str: + return self.bpe.decode( + [int(tok) if tok not in {"", ""} else tok for tok in x.split()] + ) + + def is_beginning_of_word(self, x: str) -> bool: + return self.decode(x).startswith(" ") diff --git a/data/fairseq/fairseq/data/encoders/gpt2_bpe_utils.py b/data/fairseq/fairseq/data/encoders/gpt2_bpe_utils.py new file mode 100644 index 0000000000000000000000000000000000000000..996d3d4a1175fb3841242c286f08faf43fc06e3e --- /dev/null +++ b/data/fairseq/fairseq/data/encoders/gpt2_bpe_utils.py @@ -0,0 +1,140 @@ +""" +Byte pair encoding utilities from GPT-2. + +Original source: https://github.com/openai/gpt-2/blob/master/src/encoder.py +Original license: MIT +""" + +import json +from functools import lru_cache + + +@lru_cache() +def bytes_to_unicode(): + """ + Returns list of utf-8 byte and a corresponding list of unicode strings. + The reversible bpe codes work on unicode strings. + This means you need a large # of unicode characters in your vocab if you want to avoid UNKs. + When you're at something like a 10B token dataset you end up needing around 5K for decent coverage. + This is a signficant percentage of your normal, say, 32K bpe vocab. + To avoid that, we want lookup tables between utf-8 bytes and unicode strings. + And avoids mapping to whitespace/control characters the bpe code barfs on. + """ + bs = ( + list(range(ord("!"), ord("~") + 1)) + + list(range(ord("¡"), ord("¬") + 1)) + + list(range(ord("®"), ord("ÿ") + 1)) + ) + cs = bs[:] + n = 0 + for b in range(2**8): + if b not in bs: + bs.append(b) + cs.append(2**8 + n) + n += 1 + cs = [chr(n) for n in cs] + return dict(zip(bs, cs)) + + +def get_pairs(word): + """Return set of symbol pairs in a word. + Word is represented as tuple of symbols (symbols being variable-length strings). + """ + pairs = set() + prev_char = word[0] + for char in word[1:]: + pairs.add((prev_char, char)) + prev_char = char + return pairs + + +class Encoder: + def __init__(self, encoder, bpe_merges, errors="replace"): + self.encoder = encoder + self.decoder = {v: k for k, v in self.encoder.items()} + self.errors = errors # how to handle errors in decoding + self.byte_encoder = bytes_to_unicode() + self.byte_decoder = {v: k for k, v in self.byte_encoder.items()} + self.bpe_ranks = dict(zip(bpe_merges, range(len(bpe_merges)))) + self.cache = {} + + try: + import regex as re + + self.re = re + except ImportError: + raise ImportError("Please install regex with: pip install regex") + + # Should haved added re.IGNORECASE so BPE merges can happen for capitalized versions of contractions + self.pat = self.re.compile( + r"""'s|'t|'re|'ve|'m|'ll|'d| ?\p{L}+| ?\p{N}+| ?[^\s\p{L}\p{N}]+|\s+(?!\S)|\s+""" + ) + + def bpe(self, token): + if token in self.cache: + return self.cache[token] + word = tuple(token) + pairs = get_pairs(word) + + if not pairs: + return token + + while True: + bigram = min(pairs, key=lambda pair: self.bpe_ranks.get(pair, float("inf"))) + if bigram not in self.bpe_ranks: + break + first, second = bigram + new_word = [] + i = 0 + while i < len(word): + try: + j = word.index(first, i) + new_word.extend(word[i:j]) + i = j + except: + new_word.extend(word[i:]) + break + + if word[i] == first and i < len(word) - 1 and word[i + 1] == second: + new_word.append(first + second) + i += 2 + else: + new_word.append(word[i]) + i += 1 + new_word = tuple(new_word) + word = new_word + if len(word) == 1: + break + else: + pairs = get_pairs(word) + word = " ".join(word) + self.cache[token] = word + return word + + def encode(self, text): + bpe_tokens = [] + for token in self.re.findall(self.pat, text): + token = "".join(self.byte_encoder[b] for b in token.encode("utf-8")) + bpe_tokens.extend( + self.encoder[bpe_token] for bpe_token in self.bpe(token).split(" ") + ) + return bpe_tokens + + def decode(self, tokens): + text = "".join([self.decoder.get(token, token) for token in tokens]) + text = bytearray([self.byte_decoder[c] for c in text]).decode( + "utf-8", errors=self.errors + ) + return text + + +def get_encoder(encoder_json_path, vocab_bpe_path): + with open(encoder_json_path, "r") as f: + encoder = json.load(f) + with open(vocab_bpe_path, "r", encoding="utf-8") as f: + bpe_data = f.read() + bpe_merges = [tuple(merge_str.split()) for merge_str in bpe_data.split("\n")[1:-1]] + return Encoder( + encoder=encoder, + bpe_merges=bpe_merges, + ) diff --git a/data/fairseq/fairseq/data/encoders/hf_bert_bpe.py b/data/fairseq/fairseq/data/encoders/hf_bert_bpe.py new file mode 100644 index 0000000000000000000000000000000000000000..a41c059343ec7e2914b2c9d2f53f526c33f9659d --- /dev/null +++ b/data/fairseq/fairseq/data/encoders/hf_bert_bpe.py @@ -0,0 +1,50 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from dataclasses import dataclass, field +from typing import Optional + +from fairseq.data.encoders import register_bpe +from fairseq.dataclass import FairseqDataclass + + +@dataclass +class BertBPEConfig(FairseqDataclass): + bpe_cased: bool = field(default=False, metadata={"help": "set for cased BPE"}) + bpe_vocab_file: Optional[str] = field( + default=None, metadata={"help": "bpe vocab file"} + ) + + +@register_bpe("bert", dataclass=BertBPEConfig) +class BertBPE(object): + def __init__(self, cfg): + try: + from transformers import BertTokenizer + except ImportError: + raise ImportError( + "Please install transformers with: pip install transformers" + ) + + if cfg.bpe_vocab_file: + self.bert_tokenizer = BertTokenizer( + cfg.bpe_vocab_file, do_lower_case=not cfg.bpe_cased + ) + else: + vocab_file_name = ( + "bert-base-cased" if cfg.bpe_cased else "bert-base-uncased" + ) + self.bert_tokenizer = BertTokenizer.from_pretrained(vocab_file_name) + + def encode(self, x: str) -> str: + return " ".join(self.bert_tokenizer.tokenize(x)) + + def decode(self, x: str) -> str: + return self.bert_tokenizer.clean_up_tokenization( + self.bert_tokenizer.convert_tokens_to_string(x.split(" ")) + ) + + def is_beginning_of_word(self, x: str) -> bool: + return not x.startswith("##") diff --git a/data/fairseq/fairseq/data/encoders/moses_tokenizer.py b/data/fairseq/fairseq/data/encoders/moses_tokenizer.py new file mode 100644 index 0000000000000000000000000000000000000000..e236dad167a037a8ed95f7fc8292b27b10d580b0 --- /dev/null +++ b/data/fairseq/fairseq/data/encoders/moses_tokenizer.py @@ -0,0 +1,49 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from dataclasses import dataclass, field + +from fairseq.data.encoders import register_tokenizer +from fairseq.dataclass import FairseqDataclass + + +@dataclass +class MosesTokenizerConfig(FairseqDataclass): + source_lang: str = field(default="en", metadata={"help": "source language"}) + target_lang: str = field(default="en", metadata={"help": "target language"}) + moses_no_dash_splits: bool = field( + default=False, metadata={"help": "don't apply dash split rules"} + ) + moses_no_escape: bool = field( + default=False, + metadata={"help": "don't perform HTML escaping on apostrophe, quotes, etc."}, + ) + + +@register_tokenizer("moses", dataclass=MosesTokenizerConfig) +class MosesTokenizer(object): + def __init__(self, cfg: MosesTokenizerConfig): + self.cfg = cfg + + try: + from sacremoses import MosesTokenizer, MosesDetokenizer + + self.tok = MosesTokenizer(cfg.source_lang) + self.detok = MosesDetokenizer(cfg.target_lang) + except ImportError: + raise ImportError( + "Please install Moses tokenizer with: pip install sacremoses" + ) + + def encode(self, x: str) -> str: + return self.tok.tokenize( + x, + aggressive_dash_splits=(not self.cfg.moses_no_dash_splits), + return_str=True, + escape=(not self.cfg.moses_no_escape), + ) + + def decode(self, x: str) -> str: + return self.detok.detokenize(x.split()) diff --git a/data/fairseq/fairseq/data/encoders/nltk_tokenizer.py b/data/fairseq/fairseq/data/encoders/nltk_tokenizer.py new file mode 100644 index 0000000000000000000000000000000000000000..0ab92377b3a23bb48384c3f7acf299612e8b0775 --- /dev/null +++ b/data/fairseq/fairseq/data/encoders/nltk_tokenizer.py @@ -0,0 +1,24 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from fairseq.data.encoders import register_tokenizer +from fairseq.dataclass import FairseqDataclass + + +@register_tokenizer("nltk", dataclass=FairseqDataclass) +class NLTKTokenizer(object): + def __init__(self, *unused): + try: + from nltk.tokenize import word_tokenize + + self.word_tokenize = word_tokenize + except ImportError: + raise ImportError("Please install nltk with: pip install nltk") + + def encode(self, x: str) -> str: + return " ".join(self.word_tokenize(x)) + + def decode(self, x: str) -> str: + return x diff --git a/data/fairseq/fairseq/data/encoders/sentencepiece_bpe.py b/data/fairseq/fairseq/data/encoders/sentencepiece_bpe.py new file mode 100644 index 0000000000000000000000000000000000000000..0aa6cd7681d0c3a91a6917640972d008db8faef7 --- /dev/null +++ b/data/fairseq/fairseq/data/encoders/sentencepiece_bpe.py @@ -0,0 +1,65 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from dataclasses import dataclass, field +from typing import Optional + +from fairseq import file_utils +from fairseq.data.encoders import register_bpe +from fairseq.dataclass import FairseqDataclass + + +@dataclass +class SentencepieceConfig(FairseqDataclass): + sentencepiece_model: str = field( + default="???", metadata={"help": "path to sentencepiece model"} + ) + sentencepiece_enable_sampling: bool = field( + default=False, metadata={"help": "enable sampling"} + ) + sentencepiece_alpha: Optional[float] = field( + default=None, + metadata={ + "help": "soothing parameter for unigram sampling, " + "and merge probability for BPE-dropout" + }, + ) + + +@register_bpe("sentencepiece", dataclass=SentencepieceConfig) +class SentencepieceBPE(object): + def __init__(self, cfg): + self.enable_sampling = cfg.sentencepiece_enable_sampling + self.alpha = cfg.sentencepiece_alpha + sentencepiece_model = file_utils.cached_path(cfg.sentencepiece_model) + try: + import sentencepiece as spm + + self.sp = spm.SentencePieceProcessor() + self.sp.Load(sentencepiece_model) + except ImportError: + raise ImportError( + "Please install sentencepiece with: pip install sentencepiece" + ) + + def encode(self, x: str) -> str: + return " ".join( + self.sp.Encode( + x, out_type=str, enable_sampling=self.enable_sampling, alpha=self.alpha + ) + ) + + def decode(self, x: str) -> str: + return x.replace(" ", "").replace("\u2581", " ").strip() + + def is_beginning_of_word(self, x: str) -> bool: + if x in ["", "", "", ""]: + # special elements are always considered beginnings + # HACK: this logic is already present in fairseq/tasks/masked_lm.py + # but these special tokens are also contained in the sentencepiece + # vocabulary which causes duplicate special tokens. This hack makes + # sure that they are all taken into account. + return True + return x.startswith("\u2581") diff --git a/data/fairseq/fairseq/data/encoders/space_tokenizer.py b/data/fairseq/fairseq/data/encoders/space_tokenizer.py new file mode 100644 index 0000000000000000000000000000000000000000..925ad41b7c1aee6738c63938c36bd3ee16dca812 --- /dev/null +++ b/data/fairseq/fairseq/data/encoders/space_tokenizer.py @@ -0,0 +1,21 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import re + +from fairseq.data.encoders import register_tokenizer +from fairseq.dataclass import FairseqDataclass + + +@register_tokenizer("space", dataclass=FairseqDataclass) +class SpaceTokenizer(object): + def __init__(self, *unused): + self.space_tok = re.compile(r"\s+") + + def encode(self, x: str) -> str: + return self.space_tok.sub(" ", x) + + def decode(self, x: str) -> str: + return x diff --git a/data/fairseq/fairseq/data/encoders/subword_nmt_bpe.py b/data/fairseq/fairseq/data/encoders/subword_nmt_bpe.py new file mode 100644 index 0000000000000000000000000000000000000000..5d724d2730a5895ca55af2998c2ced471625b516 --- /dev/null +++ b/data/fairseq/fairseq/data/encoders/subword_nmt_bpe.py @@ -0,0 +1,54 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from dataclasses import dataclass, field + +from fairseq import file_utils +from fairseq.data.encoders import register_bpe +from fairseq.dataclass import FairseqDataclass + + +@dataclass +class SubwordNMTBPEConfig(FairseqDataclass): + bpe_codes: str = field(default="???", metadata={"help": "path to subword NMT BPE"}) + bpe_separator: str = field(default="@@", metadata={"help": "BPE separator"}) + + +@register_bpe("subword_nmt", dataclass=SubwordNMTBPEConfig) +class SubwordNMTBPE(object): + def __init__(self, cfg): + if cfg.bpe_codes is None: + raise ValueError("--bpe-codes is required for --bpe=subword_nmt") + codes = file_utils.cached_path(cfg.bpe_codes) + try: + from subword_nmt import apply_bpe + + bpe_parser = apply_bpe.create_parser() + bpe_args = bpe_parser.parse_args( + [ + "--codes", + codes, + "--separator", + cfg.bpe_separator, + ] + ) + self.bpe = apply_bpe.BPE( + bpe_args.codes, + bpe_args.merges, + bpe_args.separator, + None, + bpe_args.glossaries, + ) + self.bpe_symbol = bpe_args.separator + " " + except ImportError: + raise ImportError( + "Please install subword_nmt with: pip install subword-nmt" + ) + + def encode(self, x: str) -> str: + return self.bpe.process_line(x) + + def decode(self, x: str) -> str: + return (x + " ").replace(self.bpe_symbol, "").rstrip() diff --git a/data/fairseq/fairseq/data/encoders/utils.py b/data/fairseq/fairseq/data/encoders/utils.py new file mode 100644 index 0000000000000000000000000000000000000000..d93eb532ef84f0e2bc708b777229ab2cb76ca14b --- /dev/null +++ b/data/fairseq/fairseq/data/encoders/utils.py @@ -0,0 +1,30 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import torch +from fairseq.data import encoders + + +def get_whole_word_mask(args, dictionary): + bpe = encoders.build_bpe(args) + if bpe is not None: + + def is_beginning_of_word(i): + if i < dictionary.nspecial: + # special elements are always considered beginnings + return True + tok = dictionary[i] + if tok.startswith("madeupword"): + return True + try: + return bpe.is_beginning_of_word(tok) + except ValueError: + return True + + mask_whole_words = torch.ByteTensor( + list(map(is_beginning_of_word, range(len(dictionary)))) + ) + return mask_whole_words + return None diff --git a/data/fairseq/fairseq/logging/__init__.py b/data/fairseq/fairseq/logging/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391 diff --git a/data/fairseq/fairseq/logging/__pycache__/__init__.cpython-38.pyc b/data/fairseq/fairseq/logging/__pycache__/__init__.cpython-38.pyc new file mode 100644 index 0000000000000000000000000000000000000000..6b0abf93741072ef1d2710879bb727033a3d21b4 Binary files /dev/null and b/data/fairseq/fairseq/logging/__pycache__/__init__.cpython-38.pyc differ diff --git a/data/fairseq/fairseq/logging/__pycache__/meters.cpython-38.pyc b/data/fairseq/fairseq/logging/__pycache__/meters.cpython-38.pyc new file mode 100644 index 0000000000000000000000000000000000000000..f6ceeb1d2ecee1b43950d323851206bf6104064a Binary files /dev/null and b/data/fairseq/fairseq/logging/__pycache__/meters.cpython-38.pyc differ diff --git a/data/fairseq/fairseq/logging/__pycache__/metrics.cpython-38.pyc b/data/fairseq/fairseq/logging/__pycache__/metrics.cpython-38.pyc new file mode 100644 index 0000000000000000000000000000000000000000..d6ce74173b36e9ae3d76ef9f1a84f20138f533f2 Binary files /dev/null and b/data/fairseq/fairseq/logging/__pycache__/metrics.cpython-38.pyc differ diff --git a/data/fairseq/fairseq/logging/__pycache__/progress_bar.cpython-38.pyc b/data/fairseq/fairseq/logging/__pycache__/progress_bar.cpython-38.pyc new file mode 100644 index 0000000000000000000000000000000000000000..1612c27ec3f115231036afd6b336284a7556c40a Binary files /dev/null and b/data/fairseq/fairseq/logging/__pycache__/progress_bar.cpython-38.pyc differ diff --git a/data/fairseq/fairseq/logging/meters.py b/data/fairseq/fairseq/logging/meters.py new file mode 100644 index 0000000000000000000000000000000000000000..495bd083000de9e4a05f1470228c1171c8c8bb9c --- /dev/null +++ b/data/fairseq/fairseq/logging/meters.py @@ -0,0 +1,351 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import bisect +import time +from collections import OrderedDict +from typing import Dict, Optional + +try: + import torch + + def type_as(a, b): + if torch.is_tensor(a) and torch.is_tensor(b): + return a.to(b) + else: + return a + +except ImportError: + torch = None + + def type_as(a, b): + return a + + +try: + import numpy as np +except ImportError: + np = None + + +class Meter(object): + """Base class for Meters.""" + + def __init__(self): + pass + + def state_dict(self): + return {} + + def load_state_dict(self, state_dict): + pass + + def reset(self): + raise NotImplementedError + + @property + def smoothed_value(self) -> float: + """Smoothed value used for logging.""" + raise NotImplementedError + + +def safe_round(number, ndigits): + if hasattr(number, "__round__"): + return round(number, ndigits) + elif torch is not None and torch.is_tensor(number) and number.numel() == 1: + return safe_round(number.item(), ndigits) + elif np is not None and np.ndim(number) == 0 and hasattr(number, "item"): + return safe_round(number.item(), ndigits) + else: + return number + + +class AverageMeter(Meter): + """Computes and stores the average and current value""" + + def __init__(self, round: Optional[int] = None): + self.round = round + self.reset() + + def reset(self): + self.val = None # most recent update + self.sum = 0 # sum from all updates + self.count = 0 # total n from all updates + + def update(self, val, n=1): + if val is not None: + self.val = val + if n > 0: + self.sum = type_as(self.sum, val) + (val * n) + self.count = type_as(self.count, n) + n + + def state_dict(self): + return { + "val": self.val, + "sum": self.sum, + "count": self.count, + "round": self.round, + } + + def load_state_dict(self, state_dict): + self.val = state_dict["val"] + self.sum = state_dict["sum"] + self.count = state_dict["count"] + self.round = state_dict.get("round", None) + + @property + def avg(self): + return self.sum / self.count if self.count > 0 else self.val + + @property + def smoothed_value(self) -> float: + val = self.avg + if self.round is not None and val is not None: + val = safe_round(val, self.round) + return val + + +class SumMeter(Meter): + """Computes and stores the sum""" + + def __init__(self, round: Optional[int] = None): + self.round = round + self.reset() + + def reset(self): + self.sum = 0 # sum from all updates + + def update(self, val): + if val is not None: + self.sum = type_as(self.sum, val) + val + + def state_dict(self): + return { + "sum": self.sum, + "round": self.round, + } + + def load_state_dict(self, state_dict): + self.sum = state_dict["sum"] + self.round = state_dict.get("round", None) + + @property + def smoothed_value(self) -> float: + val = self.sum + if self.round is not None and val is not None: + val = safe_round(val, self.round) + return val + + +class ConcatTensorMeter(Meter): + """Concatenates tensors""" + + def __init__(self, dim=0): + super().__init__() + self.reset() + self.dim = dim + + def reset(self): + self.tensor = None + + def update(self, val): + if self.tensor is None: + self.tensor = val + else: + self.tensor = torch.cat([self.tensor, val], dim=self.dim) + + def state_dict(self): + return { + "tensor": self.tensor, + } + + def load_state_dict(self, state_dict): + self.tensor = state_dict["tensor"] + + @property + def smoothed_value(self) -> float: + return [] # return a dummy value + + +class TimeMeter(Meter): + """Computes the average occurrence of some event per second""" + + def __init__( + self, + init: int = 0, + n: int = 0, + round: Optional[int] = None, + ): + self.round = round + self.reset(init, n) + + def reset(self, init=0, n=0): + self.init = init + self.start = time.perf_counter() + self.n = n + self.i = 0 + + def update(self, val=1): + self.n = type_as(self.n, val) + val + self.i += 1 + + def state_dict(self): + return { + "init": self.elapsed_time, + "n": self.n, + "round": self.round, + } + + def load_state_dict(self, state_dict): + if "start" in state_dict: + # backwards compatibility for old state_dicts + self.reset(init=state_dict["init"]) + else: + self.reset(init=state_dict["init"], n=state_dict["n"]) + self.round = state_dict.get("round", None) + + @property + def avg(self): + return self.n / self.elapsed_time + + @property + def elapsed_time(self): + return self.init + (time.perf_counter() - self.start) + + @property + def smoothed_value(self) -> float: + val = self.avg + if self.round is not None and val is not None: + val = safe_round(val, self.round) + return val + + +class StopwatchMeter(Meter): + """Computes the sum/avg duration of some event in seconds""" + + def __init__(self, round: Optional[int] = None): + self.round = round + self.sum = 0 + self.n = 0 + self.start_time = None + + def start(self): + self.start_time = time.perf_counter() + + def stop(self, n=1, prehook=None): + if self.start_time is not None: + if prehook is not None: + prehook() + delta = time.perf_counter() - self.start_time + self.sum = self.sum + delta + self.n = type_as(self.n, n) + n + + def reset(self): + self.sum = 0 # cumulative time during which stopwatch was active + self.n = 0 # total n across all start/stop + self.start() + + def state_dict(self): + return { + "sum": self.sum, + "n": self.n, + "round": self.round, + } + + def load_state_dict(self, state_dict): + self.sum = state_dict["sum"] + self.n = state_dict["n"] + self.start_time = None + self.round = state_dict.get("round", None) + + @property + def avg(self): + return self.sum / self.n if self.n > 0 else self.sum + + @property + def elapsed_time(self): + if self.start_time is None: + return 0.0 + return time.perf_counter() - self.start_time + + @property + def smoothed_value(self) -> float: + val = self.avg if self.sum > 0 else self.elapsed_time + if self.round is not None and val is not None: + val = safe_round(val, self.round) + return val + + +class MetersDict(OrderedDict): + """A sorted dictionary of :class:`Meters`. + + Meters are sorted according to a priority that is given when the + meter is first added to the dictionary. + """ + + def __init__(self, *args, **kwargs): + super().__init__(*args, **kwargs) + self.priorities = [] + + def __setitem__(self, key, value): + assert key not in self, "MetersDict doesn't support reassignment" + priority, value = value + bisect.insort(self.priorities, (priority, len(self.priorities), key)) + super().__setitem__(key, value) + for _, _, key in self.priorities: # reorder dict to match priorities + self.move_to_end(key) + + def add_meter(self, key, meter, priority): + self.__setitem__(key, (priority, meter)) + + def state_dict(self): + return [ + (pri, key, self[key].__class__.__name__, self[key].state_dict()) + for pri, _, key in self.priorities + # can't serialize DerivedMeter instances + if not isinstance(self[key], MetersDict._DerivedMeter) + ] + + def load_state_dict(self, state_dict): + self.clear() + self.priorities.clear() + for pri, key, meter_cls, meter_state in state_dict: + meter = globals()[meter_cls]() + meter.load_state_dict(meter_state) + self.add_meter(key, meter, pri) + + def get_smoothed_value(self, key: str) -> float: + """Get a single smoothed value.""" + meter = self[key] + if isinstance(meter, MetersDict._DerivedMeter): + return meter.fn(self) + else: + return meter.smoothed_value + + def get_smoothed_values(self) -> Dict[str, float]: + """Get all smoothed values.""" + return OrderedDict( + [ + (key, self.get_smoothed_value(key)) + for key in self.keys() + if not key.startswith("_") + ] + ) + + def reset(self): + """Reset Meter instances.""" + for meter in self.values(): + if isinstance(meter, MetersDict._DerivedMeter): + continue + meter.reset() + + class _DerivedMeter(Meter): + """A Meter whose values are derived from other Meters.""" + + def __init__(self, fn): + self.fn = fn + + def reset(self): + pass diff --git a/data/fairseq/fairseq/logging/metrics.py b/data/fairseq/fairseq/logging/metrics.py new file mode 100644 index 0000000000000000000000000000000000000000..49301f27f84351b83b8c869bac86c78ec9f126e6 --- /dev/null +++ b/data/fairseq/fairseq/logging/metrics.py @@ -0,0 +1,336 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. +""" +A standalone module for aggregating metrics. + +Metrics can be logged from anywhere using the `log_*` functions defined +in this module. The logged values will be aggregated dynamically based +on the aggregation context in which the logging occurs. See the +:func:`aggregate` context manager for more details. +""" + +import contextlib +import uuid +from collections import defaultdict +from typing import Callable, List, Optional + +from .meters import * + + +# Aggregation contexts are considered "active" when inside the scope +# created by the :func:`aggregate` context manager. +_aggregators = OrderedDict() +_active_aggregators = OrderedDict() +_active_aggregators_cnt = defaultdict(lambda: 0) + + +def reset() -> None: + """Reset all metrics aggregators.""" + _aggregators.clear() + _active_aggregators.clear() + _active_aggregators_cnt.clear() + + # The "default" aggregator observes all logged values. + _aggregators["default"] = MetersDict() + _active_aggregators["default"] = _aggregators["default"] + _active_aggregators_cnt["default"] = 1 + + +reset() + + +@contextlib.contextmanager +def aggregate(name: Optional[str] = None, new_root: bool = False): + """Context manager to aggregate metrics under a given name. + + Aggregations can be nested. If *new_root* is ``False``, then logged + metrics will be recorded along the entire stack of nested + aggregators, including a global "default" aggregator. If *new_root* + is ``True``, then this aggregator will be the root of a new + aggregation stack, thus bypassing any parent aggregators. + + Note that aggregation contexts are uniquely identified by their + *name* (e.g., train, valid). Creating a context with an existing + name will reuse the corresponding :class:`MetersDict` instance. + If no name is given, then a temporary aggregator will be created. + + Usage:: + + with metrics.aggregate("train"): + for step, batch in enumerate(epoch): + with metrics.aggregate("train_inner") as agg: + metrics.log_scalar("loss", get_loss(batch)) + if step % log_interval == 0: + print(agg.get_smoothed_value("loss")) + agg.reset() + print(metrics.get_smoothed_values("train")["loss"]) + + Args: + name (str): name of the aggregation. Defaults to a + random/temporary name if not given explicitly. + new_root (bool): make this aggregation the root of a new + aggregation stack. + """ + if name is None: + # generate a temporary name + name = str(uuid.uuid4()) + assert name not in _aggregators + agg = MetersDict() + else: + assert name != "default" + agg = _aggregators.setdefault(name, MetersDict()) + + if new_root: + backup_aggregators = _active_aggregators.copy() + _active_aggregators.clear() + backup_aggregators_cnt = _active_aggregators_cnt.copy() + _active_aggregators_cnt.clear() + + _active_aggregators[name] = agg + _active_aggregators_cnt[name] += 1 + + yield agg + + _active_aggregators_cnt[name] -= 1 + if _active_aggregators_cnt[name] == 0 and name in _active_aggregators: + del _active_aggregators[name] + + if new_root: + _active_aggregators.clear() + _active_aggregators.update(backup_aggregators) + _active_aggregators_cnt.clear() + _active_aggregators_cnt.update(backup_aggregators_cnt) + + +def get_active_aggregators() -> List[MetersDict]: + return list(_active_aggregators.values()) + + +def log_scalar( + key: str, + value: float, + weight: float = 1, + priority: int = 10, + round: Optional[int] = None, +): + """Log a scalar value. + + Args: + key (str): name of the field to log + value (float): value to log + weight (float): weight that this value contributes to the average. + A weight of 0 will always log the latest value. + priority (int): smaller values are logged earlier in the output + round (Optional[int]): number of digits to round to when displaying + """ + for agg in get_active_aggregators(): + if key not in agg: + agg.add_meter(key, AverageMeter(round=round), priority) + agg[key].update(value, weight) + + +def log_scalar_sum( + key: str, + value: float, + priority: int = 10, + round: Optional[int] = None, +): + """Log a scalar value that is summed for reporting. + + Args: + key (str): name of the field to log + value (float): value to log + priority (int): smaller values are logged earlier in the output + round (Optional[int]): number of digits to round to when displaying + """ + for agg in get_active_aggregators(): + if key not in agg: + agg.add_meter(key, SumMeter(round=round), priority) + agg[key].update(value) + + +def log_concat_tensor( + key: str, + value: torch.Tensor, + priority: int = 10, + dim: int = 0, +): + """Log a scalar value that is summed for reporting. + + Args: + key (str): name of the field to log + value (float): value to log + priority (int): smaller values are logged earlier in the output + round (Optional[int]): number of digits to round to when displaying + """ + for agg in get_active_aggregators(): + if key not in agg: + agg.add_meter(key, ConcatTensorMeter(dim=dim), priority) + agg[key].update(value) + + +def log_derived(key: str, fn: Callable[[MetersDict], float], priority: int = 20): + """Log a scalar value derived from other meters. + + Args: + key (str): name of the field to log + fn (Callable[[MetersDict], float]): function that takes a single + argument *meters* and returns the derived value + priority (int): smaller values are logged earlier in the output + """ + for agg in get_active_aggregators(): + if key not in agg: + agg.add_meter(key, MetersDict._DerivedMeter(fn), priority) + + +def log_speed( + key: str, + value: float, + priority: int = 30, + round: Optional[int] = None, +): + """Log the rate of some quantity per second. + + Args: + key (str): name of the field to log + value (float): value to log + priority (int): smaller values are logged earlier in the output + round (Optional[int]): number of digits to round to when displaying + """ + for agg in get_active_aggregators(): + if key not in agg: + agg.add_meter(key, TimeMeter(round=round), priority) + agg[key].reset() # reset meter on the first call + else: + agg[key].update(value) + + +def log_start_time(key: str, priority: int = 40, round: Optional[int] = None): + """Log the duration of some event in seconds. + + The duration will be computed once :func:`log_stop_time` is called. + + Args: + key (str): name of the field to log + priority (int): smaller values are logged earlier in the output + round (Optional[int]): number of digits to round to when displaying + """ + for agg in get_active_aggregators(): + if key not in agg: + agg.add_meter(key, StopwatchMeter(round=round), priority) + agg[key].start() + + +def log_stop_time(key: str, weight: float = 0.0, prehook=None): + """Log the duration of some event in seconds. + + The duration will be computed since :func:`log_start_time` was called. + Set weight > 0 to report the average time instead of the sum. + + Args: + key (str): name of the field to log + weight (float): weight that this time contributes to the average + prehook (function, no arguments): will be called before the timer + is stopped. For example, use prehook=torch.cuda.synchronize to + make sure all gpu operations are done before timer is stopped. + """ + for agg in get_active_aggregators(): + if key in agg: + agg[key].stop(weight, prehook) + + +def log_custom( + new_meter_fn: Callable[[], Meter], + key: str, + *args, + priority: int = 50, + **kwargs, +): + """Log using a custom Meter. + + Any extra *args* or *kwargs* will be passed through to the Meter's + *update* method. + + Args: + new_meter_fn (Callable[[], Meter]): function that returns a new + Meter instance + key (str): name of the field to log + priority (int): smaller values are logged earlier in the output + """ + for agg in get_active_aggregators(): + if key not in agg: + agg.add_meter(key, new_meter_fn(), priority) + agg[key].update(*args, **kwargs) + + +def reset_meter(name: str, key: str) -> None: + """Reset Meter instance aggregated under a given *name* and *key*.""" + meter = get_meter(name, key) + if meter is not None: + meter.reset() + + +def reset_meters(name: str) -> None: + """Reset Meter instances aggregated under a given *name*.""" + meters = get_meters(name) + if meters is not None: + meters.reset() + + +def get_meter(name: str, key: str) -> Meter: + """Get a single Meter instance aggregated under *name* and *key*. + + Returns: + Meter or None if no metrics have been logged under *name* and *key*. + """ + if name not in _aggregators: + return None + return _aggregators[name].get(key, None) + + +def get_meters(name: str) -> MetersDict: + """Get Meter instances aggregated under a given *name*. + + Returns: + MetersDict or None if no metrics have been logged under *name*. + """ + return _aggregators.get(name, None) + + +def get_smoothed_value(name: str, key: str) -> float: + """Get a single smoothed value. + + Raises: + KeyError: if no metrics have been logged under *name* and *key*. + """ + return _aggregators[name].get_smoothed_value(key) + + +def get_smoothed_values(name: str) -> Dict[str, float]: + """Get smoothed values aggregated under a given *name*. + + Raises: + KeyError: if no metrics have been logged under *name*. + """ + return _aggregators[name].get_smoothed_values() + + +def state_dict(): + return OrderedDict([(name, agg.state_dict()) for name, agg in _aggregators.items()]) + + +def load_state_dict(state_dict): + for name, agg_state in state_dict.items(): + _aggregators[name] = MetersDict() + _aggregators[name].load_state_dict(agg_state) + + +def xla_metrics_report(): + try: + import torch_xla.debug.metrics as met + + print(met.metrics_report()) + except ImportError: + return diff --git a/data/fairseq/fairseq/logging/progress_bar.py b/data/fairseq/fairseq/logging/progress_bar.py new file mode 100644 index 0000000000000000000000000000000000000000..4c64b61bad6edbf4b9ff5bcc2f26952e8b1bfc9c --- /dev/null +++ b/data/fairseq/fairseq/logging/progress_bar.py @@ -0,0 +1,582 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +""" +Wrapper around various loggers and progress bars (e.g., tqdm). +""" + +import atexit +import json +import logging +import os +import sys +from collections import OrderedDict +from contextlib import contextmanager +from numbers import Number +from typing import Optional + +import torch + +from .meters import AverageMeter, StopwatchMeter, TimeMeter + +logger = logging.getLogger(__name__) + + +def progress_bar( + iterator, + log_format: Optional[str] = None, + log_interval: int = 100, + log_file: Optional[str] = None, + epoch: Optional[int] = None, + prefix: Optional[str] = None, + aim_repo: Optional[str] = None, + aim_run_hash: Optional[str] = None, + aim_param_checkpoint_dir: Optional[str] = None, + tensorboard_logdir: Optional[str] = None, + default_log_format: str = "tqdm", + wandb_project: Optional[str] = None, + wandb_run_name: Optional[str] = None, + azureml_logging: Optional[bool] = False, +): + if log_format is None: + log_format = default_log_format + if log_file is not None: + handler = logging.FileHandler(filename=log_file) + logger.addHandler(handler) + + if log_format == "tqdm" and not sys.stderr.isatty(): + log_format = "simple" + + if log_format == "json": + bar = JsonProgressBar(iterator, epoch, prefix, log_interval) + elif log_format == "none": + bar = NoopProgressBar(iterator, epoch, prefix) + elif log_format == "simple": + bar = SimpleProgressBar(iterator, epoch, prefix, log_interval) + elif log_format == "tqdm": + bar = TqdmProgressBar(iterator, epoch, prefix) + else: + raise ValueError("Unknown log format: {}".format(log_format)) + + if aim_repo: + bar = AimProgressBarWrapper( + bar, + aim_repo=aim_repo, + aim_run_hash=aim_run_hash, + aim_param_checkpoint_dir=aim_param_checkpoint_dir, + ) + + if tensorboard_logdir: + try: + # [FB only] custom wrapper for TensorBoard + import palaas # noqa + + from .fb_tbmf_wrapper import FbTbmfWrapper + + bar = FbTbmfWrapper(bar, log_interval) + except ImportError: + bar = TensorboardProgressBarWrapper(bar, tensorboard_logdir) + + if wandb_project: + bar = WandBProgressBarWrapper(bar, wandb_project, run_name=wandb_run_name) + + if azureml_logging: + bar = AzureMLProgressBarWrapper(bar) + + return bar + + +def build_progress_bar( + args, + iterator, + epoch: Optional[int] = None, + prefix: Optional[str] = None, + default: str = "tqdm", + no_progress_bar: str = "none", +): + """Legacy wrapper that takes an argparse.Namespace.""" + if getattr(args, "no_progress_bar", False): + default = no_progress_bar + if getattr(args, "distributed_rank", 0) == 0: + tensorboard_logdir = getattr(args, "tensorboard_logdir", None) + else: + tensorboard_logdir = None + return progress_bar( + iterator, + log_format=args.log_format, + log_interval=args.log_interval, + epoch=epoch, + prefix=prefix, + tensorboard_logdir=tensorboard_logdir, + default_log_format=default, + ) + + +def format_stat(stat): + if isinstance(stat, Number): + stat = "{:g}".format(stat) + elif isinstance(stat, AverageMeter): + stat = "{:.3f}".format(stat.avg) + elif isinstance(stat, TimeMeter): + stat = "{:g}".format(round(stat.avg)) + elif isinstance(stat, StopwatchMeter): + stat = "{:g}".format(round(stat.sum)) + elif torch.is_tensor(stat): + stat = stat.tolist() + return stat + + +class BaseProgressBar(object): + """Abstract class for progress bars.""" + + def __init__(self, iterable, epoch=None, prefix=None): + self.iterable = iterable + self.n = getattr(iterable, "n", 0) + self.epoch = epoch + self.prefix = "" + if epoch is not None: + self.prefix += "epoch {:03d}".format(epoch) + if prefix is not None: + self.prefix += (" | " if self.prefix != "" else "") + prefix + + def __len__(self): + return len(self.iterable) + + def __enter__(self): + return self + + def __exit__(self, *exc): + return False + + def __iter__(self): + raise NotImplementedError + + def log(self, stats, tag=None, step=None): + """Log intermediate stats according to log_interval.""" + raise NotImplementedError + + def print(self, stats, tag=None, step=None): + """Print end-of-epoch stats.""" + raise NotImplementedError + + def update_config(self, config): + """Log latest configuration.""" + pass + + def _str_commas(self, stats): + return ", ".join(key + "=" + stats[key].strip() for key in stats.keys()) + + def _str_pipes(self, stats): + return " | ".join(key + " " + stats[key].strip() for key in stats.keys()) + + def _format_stats(self, stats): + postfix = OrderedDict(stats) + # Preprocess stats according to datatype + for key in postfix.keys(): + postfix[key] = str(format_stat(postfix[key])) + return postfix + + +@contextmanager +def rename_logger(logger, new_name): + old_name = logger.name + if new_name is not None: + logger.name = new_name + yield logger + logger.name = old_name + + +class JsonProgressBar(BaseProgressBar): + """Log output in JSON format.""" + + def __init__(self, iterable, epoch=None, prefix=None, log_interval=1000): + super().__init__(iterable, epoch, prefix) + self.log_interval = log_interval + self.i = None + self.size = None + + def __iter__(self): + self.size = len(self.iterable) + for i, obj in enumerate(self.iterable, start=self.n): + self.i = i + yield obj + + def log(self, stats, tag=None, step=None): + """Log intermediate stats according to log_interval.""" + step = step or self.i or 0 + if step > 0 and self.log_interval is not None and step % self.log_interval == 0: + update = ( + self.epoch - 1 + (self.i + 1) / float(self.size) + if self.epoch is not None + else None + ) + stats = self._format_stats(stats, epoch=self.epoch, update=update) + with rename_logger(logger, tag): + logger.info(json.dumps(stats)) + + def print(self, stats, tag=None, step=None): + """Print end-of-epoch stats.""" + self.stats = stats + if tag is not None: + self.stats = OrderedDict( + [(tag + "_" + k, v) for k, v in self.stats.items()] + ) + stats = self._format_stats(self.stats, epoch=self.epoch) + with rename_logger(logger, tag): + logger.info(json.dumps(stats)) + + def _format_stats(self, stats, epoch=None, update=None): + postfix = OrderedDict() + if epoch is not None: + postfix["epoch"] = epoch + if update is not None: + postfix["update"] = round(update, 3) + # Preprocess stats according to datatype + for key in stats.keys(): + postfix[key] = format_stat(stats[key]) + return postfix + + +class NoopProgressBar(BaseProgressBar): + """No logging.""" + + def __init__(self, iterable, epoch=None, prefix=None): + super().__init__(iterable, epoch, prefix) + + def __iter__(self): + for obj in self.iterable: + yield obj + + def log(self, stats, tag=None, step=None): + """Log intermediate stats according to log_interval.""" + pass + + def print(self, stats, tag=None, step=None): + """Print end-of-epoch stats.""" + pass + + +class SimpleProgressBar(BaseProgressBar): + """A minimal logger for non-TTY environments.""" + + def __init__(self, iterable, epoch=None, prefix=None, log_interval=1000): + super().__init__(iterable, epoch, prefix) + self.log_interval = log_interval + self.i = None + self.size = None + + def __iter__(self): + self.size = len(self.iterable) + for i, obj in enumerate(self.iterable, start=self.n): + self.i = i + yield obj + + def log(self, stats, tag=None, step=None): + """Log intermediate stats according to log_interval.""" + step = step or self.i or 0 + if step > 0 and self.log_interval is not None and step % self.log_interval == 0: + stats = self._format_stats(stats) + postfix = self._str_commas(stats) + with rename_logger(logger, tag): + logger.info( + "{}: {:5d} / {:d} {}".format( + self.prefix, self.i + 1, self.size, postfix + ) + ) + + def print(self, stats, tag=None, step=None): + """Print end-of-epoch stats.""" + postfix = self._str_pipes(self._format_stats(stats)) + with rename_logger(logger, tag): + logger.info("{} | {}".format(self.prefix, postfix)) + + +class TqdmProgressBar(BaseProgressBar): + """Log to tqdm.""" + + def __init__(self, iterable, epoch=None, prefix=None): + super().__init__(iterable, epoch, prefix) + from tqdm import tqdm + + self.tqdm = tqdm( + iterable, + self.prefix, + leave=False, + disable=(logger.getEffectiveLevel() > logging.INFO), + ) + + def __iter__(self): + return iter(self.tqdm) + + def log(self, stats, tag=None, step=None): + """Log intermediate stats according to log_interval.""" + self.tqdm.set_postfix(self._format_stats(stats), refresh=False) + + def print(self, stats, tag=None, step=None): + """Print end-of-epoch stats.""" + postfix = self._str_pipes(self._format_stats(stats)) + with rename_logger(logger, tag): + logger.info("{} | {}".format(self.prefix, postfix)) + + +try: + import functools + + from aim import Repo as AimRepo + + @functools.lru_cache() + def get_aim_run(repo, run_hash): + from aim import Run + + return Run(run_hash=run_hash, repo=repo) + +except ImportError: + get_aim_run = None + AimRepo = None + + +class AimProgressBarWrapper(BaseProgressBar): + """Log to Aim.""" + + def __init__(self, wrapped_bar, aim_repo, aim_run_hash, aim_param_checkpoint_dir): + self.wrapped_bar = wrapped_bar + + if get_aim_run is None: + self.run = None + logger.warning("Aim not found, please install with: pip install aim") + else: + logger.info(f"Storing logs at Aim repo: {aim_repo}") + + if not aim_run_hash: + # Find run based on save_dir parameter + query = f"run.checkpoint.save_dir == '{aim_param_checkpoint_dir}'" + try: + runs_generator = AimRepo(aim_repo).query_runs(query) + run = next(runs_generator.iter_runs()) + aim_run_hash = run.run.hash + except Exception: + pass + + if aim_run_hash: + logger.info(f"Appending to run: {aim_run_hash}") + + self.run = get_aim_run(aim_repo, aim_run_hash) + + def __iter__(self): + return iter(self.wrapped_bar) + + def log(self, stats, tag=None, step=None): + """Log intermediate stats to Aim.""" + self._log_to_aim(stats, tag, step) + self.wrapped_bar.log(stats, tag=tag, step=step) + + def print(self, stats, tag=None, step=None): + """Print end-of-epoch stats.""" + self._log_to_aim(stats, tag, step) + self.wrapped_bar.print(stats, tag=tag, step=step) + + def update_config(self, config): + """Log latest configuration.""" + if self.run is not None: + for key in config: + self.run.set(key, config[key], strict=False) + self.wrapped_bar.update_config(config) + + def _log_to_aim(self, stats, tag=None, step=None): + if self.run is None: + return + + if step is None: + step = stats["num_updates"] + + if "train" in tag: + context = {"tag": tag, "subset": "train"} + elif "val" in tag: + context = {"tag": tag, "subset": "val"} + else: + context = {"tag": tag} + + for key in stats.keys() - {"num_updates"}: + self.run.track(stats[key], name=key, step=step, context=context) + + +try: + _tensorboard_writers = {} + from torch.utils.tensorboard import SummaryWriter +except ImportError: + try: + from tensorboardX import SummaryWriter + except ImportError: + SummaryWriter = None + + +def _close_writers(): + for w in _tensorboard_writers.values(): + w.close() + + +atexit.register(_close_writers) + + +class TensorboardProgressBarWrapper(BaseProgressBar): + """Log to tensorboard.""" + + def __init__(self, wrapped_bar, tensorboard_logdir): + self.wrapped_bar = wrapped_bar + self.tensorboard_logdir = tensorboard_logdir + + if SummaryWriter is None: + logger.warning( + "tensorboard not found, please install with: pip install tensorboard" + ) + + def _writer(self, key): + if SummaryWriter is None: + return None + _writers = _tensorboard_writers + if key not in _writers: + _writers[key] = SummaryWriter(os.path.join(self.tensorboard_logdir, key)) + _writers[key].add_text("sys.argv", " ".join(sys.argv)) + return _writers[key] + + def __iter__(self): + return iter(self.wrapped_bar) + + def log(self, stats, tag=None, step=None): + """Log intermediate stats to tensorboard.""" + self._log_to_tensorboard(stats, tag, step) + self.wrapped_bar.log(stats, tag=tag, step=step) + + def print(self, stats, tag=None, step=None): + """Print end-of-epoch stats.""" + self._log_to_tensorboard(stats, tag, step) + self.wrapped_bar.print(stats, tag=tag, step=step) + + def update_config(self, config): + """Log latest configuration.""" + # TODO add hparams to Tensorboard + self.wrapped_bar.update_config(config) + + def _log_to_tensorboard(self, stats, tag=None, step=None): + writer = self._writer(tag or "") + if writer is None: + return + if step is None: + step = stats["num_updates"] + for key in stats.keys() - {"num_updates"}: + if isinstance(stats[key], AverageMeter): + writer.add_scalar(key, stats[key].val, step) + elif isinstance(stats[key], Number): + writer.add_scalar(key, stats[key], step) + elif torch.is_tensor(stats[key]) and stats[key].numel() == 1: + writer.add_scalar(key, stats[key].item(), step) + writer.flush() + + +try: + import wandb +except ImportError: + wandb = None + + +class WandBProgressBarWrapper(BaseProgressBar): + """Log to Weights & Biases.""" + + def __init__(self, wrapped_bar, wandb_project, run_name=None): + self.wrapped_bar = wrapped_bar + if wandb is None: + logger.warning("wandb not found, pip install wandb") + return + + # reinit=False to ensure if wandb.init() is called multiple times + # within one process it still references the same run + wandb.init(project=wandb_project, reinit=False, name=run_name) + + def __iter__(self): + return iter(self.wrapped_bar) + + def log(self, stats, tag=None, step=None): + """Log intermediate stats to tensorboard.""" + self._log_to_wandb(stats, tag, step) + self.wrapped_bar.log(stats, tag=tag, step=step) + + def print(self, stats, tag=None, step=None): + """Print end-of-epoch stats.""" + self._log_to_wandb(stats, tag, step) + self.wrapped_bar.print(stats, tag=tag, step=step) + + def update_config(self, config): + """Log latest configuration.""" + if wandb is not None: + wandb.config.update(config) + self.wrapped_bar.update_config(config) + + def _log_to_wandb(self, stats, tag=None, step=None): + if wandb is None: + return + if step is None: + step = stats["num_updates"] + + prefix = "" if tag is None else tag + "/" + + for key in stats.keys() - {"num_updates"}: + if isinstance(stats[key], AverageMeter): + wandb.log({prefix + key: stats[key].val}, step=step) + elif isinstance(stats[key], Number): + wandb.log({prefix + key: stats[key]}, step=step) + + +try: + from azureml.core import Run +except ImportError: + Run = None + + +class AzureMLProgressBarWrapper(BaseProgressBar): + """Log to Azure ML""" + + def __init__(self, wrapped_bar): + self.wrapped_bar = wrapped_bar + if Run is None: + logger.warning("azureml.core not found, pip install azureml-core") + return + self.run = Run.get_context() + + def __exit__(self, *exc): + if Run is not None: + self.run.complete() + return False + + def __iter__(self): + return iter(self.wrapped_bar) + + def log(self, stats, tag=None, step=None): + """Log intermediate stats to AzureML""" + self._log_to_azureml(stats, tag, step) + self.wrapped_bar.log(stats, tag=tag, step=step) + + def print(self, stats, tag=None, step=None): + """Print end-of-epoch stats""" + self._log_to_azureml(stats, tag, step) + self.wrapped_bar.print(stats, tag=tag, step=step) + + def update_config(self, config): + """Log latest configuration.""" + self.wrapped_bar.update_config(config) + + def _log_to_azureml(self, stats, tag=None, step=None): + if Run is None: + return + if step is None: + step = stats["num_updates"] + + prefix = "" if tag is None else tag + "/" + + for key in stats.keys() - {"num_updates"}: + name = prefix + key + if isinstance(stats[key], AverageMeter): + self.run.log_row(name=name, **{"step": step, key: stats[key].val}) + elif isinstance(stats[key], Number): + self.run.log_row(name=name, **{"step": step, key: stats[key]}) diff --git a/data/fairseq/fairseq/model_parallel/__init__.py b/data/fairseq/fairseq/model_parallel/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..69f21684872f72ae8ee26d9ff7d2d2b6e6d526c3 --- /dev/null +++ b/data/fairseq/fairseq/model_parallel/__init__.py @@ -0,0 +1,6 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from . import criterions, models, modules # noqa diff --git a/data/fairseq/fairseq/model_parallel/__pycache__/__init__.cpython-38.pyc b/data/fairseq/fairseq/model_parallel/__pycache__/__init__.cpython-38.pyc new file mode 100644 index 0000000000000000000000000000000000000000..83d81e9c9547ca0d4200b62f393b28205785d8a9 Binary files /dev/null and b/data/fairseq/fairseq/model_parallel/__pycache__/__init__.cpython-38.pyc differ diff --git a/data/fairseq/fairseq/model_parallel/__pycache__/megatron_trainer.cpython-38.pyc b/data/fairseq/fairseq/model_parallel/__pycache__/megatron_trainer.cpython-38.pyc new file mode 100644 index 0000000000000000000000000000000000000000..cd64c8712cdb2455982c82353b12e87b63389f2c Binary files /dev/null and b/data/fairseq/fairseq/model_parallel/__pycache__/megatron_trainer.cpython-38.pyc differ diff --git a/data/fairseq/fairseq/model_parallel/megatron_trainer.py b/data/fairseq/fairseq/model_parallel/megatron_trainer.py new file mode 100644 index 0000000000000000000000000000000000000000..aedf608bce21d11d0a1e9646d9c373aae198dce6 --- /dev/null +++ b/data/fairseq/fairseq/model_parallel/megatron_trainer.py @@ -0,0 +1,75 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +""" +Train a network across multiple GPUs. +""" + +from fairseq.dataclass.configs import FairseqConfig +from fairseq.distributed import utils as distributed_utils +from fairseq.trainer import Trainer + +try: + from fairseq.model_parallel.megatron.mpu import ( + get_data_parallel_rank, + get_data_parallel_world_size, + get_model_parallel_src_rank, + get_cuda_rng_tracker, + ) + + has_megatron_submodule = True +except (ImportError, ModuleNotFoundError): + has_megatron_submodule = False + + +class MegatronTrainer(Trainer): + """Main class for model parallel with data parallel training.""" + + def __init__(self, cfg: FairseqConfig, task, model, criterion, **kwargs): + if not has_megatron_submodule: + raise ImportError( + "\n\nPlease install the megatron submodule:" + "\n\n git submodule update --init " + "fairseq/model_parallel/megatron" + ) + super().__init__(cfg, task, model, criterion, **kwargs) + + def clip_grad_norm(self, clip_norm): + def _aggregate_model_parallel_grad_norm(total_norm): + total_norm = total_norm**2 + distributed_utils.all_reduce( + total_norm, group=distributed_utils.get_model_parallel_group() + ) + total_norm = total_norm**0.5 + return total_norm + + return self.optimizer.clip_grad_norm( + clip_norm, + aggregate_norm_fn=_aggregate_model_parallel_grad_norm, + ) + + def save_checkpoint(self, filename, extra_state): + """Save all training state in a checkpoint file.""" + extra_state["rng_tracker_states"] = get_cuda_rng_tracker().get_states() + super().save_checkpoint(filename, extra_state) + + def load_checkpoint( + self, + filename, + reset_optimizer=False, + reset_lr_scheduler=False, + optimizer_overrides=None, + reset_meters=False, + ): + extra_state = super().load_checkpoint( + filename, + reset_optimizer=reset_optimizer, + reset_lr_scheduler=reset_lr_scheduler, + optimizer_overrides=optimizer_overrides, + reset_meters=reset_meters, + ) + if extra_state is not None and "rng_tracker_states" in extra_state: + get_cuda_rng_tracker().set_states(extra_state["rng_tracker_states"]) + return extra_state diff --git a/data/fairseq/fairseq/model_parallel/models/__init__.py b/data/fairseq/fairseq/model_parallel/models/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..3532479e52a0e1f1ba204c6f5d51c71c98ee5df0 --- /dev/null +++ b/data/fairseq/fairseq/model_parallel/models/__init__.py @@ -0,0 +1,20 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import importlib +import os + + +# automatically import any Python files in the models/ directory +models_dir = os.path.dirname(__file__) +for file in os.listdir(models_dir): + path = os.path.join(models_dir, file) + if ( + not file.startswith("_") + and not file.startswith(".") + and (file.endswith(".py") or os.path.isdir(path)) + ): + model_name = file[: file.find(".py")] if file.endswith(".py") else file + module = importlib.import_module("fairseq.model_parallel.models." + model_name) diff --git a/data/fairseq/fairseq/model_parallel/models/transformer.py b/data/fairseq/fairseq/model_parallel/models/transformer.py new file mode 100644 index 0000000000000000000000000000000000000000..cf3b2e8baf01389a34056cc68cbf6ad1d4475707 --- /dev/null +++ b/data/fairseq/fairseq/model_parallel/models/transformer.py @@ -0,0 +1,121 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import logging + +import torch.nn as nn + +from fairseq.model_parallel.modules import ( + ModelParallelTransformerDecoderLayer, + ModelParallelTransformerEncoderLayer, +) +from fairseq.models import register_model +from fairseq.models.transformer import ( + TransformerDecoder, + TransformerEncoder, + TransformerModel, +) + +try: + from fairseq.model_parallel.megatron.mpu import ( + VocabParallelEmbedding, + copy_to_model_parallel_region, + gather_from_model_parallel_region, + ) + + has_megatron_submodule = True +except (ImportError, ModuleNotFoundError): + has_megatron_submodule = False + + +logger = logging.getLogger(__name__) + + +@register_model("model_parallel_transformer") +class ModelParallelTransformerModel(TransformerModel): + """ + Model parallel Transformer model. + """ + + @classmethod + def build_embedding(cls, args, dictionary, embed_dim, path=None): + if not has_megatron_submodule: + raise ImportError( + "\n\nPlease install the megatron submodule:" + "\n\n git submodule update --init " + "fairseq/model_parallel/megatron" + ) + dictionary.pad_to_multiple_(args.model_parallel_size * 8) + num_embeddings = len(dictionary) + padding_idx = dictionary.pad() + + def _vocab_init(tensor, **kwargs): + nn.init.normal_(tensor, mean=0, std=num_embeddings**-0.5) + nn.init.constant_(tensor[1], 0) + + emb = VocabParallelEmbedding( + num_embeddings, embed_dim, padding_idx, init_method=_vocab_init + ) + # if provided, load from preloaded dictionaries + if path: + raise NotImplementedError( + "Loading of embedding from path is not supported for model parallel" + ) + return emb + + @classmethod + def build_encoder(cls, args, src_dict, embed_tokens): + return ModelParallelTransformerEncoder(args, src_dict, embed_tokens) + + @classmethod + def build_decoder(cls, args, tgt_dict, embed_tokens): + return ModelParallelTransformerDecoder( + args, + tgt_dict, + embed_tokens, + no_encoder_attn=getattr(args, "no_cross_attention", False), + ) + + +class ModelParallelTransformerEncoder(TransformerEncoder): + """ + Model parallel Transformer encoder consisting of *args.encoder_layers* layers. Each layer + is a :class:`ModelParallelTransformerEncoderLayer`. + """ + + def __init__(self, args, dictionary, embed_tokens): + super().__init__(args, dictionary, embed_tokens) + + if args.no_final_layer_norm: + self.layer_norm = None + + def build_encoder_layer(self, args): + return ModelParallelTransformerEncoderLayer(args) + + +class ModelParallelTransformerDecoder(TransformerDecoder): + """ + Model Parallel Transformer decoder consisting of *args.decoder_layers* layers. Each layer + is a :class:`ModelParallelTransformerDecoderLayer`. + """ + + def build_decoder_layer(self, args, no_encoder_attn=False): + return ModelParallelTransformerDecoderLayer(args, no_encoder_attn) + + def output_layer(self, features, **kwargs): + """Project features to the vocabulary size.""" + if not self.share_input_output_embed: + raise NotImplementedError( + "Model parallel training currently requires --share-decoder-input-output-embed" + ) + + features = copy_to_model_parallel_region(features) + + # project back to size of vocabulary + x = self.output_projection(features) + + if getattr(self.args, "criterion") != "vocab_parallel_cross_entropy": + x = gather_from_model_parallel_region(x).contiguous() + return x diff --git a/data/fairseq/fairseq/model_parallel/models/transformer_lm.py b/data/fairseq/fairseq/model_parallel/models/transformer_lm.py new file mode 100644 index 0000000000000000000000000000000000000000..03e4dbe26393eedfb71da94d6675b08cbdb8626d --- /dev/null +++ b/data/fairseq/fairseq/model_parallel/models/transformer_lm.py @@ -0,0 +1,169 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import torch.nn as nn + +from fairseq.model_parallel.models.transformer import ModelParallelTransformerDecoder +from fairseq.models import register_model, register_model_architecture +from fairseq.models.transformer_lm import TransformerLanguageModel + +try: + from fairseq.model_parallel.megatron.mpu import VocabParallelEmbedding + + has_megatron_submodule = True +except (ImportError, ModuleNotFoundError): + has_megatron_submodule = False + + +DEFAULT_MAX_TARGET_POSITIONS = 1024 + + +@register_model("model_parallel_transformer_lm") +class ModelParallelTransformerLanguageModel(TransformerLanguageModel): + @staticmethod + def add_args(parser): + TransformerLanguageModel.add_args(parser) + + @classmethod + def build_model(cls, args, task): + """Build a new model instance.""" + if not has_megatron_submodule: + raise ImportError( + "\n\nPlease install the megatron submodule:" + "\n\n git submodule update --init " + "fairseq/model_parallel/megatron" + ) + + # make sure all arguments are present in older models + base_lm_architecture(args) + + task.source_dictionary.pad_to_multiple_(args.model_parallel_size * 8) + task.target_dictionary.pad_to_multiple_(args.model_parallel_size * 8) + + if args.decoder_layers_to_keep: + args.decoder_layers = len(args.decoder_layers_to_keep.split(",")) + + if getattr(args, "max_target_positions", None) is None: + args.max_target_positions = getattr( + args, "tokens_per_sample", DEFAULT_MAX_TARGET_POSITIONS + ) + + if args.character_embeddings: + raise NotImplementedError( + "Character embeddings is not supported for model parallel" + ) + elif args.adaptive_input: + raise NotImplementedError( + "Adaptive input is not supported for model parallel" + ) + else: + embed_tokens = cls.build_embedding( + args, task.source_dictionary, args.decoder_input_dim + ) + + decoder = ModelParallelTransformerDecoder( + args, + task.target_dictionary, + embed_tokens, + no_encoder_attn=True, + ) + return cls(decoder) + + @classmethod + def build_embedding(cls, args, dictionary, embed_dim, path=None): + def _vocab_init(tensor, **kwargs): + nn.init.normal_(tensor, mean=0, std=embed_dim**-0.5) + nn.init.constant_(tensor[1], 0) + + embed_tokens = VocabParallelEmbedding( + len(dictionary), embed_dim, dictionary.pad(), init_method=_vocab_init + ) + return embed_tokens + + +def base_lm_architecture(args): + # backward compatibility for older model checkpoints + if hasattr(args, "no_tie_adaptive_proj"): + # previous models defined --no-tie-adaptive-proj, so use the existence of + # that option to determine if this is an "old" model checkpoint + args.no_decoder_final_norm = True # old models always set this to True + if args.no_tie_adaptive_proj is False: + args.tie_adaptive_proj = True + if hasattr(args, "decoder_final_norm"): + args.no_decoder_final_norm = not args.decoder_final_norm + + args.activation_fn = getattr(args, "activation_fn", "relu") + args.dropout = getattr(args, "dropout", 0.1) + args.attention_dropout = getattr(args, "attention_dropout", 0.0) + args.activation_dropout = getattr(args, "activation_dropout", 0.0) + args.relu_dropout = getattr(args, "relu_dropout", 0.0) + args.decoder_embed_dim = getattr(args, "decoder_embed_dim", 512) + args.decoder_output_dim = getattr( + args, "decoder_output_dim", args.decoder_embed_dim + ) + args.decoder_input_dim = getattr(args, "decoder_input_dim", args.decoder_embed_dim) + args.decoder_ffn_embed_dim = getattr(args, "decoder_ffn_embed_dim", 2048) + args.decoder_layers = getattr(args, "decoder_layers", 6) + args.decoder_attention_heads = getattr(args, "decoder_attention_heads", 8) + # Model training is not stable without this + args.decoder_normalize_before = True + args.no_decoder_final_norm = getattr(args, "no_decoder_final_norm", False) + args.adaptive_softmax_cutoff = getattr(args, "adaptive_softmax_cutoff", None) + args.adaptive_softmax_dropout = getattr(args, "adaptive_softmax_dropout", 0) + args.adaptive_softmax_factor = getattr(args, "adaptive_softmax_factor", 4) + args.no_token_positional_embeddings = getattr( + args, "no_token_positional_embeddings", False + ) + args.share_decoder_input_output_embed = getattr( + args, "share_decoder_input_output_embed", False + ) + args.character_embeddings = getattr(args, "character_embeddings", False) + args.character_filters = getattr( + args, + "character_filters", + "[(1, 64), (2, 128), (3, 192), (4, 256), (5, 256), (6, 256), (7, 256)]", + ) + args.character_embedding_dim = getattr(args, "character_embedding_dim", 4) + args.char_embedder_highway_layers = getattr(args, "char_embedder_highway_layers", 2) + args.adaptive_input = getattr(args, "adaptive_input", False) + args.adaptive_input_factor = getattr(args, "adaptive_input_factor", 4) + args.adaptive_input_cutoff = getattr(args, "adaptive_input_cutoff", None) + args.tie_adaptive_weights = getattr(args, "tie_adaptive_weights", False) + args.tie_adaptive_proj = getattr(args, "tie_adaptive_proj", False) + args.decoder_learned_pos = getattr(args, "decoder_learned_pos", False) + args.decoder_layerdrop = getattr(args, "decoder_layerdrop", 0.0) + args.decoder_layers_to_keep = getattr(args, "decoder_layers_to_keep", None) + args.layernorm_embedding = getattr(args, "layernorm_embedding", False) + args.no_scale_embedding = getattr(args, "no_scale_embedding", False) + args.quant_noise_pq = getattr(args, "quant_noise_pq", 0.0) + args.quant_noise_pq_block_size = getattr(args, "quant_noise_pq_block_size", 8) + args.quant_noise_scalar = getattr(args, "quant_noise_scalar", 0.0) + args.add_bos_token = getattr(args, "add_bos_token", False) + + +@register_model_architecture("model_parallel_transformer_lm", "transformer_lm_megatron") +def transformer_lm_megatron(args): + args.decoder_embed_dim = getattr(args, "decoder_embed_dim", 3072) + args.decoder_ffn_embed_dim = getattr(args, "decoder_ffn_embed_dim", 3072 * 4) + args.decoder_layers = getattr(args, "decoder_layers", 72) + args.decoder_attention_heads = getattr(args, "decoder_attention_heads", 32) + args.dropout = getattr(args, "dropout", 0.1) + args.attention_dropout = getattr(args, "attention_dropout", 0.1) + args.activation_fn = getattr(args, "activation_fn", "gelu") + base_lm_architecture(args) + + +@register_model_architecture( + "model_parallel_transformer_lm", "transformer_lm_megatron_11b" +) +def transformer_lm_megatron_11b(args): + args.decoder_embed_dim = getattr(args, "decoder_embed_dim", 3072) + args.decoder_ffn_embed_dim = getattr(args, "decoder_ffn_embed_dim", 3072 * 6) + args.decoder_layers = getattr(args, "decoder_layers", 72) + args.decoder_attention_heads = getattr(args, "decoder_attention_heads", 32) + args.dropout = getattr(args, "dropout", 0.1) + args.attention_dropout = getattr(args, "attention_dropout", 0.1) + args.activation_fn = getattr(args, "activation_fn", "gelu") + base_lm_architecture(args) diff --git a/data/fairseq/fairseq/models/__pycache__/__init__.cpython-38.pyc b/data/fairseq/fairseq/models/__pycache__/__init__.cpython-38.pyc new file mode 100644 index 0000000000000000000000000000000000000000..f22b65e8db7775aa49ecb02e754386145466871f Binary files /dev/null and b/data/fairseq/fairseq/models/__pycache__/__init__.cpython-38.pyc differ diff --git a/data/fairseq/fairseq/models/__pycache__/composite_encoder.cpython-38.pyc b/data/fairseq/fairseq/models/__pycache__/composite_encoder.cpython-38.pyc new file mode 100644 index 0000000000000000000000000000000000000000..42967ea9f0c4ec28162cb9a72602c87dd269aa24 Binary files /dev/null and b/data/fairseq/fairseq/models/__pycache__/composite_encoder.cpython-38.pyc differ diff --git a/data/fairseq/fairseq/models/__pycache__/distributed_fairseq_model.cpython-38.pyc b/data/fairseq/fairseq/models/__pycache__/distributed_fairseq_model.cpython-38.pyc new file mode 100644 index 0000000000000000000000000000000000000000..67d119f41ee972d5d56e31556d24f55ad6f6e333 Binary files /dev/null and b/data/fairseq/fairseq/models/__pycache__/distributed_fairseq_model.cpython-38.pyc differ diff --git a/data/fairseq/fairseq/models/__pycache__/fairseq_decoder.cpython-38.pyc b/data/fairseq/fairseq/models/__pycache__/fairseq_decoder.cpython-38.pyc new file mode 100644 index 0000000000000000000000000000000000000000..1c66df836729eedbaea0a409b9e8f54e4b795ad2 Binary files /dev/null and b/data/fairseq/fairseq/models/__pycache__/fairseq_decoder.cpython-38.pyc differ diff --git a/data/fairseq/fairseq/models/__pycache__/fairseq_encoder.cpython-38.pyc b/data/fairseq/fairseq/models/__pycache__/fairseq_encoder.cpython-38.pyc new file mode 100644 index 0000000000000000000000000000000000000000..62af93c9c71df92d8ae8dbb980626028d0ca6052 Binary files /dev/null and b/data/fairseq/fairseq/models/__pycache__/fairseq_encoder.cpython-38.pyc differ diff --git a/data/fairseq/fairseq/models/__pycache__/fairseq_incremental_decoder.cpython-38.pyc b/data/fairseq/fairseq/models/__pycache__/fairseq_incremental_decoder.cpython-38.pyc new file mode 100644 index 0000000000000000000000000000000000000000..7f11b8e26e35f2fe4b50f9d7e7c2c14084d408fb Binary files /dev/null and b/data/fairseq/fairseq/models/__pycache__/fairseq_incremental_decoder.cpython-38.pyc differ diff --git a/data/fairseq/fairseq/models/__pycache__/fairseq_model.cpython-38.pyc b/data/fairseq/fairseq/models/__pycache__/fairseq_model.cpython-38.pyc new file mode 100644 index 0000000000000000000000000000000000000000..2069fa70566f2bf4087edfdac85117e60b28d1ad Binary files /dev/null and b/data/fairseq/fairseq/models/__pycache__/fairseq_model.cpython-38.pyc differ diff --git a/data/fairseq/fairseq/models/__pycache__/fconv.cpython-38.pyc b/data/fairseq/fairseq/models/__pycache__/fconv.cpython-38.pyc new file mode 100644 index 0000000000000000000000000000000000000000..7aa7c50936c465277516456164008f61eed2b217 Binary files /dev/null and b/data/fairseq/fairseq/models/__pycache__/fconv.cpython-38.pyc differ diff --git a/data/fairseq/fairseq/models/__pycache__/fconv_lm.cpython-38.pyc b/data/fairseq/fairseq/models/__pycache__/fconv_lm.cpython-38.pyc new file mode 100644 index 0000000000000000000000000000000000000000..59293073587604bd15e454c0cad6687bf962506c Binary files /dev/null and b/data/fairseq/fairseq/models/__pycache__/fconv_lm.cpython-38.pyc differ diff --git a/data/fairseq/fairseq/models/__pycache__/fconv_self_att.cpython-38.pyc b/data/fairseq/fairseq/models/__pycache__/fconv_self_att.cpython-38.pyc new file mode 100644 index 0000000000000000000000000000000000000000..24781924dc9cbf67b5e53d910f5a3d76bf94bde0 Binary files /dev/null and b/data/fairseq/fairseq/models/__pycache__/fconv_self_att.cpython-38.pyc differ diff --git a/data/fairseq/fairseq/models/__pycache__/lightconv.cpython-38.pyc b/data/fairseq/fairseq/models/__pycache__/lightconv.cpython-38.pyc new file mode 100644 index 0000000000000000000000000000000000000000..6b1493b860489d1682e60a72f16cf8d1d3980792 Binary files /dev/null and b/data/fairseq/fairseq/models/__pycache__/lightconv.cpython-38.pyc differ diff --git a/data/fairseq/fairseq/models/__pycache__/lightconv_lm.cpython-38.pyc b/data/fairseq/fairseq/models/__pycache__/lightconv_lm.cpython-38.pyc new file mode 100644 index 0000000000000000000000000000000000000000..0e26a76b6065d8405e521fbc7796c137aa9dfefe Binary files /dev/null and b/data/fairseq/fairseq/models/__pycache__/lightconv_lm.cpython-38.pyc differ diff --git a/data/fairseq/fairseq/models/__pycache__/lstm.cpython-38.pyc b/data/fairseq/fairseq/models/__pycache__/lstm.cpython-38.pyc new file mode 100644 index 0000000000000000000000000000000000000000..a2315d78a7843a9d1398eb7f4cf0580c04a0317d Binary files /dev/null and b/data/fairseq/fairseq/models/__pycache__/lstm.cpython-38.pyc differ diff --git a/data/fairseq/fairseq/models/__pycache__/lstm_lm.cpython-38.pyc b/data/fairseq/fairseq/models/__pycache__/lstm_lm.cpython-38.pyc new file mode 100644 index 0000000000000000000000000000000000000000..35fb1ab0840c9ea7db1ca233f835441c6f12efb0 Binary files /dev/null and b/data/fairseq/fairseq/models/__pycache__/lstm_lm.cpython-38.pyc differ diff --git a/data/fairseq/fairseq/models/__pycache__/masked_lm.cpython-38.pyc b/data/fairseq/fairseq/models/__pycache__/masked_lm.cpython-38.pyc new file mode 100644 index 0000000000000000000000000000000000000000..8fdfda66014e1a940988ad0173f73a593590ecd5 Binary files /dev/null and b/data/fairseq/fairseq/models/__pycache__/masked_lm.cpython-38.pyc differ diff --git a/data/fairseq/fairseq/models/__pycache__/model_utils.cpython-38.pyc b/data/fairseq/fairseq/models/__pycache__/model_utils.cpython-38.pyc new file mode 100644 index 0000000000000000000000000000000000000000..23fd5208eeaf8df298bf395dde9cee5542e29e24 Binary files /dev/null and b/data/fairseq/fairseq/models/__pycache__/model_utils.cpython-38.pyc differ diff --git a/data/fairseq/fairseq/models/__pycache__/multilingual_transformer.cpython-38.pyc b/data/fairseq/fairseq/models/__pycache__/multilingual_transformer.cpython-38.pyc new file mode 100644 index 0000000000000000000000000000000000000000..17fd66a35ecbf58816168ef52f0f3abb22ca85ef Binary files /dev/null and b/data/fairseq/fairseq/models/__pycache__/multilingual_transformer.cpython-38.pyc differ diff --git a/data/fairseq/fairseq/models/__pycache__/transformer_align.cpython-38.pyc b/data/fairseq/fairseq/models/__pycache__/transformer_align.cpython-38.pyc new file mode 100644 index 0000000000000000000000000000000000000000..d3f602809623b6a127b6e1558657297ac4c3b335 Binary files /dev/null and b/data/fairseq/fairseq/models/__pycache__/transformer_align.cpython-38.pyc differ diff --git a/data/fairseq/fairseq/models/__pycache__/transformer_from_pretrained_xlm.cpython-38.pyc b/data/fairseq/fairseq/models/__pycache__/transformer_from_pretrained_xlm.cpython-38.pyc new file mode 100644 index 0000000000000000000000000000000000000000..f5a59e65bc733c497cd7c752c594ec3e19120c63 Binary files /dev/null and b/data/fairseq/fairseq/models/__pycache__/transformer_from_pretrained_xlm.cpython-38.pyc differ diff --git a/data/fairseq/fairseq/models/__pycache__/transformer_lm.cpython-38.pyc b/data/fairseq/fairseq/models/__pycache__/transformer_lm.cpython-38.pyc new file mode 100644 index 0000000000000000000000000000000000000000..0b8e4790ef9b641642d78b1a4291addacd149401 Binary files /dev/null and b/data/fairseq/fairseq/models/__pycache__/transformer_lm.cpython-38.pyc differ diff --git a/data/fairseq/fairseq/models/__pycache__/transformer_ulm.cpython-38.pyc b/data/fairseq/fairseq/models/__pycache__/transformer_ulm.cpython-38.pyc new file mode 100644 index 0000000000000000000000000000000000000000..4be64d945d838f6c50e8c89e1152cbb8efd459f1 Binary files /dev/null and b/data/fairseq/fairseq/models/__pycache__/transformer_ulm.cpython-38.pyc differ diff --git a/data/fairseq/fairseq/models/ema/__init__.py b/data/fairseq/fairseq/models/ema/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..503ceaa609b092e48bd32a0031f4e2ffb875483f --- /dev/null +++ b/data/fairseq/fairseq/models/ema/__init__.py @@ -0,0 +1,20 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import importlib +import os + +from .ema import EMA + + +def build_ema(model, cfg, device): + return EMA(model, cfg, device) + + +# automatically import any Python files in the models/ema/ directory +for file in sorted(os.listdir(os.path.dirname(__file__))): + if file.endswith(".py") and not file.startswith("_"): + file_name = file[: file.find(".py")] + importlib.import_module("fairseq.models.ema." + file_name) diff --git a/data/fairseq/fairseq/models/ema/__pycache__/__init__.cpython-38.pyc b/data/fairseq/fairseq/models/ema/__pycache__/__init__.cpython-38.pyc new file mode 100644 index 0000000000000000000000000000000000000000..f91bd253aaea51529363759d34b0690628eeb87a Binary files /dev/null and b/data/fairseq/fairseq/models/ema/__pycache__/__init__.cpython-38.pyc differ diff --git a/data/fairseq/fairseq/models/ema/__pycache__/ema.cpython-38.pyc b/data/fairseq/fairseq/models/ema/__pycache__/ema.cpython-38.pyc new file mode 100644 index 0000000000000000000000000000000000000000..e7b467fd013939541c51d9d3d239d04f5a70edf2 Binary files /dev/null and b/data/fairseq/fairseq/models/ema/__pycache__/ema.cpython-38.pyc differ diff --git a/data/fairseq/fairseq/models/ema/ema.py b/data/fairseq/fairseq/models/ema/ema.py new file mode 100644 index 0000000000000000000000000000000000000000..472d5d5f17d4e2f1c06b4d4ae93526297e9abd33 --- /dev/null +++ b/data/fairseq/fairseq/models/ema/ema.py @@ -0,0 +1,209 @@ +#!/usr/bin/env python3 + +""" +This module has the EMA class used to store a copy of the exponentially decayed +model params. + +Typical usage of EMA class involves initializing an object using an existing +model (random or from a seed model) and setting the config like ema_decay, +ema_start_update which determine how the EMA model is updated. After every +update of the model i.e. at the end of the train_step, the EMA should be updated +by passing the new model to the EMA.step function. The EMA model state dict +can be stored in the extra state under the key of "ema" and dumped +into a checkpoint and loaded. The EMA object can be passed to tasks +by setting task.uses_ema property. +EMA is a smoothed/ensemble model which might have better performance +when used for inference or further fine-tuning. EMA class has a +reverse function to load the EMA params into a model and use it +like a regular model. + +This implementation is used for trainer-level ema tracking. For EMA tracking +inside the model, please use fairseq/modules/ema_module.py instead. +""" + +import copy +import logging + +import torch + +from fairseq import checkpoint_utils + + +class EMA(object): + """Exponential Moving Average of Fairseq Models + EMA keeps a copy of the exponentially decayed model params. + The set of params should include both gradient-descent and + non-gradient descent params, such as batch mean/var and buffers. + This is a modified implementation of + the open source code in https://github.com/zhawe01/fairseq-gec.git, + and internal source code in + fbcode/mobile-vision/projects/classification_pytorch/lib/utils/model_ema.py. + + Similar to TF EMA. + https://www.tensorflow.org/api_docs/python/tf/train/ExponentialMovingAverage. + EMA provides a averaged and smoothed set of model weights, and has been shown to + improve vision models. EMA class does all necessary functions to update, reload, + or init EMA methods. + + EMA object is initialized from an arbitrary model. By default, it is stored in + the same device (unless device specified at initialization) and with the + same precision as the model (unless ema_fp32 is True). ema_fp32 is recommended. + This stores the EMA parameters in fp32 only for the EMA update step, and + is used at the default precision otherwise. + EMA is usually enabled using EMAConfig with store_ema=True. Some important + parameters to configure EMA are + 1) ema_decay - The decay of EMA + 2) ema_update_freq - EMA is updated every this many model updates. + 3) ema_start_update - Start EMA update after this many model updates [default 0] + + Key methods: + 1) step - One update of EMA using new model + 2) restore - Update EMA from a state dict + 3) reverse - Load EMA into a model + 4) get_decay, _set_decay - Used to get or set the decay. Note _set_decay is + called from step. + 5) build_fp32_params - Used to initialize or update the fp32 copy of EMA params. + Note this is enabled only when ema_fp32=True + """ + + def __init__(self, model, config, device=None, skip_keys=None): + """ + @param model model to initialize the EMA with + @param config EMAConfig object with configuration like + ema_decay, ema_update_freq, ema_fp32 + @param device If provided, copy EMA to this device (e.g. gpu). + Otherwise EMA is in the same device as the model. + """ + + self.decay = config.ema_decay + self.model = copy.deepcopy(model) + self.model.requires_grad_(False) + self.config = config + self.skip_keys = skip_keys or set() + self.fp32_params = {} + + if self.config.ema_seed_model is not None: + state = checkpoint_utils.load_ema_from_checkpoint( + self.config.ema_seed_model + ) + self.model.load_state_dict(state["model"], strict=True) + + if device is not None: + logging.info(f"Copying EMA model to device {device}") + self.model = self.model.to(device=device) + + if self.config.ema_fp32: + self.build_fp32_params() + + self.update_freq_counter = 0 + + def get_model(self): + return self.model + + def build_fp32_params(self, state_dict=None): + """ + Store a copy of the EMA params in fp32. + If state dict is passed, the EMA params is copied from + the provided state dict. Otherwise, it is copied from the + current EMA model parameters. + """ + if not self.config.ema_fp32: + raise RuntimeError( + "build_fp32_params should not be called if ema_fp32=False. " + "Use ema_fp32=True if this is really intended." + ) + + if state_dict is None: + state_dict = self.model.state_dict() + + def _to_float(t): + return t.float() if torch.is_floating_point(t) else t + + for param_key in state_dict: + if param_key in self.fp32_params: + self.fp32_params[param_key].copy_(state_dict[param_key]) + else: + self.fp32_params[param_key] = _to_float(state_dict[param_key]) + + def restore(self, state_dict, build_fp32_params=False): + """Load data from a model spec into EMA model""" + self.model.load_state_dict(state_dict, strict=False) + if build_fp32_params: + self.build_fp32_params(state_dict) + + def _set_decay(self, decay): + self.decay = decay + + def get_decay(self): + return self.decay + + def _step_internal(self, new_model, updates=None): + """One update of the EMA model based on new model weights""" + decay = self.decay + + ema_state_dict = {} + ema_params = ( + self.fp32_params if self.config.ema_fp32 else self.model.state_dict() + ) + for key, param in new_model.state_dict().items(): + if isinstance(param, dict): + continue + try: + ema_param = ema_params[key] + except KeyError: + ema_param = ( + param.float().clone() if param.ndim == 1 else copy.deepcopy(param) + ) + + if param.shape != ema_param.shape: + raise ValueError( + "incompatible tensor shapes between model param and ema param" + + "{} vs. {}".format(param.shape, ema_param.shape) + ) + + if "version" in key: + # Do not decay a model.version pytorch param + continue + + if key in self.skip_keys: + ema_param = param.to(dtype=ema_param.dtype).clone() + else: + ema_param.mul_(decay) + ema_param.add_(param.to(dtype=ema_param.dtype), alpha=1 - decay) + ema_state_dict[key] = ema_param + self.restore(ema_state_dict, build_fp32_params=False) + + def step(self, new_model, updates=None): + """ + One update of EMA which is done every self.config.ema_update_freq + updates of the model. + + @param updates The current number of model updates done. + Decay is set of 0 if model updates < ema_start_update, which means + the model will be simply copied over to the EMA. + When model updates >= ema_start_updates, then EMA is updated with + a decay of self.config.ema_decay. + """ + if updates is not None: + self._set_decay( + 0 if updates < self.config.ema_start_update else self.config.ema_decay + ) + if self.config.ema_update_freq > 1: + self.update_freq_counter += 1 + if self.update_freq_counter >= self.config.ema_update_freq: + self._step_internal(new_model, updates) + self.update_freq_counter = 0 + else: + self._step_internal(new_model, updates) + + def reverse(self, model): + """ + Load the model parameters from EMA model. + Useful for inference or fine-tuning from the EMA model. + """ + d = self.model.state_dict() + if "_ema" in d: + del d["_ema"] + + model.load_state_dict(d, strict=False) + return model diff --git a/data/fairseq/fairseq/models/hubert/__init__.py b/data/fairseq/fairseq/models/hubert/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..a1b0eabbdbcaf12b15bb96b329ab1e276256f79a --- /dev/null +++ b/data/fairseq/fairseq/models/hubert/__init__.py @@ -0,0 +1,7 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from .hubert import * # noqa +from .hubert_asr import * # noqa diff --git a/data/fairseq/fairseq/models/hubert/__pycache__/__init__.cpython-38.pyc b/data/fairseq/fairseq/models/hubert/__pycache__/__init__.cpython-38.pyc new file mode 100644 index 0000000000000000000000000000000000000000..fe8ba669ce6c26717f1020b5f96bf69aa6d634d8 Binary files /dev/null and b/data/fairseq/fairseq/models/hubert/__pycache__/__init__.cpython-38.pyc differ diff --git a/data/fairseq/fairseq/models/hubert/__pycache__/hubert.cpython-38.pyc b/data/fairseq/fairseq/models/hubert/__pycache__/hubert.cpython-38.pyc new file mode 100644 index 0000000000000000000000000000000000000000..1b15afb2f62767163ce25d2ca7cd2ac5c9c59cc2 Binary files /dev/null and b/data/fairseq/fairseq/models/hubert/__pycache__/hubert.cpython-38.pyc differ diff --git a/data/fairseq/fairseq/models/hubert/__pycache__/hubert_asr.cpython-38.pyc b/data/fairseq/fairseq/models/hubert/__pycache__/hubert_asr.cpython-38.pyc new file mode 100644 index 0000000000000000000000000000000000000000..67cb92819d8b8aa83adc52b8543f990490dda46c Binary files /dev/null and b/data/fairseq/fairseq/models/hubert/__pycache__/hubert_asr.cpython-38.pyc differ diff --git a/data/fairseq/fairseq/models/hubert/hubert.py b/data/fairseq/fairseq/models/hubert/hubert.py new file mode 100644 index 0000000000000000000000000000000000000000..cc3b777efdb66622514eb97dbde3642ea1574fb5 --- /dev/null +++ b/data/fairseq/fairseq/models/hubert/hubert.py @@ -0,0 +1,576 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import logging +from dataclasses import dataclass, field +from typing import Dict, List, Optional, Tuple + +import numpy as np +import torch +import torch.nn as nn +from omegaconf import II + +from fairseq import utils +from fairseq.data.data_utils import compute_mask_indices +from fairseq.data.dictionary import Dictionary +from fairseq.dataclass import ChoiceEnum, FairseqDataclass +from fairseq.models import BaseFairseqModel, register_model +from fairseq.models.wav2vec.wav2vec2 import ( + EXTRACTOR_MODE_CHOICES, + MASKING_DISTRIBUTION_CHOICES, + LAYER_TYPE_CHOICES, + ConvFeatureExtractionModel, + TransformerEncoder, +) +from fairseq.modules import GradMultiply, LayerNorm +from fairseq.tasks.hubert_pretraining import ( + HubertPretrainingConfig, + HubertPretrainingTask, +) + +logger = logging.getLogger(__name__) + + +@dataclass +class HubertConfig(FairseqDataclass): + label_rate: float = II("task.label_rate") + + extractor_mode: EXTRACTOR_MODE_CHOICES = field( + default="default", + metadata={ + "help": "mode for feature extractor. default has a single group " + "norm with d groups in the first conv block, whereas layer_norm " + "has layer norms in every block (meant to use with normalize=True)" + }, + ) + encoder_layers: int = field( + default=12, metadata={"help": "num encoder layers in the transformer"} + ) + encoder_embed_dim: int = field( + default=768, metadata={"help": "encoder embedding dimension"} + ) + encoder_ffn_embed_dim: int = field( + default=3072, metadata={"help": "encoder embedding dimension for FFN"} + ) + encoder_attention_heads: int = field( + default=12, metadata={"help": "num encoder attention heads"} + ) + activation_fn: ChoiceEnum(utils.get_available_activation_fns()) = field( + default="gelu", metadata={"help": "activation function to use"} + ) + layer_type: LAYER_TYPE_CHOICES = field( + default="transformer", metadata={"help": "layer type in encoder"} + ) + + # dropouts + dropout: float = field( + default=0.1, + metadata={"help": "dropout probability for the transformer"}, + ) + attention_dropout: float = field( + default=0.1, + metadata={"help": "dropout probability for attention weights"}, + ) + activation_dropout: float = field( + default=0.0, + metadata={"help": "dropout probability after activation in FFN"}, + ) + encoder_layerdrop: float = field( + default=0.0, + metadata={"help": "probability of dropping a tarnsformer layer"}, + ) + dropout_input: float = field( + default=0.0, + metadata={"help": "dropout to apply to the input (after feat extr)"}, + ) + dropout_features: float = field( + default=0.0, + metadata={"help": "dropout to apply to the features (after feat extr)"}, + ) + + final_dim: int = field( + default=0, + metadata={ + "help": "project final representations and targets to this many " + "dimensions. set to encoder_embed_dim is <= 0" + }, + ) + untie_final_proj: bool = field( + default=False, + metadata={"help": "use separate projection for each target"}, + ) + layer_norm_first: bool = field( + default=False, + metadata={"help": "apply layernorm first in the transformer"}, + ) + conv_feature_layers: str = field( + default="[(512,10,5)] + [(512,3,2)] * 4 + [(512,2,2)] * 2", + metadata={ + "help": "string describing convolutional feature extraction " + "layers in form of a python list that contains " + "[(dim, kernel_size, stride), ...]" + }, + ) + conv_bias: bool = field( + default=False, metadata={"help": "include bias in conv encoder"} + ) + logit_temp: float = field( + default=0.1, metadata={"help": "temperature to divide logits by"} + ) + target_glu: bool = field( + default=False, metadata={"help": "adds projection + glu to targets"} + ) + feature_grad_mult: float = field( + default=1.0, + metadata={"help": "multiply feature extractor var grads by this"}, + ) + + # masking + mask_length: int = field(default=10, metadata={"help": "mask length"}) + mask_prob: float = field( + default=0.65, + metadata={"help": "probability of replacing a token with mask"}, + ) + mask_selection: MASKING_DISTRIBUTION_CHOICES = field( + default="static", metadata={"help": "how to choose mask length"} + ) + mask_other: float = field( + default=0, + metadata={ + "help": "secondary mask argument " + "(used for more complex distributions), " + "see help in compute_mask_indicesh" + }, + ) + no_mask_overlap: bool = field( + default=False, metadata={"help": "whether to allow masks to overlap"} + ) + mask_min_space: int = field( + default=1, + metadata={"help": "min space between spans (if no overlap is enabled)"}, + ) + + # channel masking + mask_channel_length: int = field( + default=10, + metadata={"help": "length of the mask for features (channels)"}, + ) + mask_channel_prob: float = field( + default=0.0, + metadata={"help": "probability of replacing a feature with 0"}, + ) + mask_channel_selection: MASKING_DISTRIBUTION_CHOICES = field( + default="static", + metadata={"help": "how to choose mask length for channel masking"}, + ) + mask_channel_other: float = field( + default=0, + metadata={ + "help": "secondary mask argument " + "(used for more complex distributions), " + "see help in compute_mask_indicesh" + }, + ) + no_mask_channel_overlap: bool = field( + default=False, + metadata={"help": "whether to allow channel masks to overlap"}, + ) + mask_channel_min_space: int = field( + default=1, + metadata={"help": "min space between spans (if no overlap is enabled)"}, + ) + + # positional embeddings + conv_pos: int = field( + default=128, + metadata={"help": "number of filters for convolutional positional embeddings"}, + ) + conv_pos_groups: int = field( + default=16, + metadata={"help": "number of groups for convolutional positional embedding"}, + ) + conv_pos_batch_norm: bool = field( + default=False, + metadata={ + "help": "use batch norm instead of weight norm in conv_pos (for bf16 models)" + }, + ) + + latent_temp: Tuple[float, float, float] = field( + default=(2, 0.5, 0.999995), + metadata={"help": "legacy (to be removed)"}, + ) + + # loss computation + skip_masked: bool = field( + default=False, + metadata={"help": "skip computing losses over masked frames"}, + ) + skip_nomask: bool = field( + default=False, + metadata={"help": "skip computing losses over unmasked frames"}, + ) + + checkpoint_activations: bool = field( + default=False, + metadata={"help": "recompute activations and save memory for extra compute"}, + ) + + # FP16 optimization + required_seq_len_multiple: int = field( + default=2, + metadata={ + "help": "pad the input to encoder such that the sequence length is divisible by multiple" + }, + ) + + # Conformer + depthwise_conv_kernel_size: int = field( + default=31, + metadata={ + "help": "depthwise-conv-kernel-size for convolution in conformer layer" + }, + ) + attn_type: str = field( + default="", + metadata={"help": "if espnet use ESPNET MHA"}, + ) + pos_enc_type: str = field( + default="abs", + metadata={"help": "Positional encoding type to use in conformer"}, + ) + fp16: bool = field(default=False, metadata={"help": "If fp16 is being used"}) + + +@register_model("hubert", dataclass=HubertConfig) +class HubertModel(BaseFairseqModel): + def __init__( + self, + cfg: HubertConfig, + task_cfg: HubertPretrainingConfig, + dictionaries: List[Dictionary], + ) -> None: + super().__init__() + logger.info(f"HubertModel Config: {cfg}") + + feature_enc_layers = eval(cfg.conv_feature_layers) # noqa + self.embed = feature_enc_layers[-1][0] + + self.feature_extractor = ConvFeatureExtractionModel( + conv_layers=feature_enc_layers, + dropout=0.0, + mode=cfg.extractor_mode, + conv_bias=cfg.conv_bias, + ) + feature_ds_rate = np.prod([s for _, _, s in feature_enc_layers]) + self.feat2tar_ratio = cfg.label_rate * feature_ds_rate / task_cfg.sample_rate + + self.post_extract_proj = ( + nn.Linear(self.embed, cfg.encoder_embed_dim) + if self.embed != cfg.encoder_embed_dim + else None + ) + + self.mask_prob = cfg.mask_prob + self.mask_selection = cfg.mask_selection + self.mask_other = cfg.mask_other + self.mask_length = cfg.mask_length + self.no_mask_overlap = cfg.no_mask_overlap + self.mask_min_space = cfg.mask_min_space + + self.mask_channel_prob = cfg.mask_channel_prob + self.mask_channel_selection = cfg.mask_channel_selection + self.mask_channel_other = cfg.mask_channel_other + self.mask_channel_length = cfg.mask_channel_length + self.no_mask_channel_overlap = cfg.no_mask_channel_overlap + self.mask_channel_min_space = cfg.mask_channel_min_space + + self.dropout_input = nn.Dropout(cfg.dropout_input) + self.dropout_features = nn.Dropout(cfg.dropout_features) + + self.feature_grad_mult = cfg.feature_grad_mult + self.logit_temp = cfg.logit_temp + self.skip_masked = cfg.skip_masked + self.skip_nomask = cfg.skip_nomask + + final_dim = cfg.final_dim if cfg.final_dim > 0 else cfg.encoder_embed_dim + + self.mask_emb = nn.Parameter( + torch.FloatTensor(cfg.encoder_embed_dim).uniform_() + ) + + self.encoder = TransformerEncoder(cfg) + self.layer_norm = LayerNorm(self.embed) + + self.target_glu = None + if cfg.target_glu: + self.target_glu = nn.Sequential( + nn.Linear(final_dim, final_dim * 2), nn.GLU() + ) + + self.untie_final_proj = cfg.untie_final_proj + if self.untie_final_proj: + self.final_proj = nn.Linear( + cfg.encoder_embed_dim, final_dim * len(dictionaries) + ) + else: + self.final_proj = nn.Linear(cfg.encoder_embed_dim, final_dim) + + # modules below are not needed during fine-tuning + if any([d is None for d in dictionaries]): + logger.info("cannot find dictionary. assume will be used for fine-tuning") + else: + self.num_classes = [len(d) for d in dictionaries] + self.label_embs_concat = nn.Parameter( + torch.FloatTensor(sum(self.num_classes), final_dim) + ) + nn.init.uniform_(self.label_embs_concat) + + def upgrade_state_dict_named(self, state_dict, name): + """Upgrade a (possibly old) state dict for new versions of fairseq.""" + + super().upgrade_state_dict_named(state_dict, name) + return state_dict + + @classmethod + def build_model(cls, cfg: HubertConfig, task: HubertPretrainingTask): + """Build a new model instance.""" + + model = HubertModel(cfg, task.cfg, task.dictionaries) + return model + + def apply_mask(self, x, padding_mask, target_list): + B, T, C = x.shape + if self.mask_prob > 0: + mask_indices = compute_mask_indices( + (B, T), + padding_mask, + self.mask_prob, + self.mask_length, + self.mask_selection, + self.mask_other, + min_masks=2, + no_overlap=self.no_mask_overlap, + min_space=self.mask_min_space, + ) + mask_indices = torch.from_numpy(mask_indices).to(x.device) + x[mask_indices] = self.mask_emb + else: + mask_indices = None + + if self.mask_channel_prob > 0: + mask_channel_indices = compute_mask_indices( + (B, C), + None, + self.mask_channel_prob, + self.mask_channel_length, + self.mask_channel_selection, + self.mask_channel_other, + no_overlap=self.no_mask_channel_overlap, + min_space=self.mask_channel_min_space, + ) + mask_channel_indices = ( + torch.from_numpy(mask_channel_indices) + .to(x.device) + .unsqueeze(1) + .expand(-1, T, -1) + ) + x[mask_channel_indices] = 0 + + return x, mask_indices + + def compute_nce(self, x, pos, negs): + neg_is_pos = (pos == negs).all(-1) + pos = pos.unsqueeze(0) + targets = torch.cat([pos, negs], dim=0) + + logits = torch.cosine_similarity(x.float(), targets.float(), dim=-1).type_as(x) + logits /= self.logit_temp + if neg_is_pos.any(): + logits[1:][neg_is_pos] = float("-inf") + logits = logits.transpose(0, 1) # (num_x, num_cls+1) + return logits + + def forward_features(self, source: torch.Tensor) -> torch.Tensor: + if self.feature_grad_mult > 0: + features = self.feature_extractor(source) + if self.feature_grad_mult != 1.0: + features = GradMultiply.apply(features, self.feature_grad_mult) + else: + with torch.no_grad(): + features = self.feature_extractor(source) + return features + + def forward_targets( + self, + features: torch.Tensor, + target_list: List[torch.Tensor], + ) -> Tuple[torch.Tensor, torch.Tensor]: + # Trim features to ensure labels exist and then get aligned labels + feat_tsz = features.size(2) + targ_tsz = min([t.size(1) for t in target_list]) + if self.feat2tar_ratio * feat_tsz > targ_tsz: + feat_tsz = int(targ_tsz / self.feat2tar_ratio) + features = features[..., :feat_tsz] + target_inds = torch.arange(feat_tsz).float() * self.feat2tar_ratio + target_list = [t[:, target_inds.long()] for t in target_list] + return features, target_list + + def forward_padding_mask( + self, + features: torch.Tensor, + padding_mask: torch.Tensor, + ) -> torch.Tensor: + extra = padding_mask.size(1) % features.size(1) + if extra > 0: + padding_mask = padding_mask[:, :-extra] + padding_mask = padding_mask.view(padding_mask.size(0), features.size(1), -1) + padding_mask = padding_mask.all(-1) + return padding_mask + + def forward( + self, + source: torch.Tensor, + target_list: Optional[List[torch.Tensor]] = None, + padding_mask: Optional[torch.Tensor] = None, + mask: bool = True, + features_only: bool = False, + output_layer: Optional[int] = None, + ) -> Dict[str, torch.Tensor]: + """output layer is 1-based""" + features = self.forward_features(source) + if target_list is not None: + features, target_list = self.forward_targets(features, target_list) + + features_pen = features.float().pow(2).mean() + + features = features.transpose(1, 2) + features = self.layer_norm(features) + unmasked_features = features.clone() + + if padding_mask is not None: + padding_mask = self.forward_padding_mask(features, padding_mask) + + if self.post_extract_proj is not None: + features = self.post_extract_proj(features) + + features = self.dropout_input(features) + unmasked_features = self.dropout_features(unmasked_features) + + if mask: + x, mask_indices = self.apply_mask(features, padding_mask, target_list) + else: + x = features + mask_indices = None + + # feature: (B, T, D), float + # target: (B, T), long + # x: (B, T, D), float + # padding_mask: (B, T), bool + # mask_indices: (B, T), bool + x, _ = self.encoder( + x, + padding_mask=padding_mask, + layer=None if output_layer is None else output_layer - 1, + ) + + if features_only: + return {"x": x, "padding_mask": padding_mask, "features": features} + + def compute_pred(proj_x, target, label_embs): + # compute logits for the i-th label set + y = torch.index_select(label_embs, 0, target.long()) + negs = label_embs.unsqueeze(1).expand(-1, proj_x.size(0), -1) + if self.target_glu: + y = self.target_glu(y) + negs = self.target_glu(negs) + # proj_x: (S, D) + # y: (S, D) + # negs: (Neg, S, D) + return self.compute_nce(proj_x, y, negs) + + label_embs_list = self.label_embs_concat.split(self.num_classes, 0) + + if not self.skip_masked: + masked_indices = torch.logical_and(~padding_mask, mask_indices) + proj_x_m = self.final_proj(x[masked_indices]) + if self.untie_final_proj: + proj_x_m_list = proj_x_m.chunk(len(target_list), dim=-1) + else: + proj_x_m_list = [proj_x_m for _ in range(len(target_list))] + logit_m_list = [ + compute_pred(proj_x_m, t[masked_indices], label_embs_list[i]) + for i, (proj_x_m, t) in enumerate(zip(proj_x_m_list, target_list)) + ] + else: + logit_m_list = [None for _ in target_list] + + if not self.skip_nomask: + nomask_indices = torch.logical_and(~padding_mask, ~mask_indices) + proj_x_u = self.final_proj(x[nomask_indices]) + if self.untie_final_proj: + proj_x_u_list = proj_x_u.chunk(len(target_list), dim=-1) + else: + proj_x_u_list = [proj_x_u for _ in range(len(target_list))] + + logit_u_list = [ + compute_pred(proj_x_u, t[nomask_indices], label_embs_list[i]) + for i, (proj_x_u, t) in enumerate(zip(proj_x_u_list, target_list)) + ] + else: + logit_u_list = [None for _ in target_list] + + result = { + "logit_m_list": logit_m_list, + "logit_u_list": logit_u_list, + "padding_mask": padding_mask, + "features_pen": features_pen, + } + return result + + def extract_features( + self, + source: torch.Tensor, + padding_mask: Optional[torch.Tensor] = None, + mask: bool = False, + ret_conv: bool = False, + output_layer: Optional[int] = None, + ) -> Tuple[torch.Tensor, torch.Tensor]: + res = self.forward( + source, + padding_mask=padding_mask, + mask=mask, + features_only=True, + output_layer=output_layer, + ) + feature = res["features"] if ret_conv else res["x"] + return feature, res["padding_mask"] + + def get_logits(self, net_output, is_masked=True): + if is_masked: + logits_list = net_output["logit_m_list"] + else: + logits_list = net_output["logit_u_list"] + logits_list = [x.float() for x in logits_list if x is not None] + return logits_list + + def get_targets(self, net_output, is_masked=True): + logits_list = self.get_logits(net_output, is_masked) + targets_list = [x.new_zeros(x.size(0), dtype=torch.long) for x in logits_list] + return targets_list + + def get_extra_losses(self, net_output): + extra_losses = [] + names = [] + + if "features_pen" in net_output: + extra_losses.append(net_output["features_pen"]) + names.append("features_pen") + + return extra_losses, names + + def remove_pretraining_modules(self): + self.target_glu = None + self.final_proj = None diff --git a/data/fairseq/fairseq/models/hubert/hubert_asr.py b/data/fairseq/fairseq/models/hubert/hubert_asr.py new file mode 100644 index 0000000000000000000000000000000000000000..11c85ce7d11f8529d6ff7636eb467ae56bc11772 --- /dev/null +++ b/data/fairseq/fairseq/models/hubert/hubert_asr.py @@ -0,0 +1,675 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import contextlib +import copy +import logging +import math +from argparse import Namespace +from dataclasses import dataclass, field +from typing import Any, Optional +import numpy as np +import torch +import torch.nn as nn +import torch.nn.functional as F +from omegaconf import II, MISSING, open_dict + +from fairseq import checkpoint_utils, tasks, utils +from fairseq.dataclass import FairseqDataclass +from fairseq.dataclass.utils import convert_namespace_to_omegaconf +from fairseq.models import ( + BaseFairseqModel, + FairseqEncoder, + FairseqEncoderDecoderModel, + FairseqIncrementalDecoder, + register_model, +) +from fairseq.models.hubert.hubert import MASKING_DISTRIBUTION_CHOICES +from fairseq.modules import LayerNorm, PositionalEmbedding, TransformerDecoderLayer +from fairseq.tasks import FairseqTask + +logger = logging.getLogger(__name__) + + +@dataclass +class HubertAsrConfig(FairseqDataclass): + w2v_path: str = field(default=MISSING, metadata={"help": "path to hubert model"}) + no_pretrained_weights: bool = field( + default=False, + metadata={"help": "if true, does not load pretrained weights"}, + ) + dropout_input: float = field( + default=0.0, + metadata={"help": "dropout to apply to the input (after feat extr)"}, + ) + final_dropout: float = field( + default=0.0, + metadata={"help": "dropout after transformer and before final projection"}, + ) + dropout: float = field( + default=0.0, + metadata={"help": "dropout probability inside hubert model"}, + ) + attention_dropout: float = field( + default=0.0, + metadata={ + "help": "dropout probability for attention weights " "inside hubert model" + }, + ) + activation_dropout: float = field( + default=0.0, + metadata={ + "help": "dropout probability after activation in FFN " "inside hubert model" + }, + ) + encoder_embed_dim: Optional[int] = field( + default=768, metadata={"help": "encoder embedding dimension"} + ) + + # masking + apply_mask: bool = field( + default=False, metadata={"help": "apply masking during fine-tuning"} + ) + mask_length: int = field( + default=10, metadata={"help": "repeat the mask indices multiple times"} + ) + mask_prob: float = field( + default=0.5, + metadata={ + "help": "probability of replacing a token with mask " + "(normalized by length)" + }, + ) + mask_selection: MASKING_DISTRIBUTION_CHOICES = field( + default="static", metadata={"help": "how to choose masks"} + ) + mask_other: float = field( + default=0, + metadata={ + "help": "secondary mask argument " + "(used for more complex distributions), " + "see help in compute_mask_indices" + }, + ) + no_mask_overlap: bool = field( + default=False, metadata={"help": "whether to allow masks to overlap"} + ) + + # channel masking + mask_channel_length: int = field( + default=10, + metadata={"help": "length of the mask for features (channels)"}, + ) + mask_channel_prob: float = field( + default=0.0, + metadata={"help": "probability of replacing a feature with 0"}, + ) + mask_channel_selection: MASKING_DISTRIBUTION_CHOICES = field( + default="static", + metadata={"help": "how to choose mask length for channel masking"}, + ) + mask_channel_other: float = field( + default=0, + metadata={ + "help": "secondary mask argument " + "(used for more complex distributions), " + "see help in compute_mask_indices" + }, + ) + no_mask_channel_overlap: bool = field( + default=False, + metadata={"help": "whether to allow channel masks to overlap"}, + ) + freeze_finetune_updates: int = field( + default=0, + metadata={"help": "dont finetune hubert for this many updates"}, + ) + feature_grad_mult: float = field( + default=0.0, + metadata={"help": "reset feature grad mult in hubert to this"}, + ) + layerdrop: float = field( + default=0.0, + metadata={"help": "probability of dropping a layer in hubert"}, + ) + normalize: bool = II("task.normalize") + data: str = II("task.data") + + # this holds the loaded hubert args + w2v_args: Any = None + + +@dataclass +class HubertCtcConfig(HubertAsrConfig): + pass + + +@register_model("hubert_ctc", dataclass=HubertCtcConfig) +class HubertCtc(BaseFairseqModel): + def __init__(self, cfg: HubertCtcConfig, w2v_encoder: BaseFairseqModel): + super().__init__() + self.cfg = cfg + self.w2v_encoder = w2v_encoder + + def upgrade_state_dict_named(self, state_dict, name): + super().upgrade_state_dict_named(state_dict, name) + return state_dict + + @classmethod + def build_model(cls, cfg: HubertCtcConfig, task: FairseqTask): + """Build a new model instance.""" + w2v_encoder = HubertEncoder(cfg, task) + return cls(cfg, w2v_encoder) + + def get_normalized_probs(self, net_output, log_probs): + """Get normalized probabilities (or log probs) from a net's output.""" + + logits = net_output["encoder_out"] + if log_probs: + return utils.log_softmax(logits.float(), dim=-1) + else: + return utils.softmax(logits.float(), dim=-1) + + def get_logits(self, net_output): + logits = net_output["encoder_out"] + padding = net_output["encoder_padding_mask"] + if padding is not None and padding.any(): + padding = padding.T + logits[padding][..., 0] = 0 + logits[padding][..., 1:] = float("-inf") + + return logits + + def forward(self, **kwargs): + x = self.w2v_encoder(**kwargs) + return x + + +@dataclass +class HubertSeq2SeqConfig(HubertAsrConfig): + decoder_embed_dim: int = field( + default=768, metadata={"help": "decoder embedding dimension"} + ) + decoder_ffn_embed_dim: int = field( + default=3072, metadata={"help": "decoder embedding dimension for FFN"} + ) + decoder_layers: int = field(default=6, metadata={"help": "num of decoder layers"}) + decoder_layerdrop: float = field( + default=0.0, metadata={"help": "decoder layerdrop chance"} + ) + decoder_attention_heads: int = field( + default=4, metadata={"help": "num decoder attention heads"} + ) + decoder_learned_pos: bool = field( + default=False, + metadata={"help": "use learned positional embeddings in the decoder"}, + ) + decoder_normalize_before: bool = field( + default=False, metadata={"help": "apply layernorm before each decoder block"} + ) + no_token_positional_embeddings: bool = field( + default=False, + metadata={ + "help": "if set, disables positional embeddings (outside self attention)" + }, + ) + decoder_dropout: float = field( + default=0.0, metadata={"help": "dropout probability in the decoder"} + ) + decoder_attention_dropout: float = field( + default=0.0, + metadata={ + "help": "dropout probability for attention weights inside the decoder" + }, + ) + decoder_activation_dropout: float = field( + default=0.0, + metadata={ + "help": "dropout probability after activation in FFN inside the decoder" + }, + ) + max_target_positions: int = field( + default=2048, metadata={"help": "max target positions"} + ) + share_decoder_input_output_embed: bool = field( + default=False, metadata={"help": "share decoder input and output embeddings"} + ) + autoregressive: bool = II("task.autoregressive") + seq2seq_path: str = field( + default="", + metadata={"help": "reset_dict"}, + ) + reset_dict: bool = field( + default=False, + metadata={"help": "reset_dict"}, + ) + + +@register_model("hubert_seq2seq", dataclass=HubertSeq2SeqConfig) +class HubertSeq2SeqModel(FairseqEncoderDecoderModel): + def __init__(self, encoder, decoder): + super().__init__(encoder, decoder) + + @classmethod + def build_model(cls, cfg: HubertSeq2SeqConfig, task: FairseqTask): + """Build a new model instance.""" + + assert ( + cfg.autoregressive + ), "Please set task.autoregressive=true for seq2seq asr models" + + src_dict, tgt_dict = task.source_dictionary, task.target_dictionary + + def build_embedding(dictionary, embed_dim): + num_embeddings = len(dictionary) + padding_idx = dictionary.pad() + emb = Embedding(num_embeddings, embed_dim, padding_idx) + return emb + + decoder_embed_tokens = build_embedding(tgt_dict, cfg.decoder_embed_dim) + + encoder = cls.build_encoder(cfg, task) + decoder = cls.build_decoder(cfg, tgt_dict, decoder_embed_tokens) + + model = HubertSeq2SeqModel(encoder, decoder) + + if cfg["seq2seq_path"]: + state = checkpoint_utils.load_checkpoint_to_cpu(cfg.seq2seq_path) + state = state["model"] + if cfg["reset_dict"]: + del state["decoder.embed_out"] + del state["decoder.embed_tokens.weight"] + model.load_state_dict(state, strict=False) + return model + + @classmethod + def build_encoder(cls, cfg: HubertAsrConfig, task): + return HubertEncoder(cfg, task) + + @classmethod + def build_decoder(cls, cfg: HubertSeq2SeqConfig, tgt_dict, embed_tokens): + return TransformerDecoder(cfg, tgt_dict, embed_tokens) + + def forward(self, **kwargs): + encoder_out = self.encoder(**kwargs) + decoder_out = self.decoder(encoder_out=encoder_out, **kwargs) + return decoder_out + + def upgrade_state_dict_named(self, state_dict, name): + return state_dict + + def load_state_dict( + self, + state_dict, + strict=True, + model_cfg=None, + args: Optional[Namespace] = None, + ): + if model_cfg.reset_dict: + logger.warn("Overriding loading strict state dict!") + del state_dict["decoder.embed_out"] + del state_dict["decoder.embed_tokens.weight"] + return super().load_state_dict(state_dict, False, model_cfg, args) + return super().load_state_dict(state_dict, strict, model_cfg, args) + + +class HubertEncoder(FairseqEncoder): + def __init__(self, cfg: HubertAsrConfig, task): + self.apply_mask = cfg.apply_mask + + arg_overrides = { + "dropout": cfg.dropout, + "activation_dropout": cfg.activation_dropout, + "dropout_input": cfg.dropout_input, + "attention_dropout": cfg.attention_dropout, + "mask_length": cfg.mask_length, + "mask_prob": cfg.mask_prob, + "mask_selection": cfg.mask_selection, + "mask_other": cfg.mask_other, + "no_mask_overlap": cfg.no_mask_overlap, + "mask_channel_length": cfg.mask_channel_length, + "mask_channel_prob": cfg.mask_channel_prob, + "mask_channel_selection": cfg.mask_channel_selection, + "mask_channel_other": cfg.mask_channel_other, + "no_mask_channel_overlap": cfg.no_mask_channel_overlap, + "encoder_layerdrop": cfg.layerdrop, + "feature_grad_mult": cfg.feature_grad_mult, + } + + if cfg.w2v_args is None: + state = checkpoint_utils.load_checkpoint_to_cpu(cfg.w2v_path, arg_overrides) + w2v_args = state.get("cfg", None) + if w2v_args is None: + w2v_args = convert_namespace_to_omegaconf(state["args"]) + cfg.w2v_args = w2v_args + else: + state = None + w2v_args = cfg.w2v_args + if isinstance(w2v_args, Namespace): + cfg.w2v_args = w2v_args = convert_namespace_to_omegaconf(w2v_args) + + assert cfg.normalize == w2v_args.task.normalize, ( + "Fine-tuning works best when data normalization is the same. " + "Please check that --normalize is set or unset for " + "both pre-training and here" + ) + + w2v_args.task.data = cfg.data + pretrain_task = tasks.setup_task(w2v_args.task) + if state is not None and "task_state" in state: + # This will load the stored "dictionaries" object + pretrain_task.load_state_dict(state["task_state"]) + else: + pretrain_task.load_state_dict(task.state_dict()) + + model = pretrain_task.build_model(w2v_args.model, from_checkpoint=True) + if state is not None and not cfg.no_pretrained_weights: + # set strict=False because we omit some modules + model.load_state_dict(state["model"], strict=False) + + model.remove_pretraining_modules() + + super().__init__(pretrain_task.source_dictionary) + + d = w2v_args.model.encoder_embed_dim + + self.w2v_model = model + + self.final_dropout = nn.Dropout(cfg.final_dropout) + self.freeze_finetune_updates = cfg.freeze_finetune_updates + self.num_updates = 0 + + if task.target_dictionary is not None and not cfg.autoregressive: + self.proj = Linear(d, len(task.target_dictionary)) + elif getattr(cfg, "decoder_embed_dim", d) != d: + self.proj = Linear(d, cfg.decoder_embed_dim) + else: + self.proj = None + + def set_num_updates(self, num_updates): + """Set the number of parameters updates.""" + super().set_num_updates(num_updates) + self.num_updates = num_updates + + def forward(self, source, padding_mask, tbc=True, **kwargs): + + w2v_args = { + "source": source, + "padding_mask": padding_mask, + "mask": self.apply_mask and self.training, + } + + ft = self.freeze_finetune_updates <= self.num_updates + + with torch.no_grad() if not ft else contextlib.ExitStack(): + x, padding_mask = self.w2v_model.extract_features(**w2v_args) + + if tbc: + # B x T x C -> T x B x C + x = x.transpose(0, 1) + + x = self.final_dropout(x) + + if self.proj: + x = self.proj(x) + + return { + "encoder_out": x, # T x B x C + "encoder_padding_mask": padding_mask, # B x T + "padding_mask": padding_mask, + } + + def reorder_encoder_out(self, encoder_out, new_order): + if encoder_out["encoder_out"] is not None: + encoder_out["encoder_out"] = encoder_out["encoder_out"].index_select( + 1, new_order + ) + if encoder_out["encoder_padding_mask"] is not None: + encoder_out["encoder_padding_mask"] = encoder_out[ + "encoder_padding_mask" + ].index_select(0, new_order) + if encoder_out["padding_mask"] is not None: + encoder_out["padding_mask"] = encoder_out["padding_mask"].index_select( + 0, new_order + ) + return encoder_out + + def max_positions(self): + """Maximum input length supported by the encoder.""" + return None + + def upgrade_state_dict_named(self, state_dict, name): + return state_dict + + +class TransformerDecoder(FairseqIncrementalDecoder): + """ + Transformer decoder consisting of *args.decoder_layers* layers. Each layer + is a :class:`TransformerDecoderLayer`. + + Args: + args (argparse.Namespace): parsed command-line arguments + dictionary (~fairseq.data.Dictionary): decoding dictionary + embed_tokens (torch.nn.Embedding): output embedding + no_encoder_attn (bool, optional): whether to attend to encoder outputs + (default: False). + """ + + def __init__( + self, + cfg: HubertSeq2SeqConfig, + dictionary, + embed_tokens, + no_encoder_attn=False, + ): + super().__init__(dictionary) + + self.dropout = cfg.decoder_dropout + self.share_input_output_embed = cfg.share_decoder_input_output_embed + + input_embed_dim = embed_tokens.embedding_dim + embed_dim = cfg.decoder_embed_dim + self.output_embed_dim = cfg.decoder_embed_dim + + self.layerdrop = cfg.decoder_layerdrop + + self.padding_idx = embed_tokens.padding_idx + self.max_target_positions = cfg.max_target_positions + + self.embed_tokens = embed_tokens + self.embed_scale = math.sqrt(embed_dim) # todo: try with input_embed_dim + + self.project_in_dim = ( + Linear(input_embed_dim, embed_dim, bias=False) + if embed_dim != input_embed_dim + else None + ) + + self.embed_positions = ( + PositionalEmbedding( + cfg.max_target_positions, + embed_dim, + self.padding_idx, + learned=cfg.decoder_learned_pos, + ) + if not cfg.no_token_positional_embeddings + else None + ) + + # TODO: update this when transformer gets converted to dataclass configs + transformer_cfg = copy.deepcopy(cfg) + with open_dict(transformer_cfg): + transformer_cfg.dropout = transformer_cfg.decoder_dropout + transformer_cfg.attention_dropout = ( + transformer_cfg.decoder_attention_dropout + ) + transformer_cfg.activation_dropout = ( + transformer_cfg.decoder_activation_dropout + ) + + self.layers = nn.ModuleList([]) + self.layers.extend( + [ + TransformerDecoderLayer(transformer_cfg, no_encoder_attn) + for _ in range(transformer_cfg.decoder_layers) + ] + ) + + if not self.share_input_output_embed: + self.embed_out = nn.Parameter( + torch.Tensor(len(dictionary), self.output_embed_dim) + ) + nn.init.normal_(self.embed_out, mean=0, std=self.output_embed_dim**-0.5) + + if transformer_cfg.decoder_normalize_before: + self.layer_norm = LayerNorm(embed_dim) + else: + self.layer_norm = None + + def forward( + self, prev_output_tokens, encoder_out=None, incremental_state=None, **unused + ): + """ + Args: + prev_output_tokens (LongTensor): previous decoder outputs of shape + `(batch, tgt_len)`, for teacher forcing + encoder_out (Tensor, optional): output from the encoder, used for + encoder-side attention + incremental_state (dict): dictionary used for storing state during + :ref:`Incremental decoding` + + Returns: + tuple: + - the decoder's output of shape `(batch, tgt_len, vocab)` + - a dictionary with any model-specific outputs + """ + if type(prev_output_tokens) == list: + max_len = max((len(x) for x in prev_output_tokens)) + tmp = torch.zeros( + [len(prev_output_tokens), max_len], device=prev_output_tokens[0].device + ) + for (i, p) in enumerate(prev_output_tokens): + tmp[i, : len(p)] = p + prev_output_tokens = tmp + prev_output_tokens = prev_output_tokens.long() + x, extra = self.extract_features( + prev_output_tokens, encoder_out, incremental_state + ) + x = self.output_layer(x) + return x, extra + + def extract_features( + self, prev_output_tokens, encoder_out=None, incremental_state=None, **unused + ): + """ + Similar to *forward* but only return features. + + Returns: + tuple: + - the decoder's features of shape `(batch, tgt_len, embed_dim)` + - a dictionary with any model-specific outputs + """ + + # embed positions + positions = ( + self.embed_positions( + prev_output_tokens, incremental_state=incremental_state + ) + if self.embed_positions is not None + else None + ) + + if incremental_state is not None: + prev_output_tokens = prev_output_tokens[:, -1:] + if positions is not None: + positions = positions[:, -1:] + + # embed tokens and positions + x = self.embed_scale * self.embed_tokens(prev_output_tokens) + + if self.project_in_dim is not None: + x = self.project_in_dim(x) + + if positions is not None: + x += positions + x = F.dropout(x, p=self.dropout, training=self.training) + + # B x T x C -> T x B x C + x = x.transpose(0, 1) + attn = None + + inner_states = [x] + + # decoder layers + self_attn_padding_mask = None + if prev_output_tokens.eq(self.padding_idx).any(): + self_attn_padding_mask = prev_output_tokens.eq(self.padding_idx) + for layer in self.layers: + dropout_probability = np.random.random() + if not self.training or (dropout_probability > self.layerdrop): + x, attn, _ = layer( + x, + encoder_out["encoder_out"] if encoder_out is not None else None, + encoder_out["padding_mask"] if encoder_out is not None else None, + incremental_state, + self_attn_mask=self.buffered_future_mask(x) + if incremental_state is None + else None, + self_attn_padding_mask=self_attn_padding_mask, + ) + inner_states.append(x) + + if self.layer_norm: + x = self.layer_norm(x) + + # T x B x C -> B x T x C + x = x.transpose(0, 1) + + return x, {"attn": attn, "inner_states": inner_states} + + def output_layer(self, features, **kwargs): + """Project features to the vocabulary size.""" + # project back to size of vocabulary + if self.share_input_output_embed: + return F.linear(features, self.embed_tokens.weight) + else: + return F.linear(features, self.embed_out) + + def max_positions(self): + """Maximum output length supported by the decoder.""" + if self.embed_positions is None: + return self.max_target_positions + return min(self.max_target_positions, self.embed_positions.max_positions) + + def buffered_future_mask(self, tensor): + dim = tensor.size(0) + if ( + not hasattr(self, "_future_mask") + or self._future_mask is None + or self._future_mask.device != tensor.device + or self._future_mask.size(0) < dim + ): + self._future_mask = torch.triu( + utils.fill_with_neg_inf(tensor.new(dim, dim)), 1 + ) + return self._future_mask[:dim, :dim] + + def upgrade_state_dict_named(self, state_dict, name): + return state_dict + + +def Embedding(num_embeddings, embedding_dim, padding_idx): + m = nn.Embedding(num_embeddings, embedding_dim, padding_idx=padding_idx) + nn.init.normal_(m.weight, mean=0, std=embedding_dim**-0.5) + nn.init.constant_(m.weight[padding_idx], 0) + return m + + +def Linear(in_features, out_features, bias=True): + m = nn.Linear(in_features, out_features, bias) + nn.init.xavier_uniform_(m.weight) + if bias: + nn.init.constant_(m.bias, 0.0) + return m diff --git a/data/fairseq/fairseq/models/huggingface/__init__.py b/data/fairseq/fairseq/models/huggingface/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..f7911c2c8edf516855023a285b18935e5389ec02 --- /dev/null +++ b/data/fairseq/fairseq/models/huggingface/__init__.py @@ -0,0 +1,20 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import importlib +import os + + +# automatically import any Python files in the models/huggingface/ directory +models_dir = os.path.dirname(__file__) +for file in os.listdir(models_dir): + path = os.path.join(models_dir, file) + if ( + not file.startswith("_") + and not file.startswith(".") + and (file.endswith(".py") or os.path.isdir(path)) + ): + model_name = file[: file.find(".py")] if file.endswith(".py") else file + module = importlib.import_module("fairseq.models.huggingface." + model_name) diff --git a/data/fairseq/fairseq/models/huggingface/__pycache__/__init__.cpython-38.pyc b/data/fairseq/fairseq/models/huggingface/__pycache__/__init__.cpython-38.pyc new file mode 100644 index 0000000000000000000000000000000000000000..8e104f3f8fcafab7c9df58abd69bd1f43257606f Binary files /dev/null and b/data/fairseq/fairseq/models/huggingface/__pycache__/__init__.cpython-38.pyc differ diff --git a/data/fairseq/fairseq/models/huggingface/__pycache__/hf_gpt2.cpython-38.pyc b/data/fairseq/fairseq/models/huggingface/__pycache__/hf_gpt2.cpython-38.pyc new file mode 100644 index 0000000000000000000000000000000000000000..38574928fef3db3ab49a9b37e6ea5391c911473b Binary files /dev/null and b/data/fairseq/fairseq/models/huggingface/__pycache__/hf_gpt2.cpython-38.pyc differ diff --git a/data/fairseq/fairseq/models/huggingface/hf_gpt2.py b/data/fairseq/fairseq/models/huggingface/hf_gpt2.py new file mode 100644 index 0000000000000000000000000000000000000000..3a8eb78198f5808557092f814e92f1c9d72933ec --- /dev/null +++ b/data/fairseq/fairseq/models/huggingface/hf_gpt2.py @@ -0,0 +1,168 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import logging +import os +import sys +from typing import Dict, List, Optional + +import torch +from fairseq.models import ( + FairseqIncrementalDecoder, + FairseqLanguageModel, + register_model, + register_model_architecture, +) + + +logger = logging.getLogger(__name__) + + +DEFAULT_MAX_TARGET_POSITIONS = 1024 + + +@register_model("hf_gpt2") +class HuggingFaceGPT2LanguageModel(FairseqLanguageModel): + def __init__(self, decoder): + super().__init__(decoder) + + @staticmethod + def add_args(parser): + """Add model-specific arguments to the parser.""" + # fmt: off + parser.add_argument('--embed-dim', type=int, metavar='N', + help='embedding dimension') + parser.add_argument('--num-attention-heads', type=int, metavar='N', + help='num attention heads') + parser.add_argument('--num-layers', type=int, metavar='N', + help='num layers') + parser.add_argument('--dropout', type=float, metavar='D', + help='dropout probability for all fully connected layers ' + 'in the embeddings, encoder, and pooler') + parser.add_argument('--attention-dropout', type=float, metavar='D', + help='dropout probability for attention weights') + # fmt: on + + @classmethod + def build_model(cls, args, task): + """Build a new model instance.""" + default_architecture(args) + return cls(HuggingFaceGPT2Decoder(args, task)) + + +class HuggingFaceGPT2Decoder(FairseqIncrementalDecoder): + def __init__(self, args, task): + try: + from transformers import GPT2Config, GPT2LMHeadModel + except ImportError: + raise ImportError( + "\n\nPlease install huggingface/transformers with:" + "\n\n pip install transformers" + ) + + super().__init__(task.target_dictionary) + + config = GPT2Config( + vocab_size=len(task.target_dictionary), + n_positions=args.max_target_positions + 1, + n_ctx=args.max_target_positions, + n_embd=args.embed_dim, + n_layer=args.num_layers, + n_head=args.num_attention_heads, + resid_pdrop=args.dropout, + embd_pdrop=args.dropout, + attn_pdrop=args.attention_dropout, + layer_norm_epsilon=1e-6, + ) + self.model = GPT2LMHeadModel(config) + + # set zero embedding for padding symbol + self.pad_idx = task.target_dictionary.pad() + self.model.transformer.wte.weight.data[self.pad_idx].zero_() + self.model.transformer.wpe.weight.data[0].zero_() + + def forward( + self, + prev_output_tokens, + src_lengths=None, + incremental_state: Optional[Dict[str, List[torch.Tensor]]] = None, + encoder_out=None, + ): + features = self.extract_features(prev_output_tokens, incremental_state) + lm_logits = self.model.lm_head(features) + return (lm_logits,) + + def extract_features( + self, + prev_output_tokens, + incremental_state: Optional[Dict[str, List[torch.Tensor]]] = None, + ): + if incremental_state: + past = self.get_incremental_state("past") + else: + past = None + + # don't attend to padding symbols + attention_mask = prev_output_tokens.ne(self.pad_idx).int() + + # set position ids to exclude padding symbols + position_ids = attention_mask * ( + torch.arange(1, 1 + prev_output_tokens.size(1)) + .to(prev_output_tokens) + .repeat(prev_output_tokens.size(0), 1) + ) + + outputs = self.model.transformer( + input_ids=prev_output_tokens, + past=past, + attention_mask=attention_mask, + position_ids=position_ids, + ) + last_hidden_states = outputs[0] + + if incremental_state: + self.set_incremental_state(incremental_state, "past", outputs[1]) + + return last_hidden_states + + def max_positions(self): + return self.model.config.n_positions - 1 + + +@register_model_architecture("hf_gpt2", "hf_gpt2") +def default_architecture(args): + if getattr(args, "max_target_positions", None) is None: + args.max_target_positions = getattr( + args, "tokens_per_sample", DEFAULT_MAX_TARGET_POSITIONS + ) + args.embed_dim = getattr(args, "embed_dim", 768) + args.num_attention_heads = getattr(args, "num_attention_heads", 12) + args.num_layers = getattr(args, "num_layers", 12) + args.dropout = getattr(args, "dropout", 0.1) + args.attention_dropout = getattr(args, "attention_dropout", 0.1) + + +@register_model_architecture("hf_gpt2", "hf_gpt2_medium") +def hf_gpt2_medium(args): + args.embed_dim = getattr(args, "embed_dim", 1024) + args.num_attention_heads = getattr(args, "num_attention_heads", 16) + args.num_layers = getattr(args, "num_layers", 24) + default_architecture(args) + + +@register_model_architecture("hf_gpt2", "hf_gpt2_large") +def hf_gpt2_large(args): + args.embed_dim = getattr(args, "embed_dim", 1280) + args.num_attention_heads = getattr(args, "num_attention_heads", 20) + args.num_layers = getattr(args, "num_layers", 36) + default_architecture(args) + + +@register_model_architecture("hf_gpt2", "hf_gpt2_xl") +def hf_gpt2_xl(args): + args.embed_dim = getattr(args, "embed_dim", 1600) + args.num_attention_heads = getattr(args, "num_attention_heads", 25) + args.num_layers = getattr(args, "num_layers", 48) + default_architecture(args) diff --git a/data/fairseq/fairseq/models/multires_hubert/__init__.py b/data/fairseq/fairseq/models/multires_hubert/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..ec36505b0884f48d2aaeeee9b033bdf2fc38bd8b --- /dev/null +++ b/data/fairseq/fairseq/models/multires_hubert/__init__.py @@ -0,0 +1,2 @@ +from .multires_hubert import * # noqa +from .multires_hubert_asr import * # noqa diff --git a/data/fairseq/fairseq/models/multires_hubert/__pycache__/__init__.cpython-38.pyc b/data/fairseq/fairseq/models/multires_hubert/__pycache__/__init__.cpython-38.pyc new file mode 100644 index 0000000000000000000000000000000000000000..68b0b291449bbf3ea8cd2e491c063aefa2ecf63c Binary files /dev/null and b/data/fairseq/fairseq/models/multires_hubert/__pycache__/__init__.cpython-38.pyc differ diff --git a/data/fairseq/fairseq/models/multires_hubert/__pycache__/multires_hubert.cpython-38.pyc b/data/fairseq/fairseq/models/multires_hubert/__pycache__/multires_hubert.cpython-38.pyc new file mode 100644 index 0000000000000000000000000000000000000000..d16b0b3da9d3c14b74b6529df1f3fc29b91d1eaa Binary files /dev/null and b/data/fairseq/fairseq/models/multires_hubert/__pycache__/multires_hubert.cpython-38.pyc differ diff --git a/data/fairseq/fairseq/models/multires_hubert/__pycache__/multires_hubert_asr.cpython-38.pyc b/data/fairseq/fairseq/models/multires_hubert/__pycache__/multires_hubert_asr.cpython-38.pyc new file mode 100644 index 0000000000000000000000000000000000000000..d5751a3d978f17f12fb975cfd42f00ba3644c346 Binary files /dev/null and b/data/fairseq/fairseq/models/multires_hubert/__pycache__/multires_hubert_asr.cpython-38.pyc differ diff --git a/data/fairseq/fairseq/models/multires_hubert/multires_hubert.py b/data/fairseq/fairseq/models/multires_hubert/multires_hubert.py new file mode 100644 index 0000000000000000000000000000000000000000..eacb29e5fe1d23eed39eb97656e4fe305fd18c76 --- /dev/null +++ b/data/fairseq/fairseq/models/multires_hubert/multires_hubert.py @@ -0,0 +1,1231 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import logging +from dataclasses import dataclass, field +from typing import Dict, List, Optional, Tuple + +import numpy as np +import torch +import math +import torch.nn as nn +from omegaconf import II +from fairseq.models.wav2vec.wav2vec import norm_block + +from fairseq import utils +from fairseq.data.data_utils import compute_mask_indices +from fairseq.data.dictionary import Dictionary +from fairseq.dataclass import ChoiceEnum, FairseqDataclass +from fairseq.models import BaseFairseqModel, register_model +from fairseq.models.wav2vec.wav2vec2 import ( + EXTRACTOR_MODE_CHOICES, + MASKING_DISTRIBUTION_CHOICES, + LAYER_TYPE_CHOICES, + ConvFeatureExtractionModel, + TransformerEncoder, +) +from omegaconf import II, MISSING, open_dict +from fairseq.modules import GradMultiply, LayerNorm +from fairseq.tasks.multires_hubert_pretraining import ( + MultiresHubertPretrainingConfig, + MultiresHubertPretrainingTask, +) + +logger = logging.getLogger(__name__) + + +@dataclass +class MultiresHubertConfig(FairseqDataclass): + label_rate: float = II("task.label_rate") + # label_rate: 1,2,2,5 + # (imply (1,2), (2,5)) + # if base label_rate = 50 + # (1,2), (2,5) --> label rates 50, 25, 10 + label_rate_ratios: List[int] = field( + default=MISSING, metadata={"help": "tuple for label rates e.g., [(1,2), (2,5)]"} + ) + + extractor_mode: EXTRACTOR_MODE_CHOICES = field( + default="default", + metadata={ + "help": "mode for feature extractor. default has a single group " + "norm with d groups in the first conv block, whereas layer_norm " + "has layer norms in every block (meant to use with normalize=True)" + }, + ) + # the blocks for each label rate + encoder_layers: int = field( + default="2", + metadata={ + "help": "num encoder layers in the each block (one sub module of the U-net)" + }, + ) + override_encoder_layers: str = field( + default="", + metadata={ + "help": "specific layer numbers for each block (one sub module of the U-net) for the training" + }, + ) + encoder_embed_dim: int = field( + default=768, metadata={"help": "encoder embedding dimension"} + ) + encoder_ffn_embed_dim: int = field( + default=3072, metadata={"help": "encoder embedding dimension for FFN"} + ) + encoder_attention_heads: int = field( + default=12, metadata={"help": "num encoder attention heads"} + ) + activation_fn: ChoiceEnum(utils.get_available_activation_fns()) = field( + default="gelu", metadata={"help": "activation function to use"} + ) + layer_type: LAYER_TYPE_CHOICES = field( + default="transformer", metadata={"help": "layer type in encoder"} + ) + conv_adapator_kernal: int = field( + default=7, metadata={"help": "kernal size for conv adaptor"} + ) + use_plain_updownsample: bool = field( + default=False, metadata={"help": "whether to use plain up downsample"} + ) + + # dropouts + dropout: float = field( + default=0.1, + metadata={"help": "dropout probability for the transformer"}, + ) + attention_dropout: float = field( + default=0.1, + metadata={"help": "dropout probability for attention weights"}, + ) + activation_dropout: float = field( + default=0.0, + metadata={"help": "dropout probability after activation in FFN"}, + ) + encoder_layerdrop: float = field( + default=0.0, + metadata={"help": "probability of dropping a tarnsformer layer"}, + ) + dropout_input: float = field( + default=0.0, + metadata={"help": "dropout to apply to the input (after feat extr)"}, + ) + dropout_features: float = field( + default=0.0, + metadata={"help": "dropout to apply to the features (after feat extr)"}, + ) + + final_dim: int = field( + default=0, + metadata={ + "help": "project final representations and targets to this many " + "dimensions. set to encoder_embed_dim is <= 0" + }, + ) + untie_final_proj: bool = field( + default=True, + metadata={"help": "use separate projection for each target"}, + ) + layer_norm_first: bool = field( + default=False, + metadata={"help": "apply layernorm first in the transformer"}, + ) + conv_feature_layers: str = field( + default="[(512,10,5)] + [(512,3,2)] * 4 + [(512,2,2)] * 2", + metadata={ + "help": "string describing convolutional feature extraction " + "layers in form of a python list that contains " + "[(dim, kernel_size, stride), ...]" + }, + ) + conv_bias: bool = field( + default=False, metadata={"help": "include bias in conv encoder"} + ) + logit_temp: float = field( + default=0.1, metadata={"help": "temperature to divide logits by"} + ) + target_glu: bool = field( + default=False, metadata={"help": "adds projection + glu to targets"} + ) + feature_grad_mult: float = field( + default=1.0, + metadata={"help": "multiply feature extractor var grads by this"}, + ) + use_single_target: bool = field( + default=False, + metadata={ + "help": "whether to use single data (in that case, we will compute with the fixed label rate)" + }, + ) + use_single_prediction: bool = field( + default=False, + metadata={ + "help": "if true, we will not conduct mlm prediction in low resolution in the middle" + }, + ) + use_multi_stream: bool = field( + default=False, + metadata={ + "help": "whether to use multi-stream setting (in this setting, we have multiple streams with the same resolution)" + }, + ) + + # masking + mask_length: int = field(default=10, metadata={"help": "mask length"}) + mask_prob: float = field( + default=0.65, + metadata={"help": "probability of replacing a token with mask"}, + ) + mask_selection: MASKING_DISTRIBUTION_CHOICES = field( + default="static", metadata={"help": "how to choose mask length"} + ) + mask_other: float = field( + default=0, + metadata={ + "help": "secondary mask argument " + "(used for more complex distributions), " + "see help in compute_mask_indicesh" + }, + ) + no_mask_overlap: bool = field( + default=False, metadata={"help": "whether to allow masks to overlap"} + ) + mask_min_space: int = field( + default=1, + metadata={"help": "min space between spans (if no overlap is enabled)"}, + ) + + # channel masking + mask_channel_length: int = field( + default=10, + metadata={"help": "length of the mask for features (channels)"}, + ) + mask_channel_prob: float = field( + default=0.0, + metadata={"help": "probability of replacing a feature with 0"}, + ) + mask_channel_selection: MASKING_DISTRIBUTION_CHOICES = field( + default="static", + metadata={"help": "how to choose mask length for channel masking"}, + ) + mask_channel_other: float = field( + default=0, + metadata={ + "help": "secondary mask argument " + "(used for more complex distributions), " + "see help in compute_mask_indicesh" + }, + ) + no_mask_channel_overlap: bool = field( + default=False, + metadata={"help": "whether to allow channel masks to overlap"}, + ) + mask_channel_min_space: int = field( + default=1, + metadata={"help": "min space between spans (if no overlap is enabled)"}, + ) + + # positional embeddings + conv_pos: int = field( + default=128, + metadata={"help": "number of filters for convolutional positional embeddings"}, + ) + conv_pos_groups: int = field( + default=16, + metadata={"help": "number of groups for convolutional positional embedding"}, + ) + + latent_temp: Tuple[float, float, float] = field( + default=(2, 0.5, 0.999995), + metadata={"help": "legacy (to be removed)"}, + ) + + # loss computation + skip_masked: bool = field( + default=False, + metadata={"help": "skip computing losses over masked frames"}, + ) + skip_nomask: bool = field( + default=False, + metadata={"help": "skip computing losses over unmasked frames"}, + ) + + checkpoint_activations: bool = field( + default=False, + metadata={"help": "recompute activations and save memory for extra compute"}, + ) + + # FP16 optimization + required_seq_len_multiple: int = field( + default=2, + metadata={ + "help": "pad the input to encoder such that the sequence length is divisible by multiple" + }, + ) + + # Conformer + depthwise_conv_kernel_size: int = field( + default=31, + metadata={ + "help": "depthwise-conv-kernel-size for convolution in conformer layer" + }, + ) + attn_type: str = field( + default="", + metadata={"help": "if espnet use ESPNET MHA"}, + ) + pos_enc_type: str = field( + default="abs", + metadata={"help": "Positional encoding type to use in conformer"}, + ) + fp16: bool = field(default=False, metadata={"help": "If fp16 is being used"}) + + +@register_model("multires_hubert", dataclass=MultiresHubertConfig) +class MultiresHubertModel(BaseFairseqModel): + def __init__( + self, + cfg: MultiresHubertConfig, + task_cfg: MultiresHubertPretrainingConfig, + dictionaries: List[Dictionary], + ) -> None: + super().__init__() + logger.info(f"MultiresHubertModel Config: {cfg}") + + feature_enc_layers = eval(cfg.conv_feature_layers) # noqa + self.embed = feature_enc_layers[-1][0] + + self.feature_extractor = ConvFeatureExtractionModel( + conv_layers=feature_enc_layers, + dropout=0.0, + mode=cfg.extractor_mode, + conv_bias=cfg.conv_bias, + ) + self.post_extract_proj = ( + nn.Linear(self.embed, cfg.encoder_embed_dim) + if self.embed != cfg.encoder_embed_dim + else None + ) + + # Estimate label rates + assert ( + cfg.label_rate_ratios != "None" + ), "without ratios, the model is exactly as the Hubert model" + self.label_rate_ratios = [] + self.base_rate = cfg.label_rate + self.label_rates = [] + self.downsample_modules = nn.ModuleList() + self.upsample_modules = nn.ModuleList() + self.encoders = nn.ModuleList() + self.decoders = nn.ModuleList() + self.use_single_target = cfg.use_single_target + self.use_single_prediction = cfg.use_single_prediction + self.use_plain_updownsample = cfg.use_plain_updownsample + + # For decide the override encoder layers, so that the layer number is not equally distributed + if cfg.override_encoder_layers != "": + self.override_encoder_layers = eval(cfg.override_encoder_layers) + assert ( + len(self.override_encoder_layers) % 2 == 1 + ), "must be odd number of layers if specify detailed layers" + assert ( + len(self.override_encoder_layers) // 2 + == len(cfg.label_rate_ratios) // 2 + ), "number of override encoder layers must match the label rate ratios information" + self.len_encoder_modules = len(self.override_encoder_layers) + else: + self.override_encoder_layers = None + self.len_encoder_modules = None + + # use different layers instead of equally distributed ones + middle_override_encoder_layer = ( + self.override_encoder_layers[self.len_encoder_modules // 2] + if self.override_encoder_layers is not None + else None + ) + skip_middle_pos_conv = False if len(cfg.label_rate_ratios) < 2 else True + + self.middle_encoder = TransformerEncoder( + cfg, + skip_pos_conv=skip_middle_pos_conv, + override_encoder_layer=middle_override_encoder_layer, + ) + + first_pos_conv = False # only enable pos_conv for the first encoder + raw_label_rate_ratios = cfg.label_rate_ratios + for i in range(len(raw_label_rate_ratios) // 2): + # check if have override encoder layers + if self.override_encoder_layers is not None: + override_encoder_layer = self.override_encoder_layers[i] + override_decoder_layer = self.override_encoder_layers[ + self.len_encoder_modules - 1 - i + ] + else: + override_encoder_layer, override_decoder_layer = None, None + + self.label_rate_ratios.append( + (raw_label_rate_ratios[i * 2], raw_label_rate_ratios[i * 2 + 1]) + ) + if self.use_plain_updownsample: + self.downsample_modules.append( + ConvDownsampler( + k=cfg.conv_adapator_kernal, + label_rate=( + ( + raw_label_rate_ratios[i * 2], + raw_label_rate_ratios[i * 2 + 1], + ) + ), + dropout=0.0, + channels=cfg.encoder_embed_dim, + activation=nn.GELU(), + log_compression=False, + skip_connections=True, + highway=True, + residual_scale=0.4, + ) + ) + else: + self.downsample_modules.append( + ConvAdapter( + k=cfg.conv_adapator_kernal, + label_rate=( + ( + raw_label_rate_ratios[i * 2], + raw_label_rate_ratios[i * 2 + 1], + ) + ), + dropout=0.0, + channels=cfg.encoder_embed_dim, + activation=nn.GELU(), + log_compression=False, + skip_connections=True, + highway=True, + residual_scale=0.4, + ) + ) + if not first_pos_conv: + self.encoders.append( + TransformerEncoder( + cfg, override_encoder_layer=override_encoder_layer + ) + ) # TODO(jiatong): add conformer options + first_pos_conv = True + else: + self.encoders.append( + TransformerEncoder( + cfg, + skip_pos_conv=True, + override_encoder_layer=override_encoder_layer, + ) + ) + if self.use_plain_updownsample: + self.upsample_modules.append( + ConvUpsampler( + k=cfg.conv_adapator_kernal, + label_rate=( + ( + raw_label_rate_ratios[i * 2 + 1], + raw_label_rate_ratios[i * 2], + ) + ), + dropout=0.0, + channels=cfg.encoder_embed_dim, + activation=nn.GELU(), + log_compression=False, + skip_connections=True, + highway=True, + residual_scale=0.4, + ) + ) + else: + self.upsample_modules.append( + ConvAdapter( + k=cfg.conv_adapator_kernal, + label_rate=( + ( + raw_label_rate_ratios[i * 2 + 1], + raw_label_rate_ratios[i * 2], + ) + ), + dropout=0.0, + channels=cfg.encoder_embed_dim, + activation=nn.GELU(), + log_compression=False, + skip_connections=True, + highway=True, + residual_scale=0.4, + ) + ) + self.decoders.append( + TransformerEncoder( + cfg, + skip_pos_conv=True, + override_encoder_layer=override_decoder_layer, + ) + ) + + base_ds_rate = np.prod([s for _, _, s in feature_enc_layers]) + self.feature_ds_rates = [base_ds_rate] + running_rate = self.base_rate + + if cfg.use_single_target or cfg.use_multi_stream: + self.label_rates = self.base_rate + else: + self.label_rates.append(self.base_rate) + + for label_rate_ratio in self.label_rate_ratios: + upsample_rate, downsample_rate = label_rate_ratio + if (base_ds_rate * upsample_rate) % downsample_rate != 0: + logger.warning( + "base rate: {} cannot be ideally processed with downsample rate {}".format( + base_ds_rate, downsample_rate + ) + ) + + base_ds_rate = base_ds_rate * downsample_rate // upsample_rate + self.feature_ds_rates.append(base_ds_rate) + + if not cfg.use_single_target and not cfg.use_multi_stream: + running_rate = running_rate * upsample_rate // downsample_rate + self.label_rates.append(running_rate) + self.label_nums = len( + self.feature_ds_rates + ) # the number of labels for prediction (activate at iter 2) + + if type(self.label_rates) == float: + self.feat2tar_ratios = [ + self.feature_ds_rates[i] * self.label_rates / task_cfg.sample_rate + for i in range(len(self.feature_ds_rates)) + ] + else: + self.feat2tar_ratios = [ + self.feature_ds_rates[i] * self.label_rates[i] / task_cfg.sample_rate + for i in range(len(self.feature_ds_rates)) + ] + + # self.feat2tar_ratios = self.feat2tar_ratios[::-1] + + # An running example of the label rate: + # base_ds_rate = 320 + # self.label_rate_ratios = [(1, 2)] + # self.feature_ds_rates = [320, 640] + # self.label_rates = [50, 25] + # self.feat2tar_ratios = [1, 1] + + # Another running example of the label rate: + # base_ds_rate = 320 + # self.label_rate_ratios = [(1, 2)] + # self.feature_ds_rates = [320, 640] + # self.label_rates = 100 + # self.feat2tar_ratios = [4, 2] + # self.use_sinlge_target = True + + logging.info( + "ds_rates: {}, label_rates: {}, feat2tar_ratios: {}".format( + self.feature_ds_rates, self.label_rates, self.feat2tar_ratios + ) + ) + + self.mask_prob = cfg.mask_prob + self.mask_selection = cfg.mask_selection + self.mask_other = cfg.mask_other + self.mask_length = cfg.mask_length + self.no_mask_overlap = cfg.no_mask_overlap + self.mask_min_space = cfg.mask_min_space + + self.mask_channel_prob = cfg.mask_channel_prob + self.mask_channel_selection = cfg.mask_channel_selection + self.mask_channel_other = cfg.mask_channel_other + self.mask_channel_length = cfg.mask_channel_length + self.no_mask_channel_overlap = cfg.no_mask_channel_overlap + self.mask_channel_min_space = cfg.mask_channel_min_space + + self.dropout_input = nn.Dropout(cfg.dropout_input) + self.dropout_features = nn.Dropout(cfg.dropout_features) + + self.feature_grad_mult = cfg.feature_grad_mult + self.logit_temp = cfg.logit_temp + self.skip_masked = cfg.skip_masked + self.skip_nomask = cfg.skip_nomask + + # Note(jiatong): different from hubert, we just set the final dim as encoder_embed_dim + final_dim = cfg.final_dim if cfg.final_dim > 0 else cfg.encoder_embed_dim + + self.mask_emb = nn.Parameter( + torch.FloatTensor(cfg.encoder_embed_dim).uniform_() + ) + + self.layer_norm = LayerNorm(self.embed) + + self.predictor_head_num = 1 if self.use_single_prediction else self.label_nums + + self.target_glu = None + if cfg.target_glu: + self.target_glus = nn.ModuleList() + for i in range(self.predictor_head_num): + self.target_glus.append( + nn.Sequential(nn.Linear(final_dim, final_dim * 2), nn.GLU()) + ) + + self.untie_final_proj = cfg.untie_final_proj + self.final_projs = nn.ModuleList() + + # Note(jiatong): we do not have untie cases for multires hubert + for i in range(self.predictor_head_num): + self.final_projs.append(nn.Linear(cfg.encoder_embed_dim, final_dim)) + + # modules below are not needed during fine-tuning + self.multires_classes = [] + self.label_embs_concat = nn.ParameterList() + + for i in range(self.predictor_head_num): + if self.use_single_target: + num_classes = len(dictionaries[0]) + else: + num_classes = len(dictionaries[i]) + self.multires_classes.append(num_classes) + self.label_embs_concat.append( + nn.Parameter(torch.FloatTensor(num_classes, final_dim)) + ) + nn.init.uniform_(self.label_embs_concat[i]) + + def upgrade_state_dict_named(self, state_dict, name): + """Upgrade a (possibly old) state dict for new versions of fairseq.""" + + super().upgrade_state_dict_named(state_dict, name) + return state_dict + + @classmethod + def build_model( + cls, cfg: MultiresHubertConfig, task: MultiresHubertPretrainingTask + ): + """Build a new model instance.""" + + model = MultiresHubertModel(cfg, task.cfg, task.dictionaries) + return model + + def apply_mask(self, x, padding_mask, target_list): + B, T, C = x.shape + if self.mask_prob > 0: + mask_indices = compute_mask_indices( + (B, T), + padding_mask, + self.mask_prob, + self.mask_length, + self.mask_selection, + self.mask_other, + min_masks=2, + no_overlap=self.no_mask_overlap, + min_space=self.mask_min_space, + ) + mask_indices = torch.from_numpy(mask_indices).to(x.device) + x[mask_indices] = self.mask_emb + else: + mask_indices = None + + if self.mask_channel_prob > 0: + mask_channel_indices = compute_mask_indices( + (B, C), + None, + self.mask_channel_prob, + self.mask_channel_length, + self.mask_channel_selection, + self.mask_channel_other, + no_overlap=self.no_mask_channel_overlap, + min_space=self.mask_channel_min_space, + ) + mask_channel_indices = ( + torch.from_numpy(mask_channel_indices) + .to(x.device) + .unsqueeze(1) + .expand(-1, T, -1) + ) + x[mask_channel_indices] = 0 + + return x, mask_indices + + def compute_nce(self, x, pos, negs): + neg_is_pos = (pos == negs).all(-1) + pos = pos.unsqueeze(0) + targets = torch.cat([pos, negs], dim=0) + + logits = torch.cosine_similarity(x.float(), targets.float(), dim=-1).type_as(x) + logits /= self.logit_temp + if neg_is_pos.any(): + logits[1:][neg_is_pos] = float("-inf") + logits = logits.transpose(0, 1) # (num_x, num_cls+1) + return logits + + def forward_features(self, source: torch.Tensor) -> torch.Tensor: + if self.feature_grad_mult > 0: + features = self.feature_extractor(source) + if self.feature_grad_mult != 1.0: + features = GradMultiply.apply(features, self.feature_grad_mult) + else: + with torch.no_grad(): + features = self.feature_extractor(source) + return features + + def forward_targets( + self, + features: torch.Tensor, + target: torch.Tensor, + feat2tar_ratio: float, + ) -> Tuple[torch.Tensor, torch.Tensor]: + # Trim features to ensure labels exist and then get aligned labels + + feat_tsz = features.size(1) + + # skip if no target is provided + if target is None: + return features, None, None + targ_tsz = target.size(1) + if feat2tar_ratio * feat_tsz > targ_tsz: + feat_tsz = int(targ_tsz / feat2tar_ratio) + features = features[:, :feat_tsz] + target_inds = torch.arange(feat_tsz).float() * feat2tar_ratio + target = target[:, target_inds.long()] + return features, target + + def forward_padding_mask( + self, + features: torch.Tensor, + padding_mask: torch.Tensor, + ) -> torch.Tensor: + extra = padding_mask.size(1) % features.size(1) + if extra > 0: + padding_mask = padding_mask[:, :-extra] + padding_mask = padding_mask.view(padding_mask.size(0), features.size(1), -1) + padding_mask = padding_mask.all(-1) + return padding_mask + + def forward( + self, + source: torch.Tensor, + target_list: Optional[List[torch.Tensor]] = None, + padding_mask: Optional[torch.Tensor] = None, + mask: bool = True, + features_only: bool = False, + output_layer: Optional[int] = None, + ) -> Dict[str, torch.Tensor]: + """output layer is 1-based""" + features = self.forward_features(source) + + features_pen = features.float().pow(2).mean() + + features = features.transpose(1, 2) + features = self.layer_norm(features) + unmasked_features = features.clone() + + if padding_mask is not None: + padding_mask = self.forward_padding_mask(features, padding_mask) + + if self.post_extract_proj is not None: + features = self.post_extract_proj(features) + + features = self.dropout_input(features) + unmasked_features = self.dropout_features(unmasked_features) + + if mask: + x, mask_indices = self.apply_mask(features, padding_mask, target_list) + else: + x = features + mask_indices = None + + # feature: (B, T, D), float + # target: (B, T), long + # x: (B, T, D), float + # padding_mask: (B, T), bool + # mask_indices: (B, T), bool + + def align_size_sum(feat1, pad1, feat2): + assert ( + abs(feat1.size(1) - feat2.size(1)) < 10 + ), "misaligned results for feat1 and feat2 of size {} - {}".format( + feat1.size(1), feat2.size(1) + ) + common_size = min(feat1.size(1), feat2.size(1)) + + return ( + feat1[:, :common_size] + feat2[:, :common_size], + pad1[:, :common_size], + ) + + # process encoders + res_outputs = [] # final output for different resolution + multi_mask_indices = [] # mask indices for different resolution + residuals = [] # record the x in encoders + padding_masks = [] # final padding masks + # The encoder has (self.label_nums - 1) blocks + for i in range(self.label_nums - 1): + x, _ = self.encoders[i](x, padding_mask=padding_mask, layer=None) + residuals.append(x) + x, padding_mask, mask_indices = self.downsample_modules[i]( + x, padding=padding_mask, mask_indices=mask_indices + ) + + residual = self.middle_encoder(x, padding_mask=padding_mask, layer=None)[0] + x = x + residual + res_outputs.append(x) + + # process decoders + # The encoder has (self.label_nums - 1) blocks + padding_masks.append(padding_mask) + multi_mask_indices.append(mask_indices) + residuals.reverse() # NOTE(jiatong): reverse res_output to match corresponding input + for i in range(self.label_nums - 1): + x, padding_mask, mask_indices = self.upsample_modules[ + self.label_nums - 2 - i + ](x, padding=padding_mask, mask_indices=mask_indices) + x, _ = self.decoders[i](x, padding_mask=padding_mask, layer=None) + x, padding_mask = align_size_sum(x, padding_mask, residuals[i]) + res_outputs.append(x) + padding_masks.append(padding_mask) + multi_mask_indices.append(mask_indices) + + # NOTE(jiatong): need reverse of target list to allow matched target-representation + res_outputs.reverse() + padding_masks.reverse() + multi_mask_indices.reverse() + if target_list is not None: + new_target_list = [] + for i in range(self.label_nums): + if self.use_single_target: + res_outputs[i], reformat_target_list = self.forward_targets( + res_outputs[i], target_list[0], self.feat2tar_ratios[i] + ) + new_target_list.append(reformat_target_list) + else: + if target_list[i] is not None: + res_outputs[i], reformat_target_list = self.forward_targets( + res_outputs[i], target_list[i], self.feat2tar_ratios[i] + ) + new_target_list.append(reformat_target_list) + else: + # Append a None target list then it won't be used to calculate loss + new_target_list.append(None) + if padding_masks[i] is not None: + padding_masks[i] = self.forward_padding_mask( + res_outputs[i], padding_masks[i] + ) + if multi_mask_indices[i] is not None: + multi_mask_indices[i] = self.forward_padding_mask( + res_outputs[i], multi_mask_indices[i] + ) + + + if features_only: + # NOTE(jiatong): need to reverse back + res_outputs.reverse() + return { + "x": res_outputs, + "padding_mask": padding_masks[0], + "features": features, + } + + def compute_pred(proj_x, target, label_embs): + # compute logits for the i-th label set + y = torch.index_select(label_embs, 0, target.long()) + negs = label_embs.unsqueeze(1).expand(-1, proj_x.size(0), -1) + if self.target_glu: + y = self.target_glu(y) + negs = self.target_glu(negs) + # proj_x: (S, D) + # y: (S, D) + # negs: (Neg, S, D) + return self.compute_nce(proj_x, y, negs) + + logit_m_list, logit_u_list = [], [] + for j in range(self.label_nums): + if new_target_list[j] is None: + continue # skip empty targets + label_embs_list = self.label_embs_concat[j].split( + [self.multires_classes[j]], 0 + ) + # set the variables (after the set, the procedure is the same as hubert) + # all the elements are list with only one element (to simulate the normal hubert process) + x = res_outputs[j] + target = new_target_list[j] + padding_mask = padding_masks[j] + mask_indices = multi_mask_indices[j] + final_proj = self.final_projs[j] + + if not self.skip_masked: + masked_indices = torch.logical_and(~padding_mask, mask_indices) + proj_x_m = final_proj(x[masked_indices]) + logit_m_list.append( + compute_pred(proj_x_m, target[masked_indices], label_embs_list[0]) + ) + else: + logit_m_list.append(None) + + if not self.skip_nomask: + nomask_indices = torch.logical_and(~padding_mask, ~mask_indices) + proj_x_u = final_proj(x[nomask_indices]) + logit_u_list.append( + compute_pred(proj_x_u, target[nomask_indices], label_embs_list[0]) + ) + else: + logit_u_list.append(None) + + # if we only want one prediction, we can exit now + if self.predictor_head_num == 1: + break + + result = { + "logit_m_list": logit_m_list, + "logit_u_list": logit_u_list, + "padding_mask": padding_mask, + "features_pen": features_pen, + } + return result + + def extract_features( + self, + source: torch.Tensor, + padding_mask: Optional[torch.Tensor] = None, + mask: bool = False, + ret_conv: bool = False, + output_layer: Optional[int] = None, + last_layer: Optional[bool] = False, + ) -> Tuple[torch.Tensor, torch.Tensor]: + res = self.forward( + source, + padding_mask=padding_mask, + mask=mask, + features_only=True, + output_layer=output_layer, + ) + feature = res["features"] if ret_conv else res["x"] + if last_layer: + feature = feature[-1] + return feature, res["padding_mask"] + + def get_logits(self, net_output, is_masked=True): + if is_masked: + logits_list = net_output["logit_m_list"] + else: + logits_list = net_output["logit_u_list"] + logits_list = [x.float() for x in logits_list if x is not None] + return logits_list + + def get_targets(self, net_output, is_masked=True): + logits_list = self.get_logits(net_output, is_masked) + targets_list = [x.new_zeros(x.size(0), dtype=torch.long) for x in logits_list] + return targets_list + + def get_extra_losses(self, net_output): + extra_losses = [] + names = [] + + if "features_pen" in net_output: + extra_losses.append(net_output["features_pen"]) + names.append("features_pen") + + return extra_losses, names + + def remove_pretraining_modules(self): + self.target_glu = None + self.final_proj = None + + +class ConvAdapter(nn.Module): + """Conv adapter that combines two modules with different label rate with downsample or upsample. + To allow different ratios than integer, two convs are utilized with first to upsample (numerator) + and the second to downsample (denominator)""" + + def __init__( + self, + k, + label_rate, + dropout, + channels, + activation, + log_compression=False, + skip_connections=True, + highway=True, + residual_scale=0.4, + non_affine_group_norm=False, + ): + super().__init__() + + def downsample_block(channel, k, stride): + return nn.Sequential( + # with padding (k - 1) // 2 to keep the same size + nn.Conv1d( + channel, + channel, + k, + stride=stride, + bias=False, + padding=(k - 1) // 2, + ), + nn.Dropout(p=dropout), + norm_block( + is_layer_norm=False, dim=channel, affine=not non_affine_group_norm + ), + activation, + ) + + def upsample_block(channel, k, stride): + return nn.Sequential( + # with padding (k - 1) // 2 to keep the same size + nn.ConvTranspose1d( + channel, + channel, + k, + stride=stride, + bias=False, + padding=0, # padding=(k - 1) // 2, + output_padding=(stride - 1), + ), + nn.Dropout(p=dropout), + norm_block( + is_layer_norm=False, dim=channel, affine=not non_affine_group_norm + ), + activation, + ) + + assert len(label_rate) == 2, "label_rate should be sized two to apply fusion" + # Lout =(Lin~H~R1)~Wstride~H~R2~Wpadding+dilation~W(kernel_size~H~R1)+output_padding+1 + self.upsample_conv = upsample_block(channels, k, label_rate[0]) + self.downsample_conv = downsample_block(channels, k, label_rate[1]) + + self.upsample_rate, self.downsample_rate = label_rate + self.log_compression = log_compression + self.skip_connections = skip_connections + self.highway = highway + self.residual_scale = math.sqrt(residual_scale) + + def forward(self, x, padding=None, mask_indices=None): + # Assume x1 = (B, T, C) as input + x = x.permute(0, 2, 1) + residual_before_upsample = x + x = self.upsample_conv(x) + upsample_size = x.size(2) + + # conduct upsample + if self.skip_connections: + residual_upsample = torch.repeat_interleave( + residual_before_upsample, self.upsample_rate, dim=2 + ) + upsample_size = min(upsample_size, residual_upsample.size(2)) + x = ( + x[..., :upsample_size] + residual_upsample[..., :upsample_size] + ) * self.residual_scale + + residual_before_downsample = x + x = self.downsample_conv(x) + downsample_size = x.size(2) + + if self.skip_connections: + residual_downsample = residual_before_downsample[ + ..., :: self.downsample_rate + ] + downsample_size = min(x.size(2), residual_downsample.size(2)) + x = ( + x[..., :downsample_size] + residual_downsample[..., :downsample_size] + ) * self.residual_scale + + if self.highway: + residual_after_sample = residual_upsample[..., :: self.downsample_rate] + final_size = min(x.size(2), residual_after_sample.size(2)) + x = ( + x[..., :final_size] + residual_after_sample[..., :final_size] + ) * self.residual_scale + + if self.log_compression: + x = x.abs() + x = x + 1 + x = x.log() + + x = x.permute(0, 2, 1) + + # process padding + if padding is not None: + padding = torch.repeat_interleave(padding, self.upsample_rate, dim=1) + padding = padding[..., :: self.downsample_rate] + padding = padding[..., : x.size(1)] + + # process mask indices + if mask_indices is not None: + mask_indices = torch.repeat_interleave( + mask_indices, self.upsample_rate, dim=1 + ) + mask_indices = mask_indices[..., :: self.downsample_rate] + mask_indices = mask_indices[..., : x.size(1)] + return x, padding, mask_indices + + +class ConvDownsampler(nn.Module): + """Conv downsampler that combines two modules with different label rate with downsample or upsample. + To allow different ratios than integer, two convs are utilized with first to upsample (numerator) + and the second to downsample (denominator)""" + + def __init__( + self, + k, + label_rate, + dropout, + channels, + activation, + log_compression=False, + skip_connections=True, + highway=True, + residual_scale=0.4, + non_affine_group_norm=False, + ): + super().__init__() + + def downsample_block(channel, k, stride): + return nn.Sequential( + # with padding (k - 1) // 2 to keep the same size + nn.Conv1d( + channel, + channel, + k, + stride=stride, + bias=False, + padding=(k - 1) // 2, + ), + nn.Dropout(p=dropout), + norm_block( + is_layer_norm=False, dim=channel, affine=not non_affine_group_norm + ), + activation, + ) + + assert len(label_rate) == 2, "label_rate should be sized two to apply fusion" + self.downsample_conv = downsample_block(channels, k, label_rate[1]) + + upsample_rate, self.downsample_rate = label_rate + assert upsample_rate == 1, "must be 1 to perform downsample only" + self.log_compression = log_compression + self.skip_connections = skip_connections + self.highway = highway # Useless as placeholder + self.residual_scale = math.sqrt(residual_scale) + + def forward(self, x, padding=None, mask_indices=None): + # Assume x1 = (B, T, C) as input + x = x.permute(0, 2, 1) + + residual_before_downsample = x + x = self.downsample_conv(x) + downsample_size = x.size(2) + + if self.skip_connections: + residual_downsample = residual_before_downsample[ + ..., :: self.downsample_rate + ] + downsample_size = min(x.size(2), residual_downsample.size(2)) + x = ( + x[..., :downsample_size] + residual_downsample[..., :downsample_size] + ) * self.residual_scale + + if self.log_compression: + x = x.abs() + x = x + 1 + x = x.log() + + x = x.permute(0, 2, 1) + + # process padding + if padding is not None: + padding = padding[..., :: self.downsample_rate] + padding = padding[..., : x.size(1)] + + # process mask indices + if mask_indices is not None: + mask_indices = mask_indices[..., :: self.downsample_rate] + mask_indices = mask_indices[..., : x.size(1)] + return x, padding, mask_indices + + +class ConvUpsampler(nn.Module): + """Conv upsampler that combines two modules with different label rate with downsample or upsample. + To allow different ratios than integer, two convs are utilized with first to upsample (numerator) + and the second to downsample (denominator)""" + + def __init__( + self, + k, + label_rate, + dropout, + channels, + activation, + log_compression=False, + skip_connections=True, + highway=True, + residual_scale=0.4, + non_affine_group_norm=False, + ): + super().__init__() + + def upsample_block(channel, k, stride): + return nn.Sequential( + # with padding (k - 1) // 2 to keep the same size + nn.ConvTranspose1d( + channel, + channel, + k, + stride=stride, + bias=False, + padding=0, # padding=(k - 1) // 2, + output_padding=(stride - 1), + ), + nn.Dropout(p=dropout), + norm_block( + is_layer_norm=False, dim=channel, affine=not non_affine_group_norm + ), + activation, + ) + + assert len(label_rate) == 2, "label_rate should be sized two to apply fusion" + # Lout =(Lin~H~R1)~Wstride~H~R2~Wpadding+dilation~W(kernel_size~H~R1)+output_padding+1 + self.upsample_conv = upsample_block(channels, k, label_rate[0]) + + self.upsample_rate, downsample_rate = label_rate + assert downsample_rate == 1, "must be 1 to perform downsample only" + self.log_compression = log_compression + self.skip_connections = skip_connections + self.highway = highway # Useless + self.residual_scale = math.sqrt(residual_scale) + + def forward(self, x, padding=None, mask_indices=None): + # Assume x1 = (B, T, C) as input + x = x.permute(0, 2, 1) + residual_before_upsample = x + x = self.upsample_conv(x) + upsample_size = x.size(2) + + # conduct upsample + if self.skip_connections: + residual_upsample = torch.repeat_interleave( + residual_before_upsample, self.upsample_rate, dim=2 + ) + upsample_size = min(upsample_size, residual_upsample.size(2)) + x = ( + x[..., :upsample_size] + residual_upsample[..., :upsample_size] + ) * self.residual_scale + + if self.log_compression: + x = x.abs() + x = x + 1 + x = x.log() + + x = x.permute(0, 2, 1) + + # process padding + if padding is not None: + padding = torch.repeat_interleave(padding, self.upsample_rate, dim=1) + padding = padding[..., : x.size(1)] + + # process mask indices + if mask_indices is not None: + mask_indices = torch.repeat_interleave( + mask_indices, self.upsample_rate, dim=1 + ) + mask_indices = mask_indices[..., : x.size(1)] + return x, padding, mask_indices diff --git a/data/fairseq/fairseq/models/multires_hubert/multires_hubert_asr.py b/data/fairseq/fairseq/models/multires_hubert/multires_hubert_asr.py new file mode 100644 index 0000000000000000000000000000000000000000..2e7ad99ce1a0e397994c06b3ac6c193ed2d274b4 --- /dev/null +++ b/data/fairseq/fairseq/models/multires_hubert/multires_hubert_asr.py @@ -0,0 +1,376 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import contextlib +from argparse import Namespace +from dataclasses import dataclass, field +from typing import Any + +import torch +import torch.nn as nn +from omegaconf import II, MISSING + +from fairseq import checkpoint_utils, tasks, utils +from fairseq.dataclass import FairseqDataclass +from fairseq.dataclass.utils import convert_namespace_to_omegaconf +from fairseq.models import BaseFairseqModel, FairseqEncoder, register_model +from fairseq.models.hubert.hubert import MASKING_DISTRIBUTION_CHOICES +from fairseq.tasks import FairseqTask + + +@dataclass +class MultiresHubertAsrConfig(FairseqDataclass): + multires_hubert_path: str = field( + default=MISSING, metadata={"help": "path to multires_hubert model"} + ) + no_pretrained_weights: bool = field( + default=False, + metadata={"help": "if true, does not load pretrained weights"}, + ) + dropout_input: float = field( + default=0.0, + metadata={"help": "dropout to apply to the input (after feat extr)"}, + ) + final_dropout: float = field( + default=0.0, + metadata={"help": "dropout after transformer and before final projection"}, + ) + dropout: float = field( + default=0.0, + metadata={"help": "dropout probability inside hubert model"}, + ) + attention_dropout: float = field( + default=0.0, + metadata={ + "help": "dropout probability for attention weights " "inside hubert model" + }, + ) + activation_dropout: float = field( + default=0.0, + metadata={ + "help": "dropout probability after activation in FFN " "inside hubert model" + }, + ) + + # masking + apply_mask: bool = field( + default=False, metadata={"help": "apply masking during fine-tuning"} + ) + mask_length: int = field( + default=10, metadata={"help": "repeat the mask indices multiple times"} + ) + mask_prob: float = field( + default=0.5, + metadata={ + "help": "probability of replacing a token with mask " + "(normalized by length)" + }, + ) + mask_selection: MASKING_DISTRIBUTION_CHOICES = field( + default="static", metadata={"help": "how to choose masks"} + ) + mask_other: float = field( + default=0, + metadata={ + "help": "secondary mask argument " + "(used for more complex distributions), " + "see help in compute_mask_indices" + }, + ) + no_mask_overlap: bool = field( + default=False, metadata={"help": "whether to allow masks to overlap"} + ) + + # channel masking + mask_channel_length: int = field( + default=10, + metadata={"help": "length of the mask for features (channels)"}, + ) + mask_channel_prob: float = field( + default=0.0, + metadata={"help": "probability of replacing a feature with 0"}, + ) + mask_channel_selection: MASKING_DISTRIBUTION_CHOICES = field( + default="static", + metadata={"help": "how to choose mask length for channel masking"}, + ) + mask_channel_other: float = field( + default=0, + metadata={ + "help": "secondary mask argument " + "(used for more complex distributions), " + "see help in compute_mask_indices" + }, + ) + no_mask_channel_overlap: bool = field( + default=False, + metadata={"help": "whether to allow channel masks to overlap"}, + ) + freeze_finetune_updates: int = field( + default=0, + metadata={"help": "dont finetune hubert for this many updates"}, + ) + feature_grad_mult: float = field( + default=0.0, + metadata={"help": "reset feature grad mult in hubert to this"}, + ) + layerdrop: float = field( + default=0.0, + metadata={"help": "probability of dropping a layer in hubert"}, + ) + normalize: bool = II("task.normalize") + data: str = II("task.data") + + # this holds the loaded hubert args + multires_hubert_args: Any = None + + +@dataclass +class MultiresHubertCtcConfig(MultiresHubertAsrConfig): + pass + + +@register_model("multires_hubert_ctc", dataclass=MultiresHubertAsrConfig) +class MultiresHubertCtc(BaseFairseqModel): + def __init__( + self, cfg: MultiresHubertAsrConfig, multireshubert_encoder: BaseFairseqModel + ): + super().__init__() + self.cfg = cfg + self.multireshubert_encoder = multireshubert_encoder + + def upgrade_state_dict_named(self, state_dict, name): + super().upgrade_state_dict_named(state_dict, name) + return state_dict + + @classmethod + def build_model(cls, cfg: MultiresHubertAsrConfig, task: FairseqTask): + """Build a new model instance.""" + multireshubert_encoder = MultiresHubertEncoder(cfg, task) + return cls(cfg, multireshubert_encoder) + + def get_normalized_probs(self, net_output, log_probs, sample=None): + """Get normalized probabilities (or log probs) from a net's output.""" + + logits = net_output["encoder_out"] + if log_probs: + return utils.log_softmax(logits.float(), dim=-1) + else: + return utils.softmax(logits.float(), dim=-1) + + def get_logits(self, net_output): + logits = net_output["encoder_out"] + padding = net_output["encoder_padding_mask"] + if padding is not None and padding.any(): + padding = padding.T + logits[padding][..., 0] = 0 + logits[padding][..., 1:] = float("-inf") + + return logits + + def forward(self, **kwargs): + x = self.multireshubert_encoder(**kwargs) + return x + + +@dataclass +class MultiresHubertSeq2SeqConfig(MultiresHubertAsrConfig): + decoder_embed_dim: int = field( + default=768, metadata={"help": "decoder embedding dimension"} + ) + decoder_ffn_embed_dim: int = field( + default=3072, metadata={"help": "decoder embedding dimension for FFN"} + ) + decoder_layers: int = field(default=6, metadata={"help": "num of decoder layers"}) + decoder_layerdrop: float = field( + default=0.0, metadata={"help": "decoder layerdrop chance"} + ) + decoder_attention_heads: int = field( + default=4, metadata={"help": "num decoder attention heads"} + ) + decoder_learned_pos: bool = field( + default=False, + metadata={"help": "use learned positional embeddings in the decoder"}, + ) + decoder_normalize_before: bool = field( + default=False, + metadata={"help": "apply layernorm before each decoder block"}, + ) + no_token_positional_embeddings: bool = field( + default=False, + metadata={ + "help": "if set, disables positional embeddings " "(outside self attention)" + }, + ) + decoder_dropout: float = field( + default=0.0, metadata={"help": "dropout probability in the decoder"} + ) + decoder_attention_dropout: float = field( + default=0.0, + metadata={ + "help": "dropout probability for attention weights " "inside the decoder" + }, + ) + decoder_activation_dropout: float = field( + default=0.0, + metadata={ + "help": "dropout probability after activation in FFN " "inside the decoder" + }, + ) + max_target_positions: int = field( + default=2048, metadata={"help": "max target positions"} + ) + share_decoder_input_output_embed: bool = field( + default=False, + metadata={"help": "share decoder input and output embeddings"}, + ) + + +class MultiresHubertEncoder(FairseqEncoder): + def __init__(self, cfg: MultiresHubertAsrConfig, task): + self.apply_mask = cfg.apply_mask + + arg_overrides = { + "dropout": cfg.dropout, + "activation_dropout": cfg.activation_dropout, + "dropout_input": cfg.dropout_input, + "attention_dropout": cfg.attention_dropout, + "mask_length": cfg.mask_length, + "mask_prob": cfg.mask_prob, + "mask_selection": cfg.mask_selection, + "mask_other": cfg.mask_other, + "no_mask_overlap": cfg.no_mask_overlap, + "mask_channel_length": cfg.mask_channel_length, + "mask_channel_prob": cfg.mask_channel_prob, + "mask_channel_selection": cfg.mask_channel_selection, + "mask_channel_other": cfg.mask_channel_other, + "no_mask_channel_overlap": cfg.no_mask_channel_overlap, + "encoder_layerdrop": cfg.layerdrop, + "feature_grad_mult": cfg.feature_grad_mult, + } + + if cfg.multires_hubert_args is None: + state = checkpoint_utils.load_checkpoint_to_cpu( + cfg.multires_hubert_path, arg_overrides + ) + multires_hubert_args = state.get("cfg", None) + if multires_hubert_args is None: + multires_hubert_args = convert_namespace_to_omegaconf(state["args"]) + cfg.multires_hubert_args = multires_hubert_args + else: + state = None + multires_hubert_args = cfg.multires_hubert_args + if isinstance(multires_hubert_args, Namespace): + cfg.multires_hubert_args = ( + multires_hubert_args + ) = convert_namespace_to_omegaconf(multires_hubert_args) + + assert cfg.normalize == multires_hubert_args.task.normalize, ( + "Fine-tuning works best when data normalization is the same. " + "Please check that --normalize is set or unset for " + "both pre-training and here" + ) + + multires_hubert_args.task.data = cfg.data + pretrain_task = tasks.setup_task(multires_hubert_args.task) + if state is not None and "task_state" in state: + # This will load the stored "dictionaries" object + pretrain_task.load_state_dict(state["task_state"]) + else: + pretrain_task.load_state_dict(task.state_dict()) + + model = pretrain_task.build_model( + multires_hubert_args.model, from_checkpoint=True + ) + if state is not None and not cfg.no_pretrained_weights: + # set strict=False because we omit some modules + model.load_state_dict(state["model"], strict=False) + + model.remove_pretraining_modules() + + super().__init__(pretrain_task.source_dictionary) + + d = multires_hubert_args.model.encoder_embed_dim + + self.multires_hubert_model = model + + self.final_dropout = nn.Dropout(cfg.final_dropout) + self.freeze_finetune_updates = cfg.freeze_finetune_updates + self.num_updates = 0 + + if task.target_dictionary is not None: + self.proj = Linear(d, len(task.target_dictionary)) + elif getattr(cfg, "decoder_embed_dim", d) != d: + self.proj = Linear(d, cfg.decoder_embed_dim) + else: + self.proj = None + + def set_num_updates(self, num_updates): + """Set the number of parameters updates.""" + super().set_num_updates(num_updates) + self.num_updates = num_updates + + def forward(self, source, padding_mask, tbc=True, **kwargs): + multires_hubert_args = { + "source": source, + "padding_mask": padding_mask, + "mask": self.apply_mask and self.training, + "last_layer": True, + } + + ft = self.freeze_finetune_updates <= self.num_updates + + with torch.no_grad() if not ft else contextlib.ExitStack(): + x, padding_mask = self.multires_hubert_model.extract_features( + **multires_hubert_args + ) + + if tbc: + # B x T x C -> T x B x C + x = x.transpose(0, 1) + + x = self.final_dropout(x) + + if self.proj: + x = self.proj(x) + + return { + "encoder_out": x, # T x B x C + "encoder_padding_mask": padding_mask, # B x T + "padding_mask": padding_mask, + } + + def reorder_encoder_out(self, encoder_out, new_order): + if encoder_out["encoder_out"] is not None: + encoder_out["encoder_out"] = encoder_out["encoder_out"].index_select( + 1, new_order + ) + if encoder_out["encoder_padding_mask"] is not None: + encoder_out["encoder_padding_mask"] = encoder_out[ + "encoder_padding_mask" + ].index_select(0, new_order) + return encoder_out + + def max_positions(self): + """Maximum input length supported by the encoder.""" + return None + + def upgrade_state_dict_named(self, state_dict, name): + return state_dict + + +def Embedding(num_embeddings, embedding_dim, padding_idx): + m = nn.Embedding(num_embeddings, embedding_dim, padding_idx=padding_idx) + nn.init.normal_(m.weight, mean=0, std=embedding_dim**-0.5) + nn.init.constant_(m.weight[padding_idx], 0) + return m + + +def Linear(in_features, out_features, bias=True): + m = nn.Linear(in_features, out_features, bias) + nn.init.xavier_uniform_(m.weight) + if bias: + nn.init.constant_(m.bias, 0.0) + return m diff --git a/data/fairseq/fairseq/models/nat/__init__.py b/data/fairseq/fairseq/models/nat/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..05fe822487c3bcde8346648d5826f1669c6bc1ca --- /dev/null +++ b/data/fairseq/fairseq/models/nat/__init__.py @@ -0,0 +1,13 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. +"""isort:skip_file""" + +from .fairseq_nat_model import * +from .nonautoregressive_transformer import * +from .nat_crf_transformer import * +from .iterative_nonautoregressive_transformer import * +from .cmlm_transformer import * +from .levenshtein_transformer import * +from .insertion_transformer import * diff --git a/data/fairseq/fairseq/models/nat/__pycache__/__init__.cpython-38.pyc b/data/fairseq/fairseq/models/nat/__pycache__/__init__.cpython-38.pyc new file mode 100644 index 0000000000000000000000000000000000000000..6cb559c1d3c3f3cbd765cf43743c5a9ad23c048d Binary files /dev/null and b/data/fairseq/fairseq/models/nat/__pycache__/__init__.cpython-38.pyc differ diff --git a/data/fairseq/fairseq/models/nat/__pycache__/cmlm_transformer.cpython-38.pyc b/data/fairseq/fairseq/models/nat/__pycache__/cmlm_transformer.cpython-38.pyc new file mode 100644 index 0000000000000000000000000000000000000000..7687dbd084e24d9e8e3ec40416bd0c440bd9e90b Binary files /dev/null and b/data/fairseq/fairseq/models/nat/__pycache__/cmlm_transformer.cpython-38.pyc differ diff --git a/data/fairseq/fairseq/models/nat/__pycache__/fairseq_nat_model.cpython-38.pyc b/data/fairseq/fairseq/models/nat/__pycache__/fairseq_nat_model.cpython-38.pyc new file mode 100644 index 0000000000000000000000000000000000000000..9ebbd66ac756a6718c53d7eed2b769a972722f25 Binary files /dev/null and b/data/fairseq/fairseq/models/nat/__pycache__/fairseq_nat_model.cpython-38.pyc differ diff --git a/data/fairseq/fairseq/models/nat/__pycache__/insertion_transformer.cpython-38.pyc b/data/fairseq/fairseq/models/nat/__pycache__/insertion_transformer.cpython-38.pyc new file mode 100644 index 0000000000000000000000000000000000000000..b59e1fe1fd5f9ef4aa65719082b03191cf5530db Binary files /dev/null and b/data/fairseq/fairseq/models/nat/__pycache__/insertion_transformer.cpython-38.pyc differ diff --git a/data/fairseq/fairseq/models/nat/__pycache__/iterative_nonautoregressive_transformer.cpython-38.pyc b/data/fairseq/fairseq/models/nat/__pycache__/iterative_nonautoregressive_transformer.cpython-38.pyc new file mode 100644 index 0000000000000000000000000000000000000000..83219714c9c0449780db04cd05e6756c2e667801 Binary files /dev/null and b/data/fairseq/fairseq/models/nat/__pycache__/iterative_nonautoregressive_transformer.cpython-38.pyc differ diff --git a/data/fairseq/fairseq/models/nat/__pycache__/levenshtein_transformer.cpython-38.pyc b/data/fairseq/fairseq/models/nat/__pycache__/levenshtein_transformer.cpython-38.pyc new file mode 100644 index 0000000000000000000000000000000000000000..afdcd01e4a282d85519b02fa04c9057e747aa800 Binary files /dev/null and b/data/fairseq/fairseq/models/nat/__pycache__/levenshtein_transformer.cpython-38.pyc differ diff --git a/data/fairseq/fairseq/models/nat/__pycache__/levenshtein_utils.cpython-38.pyc b/data/fairseq/fairseq/models/nat/__pycache__/levenshtein_utils.cpython-38.pyc new file mode 100644 index 0000000000000000000000000000000000000000..ba864380172a77017b05878c000bbf37ba9ba725 Binary files /dev/null and b/data/fairseq/fairseq/models/nat/__pycache__/levenshtein_utils.cpython-38.pyc differ diff --git a/data/fairseq/fairseq/models/nat/__pycache__/nat_crf_transformer.cpython-38.pyc b/data/fairseq/fairseq/models/nat/__pycache__/nat_crf_transformer.cpython-38.pyc new file mode 100644 index 0000000000000000000000000000000000000000..d66545a6ddb432e401634dec709d1f5c3f2eb63a Binary files /dev/null and b/data/fairseq/fairseq/models/nat/__pycache__/nat_crf_transformer.cpython-38.pyc differ diff --git a/data/fairseq/fairseq/models/nat/__pycache__/nonautoregressive_transformer.cpython-38.pyc b/data/fairseq/fairseq/models/nat/__pycache__/nonautoregressive_transformer.cpython-38.pyc new file mode 100644 index 0000000000000000000000000000000000000000..36503d32e6c9f83571720eb745563547baad7db4 Binary files /dev/null and b/data/fairseq/fairseq/models/nat/__pycache__/nonautoregressive_transformer.cpython-38.pyc differ diff --git a/data/fairseq/fairseq/models/nat/cmlm_transformer.py b/data/fairseq/fairseq/models/nat/cmlm_transformer.py new file mode 100644 index 0000000000000000000000000000000000000000..c876e9453c101c00bd8e93e6e6f1fb48dc26f993 --- /dev/null +++ b/data/fairseq/fairseq/models/nat/cmlm_transformer.py @@ -0,0 +1,162 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +""" +This file implements: +Ghazvininejad, Marjan, et al. +"Constant-time machine translation with conditional masked language models." +arXiv preprint arXiv:1904.09324 (2019). +""" + +from fairseq.models import register_model, register_model_architecture +from fairseq.models.nat import NATransformerModel +from fairseq.utils import new_arange + + +def _skeptical_unmasking(output_scores, output_masks, p): + sorted_index = output_scores.sort(-1)[1] + boundary_len = ( + (output_masks.sum(1, keepdim=True).type_as(output_scores) - 2) * p + ).long() + skeptical_mask = new_arange(output_masks) < boundary_len + return skeptical_mask.scatter(1, sorted_index, skeptical_mask) + + +@register_model("cmlm_transformer") +class CMLMNATransformerModel(NATransformerModel): + @staticmethod + def add_args(parser): + NATransformerModel.add_args(parser) + + def forward( + self, src_tokens, src_lengths, prev_output_tokens, tgt_tokens, **kwargs + ): + assert not self.decoder.src_embedding_copy, "do not support embedding copy." + + # encoding + encoder_out = self.encoder(src_tokens, src_lengths=src_lengths, **kwargs) + # length prediction + length_out = self.decoder.forward_length( + normalize=False, encoder_out=encoder_out + ) + length_tgt = self.decoder.forward_length_prediction( + length_out, encoder_out, tgt_tokens + ) + + # decoding + word_ins_out = self.decoder( + normalize=False, + prev_output_tokens=prev_output_tokens, + encoder_out=encoder_out, + ) + word_ins_mask = prev_output_tokens.eq(self.unk) + + return { + "word_ins": { + "out": word_ins_out, + "tgt": tgt_tokens, + "mask": word_ins_mask, + "ls": self.args.label_smoothing, + "nll_loss": True, + }, + "length": { + "out": length_out, + "tgt": length_tgt, + "factor": self.decoder.length_loss_factor, + }, + } + + def forward_decoder(self, decoder_out, encoder_out, decoding_format=None, **kwargs): + + step = decoder_out.step + max_step = decoder_out.max_step + + output_tokens = decoder_out.output_tokens + output_scores = decoder_out.output_scores + history = decoder_out.history + + # execute the decoder + output_masks = output_tokens.eq(self.unk) + _scores, _tokens = self.decoder( + normalize=True, + prev_output_tokens=output_tokens, + encoder_out=encoder_out, + ).max(-1) + output_tokens.masked_scatter_(output_masks, _tokens[output_masks]) + output_scores.masked_scatter_(output_masks, _scores[output_masks]) + + if history is not None: + history.append(output_tokens.clone()) + + # skeptical decoding (depend on the maximum decoding steps.) + if (step + 1) < max_step: + skeptical_mask = _skeptical_unmasking( + output_scores, output_tokens.ne(self.pad), 1 - (step + 1) / max_step + ) + + output_tokens.masked_fill_(skeptical_mask, self.unk) + output_scores.masked_fill_(skeptical_mask, 0.0) + + if history is not None: + history.append(output_tokens.clone()) + + return decoder_out._replace( + output_tokens=output_tokens, + output_scores=output_scores, + attn=None, + history=history, + ) + + +@register_model_architecture("cmlm_transformer", "cmlm_transformer") +def cmlm_base_architecture(args): + args.encoder_embed_path = getattr(args, "encoder_embed_path", None) + args.encoder_embed_dim = getattr(args, "encoder_embed_dim", 512) + args.encoder_ffn_embed_dim = getattr(args, "encoder_ffn_embed_dim", 2048) + args.encoder_layers = getattr(args, "encoder_layers", 6) + args.encoder_attention_heads = getattr(args, "encoder_attention_heads", 8) + args.encoder_normalize_before = getattr(args, "encoder_normalize_before", False) + args.encoder_learned_pos = getattr(args, "encoder_learned_pos", False) + args.decoder_embed_path = getattr(args, "decoder_embed_path", None) + args.decoder_embed_dim = getattr(args, "decoder_embed_dim", args.encoder_embed_dim) + args.decoder_ffn_embed_dim = getattr( + args, "decoder_ffn_embed_dim", args.encoder_ffn_embed_dim + ) + args.decoder_layers = getattr(args, "decoder_layers", 6) + args.decoder_attention_heads = getattr(args, "decoder_attention_heads", 8) + args.decoder_normalize_before = getattr(args, "decoder_normalize_before", False) + args.decoder_learned_pos = getattr(args, "decoder_learned_pos", False) + args.attention_dropout = getattr(args, "attention_dropout", 0.0) + args.activation_dropout = getattr(args, "activation_dropout", 0.0) + args.activation_fn = getattr(args, "activation_fn", "relu") + args.dropout = getattr(args, "dropout", 0.1) + args.adaptive_softmax_cutoff = getattr(args, "adaptive_softmax_cutoff", None) + args.adaptive_softmax_dropout = getattr(args, "adaptive_softmax_dropout", 0) + args.share_decoder_input_output_embed = getattr( + args, "share_decoder_input_output_embed", False + ) + args.share_all_embeddings = getattr(args, "share_all_embeddings", True) + args.no_token_positional_embeddings = getattr( + args, "no_token_positional_embeddings", False + ) + args.adaptive_input = getattr(args, "adaptive_input", False) + args.apply_bert_init = getattr(args, "apply_bert_init", False) + + args.decoder_output_dim = getattr( + args, "decoder_output_dim", args.decoder_embed_dim + ) + args.decoder_input_dim = getattr(args, "decoder_input_dim", args.decoder_embed_dim) + + # --- special arguments --- + args.sg_length_pred = getattr(args, "sg_length_pred", False) + args.pred_length_offset = getattr(args, "pred_length_offset", False) + args.length_loss_factor = getattr(args, "length_loss_factor", 0.1) + args.ngram_predictor = getattr(args, "ngram_predictor", 1) + args.src_embedding_copy = getattr(args, "src_embedding_copy", False) + + +@register_model_architecture("cmlm_transformer", "cmlm_transformer_wmt_en_de") +def cmlm_wmt_en_de(args): + cmlm_base_architecture(args) diff --git a/data/fairseq/fairseq/models/nat/fairseq_nat_model.py b/data/fairseq/fairseq/models/nat/fairseq_nat_model.py new file mode 100644 index 0000000000000000000000000000000000000000..a5594a4ed9b395a1665c70fa7410d5caa16a879d --- /dev/null +++ b/data/fairseq/fairseq/models/nat/fairseq_nat_model.py @@ -0,0 +1,172 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import math + +import torch +from fairseq.models.transformer import ( + TransformerDecoder, + TransformerEncoder, + TransformerModel, +) +from fairseq.modules.transformer_sentence_encoder import init_bert_params + + +def ensemble_encoder(func): + def wrapper(self, *args, **kwargs): + if self.ensemble_models is None or len(self.ensemble_models) == 1: + return func(self, *args, **kwargs) + encoder_outs = [ + func(model, *args, **kwargs, return_all_hiddens=True) + for model in self.ensemble_models + ] + _encoder_out = encoder_outs[0].copy() + + def stack(key): + outs = [e[key][0] for e in encoder_outs] + return [torch.stack(outs, -1) if outs[0] is not None else None] + + _encoder_out["encoder_out"] = stack("encoder_out") + _encoder_out["encoder_embedding"] = stack("encoder_embedding") + + num_layers = len(_encoder_out["encoder_states"]) + if num_layers > 0: + _encoder_out["encoder_states"] = [ + torch.stack([e["encoder_states"][i] for e in encoder_outs], -1) + for i in range(num_layers) + ] + return _encoder_out + + return wrapper + + +def ensemble_decoder(func): + def wrapper(self, normalize=False, encoder_out=None, *args, **kwargs): + if self.ensemble_models is None or len(self.ensemble_models) == 1: + return func( + self, normalize=normalize, encoder_out=encoder_out, *args, **kwargs + ) + + def _replace(encoder_out, new_val): + new_encoder_out = encoder_out.copy() + new_encoder_out["encoder_out"] = [new_val] + return new_encoder_out + + action_outs = [ + func( + model, + normalize=normalize, + encoder_out=_replace( + encoder_out, encoder_out["encoder_out"][0][:, :, :, i] + ), + *args, + **kwargs + ) + for i, model in enumerate(self.ensemble_models) + ] + + if not isinstance(action_outs[0], tuple): # return multiple values + action_outs = [[a] for a in action_outs] + else: + action_outs = [list(a) for a in action_outs] + + ensembled_outs = [] + for i in range(len(action_outs[0])): + if i == 0 and normalize: + ensembled_outs += [ + torch.logsumexp( + torch.stack([a[i] for a in action_outs], -1), dim=-1 + ) + - math.log(len(self.ensemble_models)) + ] + elif action_outs[0][i] is not None: + ensembled_outs += [torch.stack([a[i] for a in action_outs], -1)] + else: + ensembled_outs += [None] + + if len(ensembled_outs) == 1: + return ensembled_outs[0] + return tuple(ensembled_outs) + + return wrapper + + +class FairseqNATModel(TransformerModel): + """ + Abstract class for all nonautoregressive-based models + """ + + def __init__(self, args, encoder, decoder): + super().__init__(args, encoder, decoder) + self.tgt_dict = decoder.dictionary + self.bos = decoder.dictionary.bos() + self.eos = decoder.dictionary.eos() + self.pad = decoder.dictionary.pad() + self.unk = decoder.dictionary.unk() + + self.ensemble_models = None + + @property + def allow_length_beam(self): + return False + + @property + def allow_ensemble(self): + return True + + def enable_ensemble(self, models): + self.encoder.ensemble_models = [m.encoder for m in models] + self.decoder.ensemble_models = [m.decoder for m in models] + + @staticmethod + def add_args(parser): + TransformerModel.add_args(parser) + parser.add_argument( + "--apply-bert-init", + action="store_true", + help="use custom param initialization for BERT", + ) + + @classmethod + def build_decoder(cls, args, tgt_dict, embed_tokens): + decoder = FairseqNATDecoder(args, tgt_dict, embed_tokens) + if getattr(args, "apply_bert_init", False): + decoder.apply(init_bert_params) + return decoder + + @classmethod + def build_encoder(cls, args, src_dict, embed_tokens): + encoder = FairseqNATEncoder(args, src_dict, embed_tokens) + if getattr(args, "apply_bert_init", False): + encoder.apply(init_bert_params) + return encoder + + def forward_encoder(self, encoder_inputs): + return self.encoder(*encoder_inputs) + + def forward_decoder(self, *args, **kwargs): + return NotImplementedError + + def initialize_output_tokens(self, *args, **kwargs): + return NotImplementedError + + def forward(self, *args, **kwargs): + return NotImplementedError + + +class FairseqNATEncoder(TransformerEncoder): + def __init__(self, args, dictionary, embed_tokens): + super().__init__(args, dictionary, embed_tokens) + self.ensemble_models = None + + @ensemble_encoder + def forward(self, *args, **kwargs): + return super().forward(*args, **kwargs) + + +class FairseqNATDecoder(TransformerDecoder): + def __init__(self, args, dictionary, embed_tokens, no_encoder_attn=False): + super().__init__(args, dictionary, embed_tokens, no_encoder_attn) + self.ensemble_models = None diff --git a/data/fairseq/fairseq/models/nat/insertion_transformer.py b/data/fairseq/fairseq/models/nat/insertion_transformer.py new file mode 100644 index 0000000000000000000000000000000000000000..bc28000f59a3b9e8098f9fe710cc8335d39eea3e --- /dev/null +++ b/data/fairseq/fairseq/models/nat/insertion_transformer.py @@ -0,0 +1,280 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import numpy as np +import torch +import torch.nn.functional as F +from fairseq.models import register_model, register_model_architecture +from fairseq.models.nat import ( + FairseqNATModel, + LevenshteinTransformerDecoder, + LevenshteinTransformerModel, + ensemble_decoder, +) +from fairseq.models.transformer import Linear +from fairseq.modules.transformer_sentence_encoder import init_bert_params +from fairseq.utils import new_arange + + +class NegativeDistanceScore(object): + def __init__(self): + + # pre-compute some values + self.scores = {} + + self.scores[0.5] = self.compute_score_full(50, 0.5) + self.scores[1.0] = self.compute_score_full(50, 1.0) + self.scores[2.0] = self.compute_score_full(50, 2.0) + + def __call__(self, i, L, tau): + if (tau is None) or (tau > 1000): + return 1 / L + + if tau in self.scores: + if L < self.scores[tau].shape[0]: + return self.scores[tau][L - 1, i] + return self.compute_score(L, tau)[i] + + def compute_score(self, L, tau): + s = np.array([-abs(L / 2 - i) / tau for i in range(L)]) + s = np.exp(s - s.max()) + return s / s.sum() + + def compute_score_full(self, L, tau): + s = -abs(np.arange(0, L - 1)[:, None] / 2 - np.arange(L)[None, :]) / tau + s = np.tril(s, 0) + np.triu(s - float("inf"), 1) + s = np.exp(s - s.max(1, keepdims=True)) + return s / s.sum(1, keepdims=True) + + +neg_scorer = NegativeDistanceScore() + + +def _get_ins_targets(in_tokens, out_tokens, padding_idx, unk_idx, vocab_size, tau=None): + try: + from fairseq import libnat + except ImportError as e: + import sys + + sys.stderr.write("ERROR: missing libnat. run `pip install --editable .`\n") + raise e + + B = in_tokens.size(0) + T = in_tokens.size(1) + V = vocab_size + + with torch.cuda.device_of(in_tokens): + in_tokens_list = [ + [t for t in s if t != padding_idx] for i, s in enumerate(in_tokens.tolist()) + ] + out_tokens_list = [ + [t for t in s if t != padding_idx] + for i, s in enumerate(out_tokens.tolist()) + ] + + full_labels = libnat.suggested_ed2_path( + in_tokens_list, out_tokens_list, padding_idx + ) + insert_labels = [a[:-1] for a in full_labels] + + # numericalize1 + insert_label_tensors = in_tokens.new_zeros(B * (T - 1) * V).float() + insert_index, insert_labels = zip( + *[ + (w + (j + i * (T - 1)) * V, neg_scorer(k, len(label), tau)) + for i, labels in enumerate(insert_labels) + for j, label in enumerate(labels[1:-1]) + for k, w in enumerate(label) + ] + ) # HACK 1:-1 + insert_index, insert_labels = [ + torch.tensor(list(a), device=in_tokens.device) + for a in [insert_index, insert_labels] + ] + insert_label_tensors.scatter_(0, insert_index.long(), insert_labels) + insert_label_tensors = insert_label_tensors.view(B, T - 1, V) + + return insert_label_tensors + + +def _apply_ins_words(in_tokens, in_scores, word_ins_pred, word_ins_scores, padding_idx): + + padding_masks = in_tokens[:, 1:].eq(padding_idx) + word_ins_scores.masked_fill_(padding_masks, 0.0) + word_ins_pred.masked_fill_(padding_masks, padding_idx) + + in_coords = new_arange(in_tokens).type_as(in_scores) + + # shift all padding predictions to infinite + out_coords = (in_coords[:, 1:] - 0.5).masked_fill( + word_ins_pred.eq(padding_idx), float("inf") + ) + out_coords = torch.cat([in_coords, out_coords], 1).sort(-1)[1] + out_tokens = torch.cat([in_tokens, word_ins_pred], 1).gather(1, out_coords) + out_scores = torch.cat([in_scores, word_ins_scores], 1).gather(1, out_coords) + return out_tokens, out_scores + + +@register_model("insertion_transformer") +class InsertionTransformerModel(LevenshteinTransformerModel): + def __init__(self, args, encoder, decoder): + super().__init__(args, encoder, decoder) + + @staticmethod + def add_args(parser): + FairseqNATModel.add_args(parser) + parser.add_argument("--label-tau", default=None, type=float) + + @classmethod + def build_decoder(cls, args, tgt_dict, embed_tokens): + decoder = InsertionTransformerDecoder(args, tgt_dict, embed_tokens) + if getattr(args, "apply_bert_init", False): + decoder.apply(init_bert_params) + return decoder + + def forward( + self, src_tokens, src_lengths, prev_output_tokens, tgt_tokens, **kwargs + ): + + assert tgt_tokens is not None, "forward function only supports training." + + # encoding + encoder_out = self.encoder(src_tokens, src_lengths=src_lengths, **kwargs) + + # generate training labels for insertion + word_ins_out = self.decoder.forward_word_ins( + normalize=False, + prev_output_tokens=prev_output_tokens, + encoder_out=encoder_out, + ) + + word_ins_tgt = _get_ins_targets( + prev_output_tokens, + tgt_tokens, + self.pad, + self.unk, + len(self.tgt_dict), + tau=self.decoder.label_tau, + ).type_as(word_ins_out) + word_ins_masks = prev_output_tokens[:, 1:].ne(self.pad) + + return { + "word_ins": { + "out": word_ins_out, + "tgt": word_ins_tgt, + "mask": word_ins_masks, + "ls": self.args.label_smoothing, + "nll_loss": True, + } + } + + def forward_decoder( + self, decoder_out, encoder_out, eos_penalty=0.0, max_ratio=None, **kwargs + ): + + output_tokens = decoder_out.output_tokens + output_scores = decoder_out.output_scores + history = decoder_out.history + + # TODO: decoding for InsertionTransformer + word_ins_score = self.decoder.forward_word_ins( + normalize=True, prev_output_tokens=output_tokens, encoder_out=encoder_out + ) + + if eos_penalty > 0.0: + word_ins_score[:, :, self.pad] -= eos_penalty + word_ins_score, word_ins_pred = word_ins_score.max(-1) + output_tokens, output_scores = _apply_ins_words( + output_tokens, output_scores, word_ins_pred, word_ins_score, self.pad + ) + + # delete some unnecessary paddings + cut_off = output_tokens.ne(self.pad).sum(1).max() + output_tokens = output_tokens[:, :cut_off] + output_scores = output_scores[:, :cut_off] + + if history is not None: + history.append(output_tokens.clone()) + + return decoder_out._replace( + output_tokens=output_tokens, + output_scores=output_scores, + attn=None, + history=history, + ) + + +class InsertionTransformerDecoder(LevenshteinTransformerDecoder): + def __init__(self, args, dictionary, embed_tokens, no_encoder_attn=False): + # use the TransformerDecoder's __init__ + super(LevenshteinTransformerDecoder, self).__init__( + args, dictionary, embed_tokens, no_encoder_attn=no_encoder_attn + ) + + self.dictionary = dictionary + self.bos = dictionary.bos() + self.unk = dictionary.unk() + self.eos = dictionary.eos() + self.pool_out = Linear(self.output_embed_dim * 2, self.output_embed_dim) + + self.label_tau = getattr(args, "label_tau", None) + + @ensemble_decoder + def forward_word_ins(self, normalize, encoder_out, prev_output_tokens): + features = self.extract_features(prev_output_tokens, encoder_out=encoder_out)[0] + features = self.pool_out( + torch.cat([features[:, :-1, :], features[:, 1:, :]], 2) + ) + decoder_out = self.output_layer(features) + return F.log_softmax(decoder_out, -1) if normalize else decoder_out + + def forward_mask_ins(self, *args, **kwargs): + raise NotImplementedError + + def forward_word_del(self, *args, **kwargs): + raise NotImplementedError + + +@register_model_architecture("insertion_transformer", "insertion_transformer") +def insertion_base_architecture(args): + args.encoder_embed_path = getattr(args, "encoder_embed_path", None) + args.encoder_embed_dim = getattr(args, "encoder_embed_dim", 512) + args.encoder_ffn_embed_dim = getattr(args, "encoder_ffn_embed_dim", 2048) + args.encoder_layers = getattr(args, "encoder_layers", 6) + args.encoder_attention_heads = getattr(args, "encoder_attention_heads", 8) + args.encoder_normalize_before = getattr(args, "encoder_normalize_before", False) + args.encoder_learned_pos = getattr(args, "encoder_learned_pos", False) + args.decoder_embed_path = getattr(args, "decoder_embed_path", None) + args.decoder_embed_dim = getattr(args, "decoder_embed_dim", args.encoder_embed_dim) + args.decoder_ffn_embed_dim = getattr( + args, "decoder_ffn_embed_dim", args.encoder_ffn_embed_dim + ) + args.decoder_layers = getattr(args, "decoder_layers", 6) + args.decoder_attention_heads = getattr(args, "decoder_attention_heads", 8) + args.decoder_normalize_before = getattr(args, "decoder_normalize_before", False) + args.decoder_learned_pos = getattr(args, "decoder_learned_pos", False) + args.attention_dropout = getattr(args, "attention_dropout", 0.0) + args.activation_dropout = getattr(args, "activation_dropout", 0.0) + args.activation_fn = getattr(args, "activation_fn", "relu") + args.dropout = getattr(args, "dropout", 0.1) + args.adaptive_softmax_cutoff = getattr(args, "adaptive_softmax_cutoff", None) + args.adaptive_softmax_dropout = getattr(args, "adaptive_softmax_dropout", 0) + args.share_decoder_input_output_embed = getattr( + args, "share_decoder_input_output_embed", False + ) + args.share_all_embeddings = getattr(args, "share_all_embeddings", False) + args.no_token_positional_embeddings = getattr( + args, "no_token_positional_embeddings", False + ) + args.adaptive_input = getattr(args, "adaptive_input", False) + args.apply_bert_init = getattr(args, "apply_bert_init", False) + + args.decoder_output_dim = getattr( + args, "decoder_output_dim", args.decoder_embed_dim + ) + args.decoder_input_dim = getattr(args, "decoder_input_dim", args.decoder_embed_dim) + + # special for insertion transformer + args.label_tau = getattr(args, "label_tau", None) diff --git a/data/fairseq/fairseq/models/nat/iterative_nonautoregressive_transformer.py b/data/fairseq/fairseq/models/nat/iterative_nonautoregressive_transformer.py new file mode 100644 index 0000000000000000000000000000000000000000..bc39509980a80eb8c21e0bfdb304649ad3acc4d0 --- /dev/null +++ b/data/fairseq/fairseq/models/nat/iterative_nonautoregressive_transformer.py @@ -0,0 +1,228 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import torch +from fairseq.models import register_model, register_model_architecture +from fairseq.models.nat import NATransformerModel + + +def _sequential_poisoning(s, V, beta=0.33, bos=2, eos=3, pad=1): + # s: input batch + # V: vocabulary size + rand_words = torch.randint(low=4, high=V, size=s.size(), device=s.device) + choices = torch.rand(size=s.size(), device=s.device) + choices.masked_fill_((s == pad) | (s == bos) | (s == eos), 1) + + replace = choices < beta / 3 + repeat = (choices >= beta / 3) & (choices < beta * 2 / 3) + swap = (choices >= beta * 2 / 3) & (choices < beta) + safe = choices >= beta + + for i in range(s.size(1) - 1): + rand_word = rand_words[:, i] + next_word = s[:, i + 1] + self_word = s[:, i] + + replace_i = replace[:, i] + swap_i = swap[:, i] & (next_word != 3) + repeat_i = repeat[:, i] & (next_word != 3) + safe_i = safe[:, i] | ((next_word == 3) & (~replace_i)) + + s[:, i] = ( + self_word * (safe_i | repeat_i).long() + + next_word * swap_i.long() + + rand_word * replace_i.long() + ) + s[:, i + 1] = ( + next_word * (safe_i | replace_i).long() + + self_word * (swap_i | repeat_i).long() + ) + return s + + +def gumbel_noise(input, TINY=1e-8): + return ( + input.new_zeros(*input.size()) + .uniform_() + .add_(TINY) + .log_() + .neg_() + .add_(TINY) + .log_() + .neg_() + ) + + +@register_model("iterative_nonautoregressive_transformer") +class IterNATransformerModel(NATransformerModel): + @staticmethod + def add_args(parser): + NATransformerModel.add_args(parser) + parser.add_argument( + "--train-step", + type=int, + help="number of refinement iterations during training", + ) + parser.add_argument( + "--dae-ratio", + type=float, + help="the probability of switching to the denoising auto-encoder loss", + ) + parser.add_argument( + "--stochastic-approx", + action="store_true", + help="sampling from the decoder as the inputs for next iteration", + ) + + @classmethod + def build_model(cls, args, task): + model = super().build_model(args, task) + model.train_step = getattr(args, "train_step", 4) + model.dae_ratio = getattr(args, "dae_ratio", 0.5) + model.stochastic_approx = getattr(args, "stochastic_approx", False) + return model + + def forward( + self, src_tokens, src_lengths, prev_output_tokens, tgt_tokens, **kwargs + ): + + B, T = prev_output_tokens.size() + + # encoding + encoder_out = self.encoder(src_tokens, src_lengths=src_lengths, **kwargs) + + # length prediction + length_out = self.decoder.forward_length( + normalize=False, encoder_out=encoder_out + ) + length_tgt = self.decoder.forward_length_prediction( + length_out, encoder_out, tgt_tokens + ) + + # decoding + word_ins_outs, word_ins_tgts, word_ins_masks = [], [], [] + for t in range(self.train_step): + word_ins_out = self.decoder( + normalize=False, + prev_output_tokens=prev_output_tokens, + encoder_out=encoder_out, + step=t, + ) + word_ins_tgt = tgt_tokens + word_ins_mask = word_ins_tgt.ne(self.pad) + + word_ins_outs.append(word_ins_out) + word_ins_tgts.append(word_ins_tgt) + word_ins_masks.append(word_ins_mask) + + if t < (self.train_step - 1): + # prediction for next iteration + if self.stochastic_approx: + word_ins_prediction = ( + word_ins_out + gumbel_noise(word_ins_out) + ).max(-1)[1] + else: + word_ins_prediction = word_ins_out.max(-1)[1] + + prev_output_tokens = prev_output_tokens.masked_scatter( + word_ins_mask, word_ins_prediction[word_ins_mask] + ) + + if self.dae_ratio > 0: + # we do not perform denoising for the first iteration + corrputed = ( + torch.rand(size=(B,), device=prev_output_tokens.device) + < self.dae_ratio + ) + corrputed_tokens = _sequential_poisoning( + tgt_tokens[corrputed], + len(self.tgt_dict), + 0.33, + self.bos, + self.eos, + self.pad, + ) + prev_output_tokens[corrputed] = corrputed_tokens + + # concat everything + word_ins_out = torch.cat(word_ins_outs, 0) + word_ins_tgt = torch.cat(word_ins_tgts, 0) + word_ins_mask = torch.cat(word_ins_masks, 0) + + return { + "word_ins": { + "out": word_ins_out, + "tgt": word_ins_tgt, + "mask": word_ins_mask, + "ls": self.args.label_smoothing, + "nll_loss": True, + }, + "length": { + "out": length_out, + "tgt": length_tgt, + "factor": self.decoder.length_loss_factor, + }, + } + + +@register_model_architecture( + "iterative_nonautoregressive_transformer", "iterative_nonautoregressive_transformer" +) +def inat_base_architecture(args): + args.encoder_embed_path = getattr(args, "encoder_embed_path", None) + args.encoder_embed_dim = getattr(args, "encoder_embed_dim", 512) + args.encoder_ffn_embed_dim = getattr(args, "encoder_ffn_embed_dim", 2048) + args.encoder_layers = getattr(args, "encoder_layers", 6) + args.encoder_attention_heads = getattr(args, "encoder_attention_heads", 8) + args.encoder_normalize_before = getattr(args, "encoder_normalize_before", False) + args.encoder_learned_pos = getattr(args, "encoder_learned_pos", False) + args.decoder_embed_path = getattr(args, "decoder_embed_path", None) + args.decoder_embed_dim = getattr(args, "decoder_embed_dim", args.encoder_embed_dim) + args.decoder_ffn_embed_dim = getattr( + args, "decoder_ffn_embed_dim", args.encoder_ffn_embed_dim + ) + args.decoder_layers = getattr(args, "decoder_layers", 6) + args.decoder_attention_heads = getattr(args, "decoder_attention_heads", 8) + args.decoder_normalize_before = getattr(args, "decoder_normalize_before", False) + args.decoder_learned_pos = getattr(args, "decoder_learned_pos", False) + args.attention_dropout = getattr(args, "attention_dropout", 0.0) + args.activation_dropout = getattr(args, "activation_dropout", 0.0) + args.activation_fn = getattr(args, "activation_fn", "relu") + args.dropout = getattr(args, "dropout", 0.1) + args.adaptive_softmax_cutoff = getattr(args, "adaptive_softmax_cutoff", None) + args.adaptive_softmax_dropout = getattr(args, "adaptive_softmax_dropout", 0) + args.share_decoder_input_output_embed = getattr( + args, "share_decoder_input_output_embed", False + ) + args.share_all_embeddings = getattr(args, "share_all_embeddings", False) + args.no_token_positional_embeddings = getattr( + args, "no_token_positional_embeddings", False + ) + args.adaptive_input = getattr(args, "adaptive_input", False) + args.apply_bert_init = getattr(args, "apply_bert_init", False) + + args.decoder_output_dim = getattr( + args, "decoder_output_dim", args.decoder_embed_dim + ) + args.decoder_input_dim = getattr(args, "decoder_input_dim", args.decoder_embed_dim) + + # --- special arguments --- + args.sg_length_pred = getattr(args, "sg_length_pred", False) + args.pred_length_offset = getattr(args, "pred_length_offset", False) + args.length_loss_factor = getattr(args, "length_loss_factor", 0.1) + args.ngram_predictor = getattr(args, "ngram_predictor", 1) + args.src_embedding_copy = getattr(args, "src_embedding_copy", False) + + args.train_step = getattr(args, "train_step", 4) + args.dae_ratio = getattr(args, "dae_ratio", 0.5) + args.stochastic_approx = getattr(args, "stochastic_approx", False) + + +@register_model_architecture( + "iterative_nonautoregressive_transformer", + "iterative_nonautoregressive_transformer_wmt_en_de", +) +def iter_nat_wmt_en_de(args): + inat_base_architecture(args) diff --git a/data/fairseq/fairseq/models/nat/levenshtein_transformer.py b/data/fairseq/fairseq/models/nat/levenshtein_transformer.py new file mode 100644 index 0000000000000000000000000000000000000000..d60d3c52d50b1f20957039a75622ffb95d5eea24 --- /dev/null +++ b/data/fairseq/fairseq/models/nat/levenshtein_transformer.py @@ -0,0 +1,510 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import torch +import torch.nn as nn +import torch.nn.functional as F +from fairseq.iterative_refinement_generator import DecoderOut +from fairseq.models import register_model, register_model_architecture +from fairseq.models.nat import FairseqNATDecoder, FairseqNATModel, ensemble_decoder +from fairseq.models.transformer import Embedding +from fairseq.modules import TransformerDecoderLayer +from fairseq.modules.transformer_sentence_encoder import init_bert_params + +from .levenshtein_utils import ( + _apply_del_words, + _apply_ins_masks, + _apply_ins_words, + _fill, + _get_del_targets, + _get_ins_targets, + _skip, + _skip_encoder_out, +) + + +@register_model("levenshtein_transformer") +class LevenshteinTransformerModel(FairseqNATModel): + @property + def allow_length_beam(self): + return False + + @staticmethod + def add_args(parser): + FairseqNATModel.add_args(parser) + parser.add_argument( + "--early-exit", + default="6,6,6", + type=str, + help="number of decoder layers before word_del, mask_ins, word_ins", + ) + parser.add_argument( + "--no-share-discriminator", + action="store_true", + help="separate parameters for discriminator", + ) + parser.add_argument( + "--no-share-maskpredictor", + action="store_true", + help="separate parameters for mask-predictor", + ) + parser.add_argument( + "--share-discriminator-maskpredictor", + action="store_true", + help="share the parameters for both mask-predictor and discriminator", + ) + parser.add_argument( + "--sampling-for-deletion", + action="store_true", + help="instead of argmax, use sampling to predict the tokens", + ) + + @classmethod + def build_decoder(cls, args, tgt_dict, embed_tokens): + decoder = LevenshteinTransformerDecoder(args, tgt_dict, embed_tokens) + if getattr(args, "apply_bert_init", False): + decoder.apply(init_bert_params) + return decoder + + def forward( + self, src_tokens, src_lengths, prev_output_tokens, tgt_tokens, **kwargs + ): + + assert tgt_tokens is not None, "forward function only supports training." + + # encoding + encoder_out = self.encoder(src_tokens, src_lengths=src_lengths, **kwargs) + + # generate training labels for insertion + masked_tgt_masks, masked_tgt_tokens, mask_ins_targets = _get_ins_targets( + prev_output_tokens, tgt_tokens, self.pad, self.unk + ) + mask_ins_targets = mask_ins_targets.clamp(min=0, max=255) # for safe prediction + mask_ins_masks = prev_output_tokens[:, 1:].ne(self.pad) + + mask_ins_out, _ = self.decoder.forward_mask_ins( + normalize=False, + prev_output_tokens=prev_output_tokens, + encoder_out=encoder_out, + ) + word_ins_out, _ = self.decoder.forward_word_ins( + normalize=False, + prev_output_tokens=masked_tgt_tokens, + encoder_out=encoder_out, + ) + + # make online prediction + if self.decoder.sampling_for_deletion: + word_predictions = torch.multinomial( + F.softmax(word_ins_out, -1).view(-1, word_ins_out.size(-1)), 1 + ).view(word_ins_out.size(0), -1) + else: + word_predictions = F.log_softmax(word_ins_out, dim=-1).max(2)[1] + + word_predictions.masked_scatter_( + ~masked_tgt_masks, tgt_tokens[~masked_tgt_masks] + ) + + # generate training labels for deletion + word_del_targets = _get_del_targets(word_predictions, tgt_tokens, self.pad) + word_del_out, _ = self.decoder.forward_word_del( + normalize=False, + prev_output_tokens=word_predictions, + encoder_out=encoder_out, + ) + word_del_masks = word_predictions.ne(self.pad) + + return { + "mask_ins": { + "out": mask_ins_out, + "tgt": mask_ins_targets, + "mask": mask_ins_masks, + "ls": 0.01, + }, + "word_ins": { + "out": word_ins_out, + "tgt": tgt_tokens, + "mask": masked_tgt_masks, + "ls": self.args.label_smoothing, + "nll_loss": True, + }, + "word_del": { + "out": word_del_out, + "tgt": word_del_targets, + "mask": word_del_masks, + }, + } + + def forward_decoder( + self, decoder_out, encoder_out, eos_penalty=0.0, max_ratio=None, **kwargs + ): + + output_tokens = decoder_out.output_tokens + output_scores = decoder_out.output_scores + attn = decoder_out.attn + history = decoder_out.history + + bsz = output_tokens.size(0) + if max_ratio is None: + max_lens = torch.zeros_like(output_tokens).fill_(255) + else: + if not encoder_out["encoder_padding_mask"]: + max_src_len = encoder_out["encoder_out"].size(0) + src_lens = encoder_out["encoder_out"].new(bsz).fill_(max_src_len) + else: + src_lens = (~encoder_out["encoder_padding_mask"][0]).sum(1) + max_lens = (src_lens * max_ratio).clamp(min=10).long() + + # delete words + # do not delete tokens if it is + can_del_word = output_tokens.ne(self.pad).sum(1) > 2 + if can_del_word.sum() != 0: # we cannot delete, skip + word_del_score, word_del_attn = self.decoder.forward_word_del( + normalize=True, + prev_output_tokens=_skip(output_tokens, can_del_word), + encoder_out=_skip_encoder_out(self.encoder, encoder_out, can_del_word), + ) + word_del_pred = word_del_score.max(-1)[1].bool() + + _tokens, _scores, _attn = _apply_del_words( + output_tokens[can_del_word], + output_scores[can_del_word], + word_del_attn, + word_del_pred, + self.pad, + self.bos, + self.eos, + ) + output_tokens = _fill(output_tokens, can_del_word, _tokens, self.pad) + output_scores = _fill(output_scores, can_del_word, _scores, 0) + attn = _fill(attn, can_del_word, _attn, 0.0) + + if history is not None: + history.append(output_tokens.clone()) + + # insert placeholders + can_ins_mask = output_tokens.ne(self.pad).sum(1) < max_lens + if can_ins_mask.sum() != 0: + mask_ins_score, _ = self.decoder.forward_mask_ins( + normalize=True, + prev_output_tokens=_skip(output_tokens, can_ins_mask), + encoder_out=_skip_encoder_out(self.encoder, encoder_out, can_ins_mask), + ) + if eos_penalty > 0.0: + mask_ins_score[:, :, 0] = mask_ins_score[:, :, 0] - eos_penalty + mask_ins_pred = mask_ins_score.max(-1)[1] + mask_ins_pred = torch.min( + mask_ins_pred, max_lens[can_ins_mask, None].expand_as(mask_ins_pred) + ) + + _tokens, _scores = _apply_ins_masks( + output_tokens[can_ins_mask], + output_scores[can_ins_mask], + mask_ins_pred, + self.pad, + self.unk, + self.eos, + ) + output_tokens = _fill(output_tokens, can_ins_mask, _tokens, self.pad) + output_scores = _fill(output_scores, can_ins_mask, _scores, 0) + + if history is not None: + history.append(output_tokens.clone()) + + # insert words + can_ins_word = output_tokens.eq(self.unk).sum(1) > 0 + if can_ins_word.sum() != 0: + word_ins_score, word_ins_attn = self.decoder.forward_word_ins( + normalize=True, + prev_output_tokens=_skip(output_tokens, can_ins_word), + encoder_out=_skip_encoder_out(self.encoder, encoder_out, can_ins_word), + ) + word_ins_score, word_ins_pred = word_ins_score.max(-1) + _tokens, _scores = _apply_ins_words( + output_tokens[can_ins_word], + output_scores[can_ins_word], + word_ins_pred, + word_ins_score, + self.unk, + ) + + output_tokens = _fill(output_tokens, can_ins_word, _tokens, self.pad) + output_scores = _fill(output_scores, can_ins_word, _scores, 0) + attn = _fill(attn, can_ins_word, word_ins_attn, 0.0) + + if history is not None: + history.append(output_tokens.clone()) + + # delete some unnecessary paddings + cut_off = output_tokens.ne(self.pad).sum(1).max() + output_tokens = output_tokens[:, :cut_off] + output_scores = output_scores[:, :cut_off] + attn = None if attn is None else attn[:, :cut_off, :] + + return decoder_out._replace( + output_tokens=output_tokens, + output_scores=output_scores, + attn=attn, + history=history, + ) + + def initialize_output_tokens(self, encoder_out, src_tokens): + initial_output_tokens = src_tokens.new_zeros(src_tokens.size(0), 2) + initial_output_tokens[:, 0] = self.bos + initial_output_tokens[:, 1] = self.eos + + initial_output_scores = initial_output_tokens.new_zeros( + *initial_output_tokens.size() + ).type_as(encoder_out["encoder_out"][0]) + + return DecoderOut( + output_tokens=initial_output_tokens, + output_scores=initial_output_scores, + attn=None, + step=0, + max_step=0, + history=None, + ) + + +class LevenshteinTransformerDecoder(FairseqNATDecoder): + def __init__(self, args, dictionary, embed_tokens, no_encoder_attn=False): + super().__init__( + args, dictionary, embed_tokens, no_encoder_attn=no_encoder_attn + ) + self.dictionary = dictionary + self.bos = dictionary.bos() + self.unk = dictionary.unk() + self.eos = dictionary.eos() + self.sampling_for_deletion = getattr(args, "sampling_for_deletion", False) + self.embed_mask_ins = Embedding(256, self.output_embed_dim * 2, None) + self.embed_word_del = Embedding(2, self.output_embed_dim, None) + + # del_word, ins_mask, ins_word + self.early_exit = [int(i) for i in args.early_exit.split(",")] + assert len(self.early_exit) == 3 + + # copy layers for mask-predict/deletion + self.layers_msk = None + if getattr(args, "no_share_maskpredictor", False): + self.layers_msk = nn.ModuleList( + [ + TransformerDecoderLayer(args, no_encoder_attn) + for _ in range(self.early_exit[1]) + ] + ) + self.layers_del = None + if getattr(args, "no_share_discriminator", False): + self.layers_del = nn.ModuleList( + [ + TransformerDecoderLayer(args, no_encoder_attn) + for _ in range(self.early_exit[0]) + ] + ) + + if getattr(args, "share_discriminator_maskpredictor", False): + assert getattr( + args, "no_share_discriminator", False + ), "must set saperate discriminator" + self.layers_msk = self.layers_del + + def extract_features( + self, + prev_output_tokens, + encoder_out=None, + early_exit=None, + layers=None, + **unused + ): + """ + Similar to *forward* but only return features. + Inputs: + prev_output_tokens: Tensor(B, T) + encoder_out: a dictionary of hidden states and masks + + Returns: + tuple: + - the decoder's features of shape `(batch, tgt_len, embed_dim)` + - a dictionary with any model-specific outputs + the LevenshteinTransformer decoder has full-attention to all generated tokens + """ + # embed positions + positions = ( + self.embed_positions(prev_output_tokens) + if self.embed_positions is not None + else None + ) + + # embed tokens and positions + x = self.embed_scale * self.embed_tokens(prev_output_tokens) + if self.project_in_dim is not None: + x = self.project_in_dim(x) + + if positions is not None: + x += positions + x = self.dropout_module(x) + + # B x T x C -> T x B x C + x = x.transpose(0, 1) + attn = None + inner_states = [x] + + # decoder layers + decoder_padding_mask = prev_output_tokens.eq(self.padding_idx) + layers = self.layers if layers is None else layers + early_exit = len(layers) if early_exit is None else early_exit + for _, layer in enumerate(layers[:early_exit]): + x, attn, _ = layer( + x, + encoder_out["encoder_out"][0] + if (encoder_out is not None and len(encoder_out["encoder_out"]) > 0) + else None, + encoder_out["encoder_padding_mask"][0] + if ( + encoder_out is not None + and len(encoder_out["encoder_padding_mask"]) > 0 + ) + else None, + self_attn_mask=None, + self_attn_padding_mask=decoder_padding_mask, + ) + inner_states.append(x) + + if self.layer_norm: + x = self.layer_norm(x) + + # T x B x C -> B x T x C + x = x.transpose(0, 1) + + if self.project_out_dim is not None: + x = self.project_out_dim(x) + + return x, {"attn": attn, "inner_states": inner_states} + + @ensemble_decoder + def forward_mask_ins(self, normalize, encoder_out, prev_output_tokens, **unused): + features, extra = self.extract_features( + prev_output_tokens, + encoder_out=encoder_out, + early_exit=self.early_exit[1], + layers=self.layers_msk, + **unused + ) + features_cat = torch.cat([features[:, :-1, :], features[:, 1:, :]], 2) + decoder_out = F.linear(features_cat, self.embed_mask_ins.weight) + if normalize: + return F.log_softmax(decoder_out, -1), extra["attn"] + return decoder_out, extra["attn"] + + @ensemble_decoder + def forward_word_ins(self, normalize, encoder_out, prev_output_tokens, **unused): + features, extra = self.extract_features( + prev_output_tokens, + encoder_out=encoder_out, + early_exit=self.early_exit[2], + layers=self.layers, + **unused + ) + decoder_out = self.output_layer(features) + if normalize: + return F.log_softmax(decoder_out, -1), extra["attn"] + return decoder_out, extra["attn"] + + @ensemble_decoder + def forward_word_del(self, normalize, encoder_out, prev_output_tokens, **unused): + features, extra = self.extract_features( + prev_output_tokens, + encoder_out=encoder_out, + early_exit=self.early_exit[0], + layers=self.layers_del, + **unused + ) + decoder_out = F.linear(features, self.embed_word_del.weight) + if normalize: + return F.log_softmax(decoder_out, -1), extra["attn"] + return decoder_out, extra["attn"] + + +@register_model_architecture("levenshtein_transformer", "levenshtein_transformer") +def levenshtein_base_architecture(args): + args.encoder_embed_path = getattr(args, "encoder_embed_path", None) + args.encoder_embed_dim = getattr(args, "encoder_embed_dim", 512) + args.encoder_ffn_embed_dim = getattr(args, "encoder_ffn_embed_dim", 2048) + args.encoder_layers = getattr(args, "encoder_layers", 6) + args.encoder_attention_heads = getattr(args, "encoder_attention_heads", 8) + args.encoder_normalize_before = getattr(args, "encoder_normalize_before", False) + args.encoder_learned_pos = getattr(args, "encoder_learned_pos", False) + args.decoder_embed_path = getattr(args, "decoder_embed_path", None) + args.decoder_embed_dim = getattr(args, "decoder_embed_dim", args.encoder_embed_dim) + args.decoder_ffn_embed_dim = getattr( + args, "decoder_ffn_embed_dim", args.encoder_ffn_embed_dim + ) + args.decoder_layers = getattr(args, "decoder_layers", 6) + args.decoder_attention_heads = getattr(args, "decoder_attention_heads", 8) + args.decoder_normalize_before = getattr(args, "decoder_normalize_before", False) + args.decoder_learned_pos = getattr(args, "decoder_learned_pos", False) + args.attention_dropout = getattr(args, "attention_dropout", 0.0) + args.activation_dropout = getattr(args, "activation_dropout", 0.0) + args.activation_fn = getattr(args, "activation_fn", "relu") + args.dropout = getattr(args, "dropout", 0.1) + args.adaptive_softmax_cutoff = getattr(args, "adaptive_softmax_cutoff", None) + args.adaptive_softmax_dropout = getattr(args, "adaptive_softmax_dropout", 0) + args.share_decoder_input_output_embed = getattr( + args, "share_decoder_input_output_embed", False + ) + args.share_all_embeddings = getattr(args, "share_all_embeddings", False) + args.no_token_positional_embeddings = getattr( + args, "no_token_positional_embeddings", False + ) + args.adaptive_input = getattr(args, "adaptive_input", False) + args.apply_bert_init = getattr(args, "apply_bert_init", False) + + args.decoder_output_dim = getattr( + args, "decoder_output_dim", args.decoder_embed_dim + ) + args.sampling_for_deletion = getattr(args, "sampling_for_deletion", False) + args.decoder_input_dim = getattr(args, "decoder_input_dim", args.decoder_embed_dim) + args.early_exit = getattr(args, "early_exit", "6,6,6") + args.no_share_discriminator = getattr(args, "no_share_discriminator", False) + args.no_share_maskpredictor = getattr(args, "no_share_maskpredictor", False) + args.share_discriminator_maskpredictor = getattr( + args, "share_discriminator_maskpredictor", False + ) + args.no_share_last_layer = getattr(args, "no_share_last_layer", False) + + +@register_model_architecture( + "levenshtein_transformer", "levenshtein_transformer_wmt_en_de" +) +def levenshtein_transformer_wmt_en_de(args): + levenshtein_base_architecture(args) + + +# similar parameters used in the "Attention Is All You Need" paper (Vaswani et al., 2017) +@register_model_architecture( + "levenshtein_transformer", "levenshtein_transformer_vaswani_wmt_en_de_big" +) +def levenshtein_transformer_vaswani_wmt_en_de_big(args): + args.encoder_embed_dim = getattr(args, "encoder_embed_dim", 1024) + args.encoder_ffn_embed_dim = getattr(args, "encoder_ffn_embed_dim", 4096) + args.encoder_attention_heads = getattr(args, "encoder_attention_heads", 16) + args.encoder_normalize_before = getattr(args, "encoder_normalize_before", False) + args.decoder_embed_dim = getattr(args, "decoder_embed_dim", 1024) + args.decoder_ffn_embed_dim = getattr(args, "decoder_ffn_embed_dim", 4096) + args.decoder_attention_heads = getattr(args, "decoder_attention_heads", 16) + args.dropout = getattr(args, "dropout", 0.3) + levenshtein_base_architecture(args) + + +# default parameters used in tensor2tensor implementation +@register_model_architecture( + "levenshtein_transformer", "levenshtein_transformer_wmt_en_de_big" +) +def levenshtein_transformer_wmt_en_de_big_t2t(args): + args.encoder_normalize_before = getattr(args, "encoder_normalize_before", True) + args.decoder_normalize_before = getattr(args, "decoder_normalize_before", True) + args.attention_dropout = getattr(args, "attention_dropout", 0.1) + args.activation_dropout = getattr(args, "activation_dropout", 0.1) + levenshtein_transformer_vaswani_wmt_en_de_big(args) diff --git a/data/fairseq/fairseq/models/nat/levenshtein_utils.py b/data/fairseq/fairseq/models/nat/levenshtein_utils.py new file mode 100644 index 0000000000000000000000000000000000000000..375a98c2e11354de085f0a7926f407bd1a6a2ad4 --- /dev/null +++ b/data/fairseq/fairseq/models/nat/levenshtein_utils.py @@ -0,0 +1,293 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import torch +from fairseq.utils import new_arange + + +# -------------- Helper Functions --------------------------------------------------- # + + +def load_libnat(): + try: + from fairseq import libnat_cuda + + return libnat_cuda, True + + except ImportError as e: + print(str(e) + "... fall back to CPU version") + + try: + from fairseq import libnat + + return libnat, False + + except ImportError as e: + import sys + + sys.stderr.write( + "ERROR: missing libnat_cuda. run `python setup.py build_ext --inplace`\n" + ) + raise e + + +def _get_ins_targets(in_tokens, out_tokens, padding_idx, unk_idx): + libnat, use_cuda = load_libnat() + + def _get_ins_targets_cuda(in_tokens, out_tokens, padding_idx, unk_idx): + in_masks = in_tokens.ne(padding_idx) + out_masks = out_tokens.ne(padding_idx) + mask_ins_targets, masked_tgt_masks = libnat.generate_insertion_labels( + out_tokens.int(), + libnat.levenshtein_distance( + in_tokens.int(), + out_tokens.int(), + in_masks.sum(1).int(), + out_masks.sum(1).int(), + ), + ) + masked_tgt_masks = masked_tgt_masks.bool() & out_masks + mask_ins_targets = mask_ins_targets.type_as(in_tokens)[ + :, 1 : in_masks.size(1) + ].masked_fill_(~in_masks[:, 1:], 0) + masked_tgt_tokens = out_tokens.masked_fill(masked_tgt_masks, unk_idx) + return masked_tgt_masks, masked_tgt_tokens, mask_ins_targets + + def _get_ins_targets_cpu(in_tokens, out_tokens, padding_idx, unk_idx): + in_seq_len, out_seq_len = in_tokens.size(1), out_tokens.size(1) + + in_tokens_list = [ + [t for t in s if t != padding_idx] for i, s in enumerate(in_tokens.tolist()) + ] + out_tokens_list = [ + [t for t in s if t != padding_idx] + for i, s in enumerate(out_tokens.tolist()) + ] + + full_labels = libnat.suggested_ed2_path( + in_tokens_list, out_tokens_list, padding_idx + ) + mask_inputs = [ + [len(c) if c[0] != padding_idx else 0 for c in a[:-1]] for a in full_labels + ] + + # generate labels + masked_tgt_masks = [] + for mask_input in mask_inputs: + mask_label = [] + for beam_size in mask_input[1:-1]: # HACK 1:-1 + mask_label += [0] + [1 for _ in range(beam_size)] + masked_tgt_masks.append( + mask_label + [0 for _ in range(out_seq_len - len(mask_label))] + ) + mask_ins_targets = [ + mask_input[1:-1] + + [0 for _ in range(in_seq_len - 1 - len(mask_input[1:-1]))] + for mask_input in mask_inputs + ] + + # transform to tensor + masked_tgt_masks = torch.tensor( + masked_tgt_masks, device=out_tokens.device + ).bool() + mask_ins_targets = torch.tensor(mask_ins_targets, device=in_tokens.device) + masked_tgt_tokens = out_tokens.masked_fill(masked_tgt_masks, unk_idx) + return masked_tgt_masks, masked_tgt_tokens, mask_ins_targets + + if use_cuda: + return _get_ins_targets_cuda(in_tokens, out_tokens, padding_idx, unk_idx) + return _get_ins_targets_cpu(in_tokens, out_tokens, padding_idx, unk_idx) + + +def _get_del_targets(in_tokens, out_tokens, padding_idx): + libnat, use_cuda = load_libnat() + + def _get_del_targets_cuda(in_tokens, out_tokens, padding_idx): + in_masks = in_tokens.ne(padding_idx) + out_masks = out_tokens.ne(padding_idx) + + word_del_targets = libnat.generate_deletion_labels( + in_tokens.int(), + libnat.levenshtein_distance( + in_tokens.int(), + out_tokens.int(), + in_masks.sum(1).int(), + out_masks.sum(1).int(), + ), + ) + word_del_targets = word_del_targets.type_as(in_tokens).masked_fill_( + ~in_masks, 0 + ) + return word_del_targets + + def _get_del_targets_cpu(in_tokens, out_tokens, padding_idx): + out_seq_len = out_tokens.size(1) + with torch.cuda.device_of(in_tokens): + in_tokens_list = [ + [t for t in s if t != padding_idx] + for i, s in enumerate(in_tokens.tolist()) + ] + out_tokens_list = [ + [t for t in s if t != padding_idx] + for i, s in enumerate(out_tokens.tolist()) + ] + + full_labels = libnat.suggested_ed2_path( + in_tokens_list, out_tokens_list, padding_idx + ) + word_del_targets = [b[-1] for b in full_labels] + word_del_targets = [ + labels + [0 for _ in range(out_seq_len - len(labels))] + for labels in word_del_targets + ] + + # transform to tensor + word_del_targets = torch.tensor(word_del_targets, device=out_tokens.device) + return word_del_targets + + if use_cuda: + return _get_del_targets_cuda(in_tokens, out_tokens, padding_idx) + return _get_del_targets_cpu(in_tokens, out_tokens, padding_idx) + + +def _apply_ins_masks( + in_tokens, in_scores, mask_ins_pred, padding_idx, unk_idx, eos_idx +): + + in_masks = in_tokens.ne(padding_idx) + in_lengths = in_masks.sum(1) + + # HACK: hacky way to shift all the paddings to eos first. + in_tokens.masked_fill_(~in_masks, eos_idx) + mask_ins_pred.masked_fill_(~in_masks[:, 1:], 0) + + out_lengths = in_lengths + mask_ins_pred.sum(1) + out_max_len = out_lengths.max() + out_masks = new_arange(out_lengths, out_max_len)[None, :] < out_lengths[:, None] + + reordering = (mask_ins_pred + in_masks[:, 1:].long()).cumsum(1) + out_tokens = ( + in_tokens.new_zeros(in_tokens.size(0), out_max_len) + .fill_(padding_idx) + .masked_fill_(out_masks, unk_idx) + ) + out_tokens[:, 0] = in_tokens[:, 0] + out_tokens.scatter_(1, reordering, in_tokens[:, 1:]) + + out_scores = None + if in_scores is not None: + in_scores.masked_fill_(~in_masks, 0) + out_scores = in_scores.new_zeros(*out_tokens.size()) + out_scores[:, 0] = in_scores[:, 0] + out_scores.scatter_(1, reordering, in_scores[:, 1:]) + + return out_tokens, out_scores + + +def _apply_ins_words(in_tokens, in_scores, word_ins_pred, word_ins_scores, unk_idx): + word_ins_masks = in_tokens.eq(unk_idx) + out_tokens = in_tokens.masked_scatter(word_ins_masks, word_ins_pred[word_ins_masks]) + + if in_scores is not None: + out_scores = in_scores.masked_scatter( + word_ins_masks, word_ins_scores[word_ins_masks] + ) + else: + out_scores = None + + return out_tokens, out_scores + + +def _apply_del_words( + in_tokens, in_scores, in_attn, word_del_pred, padding_idx, bos_idx, eos_idx +): + # apply deletion to a tensor + in_masks = in_tokens.ne(padding_idx) + bos_eos_masks = in_tokens.eq(bos_idx) | in_tokens.eq(eos_idx) + + max_len = in_tokens.size(1) + word_del_pred.masked_fill_(~in_masks, 1) + word_del_pred.masked_fill_(bos_eos_masks, 0) + + reordering = new_arange(in_tokens).masked_fill_(word_del_pred, max_len).sort(1)[1] + + out_tokens = in_tokens.masked_fill(word_del_pred, padding_idx).gather(1, reordering) + + out_scores = None + if in_scores is not None: + out_scores = in_scores.masked_fill(word_del_pred, 0).gather(1, reordering) + + out_attn = None + if in_attn is not None: + _mask = word_del_pred[:, :, None].expand_as(in_attn) + _reordering = reordering[:, :, None].expand_as(in_attn) + out_attn = in_attn.masked_fill(_mask, 0.0).gather(1, _reordering) + + return out_tokens, out_scores, out_attn + + +def _skip(x, mask): + """ + Getting sliced (dim=0) tensor by mask. Supporting tensor and list/dict of tensors. + """ + if isinstance(x, int): + return x + + if x is None: + return None + + if isinstance(x, torch.Tensor): + if x.size(0) == mask.size(0): + return x[mask] + elif x.size(1) == mask.size(0): + return x[:, mask] + + if isinstance(x, list): + return [_skip(x_i, mask) for x_i in x] + + if isinstance(x, dict): + return {k: _skip(v, mask) for k, v in x.items()} + + raise NotImplementedError + + +def _skip_encoder_out(encoder, encoder_out, mask): + if not mask.any(): + return encoder_out + else: + return encoder.reorder_encoder_out( + encoder_out, mask.nonzero(as_tuple=False).squeeze() + ) + + +def _fill(x, mask, y, padding_idx): + """ + Filling tensor x with y at masked positions (dim=0). + """ + if x is None: + return y + assert x.dim() == y.dim() and mask.size(0) == x.size(0) + assert x.dim() == 2 or (x.dim() == 3 and x.size(2) == y.size(2)) + n_selected = mask.sum() + assert n_selected == y.size(0) + + if n_selected == x.size(0): + return y + + if x.size(1) < y.size(1): + dims = [x.size(0), y.size(1) - x.size(1)] + if x.dim() == 3: + dims.append(x.size(2)) + x = torch.cat([x, x.new_zeros(*dims).fill_(padding_idx)], 1) + x[mask] = y + elif x.size(1) > y.size(1): + x[mask] = padding_idx + if x.dim() == 2: + x[mask, : y.size(1)] = y + else: + x[mask, : y.size(1), :] = y + else: + x[mask] = y + return x diff --git a/data/fairseq/fairseq/models/nat/nat_crf_transformer.py b/data/fairseq/fairseq/models/nat/nat_crf_transformer.py new file mode 100644 index 0000000000000000000000000000000000000000..d4b3cd931ceb077eb30db73df1d5d6cd714a86c2 --- /dev/null +++ b/data/fairseq/fairseq/models/nat/nat_crf_transformer.py @@ -0,0 +1,121 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + + +from fairseq.models import register_model, register_model_architecture +from fairseq.models.nat import NATransformerModel, base_architecture +from fairseq.modules import DynamicCRF + + +@register_model("nacrf_transformer") +class NACRFTransformerModel(NATransformerModel): + def __init__(self, args, encoder, decoder): + super().__init__(args, encoder, decoder) + self.crf_layer = DynamicCRF( + num_embedding=len(self.tgt_dict), + low_rank=args.crf_lowrank_approx, + beam_size=args.crf_beam_approx, + ) + + @property + def allow_ensemble(self): + return False + + @staticmethod + def add_args(parser): + NATransformerModel.add_args(parser) + parser.add_argument( + "--crf-lowrank-approx", + type=int, + help="the dimension of low-rank approximation of transition", + ) + parser.add_argument( + "--crf-beam-approx", + type=int, + help="the beam size for apporixmating the normalizing factor", + ) + parser.add_argument( + "--word-ins-loss-factor", + type=float, + help="weights on NAT loss used to co-training with CRF loss.", + ) + + def forward( + self, src_tokens, src_lengths, prev_output_tokens, tgt_tokens, **kwargs + ): + # encoding + encoder_out = self.encoder(src_tokens, src_lengths=src_lengths, **kwargs) + + # length prediction + length_out = self.decoder.forward_length( + normalize=False, encoder_out=encoder_out + ) + length_tgt = self.decoder.forward_length_prediction( + length_out, encoder_out, tgt_tokens + ) + + # decoding + word_ins_out = self.decoder( + normalize=False, + prev_output_tokens=prev_output_tokens, + encoder_out=encoder_out, + ) + word_ins_tgt, word_ins_mask = tgt_tokens, tgt_tokens.ne(self.pad) + + # compute the log-likelihood of CRF + crf_nll = -self.crf_layer(word_ins_out, word_ins_tgt, word_ins_mask) + crf_nll = (crf_nll / word_ins_mask.type_as(crf_nll).sum(-1)).mean() + + return { + "word_ins": { + "out": word_ins_out, + "tgt": word_ins_tgt, + "mask": word_ins_mask, + "ls": self.args.label_smoothing, + "nll_loss": True, + "factor": self.args.word_ins_loss_factor, + }, + "word_crf": {"loss": crf_nll}, + "length": { + "out": length_out, + "tgt": length_tgt, + "factor": self.decoder.length_loss_factor, + }, + } + + def forward_decoder(self, decoder_out, encoder_out, decoding_format=None, **kwargs): + output_tokens = decoder_out.output_tokens + output_scores = decoder_out.output_scores + history = decoder_out.history + + # execute the decoder and get emission scores + output_masks = output_tokens.ne(self.pad) + word_ins_out = self.decoder( + normalize=False, prev_output_tokens=output_tokens, encoder_out=encoder_out + ) + + # run viterbi decoding through CRF + _scores, _tokens = self.crf_layer.forward_decoder(word_ins_out, output_masks) + output_tokens.masked_scatter_(output_masks, _tokens[output_masks]) + output_scores.masked_scatter_(output_masks, _scores[output_masks]) + if history is not None: + history.append(output_tokens.clone()) + + return decoder_out._replace( + output_tokens=output_tokens, + output_scores=output_scores, + attn=None, + history=history, + ) + + +@register_model_architecture("nacrf_transformer", "nacrf_transformer") +def nacrf_base_architecture(args): + args.crf_lowrank_approx = getattr(args, "crf_lowrank_approx", 32) + args.crf_beam_approx = getattr(args, "crf_beam_approx", 64) + args.word_ins_loss_factor = getattr(args, "word_ins_loss_factor", 0.5) + args.encoder_normalize_before = getattr(args, "encoder_normalize_before", True) + args.decoder_normalize_before = getattr(args, "decoder_normalize_before", True) + base_architecture(args) diff --git a/data/fairseq/fairseq/models/nat/nonautoregressive_ensembles.py b/data/fairseq/fairseq/models/nat/nonautoregressive_ensembles.py new file mode 100644 index 0000000000000000000000000000000000000000..0a0221f9c437afc547afa1ad019e5e4a6aaaef17 --- /dev/null +++ b/data/fairseq/fairseq/models/nat/nonautoregressive_ensembles.py @@ -0,0 +1,254 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import math + +import torch +import torch.nn.functional as F +from fairseq.models.nat import ( + _apply_del_words, + _apply_ins_masks, + _apply_ins_words, + _fill, + _skip, + _skip_encoder_out, +) + + +class _EnsembleModelEncoder(object): + def __init__(self, models): + self.models = models + + def reorder_encoder_out(self, encoder_outs, new_order): + encoder_outs = [ + model.encoder.reorder_encoder_out(encoder_out, new_order) + for model, encoder_out in zip(self.models, encoder_outs) + ] + return encoder_outs + + +class BasicEnsembleModel(torch.nn.Module): + """A wrapper around an ensemble of models.""" + + def __init__(self, models): + super().__init__() + self.models = torch.nn.ModuleList(models) + self.bos = self.models[0].decoder.dictionary.bos() + self.eos = self.models[0].decoder.dictionary.eos() + self.pad = self.models[0].decoder.dictionary.pad() + self.unk = self.models[0].decoder.dictionary.unk() + self.encoder = _EnsembleModelEncoder(self.models) + + def has_encoder(self): + return hasattr(self.models[0], "encoder") + + def max_decoder_positions(self): + return min(m.max_decoder_positions() for m in self.models) + + @torch.no_grad() + def forward_encoder(self, encoder_input): + if not self.has_encoder(): + return None + return [model.forward_encoder(encoder_input) for model in self.models] + + @torch.no_grad() + def forward_decoder(self, *inputs): + raise NotImplementedError + + def initialize_output_tokens(self, *inputs): + raise NotImplementedError + + +class EnsembleLevT(BasicEnsembleModel): + """A wrapper around an ensemble of models.""" + + def __init__(self, models): + super().__init__(models) + + @torch.no_grad() + def forward_decoder( + self, decoder_out, encoder_outs, eos_penalty=0.0, max_ratio=None, **kwargs + ): + # LevT ensembling + # A pipeline of three steps: deletion, placeholder, and word insertion. + # We need to average scores in each step in a pipeline way because of dependence. + # deletion + output_tokens = decoder_out.output_tokens + output_scores = decoder_out.output_scores + attn = decoder_out.attn + + bsz = output_tokens.size(0) + if max_ratio is None: + max_lens = output_tokens.new().fill_(255) + else: + if not encoder_outs[0]["encoder_padding_mask"]: + src_lens = ( + encoder_outs[0]["encoder_out"][0] + .new(bsz) + .fill_(encoder_outs[0]["encoder_out"][0].size(1)) + ) + else: + src_lens = (~encoder_outs[0]["encoder_padding_mask"][0]).sum(1) + max_lens = (src_lens * max_ratio).clamp(min=10).long() + + # delete words + # do not delete tokens if it is + can_del_word = output_tokens.ne(self.pad).sum(1) > 2 + if can_del_word.sum() != 0: # we cannot delete, skip + output_tokens, output_scores, attn = self.forward_word_del( + encoder_outs, + output_tokens, + output_scores, + attn, + can_del_word, + ) + + # insert placeholders + can_ins_mask = output_tokens.ne(self.pad).sum(1) < max_lens + if can_ins_mask.sum() != 0: + output_tokens, output_scores = self.forward_mask_ins( + encoder_outs, + output_tokens, + output_scores, + can_ins_mask, + eos_penalty, + max_lens, + ) + + # insert words + can_ins_word = output_tokens.eq(self.unk).sum(1) > 0 + if can_ins_word.sum() != 0: + output_tokens, output_scores, attn = self.forward_word_ins( + encoder_outs, + output_tokens, + output_scores, + attn, + can_ins_word, + ) + + # delete some unnecessary paddings + cut_off = output_tokens.ne(self.pad).sum(1).max() + output_tokens = output_tokens[:, :cut_off] + output_scores = output_scores[:, :cut_off] + attn = None if attn is None else attn[:, :cut_off, :] + return decoder_out._replace( + output_tokens=output_tokens, + output_scores=output_scores, + attn=attn, + history=None, + ) + + def forward_word_del( + self, encoder_outs, output_tokens, output_scores, attn, can_del_word + ): + word_del_score_avg = [] + word_del_attn_avg = [] + for model, encoder_out in zip(self.models, encoder_outs): + word_del_out, word_del_attn = model.decoder.forward_word_del( + _skip(output_tokens, can_del_word), + _skip_encoder_out(model.encoder, encoder_out, can_del_word), + ) + word_del_score = F.log_softmax(word_del_out, 2) + word_del_score_avg.append(word_del_score) + word_del_attn_avg.append(word_del_attn) + word_del_score_avg = torch.logsumexp( + torch.stack(word_del_score_avg, dim=0), dim=0 + ) - math.log(len(self.models)) + word_del_pred = word_del_score_avg.max(-1)[1].bool() + if word_del_attn_avg[0] is not None: + word_del_attn_avg = torch.stack(word_del_attn_avg, dim=0) / len(self.models) + else: + word_del_attn_avg = None + + _tokens, _scores, _attn = _apply_del_words( + output_tokens[can_del_word], + output_scores[can_del_word], + word_del_attn_avg, + word_del_pred, + self.pad, + self.bos, + self.eos, + ) + output_tokens = _fill(output_tokens, can_del_word, _tokens, self.pad) + output_scores = _fill(output_scores, can_del_word, _scores, 0) + attn = _fill(attn, can_del_word, _attn, 0.0) + return output_tokens, output_scores, attn + + def forward_mask_ins( + self, + encoder_outs, + output_tokens, + output_scores, + can_ins_mask, + eos_penalty, + max_lens, + ): + mask_ins_score_avg = [] + for model, encoder_out in zip(self.models, encoder_outs): + mask_ins_out, _ = model.decoder.forward_mask_ins( + _skip(output_tokens, can_ins_mask), + _skip_encoder_out(model.encoder, encoder_out, can_ins_mask), + ) + mask_ins_score = F.log_softmax(mask_ins_out, 2) + if eos_penalty > 0.0: + mask_ins_score[:, :, 0] -= eos_penalty + mask_ins_score_avg.append(mask_ins_score) + mask_ins_score_avg = torch.logsumexp( + torch.stack(mask_ins_score_avg, dim=0), dim=0 + ) - math.log(len(self.models)) + mask_ins_pred = mask_ins_score_avg.max(-1)[1] + mask_ins_pred = torch.min( + mask_ins_pred, max_lens[can_ins_mask, None].expand_as(mask_ins_pred) + ) + _tokens, _scores = _apply_ins_masks( + output_tokens[can_ins_mask], + output_scores[can_ins_mask], + mask_ins_pred, + self.pad, + self.unk, + self.eos, + ) + output_tokens = _fill(output_tokens, can_ins_mask, _tokens, self.pad) + output_scores = _fill(output_scores, can_ins_mask, _scores, 0) + return output_tokens, output_scores + + def forward_word_ins( + self, encoder_outs, output_tokens, output_scores, attn, can_ins_word + ): + word_ins_score_avg = [] + word_ins_attn_avg = [] + for model, encoder_out in zip(self.models, encoder_outs): + word_ins_out, word_ins_attn = model.decoder.forward_word_ins( + _skip(output_tokens, can_ins_word), + _skip_encoder_out(model.encoder, encoder_out, can_ins_word), + ) + word_ins_score = F.log_softmax(word_ins_out, 2) + word_ins_score_avg.append(word_ins_score) + word_ins_attn_avg.append(word_ins_attn) + word_ins_score_avg = torch.logsumexp( + torch.stack(word_ins_score_avg, dim=0), dim=0 + ) - math.log(len(self.models)) + if word_ins_attn_avg[0] is not None: + word_ins_attn_avg = torch.stack(word_ins_attn_avg, dim=0) / len(self.models) + else: + word_ins_attn_avg = None + word_ins_score_max, word_ins_pred = word_ins_score_avg.max(-1) + + _tokens, _scores = _apply_ins_words( + output_tokens[can_ins_word], + output_scores[can_ins_word], + word_ins_pred, + word_ins_score_max, + self.unk, + ) + + output_tokens = _fill(output_tokens, can_ins_word, _tokens, self.pad) + output_scores = _fill(output_scores, can_ins_word, _scores, 0) + attn = _fill(attn, can_ins_word, word_ins_attn, 0.0) + return output_tokens, output_scores, attn + + def initialize_output_tokens(self, encoder_outs, src_tokens): + # LevT doesn't do length prediction. + return self.models[0].initialize_output_tokens(encoder_outs[0], src_tokens) diff --git a/data/fairseq/fairseq/models/nat/nonautoregressive_transformer.py b/data/fairseq/fairseq/models/nat/nonautoregressive_transformer.py new file mode 100644 index 0000000000000000000000000000000000000000..d114202d25fbd1dca66c7abebb0b0a8bffbe094d --- /dev/null +++ b/data/fairseq/fairseq/models/nat/nonautoregressive_transformer.py @@ -0,0 +1,456 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import torch +import torch.nn.functional as F +from fairseq import utils +from fairseq.iterative_refinement_generator import DecoderOut +from fairseq.models import register_model, register_model_architecture +from fairseq.models.nat import FairseqNATDecoder, FairseqNATModel, ensemble_decoder +from fairseq.models.transformer import Embedding +from fairseq.modules.transformer_sentence_encoder import init_bert_params + + +def _mean_pooling(enc_feats, src_masks): + # enc_feats: T x B x C + # src_masks: B x T or None + if src_masks is None: + enc_feats = enc_feats.mean(0) + else: + src_masks = (~src_masks).transpose(0, 1).type_as(enc_feats) + enc_feats = ( + (enc_feats / src_masks.sum(0)[None, :, None]) * src_masks[:, :, None] + ).sum(0) + return enc_feats + + +def _argmax(x, dim): + return (x == x.max(dim, keepdim=True)[0]).type_as(x) + + +def _uniform_assignment(src_lens, trg_lens): + max_trg_len = trg_lens.max() + steps = (src_lens.float() - 1) / (trg_lens.float() - 1) # step-size + # max_trg_len + index_t = utils.new_arange(trg_lens, max_trg_len).float() + index_t = steps[:, None] * index_t[None, :] # batch_size X max_trg_len + index_t = torch.round(index_t).long().detach() + return index_t + + +@register_model("nonautoregressive_transformer") +class NATransformerModel(FairseqNATModel): + @property + def allow_length_beam(self): + return True + + @staticmethod + def add_args(parser): + FairseqNATModel.add_args(parser) + + # length prediction + parser.add_argument( + "--src-embedding-copy", + action="store_true", + help="copy encoder word embeddings as the initial input of the decoder", + ) + parser.add_argument( + "--pred-length-offset", + action="store_true", + help="predicting the length difference between the target and source sentences", + ) + parser.add_argument( + "--sg-length-pred", + action="store_true", + help="stop the gradients back-propagated from the length predictor", + ) + parser.add_argument( + "--length-loss-factor", + type=float, + help="weights on the length prediction loss", + ) + + @classmethod + def build_decoder(cls, args, tgt_dict, embed_tokens): + decoder = NATransformerDecoder(args, tgt_dict, embed_tokens) + if getattr(args, "apply_bert_init", False): + decoder.apply(init_bert_params) + return decoder + + def forward( + self, src_tokens, src_lengths, prev_output_tokens, tgt_tokens, **kwargs + ): + # encoding + encoder_out = self.encoder(src_tokens, src_lengths=src_lengths, **kwargs) + + # length prediction + length_out = self.decoder.forward_length( + normalize=False, encoder_out=encoder_out + ) + length_tgt = self.decoder.forward_length_prediction( + length_out, encoder_out, tgt_tokens + ) + + # decoding + word_ins_out = self.decoder( + normalize=False, + prev_output_tokens=prev_output_tokens, + encoder_out=encoder_out, + ) + + return { + "word_ins": { + "out": word_ins_out, + "tgt": tgt_tokens, + "mask": tgt_tokens.ne(self.pad), + "ls": self.args.label_smoothing, + "nll_loss": True, + }, + "length": { + "out": length_out, + "tgt": length_tgt, + "factor": self.decoder.length_loss_factor, + }, + } + + def forward_decoder(self, decoder_out, encoder_out, decoding_format=None, **kwargs): + step = decoder_out.step + output_tokens = decoder_out.output_tokens + output_scores = decoder_out.output_scores + history = decoder_out.history + + # execute the decoder + output_masks = output_tokens.ne(self.pad) + _scores, _tokens = self.decoder( + normalize=True, + prev_output_tokens=output_tokens, + encoder_out=encoder_out, + step=step, + ).max(-1) + + output_tokens.masked_scatter_(output_masks, _tokens[output_masks]) + output_scores.masked_scatter_(output_masks, _scores[output_masks]) + if history is not None: + history.append(output_tokens.clone()) + + return decoder_out._replace( + output_tokens=output_tokens, + output_scores=output_scores, + attn=None, + history=history, + ) + + def initialize_output_tokens(self, encoder_out, src_tokens): + # length prediction + length_tgt = self.decoder.forward_length_prediction( + self.decoder.forward_length(normalize=True, encoder_out=encoder_out), + encoder_out=encoder_out, + ) + + max_length = length_tgt.clamp_(min=2).max() + idx_length = utils.new_arange(src_tokens, max_length) + + initial_output_tokens = src_tokens.new_zeros( + src_tokens.size(0), max_length + ).fill_(self.pad) + initial_output_tokens.masked_fill_( + idx_length[None, :] < length_tgt[:, None], self.unk + ) + initial_output_tokens[:, 0] = self.bos + initial_output_tokens.scatter_(1, length_tgt[:, None] - 1, self.eos) + + initial_output_scores = initial_output_tokens.new_zeros( + *initial_output_tokens.size() + ).type_as(encoder_out["encoder_out"][0]) + + return DecoderOut( + output_tokens=initial_output_tokens, + output_scores=initial_output_scores, + attn=None, + step=0, + max_step=0, + history=None, + ) + + def regenerate_length_beam(self, decoder_out, beam_size): + output_tokens = decoder_out.output_tokens + length_tgt = output_tokens.ne(self.pad).sum(1) + length_tgt = ( + length_tgt[:, None] + + utils.new_arange(length_tgt, 1, beam_size) + - beam_size // 2 + ) + length_tgt = length_tgt.view(-1).clamp_(min=2) + max_length = length_tgt.max() + idx_length = utils.new_arange(length_tgt, max_length) + + initial_output_tokens = output_tokens.new_zeros( + length_tgt.size(0), max_length + ).fill_(self.pad) + initial_output_tokens.masked_fill_( + idx_length[None, :] < length_tgt[:, None], self.unk + ) + initial_output_tokens[:, 0] = self.bos + initial_output_tokens.scatter_(1, length_tgt[:, None] - 1, self.eos) + + initial_output_scores = initial_output_tokens.new_zeros( + *initial_output_tokens.size() + ).type_as(decoder_out.output_scores) + + return decoder_out._replace( + output_tokens=initial_output_tokens, output_scores=initial_output_scores + ) + + +class NATransformerDecoder(FairseqNATDecoder): + def __init__(self, args, dictionary, embed_tokens, no_encoder_attn=False): + super().__init__( + args, dictionary, embed_tokens, no_encoder_attn=no_encoder_attn + ) + self.dictionary = dictionary + self.bos = dictionary.bos() + self.unk = dictionary.unk() + self.eos = dictionary.eos() + + self.encoder_embed_dim = args.encoder_embed_dim + self.sg_length_pred = getattr(args, "sg_length_pred", False) + self.pred_length_offset = getattr(args, "pred_length_offset", False) + self.length_loss_factor = getattr(args, "length_loss_factor", 0.1) + self.src_embedding_copy = getattr(args, "src_embedding_copy", False) + self.embed_length = Embedding(256, self.encoder_embed_dim, None) + + @ensemble_decoder + def forward(self, normalize, encoder_out, prev_output_tokens, step=0, **unused): + features, _ = self.extract_features( + prev_output_tokens, + encoder_out=encoder_out, + embedding_copy=(step == 0) & self.src_embedding_copy, + ) + decoder_out = self.output_layer(features) + return F.log_softmax(decoder_out, -1) if normalize else decoder_out + + @ensemble_decoder + def forward_length(self, normalize, encoder_out): + enc_feats = encoder_out["encoder_out"][0] # T x B x C + if len(encoder_out["encoder_padding_mask"]) > 0: + src_masks = encoder_out["encoder_padding_mask"][0] # B x T + else: + src_masks = None + enc_feats = _mean_pooling(enc_feats, src_masks) + if self.sg_length_pred: + enc_feats = enc_feats.detach() + length_out = F.linear(enc_feats, self.embed_length.weight) + return F.log_softmax(length_out, -1) if normalize else length_out + + def extract_features( + self, + prev_output_tokens, + encoder_out=None, + early_exit=None, + embedding_copy=False, + **unused + ): + """ + Similar to *forward* but only return features. + + Inputs: + prev_output_tokens: Tensor(B, T) + encoder_out: a dictionary of hidden states and masks + + Returns: + tuple: + - the decoder's features of shape `(batch, tgt_len, embed_dim)` + - a dictionary with any model-specific outputs + the LevenshteinTransformer decoder has full-attention to all generated tokens + """ + # embedding + if embedding_copy: + src_embd = encoder_out["encoder_embedding"][0] + if len(encoder_out["encoder_padding_mask"]) > 0: + src_mask = encoder_out["encoder_padding_mask"][0] + else: + src_mask = None + src_mask = ( + ~src_mask + if src_mask is not None + else prev_output_tokens.new_ones(*src_embd.size()[:2]).bool() + ) + + x, decoder_padding_mask = self.forward_embedding( + prev_output_tokens, + self.forward_copying_source( + src_embd, src_mask, prev_output_tokens.ne(self.padding_idx) + ), + ) + + else: + + x, decoder_padding_mask = self.forward_embedding(prev_output_tokens) + + # B x T x C -> T x B x C + x = x.transpose(0, 1) + attn = None + inner_states = [x] + + # decoder layers + for i, layer in enumerate(self.layers): + + # early exit from the decoder. + if (early_exit is not None) and (i >= early_exit): + break + + x, attn, _ = layer( + x, + encoder_out["encoder_out"][0] + if (encoder_out is not None and len(encoder_out["encoder_out"]) > 0) + else None, + encoder_out["encoder_padding_mask"][0] + if ( + encoder_out is not None + and len(encoder_out["encoder_padding_mask"]) > 0 + ) + else None, + self_attn_mask=None, + self_attn_padding_mask=decoder_padding_mask, + ) + inner_states.append(x) + + if self.layer_norm: + x = self.layer_norm(x) + + # T x B x C -> B x T x C + x = x.transpose(0, 1) + + if self.project_out_dim is not None: + x = self.project_out_dim(x) + + return x, {"attn": attn, "inner_states": inner_states} + + def forward_embedding(self, prev_output_tokens, states=None): + # embed positions + positions = ( + self.embed_positions(prev_output_tokens) + if self.embed_positions is not None + else None + ) + + # embed tokens and positions + if states is None: + x = self.embed_scale * self.embed_tokens(prev_output_tokens) + if self.project_in_dim is not None: + x = self.project_in_dim(x) + else: + x = states + + if positions is not None: + x += positions + x = self.dropout_module(x) + decoder_padding_mask = prev_output_tokens.eq(self.padding_idx) + return x, decoder_padding_mask + + def forward_copying_source(self, src_embeds, src_masks, tgt_masks): + length_sources = src_masks.sum(1) + length_targets = tgt_masks.sum(1) + mapped_inputs = _uniform_assignment(length_sources, length_targets).masked_fill( + ~tgt_masks, 0 + ) + copied_embedding = torch.gather( + src_embeds, + 1, + mapped_inputs.unsqueeze(-1).expand( + *mapped_inputs.size(), src_embeds.size(-1) + ), + ) + return copied_embedding + + def forward_length_prediction(self, length_out, encoder_out, tgt_tokens=None): + enc_feats = encoder_out["encoder_out"][0] # T x B x C + if len(encoder_out["encoder_padding_mask"]) > 0: + src_masks = encoder_out["encoder_padding_mask"][0] # B x T + else: + src_masks = None + if self.pred_length_offset: + if src_masks is None: + src_lengs = enc_feats.new_ones(enc_feats.size(1)).fill_( + enc_feats.size(0) + ) + else: + src_lengs = (~src_masks).transpose(0, 1).type_as(enc_feats).sum(0) + src_lengs = src_lengs.long() + + if tgt_tokens is not None: + # obtain the length target + tgt_lengs = tgt_tokens.ne(self.padding_idx).sum(1).long() + if self.pred_length_offset: + length_tgt = tgt_lengs - src_lengs + 128 + else: + length_tgt = tgt_lengs + length_tgt = length_tgt.clamp(min=0, max=255) + + else: + # predict the length target (greedy for now) + # TODO: implementing length-beam + pred_lengs = length_out.max(-1)[1] + if self.pred_length_offset: + length_tgt = pred_lengs - 128 + src_lengs + else: + length_tgt = pred_lengs + + return length_tgt + + +@register_model_architecture( + "nonautoregressive_transformer", "nonautoregressive_transformer" +) +def base_architecture(args): + args.encoder_embed_path = getattr(args, "encoder_embed_path", None) + args.encoder_embed_dim = getattr(args, "encoder_embed_dim", 512) + args.encoder_ffn_embed_dim = getattr(args, "encoder_ffn_embed_dim", 2048) + args.encoder_layers = getattr(args, "encoder_layers", 6) + args.encoder_attention_heads = getattr(args, "encoder_attention_heads", 8) + args.encoder_normalize_before = getattr(args, "encoder_normalize_before", False) + args.encoder_learned_pos = getattr(args, "encoder_learned_pos", False) + args.decoder_embed_path = getattr(args, "decoder_embed_path", None) + args.decoder_embed_dim = getattr(args, "decoder_embed_dim", args.encoder_embed_dim) + args.decoder_ffn_embed_dim = getattr( + args, "decoder_ffn_embed_dim", args.encoder_ffn_embed_dim + ) + args.decoder_layers = getattr(args, "decoder_layers", 6) + args.decoder_attention_heads = getattr(args, "decoder_attention_heads", 8) + args.decoder_normalize_before = getattr(args, "decoder_normalize_before", False) + args.decoder_learned_pos = getattr(args, "decoder_learned_pos", False) + args.attention_dropout = getattr(args, "attention_dropout", 0.0) + args.activation_dropout = getattr(args, "activation_dropout", 0.0) + args.activation_fn = getattr(args, "activation_fn", "relu") + args.dropout = getattr(args, "dropout", 0.1) + args.adaptive_softmax_cutoff = getattr(args, "adaptive_softmax_cutoff", None) + args.adaptive_softmax_dropout = getattr(args, "adaptive_softmax_dropout", 0) + args.share_decoder_input_output_embed = getattr( + args, "share_decoder_input_output_embed", False + ) + args.share_all_embeddings = getattr(args, "share_all_embeddings", False) + args.no_token_positional_embeddings = getattr( + args, "no_token_positional_embeddings", False + ) + args.adaptive_input = getattr(args, "adaptive_input", False) + args.apply_bert_init = getattr(args, "apply_bert_init", False) + + args.decoder_output_dim = getattr( + args, "decoder_output_dim", args.decoder_embed_dim + ) + args.decoder_input_dim = getattr(args, "decoder_input_dim", args.decoder_embed_dim) + + # --- special arguments --- + args.sg_length_pred = getattr(args, "sg_length_pred", False) + args.pred_length_offset = getattr(args, "pred_length_offset", False) + args.length_loss_factor = getattr(args, "length_loss_factor", 0.1) + args.src_embedding_copy = getattr(args, "src_embedding_copy", False) + + +@register_model_architecture( + "nonautoregressive_transformer", "nonautoregressive_transformer_wmt_en_de" +) +def nonautoregressive_transformer_wmt_en_de(args): + base_architecture(args) diff --git a/data/fairseq/fairseq/models/speech_dlm/__init__.py b/data/fairseq/fairseq/models/speech_dlm/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..6ea914d6a578651fecd18cc7f352382623de303a --- /dev/null +++ b/data/fairseq/fairseq/models/speech_dlm/__init__.py @@ -0,0 +1,7 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from .speech_dlm import * # noqa +from .hub_interface import * # noqa diff --git a/data/fairseq/fairseq/models/speech_dlm/__pycache__/__init__.cpython-38.pyc b/data/fairseq/fairseq/models/speech_dlm/__pycache__/__init__.cpython-38.pyc new file mode 100644 index 0000000000000000000000000000000000000000..361c5345bb98916c63bec05fe0457d061703a31d Binary files /dev/null and b/data/fairseq/fairseq/models/speech_dlm/__pycache__/__init__.cpython-38.pyc differ diff --git a/data/fairseq/fairseq/models/speech_dlm/__pycache__/hub_interface.cpython-38.pyc b/data/fairseq/fairseq/models/speech_dlm/__pycache__/hub_interface.cpython-38.pyc new file mode 100644 index 0000000000000000000000000000000000000000..6ce99801d2da897bf0a5d8d40caa1d0764a99167 Binary files /dev/null and b/data/fairseq/fairseq/models/speech_dlm/__pycache__/hub_interface.cpython-38.pyc differ diff --git a/data/fairseq/fairseq/models/speech_dlm/__pycache__/speech_dlm.cpython-38.pyc b/data/fairseq/fairseq/models/speech_dlm/__pycache__/speech_dlm.cpython-38.pyc new file mode 100644 index 0000000000000000000000000000000000000000..e6d97379b5b66904b2fb997d3e9cdd61afe62628 Binary files /dev/null and b/data/fairseq/fairseq/models/speech_dlm/__pycache__/speech_dlm.cpython-38.pyc differ diff --git a/data/fairseq/fairseq/models/speech_dlm/hub_interface.py b/data/fairseq/fairseq/models/speech_dlm/hub_interface.py new file mode 100644 index 0000000000000000000000000000000000000000..11bc0f50bbb8cf3d146741c1f02b1ebfe8d4b7f6 --- /dev/null +++ b/data/fairseq/fairseq/models/speech_dlm/hub_interface.py @@ -0,0 +1,192 @@ +#!/usr/bin/env python3 -u +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import copy +import logging +from typing import Any, Dict, Iterator, List + +import torch +from fairseq import utils +from omegaconf import open_dict +from torch import nn + +from tqdm import tqdm + +from fairseq.hub_utils import GeneratorHubInterface + + +logger = logging.getLogger(__name__) + + +class MultichannelGeneratorHubInterface(GeneratorHubInterface): + """Pytorch Hub interface for generating sequences from a pre-trained + multichannel language model. + """ + + def __init__(self, cfg, task, models): + super().__init__(cfg, task, models) + self.cfg = cfg + self.task = task + self.models = nn.ModuleList(models) + self.src_dicts = task.source_dictionaries + self.tgt_dicts = task.target_dictionaries + self.channels = task.channels + + # optimize model for generation + for model in self.models: + model.prepare_for_inference_(cfg) + + def sample( + self, + sentences: List[Dict[str, str]], + beam: int = 1, + verbose: bool = False, + **kwargs + ) -> List[str]: + if isinstance(sentences, dict): + return self.sample([sentences], beam=beam, verbose=verbose, **kwargs)[0] + tokenized_sentences = [self.encode(sentence) for sentence in sentences] + batched_hypos = self.generate(tokenized_sentences, beam, verbose, **kwargs) + return [self.decode(hypos[0]["tokens"]) for hypos in batched_hypos] + + def score(self, sentences: List[Dict[str, str]], **kwargs): + raise NotImplementedError( + "MultichannelGeneratorHubInterface doesn't support score() method" + ) + + def generate( + self, + tokenized_sentences: List[Dict[str, torch.LongTensor]], + beam: int = 5, + verbose: bool = False, + skip_invalid_size_inputs=False, + inference_step_args=None, + **kwargs + ) -> List[List[Dict[str, torch.Tensor]]]: + if isinstance(tokenized_sentences, dict): + return self.generate( + [tokenized_sentences], beam=beam, verbose=verbose, **kwargs + )[0] + + # build generator using current args as well as any kwargs + gen_args = copy.deepcopy(self.cfg.generation) + with open_dict(gen_args): + gen_args.beam = beam + for k, v in kwargs.items(): + setattr(gen_args, k, v) + generator = self.task.build_generator(self.models, gen_args) + + inference_step_args = inference_step_args or {} + results = [] + for batch in tqdm( + self._build_batches(tokenized_sentences, skip_invalid_size_inputs) + ): + batch = utils.apply_to_sample(lambda t: t.to(self.device), batch) + translations = self.task.inference_step( + generator, self.models, batch, **inference_step_args + ) + for id, hypos in zip(batch["id"].tolist(), translations): + # The output of the generator is supposed to be a tensor of size (bsz x max_len x n_channels) + # So we need to convert it to dictionary form + for i in range(len(hypos)): + hypos[i]["tokens"] = { + channel: hypos[i]["tokens"][..., j] + for j, channel in enumerate(self.channels) + } + results.append((id, hypos)) + + # sort output to match input order + outputs = [hypos for _, hypos in sorted(results, key=lambda x: x[0])] + + if verbose: + + def getarg(name, default): + return getattr(gen_args, name, getattr(self.cfg, name, default)) + + for source_tokens, target_hypotheses in zip(tokenized_sentences, outputs): + src_str_with_unk = { + channel: self.string(source_tokens[channel], channel) + for channel in source_tokens + } + logger.info("S\t{}".format(src_str_with_unk)) + for hypo in target_hypotheses: + hypo_str = self.decode(hypo["tokens"]) + logger.info("H\t{}\t{}".format(hypo["score"], hypo_str)) + # hypo["positional_scores"]: T x n_channels + pos_scores = {} + for c, channel in enumerate(source_tokens): + pos_scores[channel] = " ".join( + map( + lambda x: "{:.4f}".format(x), + hypo["positional_scores"][:, c].tolist(), + ) + ) + logger.info("P\t{}".format(pos_scores)) + + return outputs + + def encode(self, sentence: Dict[str, str]) -> Dict[str, torch.LongTensor]: + assert isinstance( + sentence, dict + ), "Input sentence is expected to be a dictionary over channels" + assert set(sentence.keys()) == set( + self.channels + ), "Mismatch between input sentence keys and model channels ({} vs {})".format( + set(sentence.keys()), set(self.channels) + ) + encoded_sentence = {} + for channel in sentence: + sentence_channel = sentence[channel] + sentence_channel = self.tokenize(sentence_channel) + sentence_channel = self.apply_bpe(sentence_channel) + sentence_channel = self.binarize(sentence_channel, channel) + encoded_sentence[channel] = sentence_channel + sentence_size = encoded_sentence[self.channels[0]].size() + assert all( + encoded_sentence[channel].size() == sentence_size + for channel in encoded_sentence + ), "Input tensors are expected to have the same size in all channels" + return encoded_sentence + + def decode(self, tokens: Dict[str, torch.LongTensor]) -> Dict[str, str]: + assert isinstance( + tokens, dict + ), "Input tokens are expected to be a dictionary over channels" + assert set(tokens.keys()) == set( + self.channels + ), "Mismatch between input tokens keys and model channels ({} vs {})".format( + set(tokens.keys()), set(self.channels) + ) + decoded_sentence = {} + for channel in tokens: + tokens_channel = tokens[channel] + sentence_channel = self.string(tokens_channel, channel) + sentence_channel = self.remove_bpe(sentence_channel) + sentence_channel = self.detokenize(sentence_channel) + decoded_sentence[channel] = sentence_channel + return decoded_sentence + + def binarize(self, sentence: str, channel: str) -> torch.LongTensor: + return ( + self.src_dicts[channel].encode_line(sentence, add_if_not_exist=False).long() + ) + + def string(self, tokens: torch.LongTensor, channel: str) -> str: + return self.tgt_dicts[channel].string(tokens) + + def _build_batches( + self, tokens: List[Dict[str, List[int]]], skip_invalid_size_inputs: bool + ) -> Iterator[Dict[str, Any]]: + lengths = torch.LongTensor([next(iter(d.values())).numel() for d in tokens]) + batch_iterator = self.task.get_batch_iterator( + dataset=self.task.build_dataset_for_inference(tokens, lengths), + max_tokens=self.cfg.dataset.max_tokens, + max_sentences=self.cfg.dataset.batch_size, + max_positions=self.max_positions, + ignore_invalid_inputs=skip_invalid_size_inputs, + disable_iterator_cache=True, + ).next_epoch_itr(shuffle=False) + return batch_iterator diff --git a/data/fairseq/fairseq/models/speech_dlm/modules/__init__.py b/data/fairseq/fairseq/models/speech_dlm/modules/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391 diff --git a/data/fairseq/fairseq/models/speech_dlm/modules/__pycache__/__init__.cpython-38.pyc b/data/fairseq/fairseq/models/speech_dlm/modules/__pycache__/__init__.cpython-38.pyc new file mode 100644 index 0000000000000000000000000000000000000000..c3f803fb5d4838913c3a6a91f7cfed6f2fe2043c Binary files /dev/null and b/data/fairseq/fairseq/models/speech_dlm/modules/__pycache__/__init__.cpython-38.pyc differ diff --git a/data/fairseq/fairseq/models/speech_dlm/modules/__pycache__/speech_dlm_decoder.cpython-38.pyc b/data/fairseq/fairseq/models/speech_dlm/modules/__pycache__/speech_dlm_decoder.cpython-38.pyc new file mode 100644 index 0000000000000000000000000000000000000000..22ca3bca52e618c9022d5d88448c1c4cb0a43f95 Binary files /dev/null and b/data/fairseq/fairseq/models/speech_dlm/modules/__pycache__/speech_dlm_decoder.cpython-38.pyc differ diff --git a/data/fairseq/fairseq/models/speech_dlm/modules/__pycache__/speech_dlm_decoder_layer.cpython-38.pyc b/data/fairseq/fairseq/models/speech_dlm/modules/__pycache__/speech_dlm_decoder_layer.cpython-38.pyc new file mode 100644 index 0000000000000000000000000000000000000000..b9c1705adfda5a96fe31ffada8fa41f7d9e483de Binary files /dev/null and b/data/fairseq/fairseq/models/speech_dlm/modules/__pycache__/speech_dlm_decoder_layer.cpython-38.pyc differ diff --git a/data/fairseq/fairseq/models/speech_dlm/modules/speech_dlm_decoder.py b/data/fairseq/fairseq/models/speech_dlm/modules/speech_dlm_decoder.py new file mode 100644 index 0000000000000000000000000000000000000000..a14a1d64a87055666b017e885f0a370536f823f8 --- /dev/null +++ b/data/fairseq/fairseq/models/speech_dlm/modules/speech_dlm_decoder.py @@ -0,0 +1,572 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import math +from typing import Any, Dict, List, Optional, Tuple + +import torch +import torch.nn as nn +from fairseq import utils +from fairseq.models import FairseqIncrementalDecoder +from fairseq.modules import ( + FairseqDropout, + LayerDropModuleList, + LayerNorm, + PositionalEmbedding, +) +from .speech_dlm_decoder_layer import ( + CrossChannelTransformerDecoderLayer, + StandardTransformerDecoderLayer, +) +from fairseq.modules.checkpoint_activations import checkpoint_wrapper +from fairseq.modules.quant_noise import quant_noise as apply_quant_noise_ +from torch import Tensor + + +class CrossChannelTransformerDecoder(FairseqIncrementalDecoder): + """ + Cross-channel Transformer Decoder Block for parallel spoken dialogue units + as described in the paper: https://arxiv.org/pdf/2203.16502.pdf; + consisting of *args.decoder_layers* layers. Each layer is a + :class:`StandardTransformerDecoderLayer` or + :class:`CrossChannelTransformerDecoderLayer`. + + Args: + args (argparse.Namespace): parsed command-line arguments + dictionary (~fairseq.data.Dictionary): decoding dictionary + embed_tokens (torch.nn.Embedding): output embedding + channels (list): list of channel names (string) + no_encoder_attn (bool, optional): whether to attend to encoder outputs + (default: False). + """ + + def __init__(self, args, dictionary, embed_tokens, channels, no_encoder_attn=False): + self.args = args + super().__init__(dictionary) + self.register_buffer("version", torch.Tensor([3])) + self._future_mask = torch.empty(0) + + self.dropout_module = FairseqDropout( + args.dropout, module_name=self.__class__.__name__ + ) + self.decoder_layerdrop = args.decoder_layerdrop + self.share_input_output_embed = args.share_decoder_input_output_embed + self.channels = channels + + input_embed_dim = embed_tokens.embedding_dim + embed_dim = args.decoder_embed_dim + self.embed_dim = embed_dim + self.output_embed_dim = args.decoder_output_dim + + self.padding_idx = embed_tokens.padding_idx + self.max_target_positions = args.max_target_positions + + self.embed_tokens = embed_tokens + + self.embed_scale = 1.0 if args.no_scale_embedding else math.sqrt(embed_dim) + + if args.quant_noise_pq > 0: + self.quant_noise = apply_quant_noise_( + nn.Linear(embed_dim, embed_dim, bias=False), + args.quant_noise_pq, + args.quant_noise_pq_block_size, + ) + else: + self.quant_noise = None + + self.project_in_dim = ( + nn.Linear(input_embed_dim, embed_dim, bias=False) + if embed_dim != input_embed_dim + else None + ) + self.embed_positions = ( + PositionalEmbedding( + self.max_target_positions, + embed_dim, + self.padding_idx, + learned=args.decoder_learned_pos, + ) + if not args.no_token_positional_embeddings + else None + ) + + if getattr(args, "layernorm_embedding", False): + self.layernorm_embedding = LayerNorm(embed_dim) + else: + self.layernorm_embedding = None + + self.cross_self_attention = getattr(args, "cross_self_attention", False) + + assert 0 <= args.decoder_cross_layers <= args.decoder_layers, ( + "The number of cross-channel attention decoder layers must be non-negative" + f"and not exceeds the number of decoder layers (found {args.decoder_cross_layers})" + ) + + if self.decoder_layerdrop > 0.0: + self.layers = LayerDropModuleList(p=self.decoder_layerdrop) + else: + self.layers = nn.ModuleList([]) + self.layers.extend( + [ + self.build_decoder_layer(args, no_encoder_attn) + if i < args.decoder_layers - args.decoder_cross_layers + else self.build_cross_decoder_layer(args, no_encoder_attn) + for i in range(args.decoder_layers) + ] + ) + self.num_layers = len(self.layers) + self.non_cross_layers = args.decoder_layers - args.decoder_cross_layers + + if args.decoder_normalize_before and not getattr( + args, "no_decoder_final_norm", False + ): + self.layer_norm = LayerNorm(embed_dim) + else: + self.layer_norm = None + + self.project_out_dim = ( + nn.Linear(embed_dim, self.output_embed_dim, bias=False) + if embed_dim != self.output_embed_dim + else None + ) + + self.output_projection = None + self.is_cross_prediction = bool( + float(args.main_and_cross_weights.split(",")[1]) != 0 + ) + self.n_output_projections = ( + 1 if not self.is_cross_prediction else len(self.channels) + ) + + if self.share_input_output_embed: + # Output projection is a list of projections + # where the first proj is for the main-channel, + # then roll in a cicular way. + # For example: if the main channel has index i + # the second proj is for channel i+1 (mod N_channels), etc. + self.output_projection = nn.ModuleList( + [ + nn.Linear( + embed_tokens.weight.shape[1], # embed_dim + embed_tokens.weight.shape[0], # n_dictionaries + bias=False, + ) + for _ in range(self.n_output_projections) + ] + ) + # Only share the main-channel projection + self.output_projection[0].weight = embed_tokens.weight + for i in range(1, self.n_output_projections): + nn.init.normal_( + self.output_projection[i].weight, + mean=0, + std=embed_tokens.weight.shape[1] ** -0.5, + ) + else: + self.output_projection = nn.ModuleList( + [ + nn.Linear(self.output_embed_dim, len(dictionary), bias=False) + for _ in range(self.n_output_projections) + ] + ) + for i in range(self.n_output_projections): + nn.init.normal_( + self.output_projection[i].weight, + mean=0, + std=self.output_embed_dim**-0.5, + ) + self.output_duration_prediction = ( + None + if str(args.duration_prediction).lower() == "false" + else nn.ModuleList( + [ + nn.Linear(self.output_embed_dim, 1) + for _ in range(self.n_output_projections) + ] + ) + ) + + def build_decoder_layer(self, args, no_encoder_attn=False): + layer = StandardTransformerDecoderLayer(args, no_encoder_attn) + if getattr(args, "checkpoint_activations", False): + offload_to_cpu = getattr(args, "offload_activations", False) + layer = checkpoint_wrapper(layer, offload_to_cpu=offload_to_cpu) + return layer + + def build_cross_decoder_layer(self, args, no_encoder_attn=False): + layer = CrossChannelTransformerDecoderLayer(args, no_encoder_attn) + if getattr(args, "checkpoint_activations", False): + offload_to_cpu = getattr(args, "offload_activations", False) + layer = checkpoint_wrapper(layer, offload_to_cpu=offload_to_cpu) + return layer + + def forward( + self, + prev_output_tokens: Dict[str, Tensor], + encoder_out: Optional[Dict[str, List[Tensor]]] = None, + incremental_state: Optional[ + List[Dict[str, Dict[str, Optional[Tensor]]]] + ] = None, + features_only: bool = False, + full_context_alignment: bool = False, + alignment_layer: Optional[int] = None, + alignment_heads: Optional[int] = None, + src_lengths: Optional[Any] = None, + # return_all_hiddens: bool = False, + ): + """ + Args: + prev_output_tokens (dict[str, LongTensor]): previous decoder outputs, + dictionary over all channels with the values being the tensors + of shape `(batch, tgt_len)`, for teacher forcing + encoder_out (optional): output from the encoder, used for + encoder-side attention + incremental_state (dict): list of dictionaries used for storing state + during :ref:`Incremental decoding` + features_only (bool, optional): only return features without + applying output layer (default: False). + full_context_alignment (bool, optional): don't apply + auto-regressive mask to self-attention (default: False). + + Returns: + tuple: + - the decoder's output, dict over channels of tensors + of shape `(batch, tgt_len, vocab)` + - a dictionary with any model-specific outputs + """ + x, extra = self.extract_features( + prev_output_tokens, + encoder_out=encoder_out, + incremental_state=incremental_state, + full_context_alignment=full_context_alignment, + alignment_layer=alignment_layer, + alignment_heads=alignment_heads, + ) + if not features_only: + x = self.output_layer(x) + return x, extra + + def extract_features( + self, + prev_output_tokens: Dict[str, Tensor], + encoder_out: Optional[Dict[str, List[Tensor]]], + incremental_state: Optional[ + List[Dict[str, Dict[str, Optional[Tensor]]]] + ] = None, + full_context_alignment: bool = False, + alignment_layer: Optional[int] = None, + alignment_heads: Optional[int] = None, + ): + return self.extract_features_scriptable( + prev_output_tokens, + encoder_out, + incremental_state, + full_context_alignment, + alignment_layer, + alignment_heads, + ) + + """ + A scriptable subclass of this class has an extract_features method and calls + super().extract_features, but super() is not supported in torchscript. A copy of + this function is made to be used in the subclass instead. + """ + + def extract_features_scriptable( + self, + prev_output_tokens: Dict[str, Tensor], + encoder_out: Optional[Dict[str, List[Tensor]]], + incremental_state: Optional[ + List[Dict[str, Dict[str, Optional[Tensor]]]] + ] = None, + full_context_alignment: bool = False, + alignment_layer: Optional[int] = None, + alignment_heads: Optional[int] = None, + ): + """ + The core function of *forward* but only return features. + + The input (prev_output_tokens) is a dictionary over all channels, + expected to have the following form: + { + 'channel1' : Tensor((batch x tgt_len)), + 'channel2' : Tensor((batch x tgt_len)), + } + + Args: + full_context_alignment (bool, optional): don't apply + auto-regressive mask to self-attention (default: False). + alignment_layer (int, optional): return mean alignment over + heads at this layer (default: last layer). + alignment_heads (int, optional): only average alignment over + this many heads (default: all heads). + + Returns: + tuple: + - the decoder's features, dict over channels of tensors + of shape `(batch, tgt_len, embed_dim)` + - a dictionary with any model-specific outputs + """ + if alignment_layer is None: + alignment_layer = self.num_layers - 1 + + x_list = [] + for i, channel in enumerate(self.channels): + # embed positions + positions = None + if self.embed_positions is not None: + positions = self.embed_positions( + prev_output_tokens[channel], + incremental_state=incremental_state[i] + if incremental_state is not None + else None, + ) + + if incremental_state is not None: + prev_output_tokens[channel] = prev_output_tokens[channel][:, -1:] + if positions is not None: + positions = positions[:, -1:] + + # embed tokens and positions + x = self.embed_tokens(prev_output_tokens[channel]) + + if self.project_in_dim is not None: + x = self.project_in_dim(x) + + x = self.embed_scale * x + + if self.quant_noise is not None: + x = self.quant_noise(x) + + if positions is not None: + x += positions + + if self.layernorm_embedding is not None: + x = self.layernorm_embedding(x) + + x = self.dropout_module(x) + + # B x T x C -> T x B x C + x = x.transpose(0, 1) + + x_list.append(x) + + self_attn_padding_mask: Optional[Tensor] = None + if ( + self.cross_self_attention + or prev_output_tokens[self.channels[0]].eq(self.padding_idx).any() + ): + self_attn_padding_mask = prev_output_tokens[self.channels[0]].eq( + self.padding_idx + ) + + # decoder layers + attn: Optional[Dict[Tensor]] = None + inner_states: List[Optional[Dict[str, Tensor]]] = [ + {channel: x_list[i] for i, channel in enumerate(self.channels)} + ] + for idx, layer in enumerate(self.layers): + if incremental_state is None and not full_context_alignment: + self_attn_mask = self.buffered_future_mask(x_list[0]) + else: + self_attn_mask = None + + # need to change to tensor for the checkpoint activation to work + if isinstance(x_list, list): + x_list = torch.stack(x_list) + x_list, layer_attn_list, _ = layer( + x_list, + encoder_out["encoder_out"][0] + if (encoder_out is not None and len(encoder_out["encoder_out"]) > 0) + else None, + encoder_out["encoder_padding_mask"][0] + if ( + encoder_out is not None + and len(encoder_out["encoder_padding_mask"]) > 0 + ) + else None, + incremental_state, + self_attn_mask=self_attn_mask, + self_attn_padding_mask=self_attn_padding_mask, + need_attn=bool((idx == alignment_layer)), + need_head_weights=bool((idx == alignment_layer)), + ) + + inner_states.append( + {channel: x_list[i] for i, channel in enumerate(self.channels)} + ) + if idx == alignment_layer and all( + layer_attn is not None for layer_attn in layer_attn_list + ): + attn = { + channel: layer_attn_list[i].float().to(x_list[0]) + for i, channel in enumerate(self.channels) + } + # change back from tensor to list + if not isinstance(x_list, list): + x_list = list(torch.unbind(x_list)) + + if attn is not None: + for channel in attn: + if alignment_heads is not None: + attn[channel] = attn[channel][:alignment_heads] + + # average probabilities over heads + attn[channel] = attn[channel].mean(dim=0) + + for i, x in enumerate(x_list): + if self.layer_norm is not None: + x = self.layer_norm(x) + + # T x B x C -> B x T x C + x = x.transpose(0, 1) + + if self.project_out_dim is not None: + x = self.project_out_dim(x) + + x_list[i] = x + + x = {channel: x_list[i] for i, channel in enumerate(self.channels)} + + return x, {"attn": [attn], "inner_states": inner_states} + + def output_layer(self, features): + """Project features to the vocabulary size. + Return a dictionary of the form: + { + 'input-channel': { + 'predicted-channel': token prediction tensor of shape `(batch, tgt_len, vocab)`, + } + } + + if duration_prediction is enabled + { + 'input-channel': { + 'predicted-channel': { + 'pred_token': token prediction tensor of shape `(batch, tgt_len, vocab)`, + 'pred_duration': duration prediction tensor + } + } + } + """ + # project back to size of vocabulary + if self.output_duration_prediction is None: + if self.is_cross_prediction: + return { + channel: { + pred_channel: self.output_projection[j - i](features[channel]) + for j, pred_channel in enumerate(self.channels) + } + for i, channel in enumerate(self.channels) + } + else: + return { + channel: {channel: self.output_projection[0](features[channel])} + for i, channel in enumerate(self.channels) + } + else: + if self.is_cross_prediction: + return { + channel: { + pred_channel: { + "pred_token": self.output_projection[j - i]( + features[channel] + ), + "pred_duration": self.output_duration_prediction[j - i]( + features[channel] + ), + } + for j, pred_channel in enumerate(self.channels) + } + for i, channel in enumerate(self.channels) + } + else: + return { + channel: { + channel: { + "pred_token": self.output_projection[0](features[channel]), + "pred_duration": self.output_duration_prediction[0]( + features[channel] + ), + } + } + for i, channel in enumerate(self.channels) + } + + def max_positions(self): + """Maximum output length supported by the decoder.""" + if self.embed_positions is None: + return self.max_target_positions + return min(self.max_target_positions, self.embed_positions.max_positions) + + def buffered_future_mask(self, tensor): + dim = tensor.size(0) + # self._future_mask.device != tensor.device is not working in TorchScript. This is a workaround. + if ( + self._future_mask.size(0) == 0 + or (not self._future_mask.device == tensor.device) + or self._future_mask.size(0) < dim + ): + self._future_mask = torch.triu( + utils.fill_with_neg_inf(torch.zeros([dim, dim])), 1 + ) + self._future_mask = self._future_mask.to(tensor) + return self._future_mask[:dim, :dim] + + def get_normalized_probs_scriptable( + self, + net_output: Tuple[Tensor, Optional[Dict[str, List[Optional[Tensor]]]]], + log_probs: bool, + sample: Optional[Dict[str, Tensor]] = None, + ): + """Get normalized probabilities (or log probs) from a net's output.""" + + logits_dict = net_output[0] + out_dict = {} + for channel in logits_dict: + out_dict[channel] = {} + for pred_channel in logits_dict[channel]: + if isinstance(logits_dict[channel][pred_channel], dict): + pred_token_logits = logits_dict[channel][pred_channel]["pred_token"] + else: + pred_token_logits = logits_dict[channel][pred_channel] + if log_probs: + out = utils.log_softmax( + pred_token_logits, dim=-1, onnx_trace=self.onnx_trace + ) + else: + out = utils.softmax( + pred_token_logits, dim=-1, onnx_trace=self.onnx_trace + ) + if isinstance(logits_dict[channel][pred_channel], dict): + out_dict[channel][pred_channel] = { + "pred_token": out, + "pred_duration": logits_dict[channel][pred_channel][ + "pred_duration" + ].float(), + } # move to float32 to avoid inf loss + else: + out_dict[channel][pred_channel] = out + return out_dict + + def reorder_incremental_state_scripting( + self, + incremental_state: List[Dict[str, Dict[str, Optional[Tensor]]]], + new_order: Tensor, + ): + """Main entry point for reordering the incremental state. + + Due to limitations in TorchScript, we call this function in + :class:`fairseq.sequence_generator.SequenceGenerator` instead of + calling :func:`reorder_incremental_state` directly. + """ + for module in self.modules(): + if hasattr(module, "reorder_incremental_state"): + for i, incremental_state_channel in enumerate(incremental_state): + result = module.reorder_incremental_state( + incremental_state_channel, new_order + ) + if result is not None: + incremental_state[i] = result diff --git a/data/fairseq/fairseq/models/speech_dlm/modules/speech_dlm_decoder_layer.py b/data/fairseq/fairseq/models/speech_dlm/modules/speech_dlm_decoder_layer.py new file mode 100644 index 0000000000000000000000000000000000000000..fb65fdf810d613b7d8615d3cbe56b633833367f9 --- /dev/null +++ b/data/fairseq/fairseq/models/speech_dlm/modules/speech_dlm_decoder_layer.py @@ -0,0 +1,717 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from typing import Dict, List, Tuple, Optional + +import torch +import torch.nn as nn +from fairseq import utils +from fairseq.modules import LayerNorm, MultiheadAttention +from fairseq.modules.fairseq_dropout import FairseqDropout +from fairseq.modules.quant_noise import quant_noise +from torch import Tensor + + +class CrossChannelTransformerDecoderLayer(nn.Module): + """Cross-Attention Transformer Decoder Layer block as described + in the paper: https://arxiv.org/pdf/2203.16502.pdf + + Composed of a Multi-head Self Attention block followed by a + Multi-head Cross-Attention block which attends to the self-attention + outputs of the other channels. The weights of the attention blocks + in all channels are shared. + + Args: + args (argparse.Namespace): parsed command-line arguments + no_encoder_attn (bool, optional): whether to attend to encoder outputs + (default: False). + """ + + def __init__( + self, args, no_encoder_attn=False, add_bias_kv=False, add_zero_attn=False + ): + super().__init__() + self.embed_dim = args.decoder_embed_dim + self.dropout_module = FairseqDropout( + args.dropout, module_name=self.__class__.__name__ + ) + self.quant_noise = getattr(args, "quant_noise_pq", 0) + self.quant_noise_block_size = getattr(args, "quant_noise_pq_block_size", 8) + + # This cross_self_attention is used for encoder-decoder systems, + # It's not the cross-channel attention (defined below as cross_channel_attn) + self.cross_self_attention = getattr(args, "cross_self_attention", False) + + self.self_attn = self.build_self_attention( + self.embed_dim, + args, + add_bias_kv=add_bias_kv, + add_zero_attn=add_zero_attn, + ) + self.cross_channel_attn = self.build_cross_channel_attention( + self.embed_dim, + args, + add_bias_kv=add_bias_kv, + add_zero_attn=add_zero_attn, + ) + + self.activation_fn = utils.get_activation_fn( + activation=str(args.activation_fn) + if getattr(args, "activation_fn", None) is not None + else "relu" + ) + activation_dropout_p = getattr(args, "activation_dropout", 0) or 0 + if activation_dropout_p == 0: + # for backwards compatibility with models that use args.relu_dropout + activation_dropout_p = getattr(args, "relu_dropout", 0) or 0 + self.activation_dropout_module = FairseqDropout( + float(activation_dropout_p), module_name=self.__class__.__name__ + ) + self.normalize_before = args.decoder_normalize_before + + # use layerNorm rather than FusedLayerNorm for exporting. + # char_inputs can be used to determint this. + # TODO remove this once we update apex with the fix + export = getattr(args, "char_inputs", False) + self.self_attn_layer_norm = LayerNorm(self.embed_dim, export=export) + self.cross_channel_attn_layer_norm = LayerNorm(self.embed_dim, export=export) + + if no_encoder_attn: + self.encoder_attn = None + self.encoder_attn_layer_norm = None + else: + self.encoder_attn = self.build_encoder_attention(self.embed_dim, args) + self.encoder_attn_layer_norm = LayerNorm(self.embed_dim, export=export) + + self.fc1 = self.build_fc1( + self.embed_dim, + args.decoder_ffn_embed_dim, + self.quant_noise, + self.quant_noise_block_size, + ) + self.fc2 = self.build_fc2( + args.decoder_ffn_embed_dim, + self.embed_dim, + self.quant_noise, + self.quant_noise_block_size, + ) + + self.final_layer_norm = LayerNorm(self.embed_dim, export=export) + self.need_attn = True + + self.onnx_trace = False + + def build_fc1(self, input_dim, output_dim, q_noise, qn_block_size): + return quant_noise(nn.Linear(input_dim, output_dim), q_noise, qn_block_size) + + def build_fc2(self, input_dim, output_dim, q_noise, qn_block_size): + return quant_noise(nn.Linear(input_dim, output_dim), q_noise, qn_block_size) + + def build_self_attention( + self, embed_dim, args, add_bias_kv=False, add_zero_attn=False + ): + return MultiheadAttention( + embed_dim, + args.decoder_attention_heads, + dropout=args.attention_dropout, + add_bias_kv=add_bias_kv, + add_zero_attn=add_zero_attn, + self_attention=not getattr(args, "cross_self_attention", False), + q_noise=self.quant_noise, + qn_block_size=self.quant_noise_block_size, + ) + + def build_cross_channel_attention( + self, embed_dim, args, add_bias_kv=False, add_zero_attn=False + ): + return MultiheadAttention( + embed_dim, + args.decoder_attention_heads, + dropout=args.attention_dropout, + add_bias_kv=add_bias_kv, + add_zero_attn=add_zero_attn, + self_attention=False, + q_noise=self.quant_noise, + qn_block_size=self.quant_noise_block_size, + ) + + def build_encoder_attention(self, embed_dim, args): + return MultiheadAttention( + embed_dim, + args.decoder_attention_heads, + kdim=getattr(args, "encoder_embed_dim", None), + vdim=getattr(args, "encoder_embed_dim", None), + dropout=args.attention_dropout, + encoder_decoder_attention=True, + q_noise=self.quant_noise, + qn_block_size=self.quant_noise_block_size, + ) + + def prepare_for_onnx_export_(self): + self.onnx_trace = True + + def residual_connection(self, x, residual): + return residual + x + + def forward( + self, + x_list_tensor: List[torch.Tensor], + encoder_out: Optional[torch.Tensor] = None, + encoder_padding_mask: Optional[torch.Tensor] = None, + incremental_state: Optional[ + List[Dict[str, Dict[str, Optional[Tensor]]]] + ] = None, + prev_self_attn_state: Optional[List[Tuple[torch.Tensor, torch.Tensor]]] = None, + prev_attn_state: Optional[List[torch.Tensor]] = None, + self_attn_mask: Optional[torch.Tensor] = None, + self_attn_padding_mask: Optional[torch.Tensor] = None, + need_attn: bool = False, + need_head_weights: bool = False, + ): + """ + Args: + x_list_tensor (List[Tensor]): list of input tensors in different channels, + each tensor is of shape `(seq_len, batch, embed_dim)` + encoder_padding_mask (ByteTensor, optional): binary + ByteTensor of shape `(batch, src_len)` where padding + elements are indicated by ``1``. + incremental_state (optional): list of incremental_state dictionaries over + different channels (sequence generation mode) + prev_self_attn_state (List[Tuple[Tensor, Tensor]], optional): list of tuples + (self_attn_state, cross_channel_attn_state) over different channels + need_attn (bool, optional): return attention weights + need_head_weights (bool, optional): return attention weights + for each head (default: return average over heads). + + Returns: + list of encoded output of shape `(seq_len, batch, embed_dim)` + """ + n_channels = len(x_list_tensor) + if need_head_weights: + need_attn = True + + # incremental_state is a list of dictionaries over different channels + if incremental_state is not None: + assert isinstance(incremental_state, list) + assert len(incremental_state) == n_channels + + # prev_self_attn_state is a list of tuples (self_attn_state, cross_channel_attn_state) over different channels + if prev_self_attn_state is not None: + assert isinstance(prev_self_attn_state, list) + assert len(prev_self_attn_state) == n_channels + for prev_self_attn_state_channel in prev_self_attn_state: + assert isinstance(prev_self_attn_state_channel, tuple) + assert len(prev_self_attn_state_channel) == 2 + + # Backup for other channels & cross channel attention + self_attn_mask_orin = self_attn_mask + self_attn_padding_mask_orin = self_attn_padding_mask + + x_list = [] + attn_list = [] + for i, x in enumerate(x_list_tensor): + residual = x + + if self.normalize_before: + x = self.self_attn_layer_norm(x) + + if prev_self_attn_state is not None: + prev_key, prev_value = prev_self_attn_state[i][0][:2] + saved_state: Dict[str, Optional[Tensor]] = { + "prev_key": prev_key, + "prev_value": prev_value, + } + if len(prev_self_attn_state[i][0]) >= 3: + saved_state["prev_key_padding_mask"] = prev_self_attn_state[i][0][2] + assert incremental_state is not None + self.self_attn._set_input_buffer(incremental_state[i], saved_state) + _self_attn_input_buffer = self.self_attn._get_input_buffer( + incremental_state[i] if incremental_state is not None else None + ) + if self.cross_self_attention and not ( + incremental_state is not None + and _self_attn_input_buffer is not None + and "prev_key" in _self_attn_input_buffer + ): + if self_attn_mask_orin is not None: + assert encoder_out is not None + self_attn_mask = torch.cat( + ( + x.new_zeros(x.size(0), encoder_out.size(0)), + self_attn_mask_orin, + ), + dim=1, + ) + if self_attn_padding_mask_orin is not None: + if encoder_padding_mask is None: + assert encoder_out is not None + encoder_padding_mask = self_attn_padding_mask_orin.new_zeros( + encoder_out.size(1), encoder_out.size(0) + ) + self_attn_padding_mask = torch.cat( + (encoder_padding_mask, self_attn_padding_mask_orin), dim=1 + ) + assert encoder_out is not None + y = torch.cat((encoder_out, x), dim=0) + else: + y = x + + x, attn = self.self_attn( + query=x, + key=y, + value=y, + key_padding_mask=self_attn_padding_mask, + incremental_state=incremental_state[i] + if incremental_state is not None + else None, + need_weights=False, + attn_mask=self_attn_mask, + ) + + x = self.dropout_module(x) + x = self.residual_connection(x, residual) + if not self.normalize_before: + x = self.self_attn_layer_norm(x) + + if self.encoder_attn is not None and encoder_out is not None: + residual = x + if self.normalize_before: + x = self.encoder_attn_layer_norm(x) + if prev_attn_state is not None: + prev_key, prev_value = prev_attn_state[:2] + saved_state: Dict[str, Optional[Tensor]] = { + "prev_key": prev_key, + "prev_value": prev_value, + } + if len(prev_attn_state) >= 3: + saved_state["prev_key_padding_mask"] = prev_attn_state[2] + assert incremental_state is not None + self.encoder_attn._set_input_buffer( + incremental_state[i], saved_state + ) + + x, attn = self.encoder_attn( + query=x, + key=encoder_out, + value=encoder_out, + key_padding_mask=encoder_padding_mask, + incremental_state=incremental_state[i] + if incremental_state is not None + else None, + static_kv=True, + need_weights=need_attn or (not self.training and self.need_attn), + need_head_weights=need_head_weights, + ) + x = self.dropout_module(x) + x = self.residual_connection(x, residual) + if not self.normalize_before: + x = self.encoder_attn_layer_norm(x) + + x_list.append(x) + attn_list.append(attn) + + # Store attentions & new x(s) (bc the old x(s) are used in other channels) + x_list_new = [] + # Here comes the cross channel attention + for i, x in enumerate(x_list): + residual = x + if self.normalize_before: + x = self.cross_channel_attn_layer_norm(x) + + if prev_self_attn_state is not None: + prev_key, prev_value = prev_self_attn_state[i][1][:2] + saved_state: Dict[str, Optional[Tensor]] = { + "prev_key": prev_key, + "prev_value": prev_value, + } + if len(prev_self_attn_state[i][1]) >= 3: + saved_state["prev_key_padding_mask"] = prev_self_attn_state[i][1][2] + assert incremental_state is not None + self.cross_channel_attn._set_input_buffer( + incremental_state[i], saved_state + ) + + # The cross attention is computed with the concatenation of attentions from other channels + if len(x_list) > 1: + x_other = torch.cat( + [x_list[(i + j) % len(x_list)] for j in range(1, len(x_list))], + dim=0, + ) + else: + # Self-attention when having only one channel + x_other = x_list[i] + + x, attn = self.cross_channel_attn( + query=x, + key=x_other, + value=x_other, + key_padding_mask=self_attn_padding_mask_orin, + incremental_state=incremental_state[i] + if incremental_state is not None + else None, + need_weights=False, + attn_mask=self_attn_mask_orin, + ) + + x = self.dropout_module(x) + x = self.residual_connection(x, residual) + if not self.normalize_before: + x = self.cross_channel_attn_layer_norm(x) + + x_list_new.append(x) + x_list = x_list_new + + for i, x in enumerate(x_list): + residual = x + if self.normalize_before: + x = self.final_layer_norm(x) + + x = self.activation_fn(self.fc1(x)) + x = self.activation_dropout_module(x) + x = self.fc2(x) + x = self.dropout_module(x) + x = self.residual_connection(x, residual) + if not self.normalize_before: + x = self.final_layer_norm(x) + + x_list[i] = x + # Trick for the checkpoint activation + x_list_tensor = torch.stack(x_list) + if self.onnx_trace and incremental_state is not None: + self_and_cross_attn_state_list = [] + for i in range(n_channels): + self_and_cross_attn_state = [] + for self_attn_module in [self.self_attn, self.cross_channel_attn]: + saved_state = self_attn_module._get_input_buffer( + incremental_state[i] + ) + assert saved_state is not None + if self_attn_padding_mask is not None: + self_attn_module_state = [ + saved_state["prev_key"], + saved_state["prev_value"], + saved_state["prev_key_padding_mask"], + ] + else: + self_attn_module_state = [ + saved_state["prev_key"], + saved_state["prev_value"], + ] + self_and_cross_attn_state.append(self_attn_module_state) + self_and_cross_attn_state_list.append(tuple(self_and_cross_attn_state)) + return x_list_tensor, attn_list, self_and_cross_attn_state_list + return x_list_tensor, attn_list, None + + def make_generation_fast_(self, need_attn: bool = False, **kwargs): + self.need_attn = need_attn + + +# Rewrite fairseq.modules.TransformerDecoderLayer +# to be compatible with checkpoint_activations +# (avoid forwarding model multiple times) +class StandardTransformerDecoderLayer(nn.Module): + """Rewrite fairseq.modules.TransformerDecoderLayer to avoid forwarding + model multiple times and be compatible with checkpoint_activations. + + The input is expected to be a list of tensors from different channels, + each is forwarded to the same model (shared attention weights). + + In the original paper each operation (multi-head attention, encoder + attention or FFN) is postprocessed with: `dropout -> add residual -> + layernorm`. In the tensor2tensor code they suggest that learning is more + robust when preprocessing each layer with layernorm and postprocessing with: + `dropout -> add residual`. We default to the approach in the paper, but the + tensor2tensor approach can be enabled by setting + *args.decoder_normalize_before* to ``True``. + + Args: + args (argparse.Namespace): parsed command-line arguments + no_encoder_attn (bool, optional): whether to attend to encoder outputs + (default: False). + """ + + def __init__( + self, args, no_encoder_attn=False, add_bias_kv=False, add_zero_attn=False + ): + super().__init__() + self.embed_dim = args.decoder_embed_dim + self.dropout_module = FairseqDropout( + args.dropout, module_name=self.__class__.__name__ + ) + self.quant_noise = getattr(args, "quant_noise_pq", 0) + self.quant_noise_block_size = getattr(args, "quant_noise_pq_block_size", 8) + + self.cross_self_attention = getattr(args, "cross_self_attention", False) + + self.self_attn = self.build_self_attention( + self.embed_dim, + args, + add_bias_kv=add_bias_kv, + add_zero_attn=add_zero_attn, + ) + + self.activation_fn = utils.get_activation_fn( + activation=str(args.activation_fn) + if getattr(args, "activation_fn", None) is not None + else "relu" + ) + activation_dropout_p = getattr(args, "activation_dropout", 0) or 0 + if activation_dropout_p == 0: + # for backwards compatibility with models that use args.relu_dropout + activation_dropout_p = getattr(args, "relu_dropout", 0) or 0 + self.activation_dropout_module = FairseqDropout( + float(activation_dropout_p), module_name=self.__class__.__name__ + ) + self.normalize_before = args.decoder_normalize_before + + # use layerNorm rather than FusedLayerNorm for exporting. + # char_inputs can be used to determint this. + # TODO remove this once we update apex with the fix + export = getattr(args, "char_inputs", False) + self.self_attn_layer_norm = LayerNorm(self.embed_dim, export=export) + + if no_encoder_attn: + self.encoder_attn = None + self.encoder_attn_layer_norm = None + else: + self.encoder_attn = self.build_encoder_attention(self.embed_dim, args) + self.encoder_attn_layer_norm = LayerNorm(self.embed_dim, export=export) + + self.fc1 = self.build_fc1( + self.embed_dim, + args.decoder_ffn_embed_dim, + self.quant_noise, + self.quant_noise_block_size, + ) + self.fc2 = self.build_fc2( + args.decoder_ffn_embed_dim, + self.embed_dim, + self.quant_noise, + self.quant_noise_block_size, + ) + + self.final_layer_norm = LayerNorm(self.embed_dim, export=export) + self.need_attn = True + + self.onnx_trace = False + + def build_fc1(self, input_dim, output_dim, q_noise, qn_block_size): + return quant_noise(nn.Linear(input_dim, output_dim), q_noise, qn_block_size) + + def build_fc2(self, input_dim, output_dim, q_noise, qn_block_size): + return quant_noise(nn.Linear(input_dim, output_dim), q_noise, qn_block_size) + + def build_self_attention( + self, embed_dim, args, add_bias_kv=False, add_zero_attn=False + ): + return MultiheadAttention( + embed_dim, + args.decoder_attention_heads, + dropout=args.attention_dropout, + add_bias_kv=add_bias_kv, + add_zero_attn=add_zero_attn, + self_attention=not getattr(args, "cross_self_attention", False), + q_noise=self.quant_noise, + qn_block_size=self.quant_noise_block_size, + ) + + def build_encoder_attention(self, embed_dim, args): + return MultiheadAttention( + embed_dim, + args.decoder_attention_heads, + kdim=getattr(args, "encoder_embed_dim", None), + vdim=getattr(args, "encoder_embed_dim", None), + dropout=args.attention_dropout, + encoder_decoder_attention=True, + q_noise=self.quant_noise, + qn_block_size=self.quant_noise_block_size, + ) + + def prepare_for_onnx_export_(self): + self.onnx_trace = True + + def residual_connection(self, x, residual): + return residual + x + + def forward( + self, + x_list_tensor: List[torch.Tensor], + encoder_out: Optional[torch.Tensor] = None, + encoder_padding_mask: Optional[torch.Tensor] = None, + incremental_state: Optional[ + List[Dict[str, Dict[str, Optional[Tensor]]]] + ] = None, + prev_self_attn_state: Optional[List[Tuple[torch.Tensor, torch.Tensor]]] = None, + prev_attn_state: Optional[List[torch.Tensor]] = None, + self_attn_mask: Optional[torch.Tensor] = None, + self_attn_padding_mask: Optional[torch.Tensor] = None, + need_attn: bool = False, + need_head_weights: bool = False, + ): + """ + Args: + x_list_tensor (List[Tensor]): list of input tensors in different channels, + each tensor is of shape `(seq_len, batch, embed_dim)` + encoder_padding_mask (ByteTensor, optional): binary + ByteTensor of shape `(batch, src_len)` where padding + elements are indicated by ``1``. + incremental_state (optional): list of incremental_state dictionaries over + different channels (sequence generation mode) + prev_self_attn_state (List[Tuple[Tensor, Tensor]], optional): list of tuples + (self_attn_state, cross_channel_attn_state) over different channels + need_attn (bool, optional): return attention weights + need_head_weights (bool, optional): return attention weights + for each head (default: return average over heads). + + Returns: + list of encoded output of shape `(seq_len, batch, embed_dim)` + """ + n_channels = len(x_list_tensor) + if need_head_weights: + need_attn = True + + # incremental_state is a list of dictionaries over different channels + if incremental_state is not None: + assert isinstance(incremental_state, list) + assert len(incremental_state) == n_channels + + # prev_self_attn_state is a list of self_attn_state over different channels + if prev_self_attn_state is not None: + assert isinstance(prev_self_attn_state, list) + assert len(prev_self_attn_state) == n_channels + + x_list = [] + attn_list = [] + for i, x in enumerate(x_list_tensor): + residual = x + + if self.normalize_before: + x = self.self_attn_layer_norm(x) + + if prev_self_attn_state is not None: + prev_key, prev_value = prev_self_attn_state[i][:2] + saved_state: Dict[str, Optional[Tensor]] = { + "prev_key": prev_key, + "prev_value": prev_value, + } + if len(prev_self_attn_state[i]) >= 3: + saved_state["prev_key_padding_mask"] = prev_self_attn_state[2] + assert incremental_state is not None + self.self_attn._set_input_buffer(incremental_state[i], saved_state) + _self_attn_input_buffer = self.self_attn._get_input_buffer( + incremental_state + ) + if self.cross_self_attention and not ( + incremental_state is not None + and _self_attn_input_buffer is not None + and "prev_key" in _self_attn_input_buffer + ): + if self_attn_mask is not None: + assert encoder_out is not None + self_attn_mask = torch.cat( + (x.new_zeros(x.size(0), encoder_out.size(0)), self_attn_mask), + dim=1, + ) + if self_attn_padding_mask is not None: + if encoder_padding_mask is None: + assert encoder_out is not None + encoder_padding_mask = self_attn_padding_mask.new_zeros( + encoder_out.size(1), encoder_out.size(0) + ) + self_attn_padding_mask = torch.cat( + (encoder_padding_mask, self_attn_padding_mask), dim=1 + ) + assert encoder_out is not None + y = torch.cat((encoder_out, x), dim=0) + else: + y = x + + x, attn = self.self_attn( + query=x, + key=y, + value=y, + key_padding_mask=self_attn_padding_mask, + incremental_state=incremental_state[i] + if incremental_state is not None + else None, + need_weights=False, + attn_mask=self_attn_mask, + ) + x = self.dropout_module(x) + x = self.residual_connection(x, residual) + if not self.normalize_before: + x = self.self_attn_layer_norm(x) + + if self.encoder_attn is not None and encoder_out is not None: + residual = x + if self.normalize_before: + x = self.encoder_attn_layer_norm(x) + if prev_attn_state is not None: + prev_key, prev_value = prev_attn_state[:2] + saved_state: Dict[str, Optional[Tensor]] = { + "prev_key": prev_key, + "prev_value": prev_value, + } + if len(prev_attn_state) >= 3: + saved_state["prev_key_padding_mask"] = prev_attn_state[2] + assert incremental_state is not None + self.encoder_attn._set_input_buffer(incremental_state, saved_state) + + x, attn = self.encoder_attn( + query=x, + key=encoder_out, + value=encoder_out, + key_padding_mask=encoder_padding_mask, + incremental_state=incremental_state[i] + if incremental_state is not None + else None, + static_kv=True, + need_weights=need_attn or (not self.training and self.need_attn), + need_head_weights=need_head_weights, + ) + x = self.dropout_module(x) + x = self.residual_connection(x, residual) + if not self.normalize_before: + x = self.encoder_attn_layer_norm(x) + + residual = x + if self.normalize_before: + x = self.final_layer_norm(x) + + x = self.activation_fn(self.fc1(x)) + x = self.activation_dropout_module(x) + x = self.fc2(x) + x = self.dropout_module(x) + x = self.residual_connection(x, residual) + if not self.normalize_before: + x = self.final_layer_norm(x) + + x_list.append(x) + attn_list.append(attn) + + # Trick for the checkpoint activation + x_list_tensor = torch.stack(x_list) + if self.onnx_trace and incremental_state is not None: + self_attn_state_list = [] + for i in range(n_channels): + saved_state = self.self_attn._get_input_buffer(incremental_state[i]) + assert saved_state is not None + if self_attn_padding_mask is not None: + self_attn_state = [ + saved_state["prev_key"], + saved_state["prev_value"], + saved_state["prev_key_padding_mask"], + ] + else: + self_attn_state = [ + saved_state["prev_key"], + saved_state["prev_value"], + ] + self_attn_state_list.append(self_attn_state) + return x_list_tensor, attn_list, self_attn_state_list + return x_list_tensor, attn_list, None + + def make_generation_fast_(self, need_attn: bool = False, **kwargs): + self.need_attn = need_attn diff --git a/data/fairseq/fairseq/models/speech_dlm/sequence_generator/__init__.py b/data/fairseq/fairseq/models/speech_dlm/sequence_generator/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..a88e14457766bf30c916926f171b81fa60dd33ce --- /dev/null +++ b/data/fairseq/fairseq/models/speech_dlm/sequence_generator/__init__.py @@ -0,0 +1,6 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from .multichannel_sequence_generator import * # noqa diff --git a/data/fairseq/fairseq/models/speech_dlm/sequence_generator/multichannel_search.py b/data/fairseq/fairseq/models/speech_dlm/sequence_generator/multichannel_search.py new file mode 100644 index 0000000000000000000000000000000000000000..db4b77f3457ca2da089772e2c7e3df850c87442f --- /dev/null +++ b/data/fairseq/fairseq/models/speech_dlm/sequence_generator/multichannel_search.py @@ -0,0 +1,430 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from typing import Dict, Optional + +import torch +import torch.nn as nn +from torch import Tensor + + +class MultichannelSearch(nn.Module): + def __init__(self, tgt_dicts): + super().__init__() + tgt_dict = list(tgt_dicts.values())[0] + self.pad = tgt_dict.pad() + self.unk = tgt_dict.unk() + self.eos = tgt_dict.eos() + for tgt_dict in tgt_dicts.values(): + assert self.pad == tgt_dict.pad() + assert self.unk == tgt_dict.unk() + assert self.eos == tgt_dict.eos() + self.vocab_sizes = {channel: len(tgt_dicts[channel]) for channel in tgt_dicts} + self.src_lengths = torch.tensor(-1) + self.supports_constraints = False + self.stop_on_max_len = False + + def step( + self, step, lprobs, scores, prev_output_tokens=None, original_batch_idxs=None + ): + """Take a single search step. + + Args: + step: the current search step, starting at 0 + lprobs: dictionary of channels {channel : (bsz x input_beam_size x vocab_size_channel)} + the model's log-probabilities over the vocabulary at the current step + scores: {channel : (bsz x input_beam_size x step)} + the historical model scores of each hypothesis up to this point + prev_output_tokens: {channel : (bsz x step)} + the previously generated oputput tokens + original_batch_idxs: (bsz) + the tensor with the batch indices, in the range [0, bsz) + this is useful in case there has been applied a re-ordering + and we need to know the orignal indices + + Return: A tuple of (scores, indices, beams) where: + scores: {channel : (bsz x output_beam_size)} + the scores of the chosen elements; output_beam_size can be + larger than input_beam_size, e.g., we may return + 2*input_beam_size to account for EOS + indices: {channel : (bsz x output_beam_size)} + the indices of the chosen elements + beams: (bsz x output_beam_size) + the hypothesis ids of the chosen elements, in the range [0, input_beam_size) + """ + raise NotImplementedError + + @torch.jit.export + def set_src_lengths(self, src_lengths): + self.src_lengths = src_lengths + + @torch.jit.export + def init_constraints(self, batch_constraints: Optional[Tensor], beam_size: int): + """Initialize constraint states for constrained decoding (if supported). + + Args: + batch_constraints: (torch.Tensor, optional) + the list of constraints, in packed form + beam_size: (int) + the beam size + Returns: + *encoder_out* rearranged according to *new_order* + """ + pass + + def prune_sentences(self, batch_idxs: Tensor): + """ + Removes constraint states for completed sentences (if supported). + This is called from sequence_generator._generate() when sentences are + deleted from the batch. + + Args: + batch_idxs: Indices of *sentences* whose constraint state should be *kept*. + """ + pass + + def update_constraints(self, active_hypos: Tensor): + """ + Updates the constraint states by selecting the beam items that are retained. + This is called at each time step of sequence_generator._generate() when + the set of 2 * {beam_size} candidate hypotheses are reduced to the beam size. + + Args: + active_hypos: (batch size, beam size) + list of integers denoting, for each sentence, which beam candidate items + should be kept. + """ + pass + + +def unravel_index(index, shape): + out = [] + for dim in reversed(shape): + out.append(index % dim) + index = index // dim + return torch.stack(tuple(reversed(out)), dim=-1) + + +def topk_sum(lprobs_list, k): + """ + lprobs_list = [lprobs_1,...,lprobs_n], where: + lprobs_1 : (batch_size x beam_size x vocab_1) + ... + lprobs_n : (batch_size x beam_size x vocab_n) + + Return: + - topk_values : (batch_size x k) + values of the topk sum of the form : + lprobs_1[bsz, beam_idx, vocab_1_idx] + ... + lprobs_n[bsz, beam_idx, vocab_n_idx] + - topk_idxs : (batch_size x k x n+1) + each (n+1)-tensor being [beam_idx, vocab_1_idx, ..., vocab_n_idx] + """ + # Reduce all lprobs to k candidates first to reduce later complexity + # We may assume that k << vocab + lprobs_topk_list = [] + lprobs_topk_indices_list = [] + for lprobs in lprobs_list: + k_i = min(k, lprobs.size(-1)) + topk_values, topk_indices = torch.topk(lprobs, k=k_i) + # topk_values : (batch_size x beam_size x k_i) + # topk_indices : (batch_size x beam_size x k_i) + lprobs_topk_list.append(topk_values) + lprobs_topk_indices_list.append(topk_indices) + + # Compute all possible sums + sum_lprobs_topk = lprobs_topk_list[0] + for i in range(1, len(lprobs_topk_list)): + unsqueezed_lprobs = lprobs_topk_list[i] + for _ in range(i): + unsqueezed_lprobs = unsqueezed_lprobs.unsqueeze(-2) + sum_lprobs_topk = sum_lprobs_topk.unsqueeze(-1) + unsqueezed_lprobs + # sum_lprobs : (batch_size x beam_size x k_1 x ... x k_n) + + # Get the top k sums and the (transformed indices) + topk_sum_values, topk_sum_indices = torch.topk( + sum_lprobs_topk.view(sum_lprobs_topk.size(0), -1), k=k + ) + # topk_sum_values : (batch_size x k) + # topk_sum_indices : (batch_size x k) + topk_sum_indices = unravel_index(topk_sum_indices, tuple(sum_lprobs_topk.shape[1:])) + # topk_sum_indices : (batch_size x k x n+1) + + # Convert the transformed indices to the true indices + for i_batch in range(topk_sum_indices.size(0)): + for i_cand in range(topk_sum_indices.size(1)): + i_beam, *transformed_vocab_indices = topk_sum_indices[i_batch, i_cand] + true_vocab_indices = [i_beam] + for j, transformed_vocab_j_idx in enumerate(transformed_vocab_indices): + true_vocab_j_idx = lprobs_topk_indices_list[j][ + i_batch, i_beam, transformed_vocab_j_idx + ] + true_vocab_indices.append(true_vocab_j_idx) + topk_sum_indices[i_batch, i_cand] = torch.tensor(true_vocab_indices) + + topk_sum_beams = topk_sum_indices[:, :, 0] + topk_sum_indices = topk_sum_indices[:, :, 1:] + + return topk_sum_values, topk_sum_indices, topk_sum_beams + + +class MultichannelBeamSearch(MultichannelSearch): + def __init__(self, tgt_dicts): + super().__init__(tgt_dicts) + self.constraint_states = None + + @torch.jit.export + def step( + self, + step: int, + lprobs, + scores: Optional[Dict[str, Tensor]], + prev_output_tokens: Optional[Dict[str, Tensor]] = None, + original_batch_idxs: Optional[Tensor] = None, + ): + channels = list(lprobs.keys()) + bsz, beam_size, _ = lprobs[channels[0]].size() + + lprobs_list = [] + if step == 0: + # at the first step all hypotheses are equally likely, so use + # only the first beam + for channel in channels: + lprobs_list.append(lprobs[channel][:, ::beam_size, :].contiguous()) + else: + # make probs contain cumulative scores for each hypothesis + assert scores is not None + for channel in channels: + lprobs_list.append( + lprobs[channel] + scores[channel][:, :, step - 1].unsqueeze(-1) + ) + + topk_sum_values, topk_sum_indices, topk_sum_beams = topk_sum( + lprobs_list, k=beam_size * 2 + ) + + beams_buf = topk_sum_beams + scores_buf = {} + indices_buf = {} + for i, channel in enumerate(channels): + indices_buf[channel] = topk_sum_indices[:, :, i] + scores_buf[channel] = ( + torch.tensor( + [ + lprobs_list[i][i_batch, i_beam, i_index] + for i_batch in range(bsz) + for i_beam, i_index in zip( + beams_buf[i_batch], indices_buf[channel][i_batch] + ) + ] + ) + .view(bsz, -1) + .to(lprobs_list[i].device) + ) + + # At this point, beams_buf and indices_buf are single-dim and contain relative indices + return scores_buf, indices_buf, beams_buf + + +class ContiguousMultichannelBeamSearch(MultichannelSearch): + def __init__(self, tgt_dicts): + super().__init__(tgt_dicts) + self.constraint_states = None + + @torch.jit.export + def step( + self, + step: int, + lprobs, + scores: Optional[Tensor], + prev_output_tokens: Optional[Tensor] = None, + original_batch_idxs: Optional[Tensor] = None, + ): + n_channels = len(lprobs) + bsz, beam_size, _ = lprobs[0].size() + + lprobs_list = [] + if step == 0: + # at the first step all hypotheses are equally likely, so use + # only the first beam + for i in range(n_channels): + lprobs_list.append(lprobs[i][:, ::beam_size, :].contiguous()) + else: + # make probs contain cumulative scores for each hypothesis + assert scores is not None + for i in range(n_channels): + lprobs_list.append(lprobs[i] + scores[:, :, step - 1, i].unsqueeze(-1)) + + topk_sum_values, topk_sum_indices, topk_sum_beams = topk_sum( + lprobs_list, k=beam_size * 2 + ) + + beams_buf = topk_sum_beams + indices_buf = topk_sum_indices + scores_buf = ( + torch.tensor( + [ + lprobs_list[i][i_batch, i_beam, i_index] + for i in range(len(lprobs_list)) + for i_batch in range(bsz) + for i_beam, i_index in zip( + beams_buf[i_batch], indices_buf[i_batch, :, i] + ) + ] + ) + .view(len(lprobs_list), bsz, -1) + .permute(1, 2, 0) + .to(lprobs_list[0].device) + ) + + # At this point, beams_buf and indices_buf are single-dim and contain relative indices + return scores_buf, indices_buf, beams_buf + + +class ContiguousMultichannelSampling(MultichannelSearch): + sampling_topk: int + sampling_topp: float + + def __init__(self, tgt_dicts, sampling_topk=-1, sampling_topp=-1.0): + super().__init__(tgt_dicts) + self.sampling_topk = sampling_topk + self.sampling_topp = sampling_topp + + def _sample_topp(self, lprobs): + """Sample among the smallest set of elements whose cumulative probability mass exceeds p. + + See `"The Curious Case of Neural Text Degeneration" + (Holtzman et al., 2019) `_. + + Args: + lprobs: (bsz x input_beam_size x vocab_size) + the model's log-probabilities over the vocabulary at the current step + + Return: A tuple of (trimed_probs, truncated_indices) where: + trimed_probs: (bsz x input_beam_size x ?) + the model's probabilities over the elements selected to sample from. The + width of the third dimension is determined by top-P. + truncated_indices: (bsz x input_beam_size x ?) + the indices of the chosen elements. + """ + probs = lprobs.exp_() + + # sort the last dimension (vocab dimension) in descending order + sorted_probs, sorted_indices = probs.sort(descending=True) + + # compute a mask to indicate the words to be included in the top-P set. + cumsum_probs = sorted_probs.cumsum(dim=2) + mask = cumsum_probs.lt(self.sampling_topp) + + # note that mask was computed by 'lt'. One more word needs to be included + # so that the cumulative probability mass can exceed p. + cumsum_mask = mask.cumsum(dim=2) + last_included = cumsum_mask[:, :, -1:] + last_included.clamp_(0, mask.size()[2] - 1) + mask = mask.scatter_(2, last_included, 1) + + # truncate unnecessary dims. + max_dim = last_included.max() + truncated_mask = mask[:, :, : max_dim + 1] + truncated_probs = sorted_probs[:, :, : max_dim + 1] + truncated_indices = sorted_indices[:, :, : max_dim + 1] + + # trim the words that are not in top-P by setting their probabilities + # to 0, so that they would not be sampled later. + trim_mask = ~truncated_mask + trimed_probs = truncated_probs.masked_fill_(trim_mask, 0) + return trimed_probs, truncated_indices + + @torch.jit.export + def step( + self, + step: int, + lprobs, + scores, + prev_output_tokens: Optional[Tensor] = None, + original_batch_idxs: Optional[Tensor] = None, + ): + n_channels = len(lprobs) + bsz, beam_size, vocab_size = lprobs[0].size() + + if step == 0: + # at the first step all hypotheses are equally likely, so use + # only the first beam + for i in range(n_channels): + lprobs[i] = lprobs[i][:, ::beam_size, :].contiguous() + + probs = [] + top_indices = [] + for i in range(n_channels): + if self.sampling_topp > 0: + # only sample from the smallest set of words whose cumulative probability mass exceeds p + probs_i, top_indices_i = self._sample_topp(lprobs[i]) + elif self.sampling_topk > 0: + # only sample from top-k candidates + lprobs[i], top_indices_i = lprobs[i].topk( + min(self.sampling_topk, lprobs[i].size(-1)) + ) + probs_i = lprobs[i].exp_() + else: + probs_i = lprobs[i].exp_() + + # dummy data to be consistent with true branch for type check + top_indices_i = torch.empty(0).to(probs_i) + probs.append(probs_i) + top_indices.append(top_indices_i) + # sample + indices_buf = [] + for i in range(n_channels): + if step == 0: + indices_buf.append( + torch.multinomial( + probs[i].view(bsz, -1), + beam_size, + replacement=True, + ).view(bsz, beam_size) + ) + else: + indices_buf.append( + torch.multinomial( + probs[i].view(bsz * beam_size, -1), + 1, + replacement=True, + ).view(bsz, beam_size) + ) + + if step == 0: + for i in range(n_channels): + # expand to beam size + probs[i] = probs[i].expand(bsz, beam_size, -1) + + # gather scores + scores_buf = [] + for i in range(n_channels): + scores_buf.append( + torch.gather(probs[i], dim=2, index=indices_buf[i].unsqueeze(-1)) + ) + scores_buf[i] = scores_buf[i].log_().view(bsz, -1) + + # remap indices if using top-k or top-P sampling + if self.sampling_topk > 0 or self.sampling_topp > 0: + for i in range(n_channels): + indices_buf[i] = torch.gather( + top_indices[i].expand(bsz, beam_size, -1), + dim=2, + index=indices_buf[i].unsqueeze(-1), + ).squeeze(2) + + if step == 0: + beams_buf = indices_buf[0].new_zeros(bsz, beam_size) + else: + beams_buf = torch.arange(0, beam_size).to(indices_buf[0]).repeat(bsz, 1) + # make scores cumulative + for i in range(n_channels): + scores_buf[i].add_( + torch.gather(scores[:, :, step - 1, i], dim=1, index=beams_buf) + ) + scores_buf = torch.stack(scores_buf, dim=-1) + indices_buf = torch.stack(indices_buf, dim=-1) + + return scores_buf, indices_buf, beams_buf diff --git a/data/fairseq/fairseq/models/speech_dlm/sequence_generator/multichannel_sequence_generator.py b/data/fairseq/fairseq/models/speech_dlm/sequence_generator/multichannel_sequence_generator.py new file mode 100644 index 0000000000000000000000000000000000000000..24807b866de15661973af9e8079a4bd50bb8004b --- /dev/null +++ b/data/fairseq/fairseq/models/speech_dlm/sequence_generator/multichannel_sequence_generator.py @@ -0,0 +1,1110 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import math +from typing import Dict, List, Optional + +from omegaconf.listconfig import ListConfig +from omegaconf.dictconfig import DictConfig + +import torch +import torch.nn as nn +from fairseq.models import FairseqIncrementalDecoder +from torch import Tensor +from fairseq.ngram_repeat_block import NGramRepeatBlock +from .multichannel_search import ContiguousMultichannelBeamSearch +from fairseq.models.speech_dlm import SpeechDLM + + +class MultichannelSequenceGenerator(nn.Module): + def __init__( + self, + models, + tgt_dicts, + beam_size=1, + max_len_a=0, + max_len_b=200, + min_len=1, + normalize_scores=True, + len_penalty=1.0, + unk_penalty=0.0, + temperature=1.0, + match_source_len=False, + no_repeat_ngram_size=0, + search_strategy=None, + eos=None, + symbols_to_strip_from_output=None, + lm_model=None, + lm_weight=1.0, + duration_temperature=1.0, + ): + """Generate multi-channel parallel units with the SpeechDLM model + as described in the paper: https://arxiv.org/pdf/2203.16502.pdf; + + Args: + models (List[~fairseq.models.FairseqModel]): ensemble of models, + currently support fairseq.models.TransformerModel for scripting + beam_size (int, optional): beam width (default: 1) + max_len_a/b (int, optional): generate sequences of maximum length + ax + b, where x is the source length + min_len (int, optional): the minimum length of the generated output + (not including end-of-sentence) + normalize_scores (bool, optional): normalize scores by the length + of the output (default: True) + len_penalty (float, optional): length penalty, where <1.0 favors + shorter, >1.0 favors longer sentences (default: 1.0) + unk_penalty (float, optional): unknown word penalty, where <0 + produces more unks, >0 produces fewer (default: 0.0) + temperature (float, optional): temperature, where values + >1.0 produce more uniform samples and values <1.0 produce + sharper samples (default: 1.0) + match_source_len (bool, optional): outputs should match the source + length (default: False) + duration_temperature (float, optional): rate of the duration prediction, + higher rate induces a faster generated wav (default: 1.0) + """ + super().__init__() + if isinstance(models, MultichannelEnsembleModel): + self.model = models + else: + self.model = MultichannelEnsembleModel(models) + self.tgt_dicts = tgt_dicts + self.pad = list(tgt_dicts.values())[0].pad() + self.unk = list(tgt_dicts.values())[0].unk() + self.eos = list(tgt_dicts.values())[0].eos() if eos is None else eos + self.symbols_to_strip_from_output = ( + symbols_to_strip_from_output.union({self.eos}) + if symbols_to_strip_from_output is not None + else {self.eos} + ) + self.channels = list(tgt_dicts.keys()) + self.n_channels = len(self.channels) + self.vocab_sizes = [len(tgt_dicts[channel]) for channel in self.channels] + # the max beam size is the dictionary size - 1, since we never select pad + max_possible_beam_size = 1 + for i in self.vocab_sizes: + max_possible_beam_size *= i - 1 + self.beam_size = min(beam_size, max_possible_beam_size) + self.max_len_a = max_len_a + self.max_len_b = max_len_b + self.min_len = min_len + + self.normalize_scores = normalize_scores + self.len_penalty = len_penalty + self.unk_penalty = unk_penalty + if isinstance(temperature, (int, float)): + temperature = {channel: temperature for channel in self.channels} + elif isinstance(temperature, ListConfig) or isinstance(temperature, list): + temperature = { + channel: temperature[i] for i, channel in enumerate(self.channels) + } + assert isinstance(temperature, DictConfig) or isinstance( + temperature, dict + ), f"temperature: expected dict, but found {type(temperature)}" + self.temperature = temperature + self.match_source_len = match_source_len + + if no_repeat_ngram_size > 0: + self.repeat_ngram_blocker = NGramRepeatBlock(no_repeat_ngram_size) + else: + self.repeat_ngram_blocker = None + + for channel in temperature: + assert temperature[channel] > 0, "--temperature must be greater than 0" + + if search_strategy is None: + self.search = ContiguousMultichannelBeamSearch(tgt_dicts) + else: + self.search = search_strategy + # We only need to set src_lengths in LengthConstrainedBeamSearch. + # As a module attribute, setting it would break in multithread + # settings when the model is shared. + self.should_set_src_lengths = ( + hasattr(self.search, "needs_src_lengths") and self.search.needs_src_lengths + ) + + self.model.eval() + + self.lm_model = lm_model + self.lm_weight = lm_weight + if self.lm_model is not None: + self.lm_model.eval() + + self.duration_prediction = bool( + str(getattr(models[0].decoder.args, "duration_prediction", "false")).lower() + == "true" + ) + self.delayed_duration = bool( + str( + getattr(models[0].decoder.args, "delayed_duration_target", "false") + ).lower() + == "true" + ) + self.duration_temperature = duration_temperature + + def cuda(self): + self.model.cuda() + return self + + @torch.no_grad() + def forward( + self, + sample: Dict[str, Dict[str, Tensor]], # TODO: Modify this + prefix_tokens: Optional[Dict[str, Tensor]] = None, + bos_token: Optional[int] = None, + ): + """Generate a batch of translations. + + Args: + sample (dict): batch + prefix_tokens (dict of torch.LongTensor, optional): force decoder to begin + with these tokens + bos_token (int, optional): beginning of sentence token + (default: self.eos) + """ + return self._generate(sample, prefix_tokens, bos_token=bos_token) + + @torch.no_grad() + def generate(self, models, sample: Dict[str, Dict[str, Tensor]], **kwargs): + """Generate translations. Match the api of other fairseq generators. + + Args: + models (List[~fairseq.models.FairseqModel]): ensemble of models + sample (dict): batch + prefix_tokens (dict of torch.LongTensor, optional): force decoder to begin + with these tokens + constraints (torch.LongTensor, optional): force decoder to include + the list of constraints + bos_token (int, optional): beginning of sentence token + (default: self.eos) + """ + return self._generate(sample, **kwargs) + + def _generate( + self, + sample: Dict[str, Dict[str, Tensor]], + prefix_tokens: Optional[Dict[str, Tensor]] = None, + constraints: Optional[Tensor] = None, + bos_token: Optional[int] = None, + ): + """ + Here sample is expected to have the following form + { + 'id': index, + 'net_input': { + 'src_tokens': { + 'channel1' : tensor((batch x src_length)), + 'channel2' : tensor((batch x src_length)), + }, + ... + }, + } + and prefix_tokens + { + 'channel1' : tensor((batch x prefix_length)), + 'channel2' : tensor((batch x prefix_length)), + } + """ + if self.model.is_speech_dlm: + incremental_states = torch.jit.annotate( + List[Dict[str, Dict[str, Optional[Tensor]]]], + [ + torch.jit.annotate( + List[Dict[str, Dict[str, Optional[Tensor]]]], + [{} for _ in range(self.n_channels)], + ) + for i in range(self.model.models_size) + ], + ) + else: + incremental_states = torch.jit.annotate( + List[Dict[str, Dict[str, Optional[Tensor]]]], + [ + torch.jit.annotate(Dict[str, Dict[str, Optional[Tensor]]], {}) + for i in range(self.model.models_size) + ], + ) + net_input = sample["net_input"] + # Convert from dict to tensor form + # shape of src_tokens : (bsz x src_len x n_channels) + src_tokens = torch.stack( + [net_input["src_tokens"][channel] for channel in self.channels], dim=-1 + ) + prefix_tokens = torch.stack( + [prefix_tokens[channel] for channel in self.channels], dim=-1 + ) + # length of the source text being the character length except EndOfSentence and pad + src_lengths = ( + (src_tokens[..., 0].ne(self.eos) & src_tokens[..., 0].ne(self.pad)) + .long() + .sum(dim=1) + ) + + # bsz: total number of sentences in beam + # Note that src_tokens may have more than 2 dimensions (i.e. audio features) + bsz, src_len = src_tokens.size()[:2] + beam_size = self.beam_size + + if constraints is not None and not self.search.supports_constraints: + raise NotImplementedError( + "Target-side constraints were provided, but search method doesn't support them" + ) + + # Initialize constraints, when active + self.search.init_constraints(constraints, beam_size) + + max_len: int = -1 + if self.match_source_len: + max_len = src_lengths.max().item() + else: + max_len = min( + int(self.max_len_a * src_len + self.max_len_b), + # exclude the EOS marker + self.model.max_decoder_positions() - 1, + ) + assert ( + self.min_len <= max_len + ), "min_len cannot be larger than max_len, please adjust these!" + # compute the encoder output for each beam + encoder_outs = self.model.forward_encoder(net_input) + + # placeholder of indices for bsz * beam_size to hold tokens and accumulative scores + new_order = torch.arange(bsz).view(-1, 1).repeat(1, beam_size).view(-1) + new_order = new_order.to(src_tokens.device).long() + encoder_outs = self.model.reorder_encoder_out(encoder_outs, new_order) + # ensure encoder_outs is a List. + assert encoder_outs is not None + + # initialize buffers + # cumulative scores of hypotheses + scores = ( + torch.zeros(bsz * beam_size, max_len + 1, self.n_channels) + .to(src_tokens) + .float() + ) # +1 for eos; pad is never chosen for scoring + tokens = ( + torch.zeros(bsz * beam_size, max_len + 2, self.n_channels) + .to(src_tokens) + .long() + .fill_(self.pad) + ) # +2 for eos and pad + tokens[:, 0] = self.eos if bos_token is None else bos_token + attn: Optional[Tensor] = None + + # A list that indicates candidates that should be ignored. + # For example, suppose we're sampling and have already finalized 2/5 + # samples. Then cands_to_ignore would mark 2 positions as being ignored, + # so that we only finalize the remaining 3 samples. + cands_to_ignore = ( + torch.zeros(bsz, beam_size).to(src_tokens).eq(-1) + ) # forward and backward-compatible False mask + + # list of completed sentences + finalized = torch.jit.annotate( + List[List[Dict[str, Tensor]]], + [torch.jit.annotate(List[Dict[str, Tensor]], []) for i in range(bsz)], + ) # contains lists of dictionaries of infomation about the hypothesis being finalized at each step + + finished = [ + False for i in range(bsz) + ] # a boolean array indicating if the sentence at the index is finished or not + num_remaining_sent = bsz # number of sentences remaining + + # number of candidate hypos per step + cand_size = 2 * beam_size # 2 x beam size in case half are EOS + + # offset arrays for converting between different indexing schemes + bbsz_offsets = ( + (torch.arange(0, bsz) * beam_size) + .unsqueeze(1) + .type_as(tokens) + .to(src_tokens.device) + ) + cand_offsets = torch.arange(0, cand_size).type_as(tokens).to(src_tokens.device) + + reorder_state: Optional[Tensor] = None + batch_idxs: Optional[Tensor] = None + + original_batch_idxs: Optional[Tensor] = None + if "id" in sample and isinstance(sample["id"], Tensor): + original_batch_idxs = sample["id"] + else: + original_batch_idxs = torch.arange(0, bsz).type_as(tokens) + + if self.duration_prediction: + dur_counter = torch.ones(bsz * beam_size, self.n_channels).to(src_tokens) + # save the indice where the dur_counter just copied from dur_pred + dur_counter_jump_indices = None + + for step in range(max_len + 1): # one extra step for EOS marker + # reorder decoder internal states based on the prev choice of beams + if reorder_state is not None: + if batch_idxs is not None: + # update beam indices to take into account removed sentences + corr = batch_idxs - torch.arange(batch_idxs.numel()).type_as( + batch_idxs + ) + reorder_state.view(-1, beam_size).add_( + corr.unsqueeze(-1) * beam_size + ) + original_batch_idxs = original_batch_idxs[batch_idxs] + self.model.reorder_incremental_state(incremental_states, reorder_state) + encoder_outs = self.model.reorder_encoder_out( + encoder_outs, reorder_state + ) + + input_tokens = { + channel: tokens[:, : step + 1, i] + for i, channel in enumerate(self.channels) + } + + lprobs_dict, avg_attn_scores = self.model.forward_decoder( + input_tokens, + encoder_outs, + incremental_states, + self.temperature, + ) + + # Because the sizes of vocab is different, we cannot concat the lprobs to form a single tensor + if not self.duration_prediction: + lprobs_list = list(lprobs_dict.values()) + else: + lprobs_list = [ + net_output["pred_token"] for net_output in lprobs_dict.values() + ] + + # non-positive predicted durations + dur_preds = ( + torch.stack( + [ + net_output["pred_duration"] + for net_output in lprobs_dict.values() + ] + ) + .squeeze(-1) + .T + ) + dur_preds = dur_preds / self.duration_temperature + dur_preds = dur_preds.round().long() + dur_preds[dur_preds < 1] = 1 + + # dur_preds & dur_counter needs to be modified when there isn't an edge + if step > 0: + non_edge_indices = tokens[:, step, :] == tokens[:, step - 1, :] + if self.delayed_duration: + dur_preds[non_edge_indices] = 1 + else: + if dur_counter_jump_indices is not None: + dur_counter[dur_counter_jump_indices & non_edge_indices] = 2 + + # update dur_counter + if step > 0: + if self.delayed_duration: + dur_counter -= ( + (dur_counter == 1) + | (tokens[:, step, :] == tokens[:, step - 1, :]) + ).int() + dur_counter[dur_counter < 0] = 0 + else: + dur_counter -= ( + tokens[:, step, :] == tokens[:, step - 1, :] + ).int() + dur_counter[dur_counter < 1] = 1 + + # whether to copy previous token (ie. if the counter is still on) + # and get get the new duration + if self.delayed_duration: + dur_counter_jump_indices = dur_counter == 0 + dur_counter[dur_counter_jump_indices] = dur_preds[ + dur_counter_jump_indices + ] + + # whether to copy previous token in this step + copy_prev_token = dur_counter != 1 + if self.delayed_duration is False: + dur_counter_jump_indices = dur_counter == 1 + dur_counter[dur_counter_jump_indices] = dur_preds[ + dur_counter_jump_indices + ] + # else: + # dur_counter[dur_counter==0] = dur_preds[dur_counter==0] - 1 + # copy_prev_token = (dur_counter > 0) + + if self.lm_model is not None: + assert False, "Currently not supported in multichannelLM case" + + for i in range(self.n_channels): + lprobs_list[i][lprobs_list[i] != lprobs_list[i]] = torch.tensor( + -math.inf + ).to(lprobs_list[i]) + + lprobs_list[i][:, self.pad] = -math.inf # never select pad + lprobs_list[i][:, self.unk] -= self.unk_penalty # apply unk penalty + + # handle max length constraint + if step >= max_len: + lprobs_list[i][:, : self.eos] = -math.inf + lprobs_list[i][:, self.eos + 1 :] = -math.inf + else: + lprobs_list[i][ + :, self.eos + ] = -math.inf # quick fix for short generation + + # handle prefix tokens (possibly with different lengths) + if ( + prefix_tokens is not None + and step < prefix_tokens.size(1) + and step < max_len + ): + ( + lprobs_list[i], + tokens[..., i], + scores[..., i], + ) = self._prefix_tokens( + step, + lprobs_list[i], + scores[..., i], + tokens[..., i], + prefix_tokens[..., i], + beam_size, + ) + if self.duration_prediction: + # Can copy previous token if the prefix token is padding or unk (1-channel conditionned case) + can_copy_mask = ( + prefix_tokens[:, step, i].eq(self.pad) + | prefix_tokens[:, step, i].eq(self.unk) + ).repeat_interleave(beam_size) + copy_prev_token[:, i] &= can_copy_mask + elif step < self.min_len: + # minimum length constraint (does not apply if using prefix_tokens) + lprobs_list[i][:, self.eos] = -math.inf + + if self.duration_prediction: + if step < max_len: + for j in range(copy_prev_token.size(0)): + if copy_prev_token[j, i]: + prev_token = tokens[j, step, i] + lprobs_list[i][j, :prev_token] = -math.inf + lprobs_list[i][j, prev_token + 1 :] = -math.inf + # lprobs_list[i][j, prev_token] = 0. + # dur_counter[j,i] -= 1 + # else: + # prev_token = tokens[j, step, i] + # if not (lprobs_list[i][j,:].ne(-math.inf).nonzero() == prev_token).all(): + # lprobs_list[i][j, prev_token] = -math.inf + # dur_counter[j,i] = 0. + + # Record attention scores, only support avg_attn_scores is a Tensor + if avg_attn_scores is not None: + if attn is None: + attn = torch.empty( + bsz * beam_size, avg_attn_scores.size(1), max_len + 2 + ).to(scores) + attn[:, :, step + 1].copy_(avg_attn_scores) + + scores = scores.type_as(lprobs_list[0]) + eos_bbsz_idx = torch.empty(0).to( + tokens + ) # indices of hypothesis ending with eos (finished sentences) + eos_scores = torch.empty(0).to( + scores + ) # scores of hypothesis ending with eos (finished sentences) + + if self.should_set_src_lengths: + self.search.set_src_lengths(src_lengths) + + if self.repeat_ngram_blocker is not None: + for i in range(self.n_channels): + lprobs_list[i] = self.repeat_ngram_blocker( + tokens, lprobs_list[i], bsz, beam_size, step + ) + + # Shape: (batch, cand_size) + cand_scores, cand_indices, cand_beams = self.search.step( + step, + [ + lprobs_list[i].view(bsz, -1, self.vocab_sizes[i]) + for i in range(self.n_channels) + ], + scores.view(bsz, beam_size, -1, self.n_channels)[:, :, :step, :], + tokens[:, : step + 1], + original_batch_idxs, + ) + + # cand_bbsz_idx contains beam indices for the top candidate + # hypotheses, with a range of values: [0, bsz*beam_size), + # and dimensions: [bsz, cand_size] + cand_bbsz_idx = cand_beams.add(bbsz_offsets) + + # finalize hypotheses that end in eos + # Shape of eos_mask: (batch size, beam size) + eos_mask = cand_indices.eq(self.eos) & cand_scores.ne(-math.inf) + eos_mask = torch.any(eos_mask, dim=-1, keepdim=False) + eos_mask[:, :beam_size][cands_to_ignore] = torch.tensor(0).to(eos_mask) + + # only consider eos when it's among the top beam_size indices + # Now we know what beam item(s) to finish + # Shape: 1d list of absolute-numbered + eos_bbsz_idx = torch.masked_select( + cand_bbsz_idx[:, :beam_size], mask=eos_mask[:, :beam_size] + ) + + finalized_sents: List[int] = [] + if eos_bbsz_idx.numel() > 0: + eos_scores = torch.stack( + [ + torch.masked_select( + cand_scores[:, :beam_size, i], mask=eos_mask[:, :beam_size] + ) + for i in range(self.n_channels) + ], + dim=-1, + ) + finalized_sents = self.finalize_hypos( + step, + eos_bbsz_idx, + eos_scores, + tokens, + scores, + finalized, + finished, + beam_size, + attn, + src_lengths, + max_len, + ) + num_remaining_sent -= len(finalized_sents) + + assert num_remaining_sent >= 0 + if num_remaining_sent == 0: + break + if self.search.stop_on_max_len and step >= max_len: + break + assert step < max_len, f"{step} < {max_len}" + + # Remove finalized sentences (ones for which {beam_size} + # finished hypotheses have been generated) from the batch. + if len(finalized_sents) > 0: + new_bsz = bsz - len(finalized_sents) + + # construct batch_idxs which holds indices of batches to keep for the next pass + batch_mask = torch.ones( + bsz, dtype=torch.bool, device=cand_indices.device + ) + batch_mask[finalized_sents] = False + # TODO replace `nonzero(as_tuple=False)` after TorchScript supports it + batch_idxs = torch.arange( + bsz, device=cand_indices.device + ).masked_select(batch_mask) + + # Choose the subset of the hypothesized constraints that will continue + self.search.prune_sentences(batch_idxs) + + eos_mask = eos_mask[batch_idxs] + cand_beams = cand_beams[batch_idxs] + bbsz_offsets.resize_(new_bsz, 1) + cand_bbsz_idx = cand_beams.add(bbsz_offsets) + cand_scores = cand_scores[batch_idxs] + cand_indices = cand_indices[batch_idxs] + + if prefix_tokens is not None: + prefix_tokens = prefix_tokens[batch_idxs] + src_lengths = src_lengths[batch_idxs] + cands_to_ignore = cands_to_ignore[batch_idxs] + + scores = scores.view(bsz, -1)[batch_idxs].view( + new_bsz * beam_size, -1, self.n_channels + ) + tokens = tokens.view(bsz, -1)[batch_idxs].view( + new_bsz * beam_size, -1, self.n_channels + ) + if self.duration_prediction: + dur_counter = dur_counter.view(bsz, -1)[batch_idxs].view( + new_bsz * beam_size, self.n_channels + ) + if attn is not None: + attn = attn.view(bsz, -1)[batch_idxs].view( + new_bsz * beam_size, attn.size(1), -1 + ) + bsz = new_bsz + else: + batch_idxs = None + + # Set active_mask so that values > cand_size indicate eos hypos + # and values < cand_size indicate candidate active hypos. + # After, the min values per row are the top candidate active hypos + + # Rewrite the operator since the element wise or is not supported in torchscript. + + eos_mask[:, :beam_size] = ~((~cands_to_ignore) & (~eos_mask[:, :beam_size])) + active_mask = torch.add( + eos_mask.type_as(cand_offsets) * cand_size, + cand_offsets[: eos_mask.size(1)], + ) + + # get the top beam_size active hypotheses, which are just + # the hypos with the smallest values in active_mask. + # {active_hypos} indicates which {beam_size} hypotheses + # from the list of {2 * beam_size} candidates were + # selected. Shapes: (batch size, beam size) + new_cands_to_ignore, active_hypos = torch.topk( + active_mask, k=beam_size, dim=1, largest=False + ) + + # update cands_to_ignore to ignore any finalized hypos. + cands_to_ignore = new_cands_to_ignore.ge(cand_size)[:, :beam_size] + # Make sure there is at least one active item for each sentence in the batch. + assert (~cands_to_ignore).any(dim=1).all() + + # update cands_to_ignore to ignore any finalized hypos + # {active_bbsz_idx} denotes which beam number is continued for each new hypothesis (a beam + # can be selected more than once). + active_bbsz_idx = torch.gather(cand_bbsz_idx, dim=1, index=active_hypos) + active_bbsz_idx = active_bbsz_idx.view(-1) + + # active_scores = torch.stack([ + # torch.gather(cand_scores[...,0], dim=1, index=active_hypos) + # for i in range(self.n_channels) + # ], dim = -1) + # active_scores = active_scores.view(-1) + + # copy tokens and scores for active hypotheses + + # Set the tokens for each beam (can select the same row more than once) + tokens[:, : step + 1] = torch.index_select( + tokens[:, : step + 1], dim=0, index=active_bbsz_idx + ) + # Select the next token for each of them + for i in range(self.n_channels): + tokens.view(bsz, beam_size, -1, self.n_channels)[ + :, :, step + 1, i + ] = torch.gather(cand_indices[..., i], dim=1, index=active_hypos) + if step > 0: + scores[:, :step] = torch.index_select( + scores[:, :step], dim=0, index=active_bbsz_idx + ) + for i in range(self.n_channels): + scores.view(bsz, beam_size, -1, self.n_channels)[ + :, :, step, i + ] = torch.gather(cand_scores[..., i], dim=1, index=active_hypos) + + if self.duration_prediction: + dur_counter = torch.index_select( + dur_counter, dim=0, index=active_bbsz_idx + ) + + # Update constraints based on which candidates were selected for the next beam + self.search.update_constraints(active_hypos) + + # copy attention for active hypotheses + if attn is not None: + attn[:, :, : step + 2] = torch.index_select( + attn[:, :, : step + 2], dim=0, index=active_bbsz_idx + ) + + # reorder incremental state in decoder + reorder_state = active_bbsz_idx + + # sort by score descending + for sent in range(len(finalized)): + scores = torch.tensor( + [float(elem["score"].item()) for elem in finalized[sent]] + ) + _, sorted_scores_indices = torch.sort(scores, descending=True) + finalized[sent] = [finalized[sent][ssi] for ssi in sorted_scores_indices] + finalized[sent] = torch.jit.annotate( + List[Dict[str, Tensor]], finalized[sent] + ) + return finalized + + def _prefix_tokens( + self, step: int, lprobs, scores, tokens, prefix_tokens, beam_size: int + ): + """Handle prefix tokens""" + prefix_toks = prefix_tokens[:, step].unsqueeze(-1).repeat(1, beam_size).view(-1) + prefix_lprobs = lprobs.gather(-1, prefix_toks.unsqueeze(-1)) + prefix_mask = prefix_toks.ne(self.pad) + # used for 1-channel generation, do not force the unk token (i.e. unk tokens are changed) + prefix_mask &= prefix_toks.ne(self.unk) + # zeroing the copying tokens + # if step > 0: + # copy_mask = (prefix_tokens[:, step] == prefix_tokens[:, step-1]).unsqueeze(-1).repeat(1, beam_size).view(-1) + # prefix_lprobs[copy_mask & prefix_mask] = 0. + lprobs[prefix_mask] = torch.tensor(-math.inf).to(lprobs) + lprobs[prefix_mask] = lprobs[prefix_mask].scatter( + -1, prefix_toks[prefix_mask].unsqueeze(-1), prefix_lprobs[prefix_mask] + ) + # shouldn't stop at unk token + unk_mask = prefix_toks.eq(self.unk) + if len(lprobs[unk_mask]) > 0: + # otherwise it won't assign to lprobs, + # see: https://discuss.pytorch.org/t/how-to-mask-and-assign-a-value-to-tensor/18437 + copy_lprobs = lprobs[unk_mask][:, :] + copy_lprobs[:, self.eos] = -math.inf + lprobs[unk_mask] = copy_lprobs + # if prefix includes eos, then we should make sure tokens and + # scores are the same across all beams + eos_mask = prefix_toks.eq(self.eos) + if eos_mask.any(): + # validate that the first beam matches the prefix + first_beam = tokens[eos_mask].view(-1, beam_size, tokens.size(-1))[ + :, 0, 1 : step + 1 + ] + eos_mask_batch_dim = eos_mask.view(-1, beam_size)[:, 0] + target_prefix = prefix_tokens[eos_mask_batch_dim][:, :step] + assert (first_beam == target_prefix).all() + + # copy tokens, scores and lprobs from the first beam to all beams + tokens = self.replicate_first_beam(tokens, eos_mask_batch_dim, beam_size) + scores = self.replicate_first_beam(scores, eos_mask_batch_dim, beam_size) + lprobs = self.replicate_first_beam(lprobs, eos_mask_batch_dim, beam_size) + return lprobs, tokens, scores + + def replicate_first_beam(self, tensor, mask, beam_size: int): + tensor = tensor.view(-1, beam_size, tensor.size(-1)) + tensor[mask] = tensor[mask][:, :1, :] + return tensor.view(-1, tensor.size(-1)) + + def finalize_hypos( + self, + step: int, + bbsz_idx, + eos_scores, + tokens, + scores, + finalized: List[List[Dict[str, Tensor]]], + finished: List[bool], + beam_size: int, + attn: Optional[Tensor], + src_lengths, + max_len: int, + ): + """Finalize hypothesis, store finalized information in `finalized`, and change `finished` accordingly. + A sentence is finalized when {beam_size} finished items have been collected for it. + + Returns number of sentences (not beam items) being finalized. + These will be removed from the batch and not processed further. + Args: + bbsz_idx (Tensor): + """ + assert bbsz_idx.numel() == eos_scores.size(0) + + # clone relevant token and attention tensors. + # tokens is (batch * beam, max_len). So the index_select + # gets the newly EOS rows, then selects cols 1..{step + 2} + tokens_clone = tokens.index_select(0, bbsz_idx)[ + :, 1 : step + 2 + ] # skip the first index, which is EOS + + tokens_clone[:, step] = self.eos + attn_clone = ( + attn.index_select(0, bbsz_idx)[:, :, 1 : step + 2] + if attn is not None + else None + ) + + # compute scores per token position + pos_scores = scores.index_select(0, bbsz_idx)[:, : step + 1] + pos_scores[:, step, :] = eos_scores + # convert from cumulative to per-position scores + pos_scores[:, 1:] = pos_scores[:, 1:] - pos_scores[:, :-1] + + # normalize sentence-level scores + if self.normalize_scores: + eos_scores /= (step + 1) ** self.len_penalty + + # cum_unfin records which sentences in the batch are finished. + # It helps match indexing between (a) the original sentences + # in the batch and (b) the current, possibly-reduced set of + # sentences. + cum_unfin: List[int] = [] + prev = 0 + for f in finished: + if f: + prev += 1 + else: + cum_unfin.append(prev) + + # The keys here are of the form "{sent}_{unfin_idx}", where + # "unfin_idx" is the index in the current (possibly reduced) + # list of sentences, and "sent" is the index in the original, + # unreduced batch + # set() is not supported in script export + sents_seen: Dict[str, Optional[Tensor]] = {} + + # For every finished beam item + for i in range(bbsz_idx.size()[0]): + idx = bbsz_idx[i] + score = eos_scores[i].sum() + # sentence index in the current (possibly reduced) batch + unfin_idx = idx // beam_size + # sentence index in the original (unreduced) batch + sent = unfin_idx + cum_unfin[unfin_idx] + # Cannot create dict for key type '(int, int)' in torchscript. + # The workaround is to cast int to string + seen = str(sent.item()) + "_" + str(unfin_idx.item()) + if seen not in sents_seen: + sents_seen[seen] = None + + if self.match_source_len and step > src_lengths[unfin_idx]: + score = torch.tensor(-math.inf).to(score) + + # An input sentence (among those in a batch) is finished when + # beam_size hypotheses have been collected for it + if len(finalized[sent]) < beam_size: + if attn_clone is not None: + # remove padding tokens from attn scores + hypo_attn = attn_clone[i] + else: + hypo_attn = torch.empty(0) + + finalized[sent].append( + { + "tokens": tokens_clone[i], + "score": score, + "attention": hypo_attn, # src_len x tgt_len + "alignment": torch.empty(0), + "positional_scores": pos_scores[i], + } + ) + + newly_finished: List[int] = [] + + for seen in sents_seen.keys(): + # check termination conditions for this sentence + sent: int = int(float(seen.split("_")[0])) + unfin_idx: int = int(float(seen.split("_")[1])) + + if not finished[sent] and self.is_finished( + step, unfin_idx, max_len, len(finalized[sent]), beam_size + ): + finished[sent] = True + newly_finished.append(unfin_idx) + + return newly_finished + + def is_finished( + self, + step: int, + unfin_idx: int, + max_len: int, + finalized_sent_len: int, + beam_size: int, + ): + """ + Check whether decoding for a sentence is finished, which + occurs when the list of finalized sentences has reached the + beam size, or when we reach the maximum length. + """ + assert finalized_sent_len <= beam_size + if finalized_sent_len == beam_size or step == max_len: + return True + return False + + +class MultichannelEnsembleModel(nn.Module): + """A wrapper around an ensemble of SpeechDLM models.""" + + def __init__(self, models): + super().__init__() + self.models_size = len(models) + # method '__len__' is not supported in ModuleList for torch script + self.single_model = models[0] + self.models = nn.ModuleList(models) + + self.has_incremental: bool = False + if all( + hasattr(m, "decoder") and isinstance(m.decoder, FairseqIncrementalDecoder) + for m in models + ): + self.has_incremental = True + + if isinstance(models[0], SpeechDLM): + self.is_speech_dlm = True + # Otherwise it's a multi-channel language model (without cross-prediction outputs) + else: + self.is_speech_dlm = False + + if getattr(models[0].decoder.args, "duration_prediction", False): + self.is_duration_prediction = True + else: + self.is_duration_prediction = False + + def forward(self): + pass + + def has_encoder(self): + return hasattr(self.single_model, "encoder") + + def has_incremental_states(self): + return self.has_incremental + + def max_decoder_positions(self): + return min([m.max_decoder_positions() for m in self.models]) + + @torch.jit.export + def forward_encoder(self, net_input: Dict[str, Tensor]): + if not self.has_encoder(): + return None + return [model.encoder.forward_torchscript(net_input) for model in self.models] + + @torch.jit.export + def forward_decoder( + self, + tokens, + encoder_outs: List[Dict[str, List[Tensor]]], + incremental_states: List[Dict[str, Dict[str, Optional[Tensor]]]], + temperature: Dict[str, float] = 1.0, + ): + if isinstance(temperature, (float, int)): + temperature = {channel: temperature for channel in tokens} + log_probs = {channel: [] for channel in tokens} + avg_attn: Optional[Tensor] = None + encoder_out: Optional[Dict[str, List[Tensor]]] = None + for i, model in enumerate(self.models): + if self.has_encoder(): + encoder_out = encoder_outs[i] + # decode each model + if self.has_incremental_states(): + decoder_out = model.decoder.forward( + tokens, + encoder_out=encoder_out, + incremental_state=incremental_states[i], + ) + else: + decoder_out = model.decoder.forward(tokens, encoder_out=encoder_out) + + attn: Optional[Tensor] = None + decoder_len = len(decoder_out) + if decoder_len > 1 and decoder_out[1] is not None: + if isinstance(decoder_out[1], Tensor): + attn = decoder_out[1] + else: + attn_holder = decoder_out[1]["attn"] + if isinstance(attn_holder, Tensor): + attn = attn_holder + elif attn_holder is not None: + attn = attn_holder[0] + if attn is not None: + attn = attn[:, -1, :] + + if self.is_speech_dlm: + if self.is_duration_prediction: + decoder_out_divided_by_temperature = { + channel_src: { + channel_pred: { + "pred_token": decoder_out[0][channel_src][channel_pred][ + "pred_token" + ][:, -1:, :].div_(temperature[channel_pred]), + "pred_duration": decoder_out[0][channel_src][ + channel_pred + ]["pred_duration"][:, -1:, :], + } + for channel_pred in decoder_out[0][channel_src] + } + for channel_src in decoder_out[0] + } + else: + decoder_out_divided_by_temperature = { + channel_src: { + channel_pred: decoder_out[0][channel_src][channel_pred][ + :, -1:, : + ].div_(temperature[channel_pred]) + for channel_pred in decoder_out[0][channel_src] + } + for channel_src in decoder_out[0] + } + else: + decoder_out_divided_by_temperature = { + channel: decoder_out[0][channel][:, -1:, :].div_( + temperature[channel] + ) + for channel in decoder_out[0] + } + decoder_out_tuple = ( + decoder_out_divided_by_temperature, + None if decoder_len <= 1 else decoder_out[1], + ) + + probs = model.get_normalized_probs( + decoder_out_tuple, log_probs=True, sample=None + ) + + if self.is_speech_dlm: + if self.is_duration_prediction: + probs = { + channel: { + "pred_token": probs[channel][channel]["pred_token"][ + :, -1, : + ], + "pred_duration": probs[channel][channel]["pred_duration"][ + :, -1, : + ], + } + for channel in probs + } + else: + probs = { + channel: probs[channel][channel][:, -1, :] for channel in probs + } + else: + probs = {channel: probs[channel][:, -1, :] for channel in probs} + if self.models_size == 1: + return probs, attn + + for channel in probs: + log_probs[channel].append(probs[channel]) + if attn is not None: + if avg_attn is None: + avg_attn = attn + else: + avg_attn.add_(attn) + + avg_probs = {} + for channel in log_probs: + avg_probs[channel] = torch.logsumexp( + torch.stack(log_probs[channel], dim=0), dim=0 + ) - math.log(self.models_size) + + if avg_attn is not None: + avg_attn.div_(self.models_size) + return avg_probs, avg_attn + + @torch.jit.export + def reorder_encoder_out( + self, encoder_outs: Optional[List[Dict[str, List[Tensor]]]], new_order + ): + """ + Reorder encoder output according to *new_order*. + + Args: + encoder_out: output from the ``forward()`` method + new_order (LongTensor): desired order + + Returns: + *encoder_out* rearranged according to *new_order* + """ + new_outs: List[Dict[str, List[Tensor]]] = [] + if not self.has_encoder(): + return new_outs + for i, model in enumerate(self.models): + assert encoder_outs is not None + new_outs.append( + model.encoder.reorder_encoder_out(encoder_outs[i], new_order) + ) + return new_outs + + @torch.jit.export + def reorder_incremental_state( + self, + incremental_states: List[Dict[str, Dict[str, Optional[Tensor]]]], + new_order, + ): + if not self.has_incremental_states(): + return + for i, model in enumerate(self.models): + model.decoder.reorder_incremental_state_scripting( + incremental_states[i], new_order + ) diff --git a/data/fairseq/fairseq/models/speech_dlm/speech_dlm.py b/data/fairseq/fairseq/models/speech_dlm/speech_dlm.py new file mode 100644 index 0000000000000000000000000000000000000000..dc13f565f147229ff7d01c590184b14d52f308e3 --- /dev/null +++ b/data/fairseq/fairseq/models/speech_dlm/speech_dlm.py @@ -0,0 +1,280 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import logging +from dataclasses import dataclass, field +from typing import Optional + +from fairseq import utils +from fairseq.dataclass import ChoiceEnum, FairseqDataclass +from fairseq.models import ( + FairseqLanguageModel, + register_model, + register_model_architecture, +) +from fairseq.models.transformer import Embedding +from .modules.speech_dlm_decoder import CrossChannelTransformerDecoder +from omegaconf import II + + +DEFAULT_MAX_TARGET_POSITIONS = 1024 + +logger = logging.getLogger(__name__) + + +@dataclass +class SpeechDLMConfig(FairseqDataclass): + activation_fn: ChoiceEnum(utils.get_available_activation_fns()) = field( + default="relu", metadata={"help": "activation function to use"} + ) + dropout: float = field(default=0.1, metadata={"help": "dropout probability"}) + attention_dropout: float = field( + default=0.0, metadata={"help": "dropout probability for attention weights"} + ) + activation_dropout: float = field( + default=0.0, metadata={"help": "dropout probability after activation in FFN."} + ) + relu_dropout: float = field( + default=0.0, metadata={"help": "dropout probability after activation in FFN."} + ) + decoder_embed_dim: int = field( + default=512, metadata={"help": "decoder embedding dimension"} + ) + decoder_output_dim: int = field( + default=512, metadata={"help": "decoder output dimension"} + ) + decoder_input_dim: int = field( + default=512, metadata={"help": "decoder input dimension"} + ) + decoder_ffn_embed_dim: int = field( + default=2048, metadata={"help": "decoder embedding dimension for FFN"} + ) + decoder_layers: int = field(default=6, metadata={"help": "num decoder layers"}) + decoder_cross_layers: int = field( + default=-1, metadata={"help": "num self cross attention decoder layers"} + ) + decoder_attention_heads: int = field( + default=8, metadata={"help": "num decoder attention heads"} + ) + decoder_normalize_before: bool = field( + default=False, metadata={"help": "apply layernorm before each decoder block"} + ) + no_decoder_final_norm: bool = field( + default=False, + metadata={"help": "don't add an extra layernorm after the last decoder block"}, + ) + no_token_positional_embeddings: bool = field( + default=False, + metadata={ + "help": "if set, disables positional embeddings (outside self attention)" + }, + ) + share_decoder_input_output_embed: bool = field( + default=False, metadata={"help": "share decoder input and output embeddings"} + ) + decoder_learned_pos: bool = field( + default=False, + metadata={"help": "use learned positional embeddings in the decoder"}, + ) + decoder_layerdrop: float = field( + default=0.0, metadata={"help": "LayerDrop probability for decoder"} + ) + decoder_layers_to_keep: Optional[str] = field( + default=None, + metadata={ + "help": "which layers to *keep* when pruning as a comma-separated list" + }, + ) + layernorm_embedding: bool = field( + default=False, metadata={"help": "add layernorm to embedding"} + ) + no_scale_embedding: bool = field( + default=False, metadata={"help": "if True, dont scale embeddings"} + ) + checkpoint_activations: bool = field( + default=False, metadata={"help": "checkpoint activations at each layer"} + ) + offload_activations: bool = field( + default=False, + metadata={"help": "move checkpointed activations to CPU after they are used."}, + ) + quant_noise_pq: float = field( + default=0.0, + metadata={"help": "iterative PQ quantization noise at training time"}, + ) + quant_noise_pq_block_size: int = field( + default=8, + metadata={"help": "block size of quantization noise at training time"}, + ) + # TODO common var add to parent + quant_noise_scalar: float = field( + default=0.0, + metadata={ + "help": "scalar quantization noise and scalar quantization at training time" + }, + ) + add_bos_token: bool = II("task.add_bos_token") + tokens_per_sample: int = II("task.tokens_per_sample") + max_target_positions: Optional[int] = II("task.max_target_positions") + tpu: bool = II("common.tpu") + duration_prediction: str = II("task.duration_prediction") + delayed_duration_target: str = II("task.delayed_duration_target") + main_and_cross_weights: str = II("criterion.main_and_cross_weights") + + +@register_model("speech_dlm", dataclass=SpeechDLMConfig) +class SpeechDLM(FairseqLanguageModel): + """Spoken Unit-based Dialogue Language Model model (SpeechDLM) as described + in the paper: https://arxiv.org/pdf/2203.16502.pdf + """ + + def __init__(self, decoder): + super().__init__(decoder) + + @classmethod + def build_model(cls, args, task): + """Build a new model instance.""" + # make sure all arguments are present in older models + base_lm_architecture(args) + + if args.decoder_layers_to_keep: + args.decoder_layers = len(args.decoder_layers_to_keep.split(",")) + + if args.decoder_cross_layers < 0: + args.decoder_cross_layers = args.decoder_layers + + if getattr(args, "max_target_positions", None) is None: + args.max_target_positions = getattr( + args, "tokens_per_sample", DEFAULT_MAX_TARGET_POSITIONS + ) + + # Assert all dictionary to be the same + assert all( + task.source_dictionaries[channel] == task.source_dictionary + for channel in task.channels + ), "Source dictionaries of all channels are expected to be the same!!!" + assert all( + task.target_dictionaries[channel] == task.target_dictionary + for channel in task.channels + ), "Target dictionaries of all channels are expected to be the same!!!" + # Build the unit embeddings + embed_tokens = cls.build_embedding( + args, task.source_dictionary, args.decoder_input_dim + ) + + decoder = CrossChannelTransformerDecoder( + args, + task.target_dictionary, + embed_tokens, + channels=task.channels, + no_encoder_attn=True, + ) + return cls(decoder) + + @classmethod + def build_embedding(cls, args, dictionary, embed_dim, path=None): + embed_tokens = Embedding(len(dictionary), embed_dim, dictionary.pad()) + return embed_tokens + + @classmethod + def from_pretrained( + cls, + model_name_or_path, + checkpoint_file="model.pt", + data_name_or_path=".", + **kwargs, + ): + """ + Load a :class:`~fairseq.models.FairseqModel` from a pre-trained model + file. Downloads and caches the pre-trained model file if needed. + + The base implementation returns a + :class:`~fairseq.hub_utils.GeneratorHubInterface`, which can be used to + generate translations or sample from language models. The underlying + :class:`~fairseq.models.FairseqModel` can be accessed via the + *generator.models* attribute. + + This function return a class:`MultichannelGeneratorHubInterface` object, + which allows generation in multiple channels with a multichannel model. + + Args: + model_name_or_path (str): either the name of a pre-trained model to + load or a path/URL to a pre-trained model state dict + checkpoint_file (str, optional): colon-separated list of checkpoint + files in the model archive to ensemble (default: 'model.pt') + data_name_or_path (str, optional): point args.data to the archive + at the given path/URL. Can start with '.' or './' to reuse the + model archive path. + """ + from fairseq import hub_utils + from .hub_interface import MultichannelGeneratorHubInterface + + x = hub_utils.from_pretrained( + model_name_or_path, + checkpoint_file, + data_name_or_path, + archive_map=cls.hub_models(), + **kwargs, + ) + logger.info(x["args"]) + return MultichannelGeneratorHubInterface(x["args"], x["task"], x["models"]) + + @property + def supported_targets(self): + return {"next", "edge", "duration"} + + +def base_lm_architecture(args): + # backward compatibility for older model checkpoints + if hasattr(args, "decoder_final_norm"): + args.no_decoder_final_norm = not args.decoder_final_norm + + args.dropout = getattr(args, "dropout", 0.1) + args.attention_dropout = getattr(args, "attention_dropout", 0.0) + + args.decoder_embed_dim = getattr(args, "decoder_embed_dim", 512) + args.decoder_ffn_embed_dim = getattr(args, "decoder_ffn_embed_dim", 2048) + args.decoder_layers = getattr(args, "decoder_layers", 6) + args.decoder_cross_layers = getattr(args, "decoder_cross_layers", 6) + args.decoder_attention_heads = getattr(args, "decoder_attention_heads", 8) + args.decoder_learned_pos = getattr(args, "decoder_learned_pos", False) + args.activation_fn = getattr(args, "activation_fn", "relu") + args.decoder_layerdrop = getattr(args, "decoder_layerdrop", 0) + args.decoder_layers_to_keep = getattr(args, "decoder_layers_to_keep", None) + args.quant_noise_pq = getattr(args, "quant_noise_pq", 0) + args.quant_noise_pq_block_size = getattr(args, "quant_noise_pq_block_size", 8) + args.quant_noise_scalar = getattr(args, "quant_noise_scalar", 0) + + args.add_bos_token = getattr(args, "add_bos_token", False) + args.no_token_positional_embeddings = getattr( + args, "no_token_positional_embeddings", False + ) + args.share_decoder_input_output_embed = getattr( + args, "share_decoder_input_output_embed", False + ) + args.decoder_output_dim = getattr( + args, "decoder_output_dim", args.decoder_embed_dim + ) + args.decoder_input_dim = getattr(args, "decoder_input_dim", args.decoder_embed_dim) + + # Model training is not stable without this + args.decoder_normalize_before = True + args.no_decoder_final_norm = getattr(args, "no_decoder_final_norm", False) + args.no_scale_embedding = getattr(args, "no_scale_embedding", False) + args.layernorm_embedding = getattr(args, "layernorm_embedding", False) + args.checkpoint_activations = getattr(args, "checkpoint_activations", False) + args.offload_activations = getattr(args, "offload_activations", False) + if args.offload_activations: + args.checkpoint_activations = True + + +@register_model_architecture("speech_dlm", "speech_dlm_big") +def speech_dlm_big(args): + args.decoder_layers = getattr(args, "decoder_layers", 12) + args.decoder_cross_layers = getattr(args, "decoder_cross_layers", 12) + args.decoder_embed_dim = getattr(args, "decoder_embed_dim", 1024) + args.decoder_ffn_embed_dim = getattr(args, "decoder_ffn_embed_dim", 4096) + args.decoder_attention_heads = getattr(args, "decoder_attention_heads", 16) + base_lm_architecture(args) diff --git a/data/fairseq/fairseq/models/speech_to_speech/__init__.py b/data/fairseq/fairseq/models/speech_to_speech/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..f29215c2fe6eedd203d105703ed94576c625ba86 --- /dev/null +++ b/data/fairseq/fairseq/models/speech_to_speech/__init__.py @@ -0,0 +1,9 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from .s2s_conformer import * # noqa +from .s2s_conformer_translatotron2 import * # noqa +from .s2s_conformer_unity import * # noqa +from .s2s_transformer import * # noqa diff --git a/data/fairseq/fairseq/models/speech_to_speech/__pycache__/__init__.cpython-38.pyc b/data/fairseq/fairseq/models/speech_to_speech/__pycache__/__init__.cpython-38.pyc new file mode 100644 index 0000000000000000000000000000000000000000..5ca6f19096489998b2c5ac2a47e54c45f504c918 Binary files /dev/null and b/data/fairseq/fairseq/models/speech_to_speech/__pycache__/__init__.cpython-38.pyc differ diff --git a/data/fairseq/fairseq/models/speech_to_speech/__pycache__/s2s_conformer.cpython-38.pyc b/data/fairseq/fairseq/models/speech_to_speech/__pycache__/s2s_conformer.cpython-38.pyc new file mode 100644 index 0000000000000000000000000000000000000000..f1c64c2ceb8e08eed8d60d0c5bea94e573748a2a Binary files /dev/null and b/data/fairseq/fairseq/models/speech_to_speech/__pycache__/s2s_conformer.cpython-38.pyc differ diff --git a/data/fairseq/fairseq/models/speech_to_speech/__pycache__/s2s_conformer_translatotron2.cpython-38.pyc b/data/fairseq/fairseq/models/speech_to_speech/__pycache__/s2s_conformer_translatotron2.cpython-38.pyc new file mode 100644 index 0000000000000000000000000000000000000000..6faea8a1fa804db8c615cff311deae41347fd093 Binary files /dev/null and b/data/fairseq/fairseq/models/speech_to_speech/__pycache__/s2s_conformer_translatotron2.cpython-38.pyc differ diff --git a/data/fairseq/fairseq/models/speech_to_speech/__pycache__/s2s_conformer_unity.cpython-38.pyc b/data/fairseq/fairseq/models/speech_to_speech/__pycache__/s2s_conformer_unity.cpython-38.pyc new file mode 100644 index 0000000000000000000000000000000000000000..b7965e319b669608cc437263f52efb37464cf46a Binary files /dev/null and b/data/fairseq/fairseq/models/speech_to_speech/__pycache__/s2s_conformer_unity.cpython-38.pyc differ diff --git a/data/fairseq/fairseq/models/speech_to_speech/__pycache__/s2s_transformer.cpython-38.pyc b/data/fairseq/fairseq/models/speech_to_speech/__pycache__/s2s_transformer.cpython-38.pyc new file mode 100644 index 0000000000000000000000000000000000000000..c0550057055a9b15a92e9182815ce65af3e770f7 Binary files /dev/null and b/data/fairseq/fairseq/models/speech_to_speech/__pycache__/s2s_transformer.cpython-38.pyc differ diff --git a/data/fairseq/fairseq/models/speech_to_speech/modules/__init__.py b/data/fairseq/fairseq/models/speech_to_speech/modules/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391 diff --git a/data/fairseq/fairseq/models/speech_to_speech/modules/__pycache__/__init__.cpython-38.pyc b/data/fairseq/fairseq/models/speech_to_speech/modules/__pycache__/__init__.cpython-38.pyc new file mode 100644 index 0000000000000000000000000000000000000000..43c871eb2782ef14b08ca5ac987c01bbf8c27eb1 Binary files /dev/null and b/data/fairseq/fairseq/models/speech_to_speech/modules/__pycache__/__init__.cpython-38.pyc differ diff --git a/data/fairseq/fairseq/models/speech_to_speech/modules/__pycache__/ctc_decoder.cpython-38.pyc b/data/fairseq/fairseq/models/speech_to_speech/modules/__pycache__/ctc_decoder.cpython-38.pyc new file mode 100644 index 0000000000000000000000000000000000000000..978106d3c4128e949b0c0cee514c5eb11cf60031 Binary files /dev/null and b/data/fairseq/fairseq/models/speech_to_speech/modules/__pycache__/ctc_decoder.cpython-38.pyc differ diff --git a/data/fairseq/fairseq/models/speech_to_speech/modules/__pycache__/stacked_embedding.cpython-38.pyc b/data/fairseq/fairseq/models/speech_to_speech/modules/__pycache__/stacked_embedding.cpython-38.pyc new file mode 100644 index 0000000000000000000000000000000000000000..27f181efb64c539c6ebe08820936ef28e34b3a68 Binary files /dev/null and b/data/fairseq/fairseq/models/speech_to_speech/modules/__pycache__/stacked_embedding.cpython-38.pyc differ diff --git a/data/fairseq/fairseq/models/speech_to_speech/modules/__pycache__/transformer_decoder_aug.cpython-38.pyc b/data/fairseq/fairseq/models/speech_to_speech/modules/__pycache__/transformer_decoder_aug.cpython-38.pyc new file mode 100644 index 0000000000000000000000000000000000000000..16df731bac74aec9f5c6a6e7b2e69ee42bbe7677 Binary files /dev/null and b/data/fairseq/fairseq/models/speech_to_speech/modules/__pycache__/transformer_decoder_aug.cpython-38.pyc differ diff --git a/data/fairseq/fairseq/models/speech_to_speech/modules/__pycache__/transformer_encoder.cpython-38.pyc b/data/fairseq/fairseq/models/speech_to_speech/modules/__pycache__/transformer_encoder.cpython-38.pyc new file mode 100644 index 0000000000000000000000000000000000000000..c153a4cf6cc222becdea48fc8839e519fe7e629f Binary files /dev/null and b/data/fairseq/fairseq/models/speech_to_speech/modules/__pycache__/transformer_encoder.cpython-38.pyc differ diff --git a/data/fairseq/fairseq/models/speech_to_speech/modules/ctc_decoder.py b/data/fairseq/fairseq/models/speech_to_speech/modules/ctc_decoder.py new file mode 100644 index 0000000000000000000000000000000000000000..721efbf61ae335e13fbf40a2fecfb5bc97f9638a --- /dev/null +++ b/data/fairseq/fairseq/models/speech_to_speech/modules/ctc_decoder.py @@ -0,0 +1,18 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from torch import nn + +from fairseq.models import FairseqEncoder + + +class CTCDecoder(FairseqEncoder): + def __init__(self, dictionary, in_dim): + super().__init__(dictionary) + self.proj = nn.Linear(in_dim, len(dictionary)) + + def forward(self, src_tokens, src_lengths=None, **kwargs): + encoder_out = self.proj(src_tokens) + return {"encoder_out": encoder_out} diff --git a/data/fairseq/fairseq/models/speech_to_speech/modules/stacked_embedding.py b/data/fairseq/fairseq/models/speech_to_speech/modules/stacked_embedding.py new file mode 100644 index 0000000000000000000000000000000000000000..5955a08538f54eaac871d222378a0ed86071581f --- /dev/null +++ b/data/fairseq/fairseq/models/speech_to_speech/modules/stacked_embedding.py @@ -0,0 +1,48 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import torch +from torch import nn + +from fairseq.models.transformer import Linear + + +class StackedEmbedding(nn.Embedding): + """Embedding module that supports stacked units -> single embedding""" + + def __init__(self, num_embeddings, embed_dim, padding_idx, num_stacked=1): + super().__init__(num_embeddings, embed_dim, padding_idx) + # follow transformer.Embedding + nn.init.normal_(self.weight, mean=0, std=embed_dim**-0.5) + nn.init.constant_(self.weight[padding_idx], 0) + + self.offset = ( + 4 # skip , , , , specific to fairseq dictionary + ) + self.vocab_size = num_embeddings - self.offset + self.num_stacked = num_stacked + + if self.num_stacked > 1: + self.project_in_dim = Linear(embed_dim * num_stacked, embed_dim, bias=False) + + def forward(self, input): + if self.num_stacked == 1: + return super().forward(input) + + # expand input indices + mask = input >= self.offset + stacked_input = [] + cum_input = input.new_zeros(input.shape) + for i in range(1, self.num_stacked + 1): + div = pow(self.vocab_size, i) + next_input = torch.remainder(input - self.offset - cum_input, div) + cum_input += next_input + next_input = torch.floor_divide(next_input, div // self.vocab_size) + stacked_input.append((next_input + self.offset) * mask + input * ~mask) + + stacked_input = torch.stack(stacked_input[::-1], dim=2) + embed = super().forward(stacked_input).view(input.size(0), input.size(1), -1) + embed = self.project_in_dim(embed) + return embed diff --git a/data/fairseq/fairseq/models/speech_to_speech/modules/transformer_decoder_aug.py b/data/fairseq/fairseq/models/speech_to_speech/modules/transformer_decoder_aug.py new file mode 100644 index 0000000000000000000000000000000000000000..68f42c2b3633a330a605ed169fd527cded73d9f0 --- /dev/null +++ b/data/fairseq/fairseq/models/speech_to_speech/modules/transformer_decoder_aug.py @@ -0,0 +1,108 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from typing import Any, Dict, List, Optional + +from torch import Tensor + +from fairseq.models.transformer import Linear +from fairseq.models.transformer.transformer_decoder_aug import AugTransformerDecoder + + +class AugTransformerUnitDecoder(AugTransformerDecoder): + """Based on Transformer decoder, with support to decoding stacked units""" + + def __init__( + self, + args, + dictionary, + embed_tokens, + no_encoder_attn=False, + output_projection=None, + ): + super().__init__( + args, dictionary, embed_tokens, no_encoder_attn, output_projection + ) + self.n_frames_per_step = args.n_frames_per_step + + self.out_proj_n_frames = ( + Linear( + self.output_embed_dim, + self.output_embed_dim * self.n_frames_per_step, + bias=False, + ) + if self.n_frames_per_step > 1 + else None + ) + + def forward( + self, + prev_output_tokens, + encoder_out: Optional[Dict[str, List[Tensor]]] = None, + encoder_out_aug: Optional[Dict[str, List[Tensor]]] = None, + incremental_state: Optional[Dict[str, Dict[str, Optional[Tensor]]]] = None, + features_only: bool = False, + full_context_alignment: bool = False, + alignment_layer: Optional[int] = None, + alignment_heads: Optional[int] = None, + src_lengths: Optional[Any] = None, + return_all_hiddens: bool = False, + ): + """ + Args: + prev_output_tokens (LongTensor): previous decoder outputs of shape + `(batch, tgt_len)`, for teacher forcing + encoder_out (optional): output from the encoder, used for + encoder-side attention, should be of size T x B x C + incremental_state (dict): dictionary used for storing state during + :ref:`Incremental decoding` + features_only (bool, optional): only return features without + applying output layer (default: False). + full_context_alignment (bool, optional): don't apply + auto-regressive mask to self-attention (default: False). + + Returns: + tuple: + - the decoder's output of shape `(batch, tgt_len, vocab)` + - a dictionary with any model-specific outputs + """ + + x, extra = self.extract_features( + prev_output_tokens, + encoder_out=encoder_out, + encoder_out_aug=encoder_out_aug, + incremental_state=incremental_state, + full_context_alignment=full_context_alignment, + alignment_layer=alignment_layer, + alignment_heads=alignment_heads, + ) + + if not features_only: + bsz, seq_len, d = x.size() + if self.out_proj_n_frames: + x = self.out_proj_n_frames(x) + x = self.output_layer(x.view(bsz, seq_len, self.n_frames_per_step, d)) + x = x.view(bsz, seq_len * self.n_frames_per_step, -1) + if ( + incremental_state is None and self.n_frames_per_step > 1 + ): # teacher-forcing mode in training + x = x[ + :, : -(self.n_frames_per_step - 1), : + ] # remove extra frames after + + return x, extra + + def upgrade_state_dict_named(self, state_dict, name): + if self.n_frames_per_step > 1: + move_keys = [ + ( + f"{name}.project_in_dim.weight", + f"{name}.embed_tokens.project_in_dim.weight", + ) + ] + for from_k, to_k in move_keys: + if from_k in state_dict and to_k not in state_dict: + state_dict[to_k] = state_dict[from_k] + del state_dict[from_k] diff --git a/data/fairseq/fairseq/models/speech_to_speech/modules/transformer_encoder.py b/data/fairseq/fairseq/models/speech_to_speech/modules/transformer_encoder.py new file mode 100644 index 0000000000000000000000000000000000000000..fb1af433d81972d2190218c8b4fd18ac2e946150 --- /dev/null +++ b/data/fairseq/fairseq/models/speech_to_speech/modules/transformer_encoder.py @@ -0,0 +1,85 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import torch.nn as nn + +from fairseq.models import FairseqEncoder +from fairseq.modules import LayerNorm, TransformerEncoderLayer + + +class TransformerEncoderNoEmb(FairseqEncoder): + """Transformer encoder without token embeddings.""" + + def __init__(self, args): + super().__init__(None) + + self.layers = nn.ModuleList( + [TransformerEncoderLayer(args) for _ in range(args.encoder_layers)] + ) + if args.encoder_normalize_before: + self.layer_norm = LayerNorm(args.encoder_embed_dim) + else: + self.layer_norm = None + + def forward(self, x, encoder_padding_mask, return_all_hiddens=False): + + encoder_states = [] + + for layer in self.layers: + x = layer(x, encoder_padding_mask) + if return_all_hiddens: + encoder_states.append(x) + + if self.layer_norm is not None: + x = self.layer_norm(x) + + return { + "encoder_out": [x], # T x B x C + "encoder_padding_mask": [encoder_padding_mask] + if encoder_padding_mask is not None and encoder_padding_mask.any() + else [], # B x T + "encoder_embedding": [], # B x T x C + "encoder_states": encoder_states, # List[T x B x C] + "src_tokens": [], + "src_lengths": [], + } + + def reorder_encoder_out(self, encoder_out, new_order): + new_encoder_out = ( + [] + if len(encoder_out["encoder_out"]) == 0 + else [x.index_select(1, new_order) for x in encoder_out["encoder_out"]] + ) + + new_encoder_padding_mask = ( + [] + if len(encoder_out["encoder_padding_mask"]) == 0 + else [ + x.index_select(0, new_order) + for x in encoder_out["encoder_padding_mask"] + ] + ) + + new_encoder_embedding = ( + [] + if len(encoder_out["encoder_embedding"]) == 0 + else [ + x.index_select(0, new_order) for x in encoder_out["encoder_embedding"] + ] + ) + + encoder_states = encoder_out["encoder_states"] + if len(encoder_states) > 0: + for idx, state in enumerate(encoder_states): + encoder_states[idx] = state.index_select(1, new_order) + + return { + "encoder_out": new_encoder_out, # T x B x C + "encoder_padding_mask": new_encoder_padding_mask, # B x T + "encoder_embedding": new_encoder_embedding, # B x T x C + "encoder_states": encoder_states, # List[T x B x C] + "src_tokens": [], # B x T + "src_lengths": [], # B x 1 + } diff --git a/data/fairseq/fairseq/models/speech_to_speech/s2s_conformer.py b/data/fairseq/fairseq/models/speech_to_speech/s2s_conformer.py new file mode 100644 index 0000000000000000000000000000000000000000..636396d536689772e44c29506e4a49b683562f37 --- /dev/null +++ b/data/fairseq/fairseq/models/speech_to_speech/s2s_conformer.py @@ -0,0 +1,172 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import logging +from pathlib import Path + +import torch + +from fairseq import checkpoint_utils +from fairseq.models import register_model, register_model_architecture +from fairseq.models.speech_to_speech.s2s_transformer import ( + S2SpecTTransformerModel, + S2UTTransformerModel, + s2spect_architecture_base, + s2ut_architecture_base, +) +from fairseq.models.speech_to_text import S2TConformerEncoder +from fairseq.models.transformer import Linear + +logger = logging.getLogger(__name__) + + +def build_s2s_conformer_encoder(args): + encoder = S2SConformerEncoder(args) + pretraining_path = getattr(args, "load_pretrained_encoder_from", None) + if pretraining_path is not None: + if not Path(pretraining_path).exists(): + logger.warning( + f"skipped pretraining because {pretraining_path} does not exist" + ) + else: + encoder = checkpoint_utils.load_pretrained_component_from_model( + component=encoder, checkpoint=pretraining_path + ) + logger.info(f"loaded pretrained encoder from: {pretraining_path}") + return encoder + + +class S2SConformerEncoder(S2TConformerEncoder): + """Based on S2T transformer encoder, with support + to incorporate target speaker embedding.""" + + def __init__(self, args): + super().__init__(args) + + self.spk_emb_proj = None + if args.target_speaker_embed: + self.spk_emb_proj = Linear( + args.encoder_embed_dim + args.speaker_embed_dim, args.encoder_embed_dim + ) + + def forward( + self, src_tokens, src_lengths, tgt_speaker=None, return_all_hiddens=False + ): + out = super().forward(src_tokens, src_lengths, return_all_hiddens) + + if self.spk_emb_proj: + x = out["encoder_out"][0] + seq_len, bsz, _ = x.size() + tgt_speaker_emb = tgt_speaker.view(1, bsz, -1).expand(seq_len, bsz, -1) + x = self.spk_emb_proj(torch.cat([x, tgt_speaker_emb], dim=2)) + out["encoder_out"][0] = x + + return out + + +@register_model("s2ut_conformer") +class S2UTConformerModel(S2UTTransformerModel): + """ + Direct speech-to-speech translation model with Conformer encoder + Transformer discrete unit decoder + """ + + @staticmethod + def add_args(parser): + S2UTTransformerModel.add_args(parser) + parser.add_argument( + "--depthwise-conv-kernel-size", + type=int, + metavar="N", + help="kernel size of depthwise convolution layers", + ) + parser.add_argument( + "--attn-type", + type=str, + metavar="STR", + help="If not specified uses fairseq MHA. Other valid option is espnet for using conformer", + ) + parser.add_argument( + "--pos-enc-type", + type=str, + metavar="STR", + help="Must be specified in addition to attn-type=espnet for rel_pos and rope", + ) + + @classmethod + def build_encoder(cls, args): + return build_s2s_conformer_encoder(args) + + +@register_model("s2spect_conformer") +class S2SpecTConformerModel(S2SpecTTransformerModel): + """ + Direct speech-to-speech translation model with Conformer encoder + TTS Transformer decoder + """ + + @staticmethod + def add_args(parser): + S2SpecTTransformerModel.add_args(parser) + parser.add_argument("--depthwise-conv-kernel-size", type=int, default=31) + parser.add_argument( + "--attn-type", + type=str, + default=None, + help="If not specified uses fairseq MHA. Other valid option is espnet for using conformer", + ) + parser.add_argument( + "--pos-enc-type", + type=str, + default="abs", + help="Must be specified in addition to attn-type=espnet for rel_pos and rope", + ) + + @classmethod + def build_encoder(cls, args): + return build_s2s_conformer_encoder(args) + + +@register_model_architecture("s2ut_conformer", "s2ut_conformer") +def s2ut_conformer_architecture_base(args): + args.attn_type = getattr(args, "attn_type", None) + args.pos_enc_type = getattr(args, "pos_enc_type", "abs") + args.input_feat_per_channel = getattr(args, "input_feat_per_channel", 80) + args.input_channels = getattr(args, "input_channels", 1) + args.max_source_positions = getattr(args, "max_source_positions", 6000) + args.encoder_embed_dim = getattr(args, "encoder_embed_dim", 256) + args.encoder_ffn_embed_dim = getattr(args, "encoder_ffn_embed_dim", 2048) + args.encoder_attention_heads = getattr(args, "encoder_attention_heads", 4) + args.dropout = getattr(args, "dropout", 0.1) + args.encoder_layers = getattr(args, "encoder_layers", 16) + args.depthwise_conv_kernel_size = getattr(args, "depthwise_conv_kernel_size", 31) + s2ut_architecture_base(args) + + +@register_model_architecture("s2spect_conformer", "s2spect_conformer") +def s2spect_conformer_architecture_base(args): + args.attn_type = getattr(args, "attn_type", None) + args.pos_enc_type = getattr(args, "pos_enc_type", "abs") + args.input_feat_per_channel = getattr(args, "input_feat_per_channel", 80) + args.input_channels = getattr(args, "input_channels", 1) + args.max_source_positions = getattr(args, "max_source_positions", 6000) + args.encoder_embed_dim = getattr(args, "encoder_embed_dim", 256) + args.encoder_ffn_embed_dim = getattr(args, "encoder_ffn_embed_dim", 2048) + args.encoder_attention_heads = getattr(args, "encoder_attention_heads", 4) + args.dropout = getattr(args, "dropout", 0.1) + args.encoder_layers = getattr(args, "encoder_layers", 16) + args.depthwise_conv_kernel_size = getattr(args, "depthwise_conv_kernel_size", 31) + s2spect_architecture_base(args) + + +@register_model_architecture("s2spect_conformer", "s2spect_conformer_fisher") +def s2spect_architecture_fisher(args): + args.encoder_embed_dim = getattr(args, "encoder_embed_dim", 256) + args.encoder_ffn_embed_dim = getattr(args, "encoder_ffn_embed_dim", 256 * 8) + args.encoder_attention_heads = getattr(args, "encoder_attention_heads", 4) + args.dropout = getattr(args, "dropout", 0.1) + + # decoder + args.prenet_dim = getattr(args, "prenet_dim", 32) + + s2spect_conformer_architecture_base(args) diff --git a/data/fairseq/fairseq/models/speech_to_speech/s2s_conformer_translatotron2.py b/data/fairseq/fairseq/models/speech_to_speech/s2s_conformer_translatotron2.py new file mode 100644 index 0000000000000000000000000000000000000000..8016daee8d62c1857307807cb925d4bb26aac6ec --- /dev/null +++ b/data/fairseq/fairseq/models/speech_to_speech/s2s_conformer_translatotron2.py @@ -0,0 +1,262 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import copy +import logging + +from fairseq.models import ( + FairseqEncoderModel, + FairseqLanguageModel, + register_model, + register_model_architecture, +) +from fairseq.models.speech_to_speech.modules.ctc_decoder import CTCDecoder +from fairseq.models.speech_to_speech.modules.transformer_encoder import ( + TransformerEncoderNoEmb, +) +from fairseq.models.speech_to_speech.s2s_conformer import S2SpecTConformerModel +from fairseq.models.speech_to_speech.s2s_conformer_unity import ( + multitask_text_transformer_decoder_arch, +) +from fairseq.models.speech_to_speech.s2s_transformer import ( + base_multitask_text_transformer_decoder_arch, + s2spect_architecture_base, +) +from fairseq.models.text_to_speech import TTSTransformerDecoder +from fairseq.models.transformer import TransformerDecoder, TransformerModelBase + +logger = logging.getLogger(__name__) + + +@register_model("s2spect2_conformer") +class S2SpecT2ConformerModel(S2SpecTConformerModel): + """ + Direct speech-to-speech translation model with Conformer encoder + MT Transformer decoder + TTS Transformer decoder + """ + + @staticmethod + def add_args(parser): + S2SpecTConformerModel.add_args(parser) + parser.add_argument( + "--translation-decoder-layers", + type=int, + default=4, + metavar="N", + help="num decoder layers in the first-pass translation module", + ) + parser.add_argument( + "--synthesizer", + default="transformer", + choices=["transformer"], + help="", + ) + parser.add_argument( + "--synthesizer-encoder-layers", + type=int, + default=0, + metavar="N", + help="num encoder layers in the second-pass synthesizer module", + ) + + @classmethod + def build_multitask_decoder( + cls, + args, + tgt_dict, + in_dim, + is_mt_decoder, + decoder_layers, + decoder_embed_dim, + decoder_attention_heads, + ): + decoder_args = args.decoder_args + decoder_args.encoder_embed_dim = in_dim + if args.decoder_type == "transformer": + if is_mt_decoder: + multitask_text_transformer_decoder_arch( + decoder_args, + decoder_layers, + decoder_embed_dim, + decoder_attention_heads, + ) # 4L + else: + base_multitask_text_transformer_decoder_arch(decoder_args) # 2L + task_decoder = TransformerDecoder( + decoder_args, + tgt_dict, + embed_tokens=TransformerModelBase.build_embedding( + decoder_args, + tgt_dict, + decoder_args.decoder_embed_dim, + ), + ) + elif args.decoder_type == "ctc": + task_decoder = CTCDecoder( + dictionary=tgt_dict, + in_dim=in_dim, + ) + else: + raise NotImplementedError( + "currently only support multitask decoder_type 'transformer', 'ctc'" + ) + + return task_decoder + + @classmethod + def build_decoder(cls, args): + _args = copy.deepcopy(args) + _args.encoder_embed_dim = args.decoder_embed_dim + + if args.synthesizer == "transformer": + return TTSTransformerDecoder(_args, None, padding_idx=1) + else: + raise NotImplementedError(args.synthesizer) + + @classmethod + def build_model(cls, args, task): + encoder = cls.build_encoder(args) + decoder = cls.build_decoder(args) + base_model = cls(encoder, decoder) + + # set up multitask decoders + base_model.mt_task_name = None + base_model.multitask_decoders = {} + has_first_pass_decoder = False + for task_name, task_obj in task.multitask_tasks.items(): + if task_obj.is_first_pass_decoder: + has_first_pass_decoder = True + base_model.mt_task_name = task_name + + in_dim = ( + args.encoder_embed_dim + if task_obj.args.input_from == "encoder" + else args.decoder_embed_dim + ) + task_decoder = cls.build_multitask_decoder( + task_obj.args, + task_obj.target_dictionary, + in_dim, + task_obj.is_first_pass_decoder, + getattr(args, "translation_decoder_layers", 4), + getattr(args, "decoder_embed_dim", 256), + getattr(args, "decoder_attention_heads", 4), + ) + + setattr(base_model, f"{task_name}_decoder", task_decoder) + decoder_model_cls = ( + FairseqEncoderModel + if task_obj.args.decoder_type == "ctc" + else FairseqLanguageModel + ) + base_model.multitask_decoders[task_name] = decoder_model_cls( + getattr(base_model, f"{task_name}_decoder") + ) + + assert has_first_pass_decoder, "set at least one intermediate non-CTC decoder" + + # set up encoder on top of the auxiliary MT decoder + if getattr(args, "synthesizer_encoder_layers", 0) > 0: + base_model.synthesizer_encoder = cls.build_text_encoder(args) + else: + base_model.synthesizer_encoder = None + + return base_model + + @classmethod + def build_text_encoder(cls, args): + _args = copy.deepcopy(args) + _args.encoder_layers = args.synthesizer_encoder_layers + _args.encoder_embed_dim = args.decoder_embed_dim + _args.encoder_ffn_embed_dim = args.decoder_ffn_embed_dim + _args.encoder_attention_heads = args.decoder_attention_heads + _args.encoder_normalize_before = True + return TransformerEncoderNoEmb(_args) + + def forward( + self, + src_tokens, + src_lengths, + prev_output_tokens, + prev_output_tokens_mt, + tgt_speaker=None, + incremental_state=None, + target_lengths=None, + speaker=None, + return_all_hiddens=False, + ): + encoder_out = self.encoder( + src_tokens, + src_lengths=src_lengths, + tgt_speaker=tgt_speaker, + return_all_hiddens=return_all_hiddens, + ) + + # 1. MT decoder + mt_decoder = getattr(self, f"{self.mt_task_name}_decoder") + mt_decoder_out = mt_decoder( + prev_output_tokens_mt, + encoder_out=encoder_out, + ) + x = mt_decoder_out[1]["inner_states"][-1] + if mt_decoder.layer_norm is not None: + x = mt_decoder.layer_norm(x) + + mt_decoder_padding_mask = None + if prev_output_tokens_mt.eq(mt_decoder.padding_idx).any(): + mt_decoder_padding_mask = prev_output_tokens_mt.eq(mt_decoder.padding_idx) + + # 2. TTS encoder + if self.synthesizer_encoder is not None: + tts_encoder_out = self.synthesizer_encoder( + x, + mt_decoder_padding_mask, + return_all_hiddens=return_all_hiddens, + ) + else: + tts_encoder_out = { + "encoder_out": [x], # T x B x C + "encoder_padding_mask": [mt_decoder_padding_mask], # B x T + } + + # 3. TTS decoder + decoder_out = self.decoder( + prev_output_tokens, + encoder_out=tts_encoder_out, + incremental_state=incremental_state, + target_lengths=target_lengths, + speaker=speaker, + ) + if return_all_hiddens: + decoder_out[-1]["encoder_states"] = encoder_out["encoder_states"] + decoder_out[-1]["encoder_padding_mask"] = encoder_out[ + "encoder_padding_mask" + ] + decoder_out[-1]["mt_decoder_out"] = mt_decoder_out + return decoder_out + + +@register_model_architecture( + model_name="s2spect2_conformer", arch_name="s2spect2_conformer" +) +def s2spect2_conformer_architecture_base(args): + args.conv_version = getattr(args, "conv_version", "convtransformer") + args.attn_type = getattr(args, "attn_type", None) + args.pos_enc_type = getattr(args, "pos_enc_type", "abs") + args.max_source_positions = getattr(args, "max_source_positions", 6000) + args.encoder_embed_dim = getattr(args, "encoder_embed_dim", 256) + args.encoder_ffn_embed_dim = getattr(args, "encoder_ffn_embed_dim", 2048) + args.encoder_attention_heads = getattr(args, "encoder_attention_heads", 4) + args.dropout = getattr(args, "dropout", 0.1) + args.encoder_layers = getattr(args, "encoder_layers", 16) + args.depthwise_conv_kernel_size = getattr(args, "depthwise_conv_kernel_size", 31) + s2spect_architecture_base(args) + + +# for old naming +@register_model_architecture( + model_name="s2spect2_conformer", arch_name="s2spect_conformer_translatotron2" +) +def s2spect2_conformer_architecture_base_legacy(args): + s2spect2_conformer_architecture_base(args) diff --git a/data/fairseq/fairseq/models/speech_to_speech/s2s_conformer_unity.py b/data/fairseq/fairseq/models/speech_to_speech/s2s_conformer_unity.py new file mode 100644 index 0000000000000000000000000000000000000000..64388d6d1688c338bc0c11cc23ab985a0c652e36 --- /dev/null +++ b/data/fairseq/fairseq/models/speech_to_speech/s2s_conformer_unity.py @@ -0,0 +1,298 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import copy +import logging + +from fairseq.models import ( + FairseqEncoder, + FairseqEncoderModel, + FairseqLanguageModel, + register_model, + register_model_architecture, +) +from fairseq.models.speech_to_speech.modules.ctc_decoder import CTCDecoder +from fairseq.models.speech_to_speech.modules.stacked_embedding import StackedEmbedding +from fairseq.models.speech_to_speech.modules.transformer_decoder_aug import ( + AugTransformerUnitDecoder, +) +from fairseq.models.speech_to_speech.modules.transformer_encoder import ( + TransformerEncoderNoEmb, +) +from fairseq.models.speech_to_speech.s2s_conformer import S2UTConformerModel +from fairseq.models.speech_to_speech.s2s_transformer import ( + TransformerUnitDecoder, + base_multitask_text_transformer_decoder_arch, + s2ut_architecture_base, +) +from fairseq.models.transformer import TransformerDecoder, TransformerModelBase + +logger = logging.getLogger(__name__) + + +def multitask_text_transformer_decoder_arch( + args, decoder_layers, decoder_embed_dim=256, decoder_attention_heads=4 +): + args.decoder_layers = decoder_layers + args.decoder_embed_dim = decoder_embed_dim + args.decoder_attention_heads = decoder_attention_heads + base_multitask_text_transformer_decoder_arch(args) + + +@register_model("unity_conformer") +class UnityConformerModel(S2UTConformerModel): + """ + Direct speech-to-speech translation model with Conformer encoder + MT Transformer decoder + Transformer discrete unit decoder + """ + + @staticmethod + def add_args(parser): + S2UTConformerModel.add_args(parser) + parser.add_argument( + "--translation-decoder-layers", + type=int, + default=4, + metavar="N", + help="num decoder layers in the first-pass translation module", + ) + parser.add_argument( + "--synthesizer", + default="transformer", + choices=["transformer"], + help="", + ) + parser.add_argument( + "--synthesizer-encoder-layers", + type=int, + default=0, + metavar="N", + help="num encoder layers in the second-pass synthesizer module", + ) + parser.add_argument( + "--synthesizer-augmented-cross-attention", + action="store_true", + default=False, + help="augmented cross-attention over speech encoder output", + ) + + @classmethod + def build_multitask_decoder( + cls, + args, + tgt_dict, + in_dim, + is_first_pass_decoder, + decoder_layers, + decoder_embed_dim, + decoder_attention_heads, + ): + decoder_args = args.decoder_args + decoder_args.encoder_embed_dim = in_dim + if args.decoder_type == "transformer": + if is_first_pass_decoder: + multitask_text_transformer_decoder_arch( + decoder_args, + decoder_layers, + decoder_embed_dim, + decoder_attention_heads, + ) # 4L + else: + base_multitask_text_transformer_decoder_arch(decoder_args) # 2L + task_decoder = TransformerDecoder( + decoder_args, + tgt_dict, + embed_tokens=TransformerModelBase.build_embedding( + decoder_args, + tgt_dict, + decoder_args.decoder_embed_dim, + ), + ) + elif args.decoder_type == "ctc": + task_decoder = CTCDecoder( + dictionary=tgt_dict, + in_dim=in_dim, + ) + else: + raise NotImplementedError( + "currently only support multitask decoder_type 'transformer', 'ctc'" + ) + + return task_decoder + + @classmethod + def build_decoder(cls, args, tgt_dict, aug_attn=False): + num_embeddings = len(tgt_dict) + padding_idx = tgt_dict.pad() + embed_tokens = StackedEmbedding( + num_embeddings, + args.decoder_embed_dim, + padding_idx, + num_stacked=args.n_frames_per_step, + ) + + _args = copy.deepcopy(args) + _args.encoder_embed_dim = args.decoder_embed_dim + + decoder_cls = AugTransformerUnitDecoder if aug_attn else TransformerUnitDecoder + return decoder_cls( + _args, + tgt_dict, + embed_tokens, + ) + + @classmethod + def build_model(cls, args, task): + encoder = cls.build_encoder(args) + decoder = cls.build_decoder( + args, + task.target_dictionary, + aug_attn=getattr(args, "synthesizer_augmented_cross_attention", False), + ) + base_model = cls(encoder, decoder) + + base_model.t2u_augmented_cross_attn = getattr( + args, "synthesizer_augmented_cross_attention", False + ) + + # set up multitask decoders + base_model.mt_task_name = None + base_model.multitask_decoders = {} + has_first_pass_decoder = False + for task_name, task_obj in task.multitask_tasks.items(): + if task_obj.is_first_pass_decoder: + has_first_pass_decoder = True + base_model.mt_task_name = task_name + + in_dim = ( + args.encoder_embed_dim + if task_obj.args.input_from == "encoder" + else args.decoder_embed_dim + ) + task_decoder = cls.build_multitask_decoder( + task_obj.args, + task_obj.target_dictionary, + in_dim, + task_obj.is_first_pass_decoder, + getattr(args, "translation_decoder_layers", 4), + getattr(args, "decoder_embed_dim", 256), + getattr(args, "decoder_attention_heads", 4), + ) + + setattr(base_model, f"{task_name}_decoder", task_decoder) + decoder_model_cls = ( + FairseqEncoderModel + if task_obj.args.decoder_type == "ctc" + else FairseqLanguageModel + ) + base_model.multitask_decoders[task_name] = decoder_model_cls( + getattr(base_model, f"{task_name}_decoder") + ) + + assert has_first_pass_decoder, "set at least one intermediate non-CTC decoder" + + # set up encoder on top of the auxiliary MT decoder + if getattr(args, "synthesizer_encoder_layers", 0) > 0: + base_model.synthesizer_encoder = cls.build_text_encoder(args) + else: + base_model.synthesizer_encoder = None + + return base_model + + @classmethod + def build_text_encoder(cls, args): + _args = copy.deepcopy(args) + _args.encoder_layers = args.synthesizer_encoder_layers + _args.encoder_embed_dim = args.decoder_embed_dim + _args.encoder_ffn_embed_dim = args.decoder_ffn_embed_dim + _args.encoder_attention_heads = args.decoder_attention_heads + _args.encoder_normalize_before = True + return TransformerEncoderNoEmb(_args) + + def forward( + self, + src_tokens, + src_lengths, + prev_output_tokens, + prev_output_tokens_mt, + tgt_speaker=None, + return_all_hiddens=False, + ): + mt_decoder = getattr(self, f"{self.mt_task_name}_decoder") + + encoder_out = self.encoder( + src_tokens, + src_lengths=src_lengths, + tgt_speaker=tgt_speaker, + return_all_hiddens=return_all_hiddens, + ) + + # 1. MT decoder + mt_decoder_out = mt_decoder( + prev_output_tokens_mt, + encoder_out=encoder_out, + ) + x = mt_decoder_out[1]["inner_states"][-1] + if mt_decoder.layer_norm is not None: + x = mt_decoder.layer_norm(x) + + mt_decoder_padding_mask = None + if prev_output_tokens_mt.eq(mt_decoder.padding_idx).any(): + mt_decoder_padding_mask = prev_output_tokens_mt.eq(mt_decoder.padding_idx) + + # 2. T2U encoder + if self.synthesizer_encoder is not None: + t2u_encoder_out = self.synthesizer_encoder( + x, + mt_decoder_padding_mask, + return_all_hiddens=return_all_hiddens, + ) + else: + t2u_encoder_out = { + "encoder_out": [x], # T x B x C + "encoder_padding_mask": [mt_decoder_padding_mask], # B x T + } + + # 3. T2U decoder + if self.t2u_augmented_cross_attn: + decoder_out = self.decoder( + prev_output_tokens, + encoder_out=encoder_out, + encoder_out_aug=t2u_encoder_out, + ) + else: + decoder_out = self.decoder( + prev_output_tokens, + encoder_out=t2u_encoder_out, + ) + if return_all_hiddens: + decoder_out[-1]["encoder_states"] = encoder_out["encoder_states"] + decoder_out[-1]["encoder_padding_mask"] = encoder_out[ + "encoder_padding_mask" + ] + decoder_out[-1]["mt_decoder_out"] = mt_decoder_out + return decoder_out + + +@register_model_architecture(model_name="unity_conformer", arch_name="unity_conformer") +def unity_conformer_architecture_base(args): + args.conv_version = getattr(args, "conv_version", "convtransformer") + args.attn_type = getattr(args, "attn_type", None) + args.pos_enc_type = getattr(args, "pos_enc_type", "abs") + args.max_source_positions = getattr(args, "max_source_positions", 6000) + args.encoder_embed_dim = getattr(args, "encoder_embed_dim", 256) + args.encoder_ffn_embed_dim = getattr(args, "encoder_ffn_embed_dim", 2048) + args.encoder_attention_heads = getattr(args, "encoder_attention_heads", 4) + args.dropout = getattr(args, "dropout", 0.1) + args.encoder_layers = getattr(args, "encoder_layers", 16) + args.depthwise_conv_kernel_size = getattr(args, "depthwise_conv_kernel_size", 31) + s2ut_architecture_base(args) + + +# for old naming +@register_model_architecture( + model_name="unity_conformer", arch_name="s2ut_conformer_translatotron2" +) +def unity_conformer_architecture_base_legacy(args): + unity_conformer_architecture_base(args) diff --git a/data/fairseq/fairseq/models/speech_to_speech/s2s_transformer.py b/data/fairseq/fairseq/models/speech_to_speech/s2s_transformer.py new file mode 100644 index 0000000000000000000000000000000000000000..07393d2598c0a38ec98bc2b66abd7dfa8af18825 --- /dev/null +++ b/data/fairseq/fairseq/models/speech_to_speech/s2s_transformer.py @@ -0,0 +1,722 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import logging +from pathlib import Path +from typing import Any, Dict, List, Optional + +import torch +from torch import Tensor + +from fairseq import checkpoint_utils, utils +from fairseq.models import ( + FairseqEncoderDecoderModel, + FairseqEncoderModel, + FairseqLanguageModel, + register_model, + register_model_architecture, +) +from fairseq.models.speech_to_speech.modules.ctc_decoder import CTCDecoder +from fairseq.models.speech_to_speech.modules.stacked_embedding import StackedEmbedding +from fairseq.models.speech_to_text import S2TTransformerEncoder +from fairseq.models.text_to_speech import TTSTransformerDecoder +from fairseq.models.transformer import Linear, TransformerDecoder, TransformerModelBase + +logger = logging.getLogger(__name__) + + +class S2STransformerEncoder(S2TTransformerEncoder): + """Based on S2T transformer encoder, with support + to incorporate target speaker embedding.""" + + def __init__(self, args): + super().__init__(args) + + self.spk_emb_proj = None + if args.target_speaker_embed: + self.spk_emb_proj = Linear( + args.encoder_embed_dim + args.speaker_embed_dim, args.encoder_embed_dim + ) + + def forward( + self, src_tokens, src_lengths, tgt_speaker=None, return_all_hiddens=False + ): + out = super().forward(src_tokens, src_lengths, return_all_hiddens) + + if self.spk_emb_proj: + x = out["encoder_out"][0] + seq_len, bsz, _ = x.size() + tgt_speaker_emb = tgt_speaker.view(1, bsz, -1).expand(seq_len, bsz, -1) + x = self.spk_emb_proj(torch.cat([x, tgt_speaker_emb], dim=2)) + out["encoder_out"][0] = x + + return out + + +class TransformerUnitDecoder(TransformerDecoder): + """Based on Transformer decoder, with support to decoding stacked units""" + + def __init__( + self, + args, + dictionary, + embed_tokens, + no_encoder_attn=False, + output_projection=None, + ): + super().__init__( + args, dictionary, embed_tokens, no_encoder_attn, output_projection + ) + self.n_frames_per_step = args.n_frames_per_step + + self.out_proj_n_frames = ( + Linear( + self.output_embed_dim, + self.output_embed_dim * self.n_frames_per_step, + bias=False, + ) + if self.n_frames_per_step > 1 + else None + ) + + def forward( + self, + prev_output_tokens, + encoder_out: Optional[Dict[str, List[Tensor]]] = None, + incremental_state: Optional[Dict[str, Dict[str, Optional[Tensor]]]] = None, + features_only: bool = False, + full_context_alignment: bool = False, + alignment_layer: Optional[int] = None, + alignment_heads: Optional[int] = None, + src_lengths: Optional[Any] = None, + return_all_hiddens: bool = False, + ): + """ + Args: + prev_output_tokens (LongTensor): previous decoder outputs of shape + `(batch, tgt_len)`, for teacher forcing + encoder_out (optional): output from the encoder, used for + encoder-side attention, should be of size T x B x C + incremental_state (dict): dictionary used for storing state during + :ref:`Incremental decoding` + features_only (bool, optional): only return features without + applying output layer (default: False). + full_context_alignment (bool, optional): don't apply + auto-regressive mask to self-attention (default: False). + + Returns: + tuple: + - the decoder's output of shape `(batch, tgt_len, vocab)` + - a dictionary with any model-specific outputs + """ + + x, extra = self.extract_features( + prev_output_tokens, + encoder_out=encoder_out, + incremental_state=incremental_state, + full_context_alignment=full_context_alignment, + alignment_layer=alignment_layer, + alignment_heads=alignment_heads, + ) + + if not features_only: + bsz, seq_len, d = x.size() + if self.out_proj_n_frames: + x = self.out_proj_n_frames(x) + x = self.output_layer(x.view(bsz, seq_len, self.n_frames_per_step, d)) + x = x.view(bsz, seq_len * self.n_frames_per_step, -1) + if ( + incremental_state is None and self.n_frames_per_step > 1 + ): # teacher-forcing mode in training + x = x[ + :, : -(self.n_frames_per_step - 1), : + ] # remove extra frames after + + return x, extra + + def upgrade_state_dict_named(self, state_dict, name): + if self.n_frames_per_step > 1: + move_keys = [ + ( + f"{name}.project_in_dim.weight", + f"{name}.embed_tokens.project_in_dim.weight", + ) + ] + for from_k, to_k in move_keys: + if from_k in state_dict and to_k not in state_dict: + state_dict[to_k] = state_dict[from_k] + del state_dict[from_k] + + +class S2STransformerMultitaskModelBase(FairseqEncoderDecoderModel): + @classmethod + def build_encoder(cls, args): + encoder = S2STransformerEncoder(args) + pretraining_path = getattr(args, "load_pretrained_encoder_from", None) + if pretraining_path is not None: + if not Path(pretraining_path).exists(): + logger.warning( + f"skipped pretraining because {pretraining_path} does not exist" + ) + else: + encoder = checkpoint_utils.load_pretrained_component_from_model( + component=encoder, checkpoint=pretraining_path + ) + logger.info(f"loaded pretrained encoder from: {pretraining_path}") + return encoder + + @classmethod + def build_multitask_decoder(cls, args, tgt_dict, in_dim): + decoder_args = args.decoder_args + decoder_args.encoder_embed_dim = in_dim + if args.decoder_type == "transformer": + base_multitask_text_transformer_decoder_arch(decoder_args) + task_decoder = TransformerDecoder( + decoder_args, + tgt_dict, + embed_tokens=TransformerModelBase.build_embedding( + decoder_args, + tgt_dict, + decoder_args.decoder_embed_dim, + ), + ) + elif args.decoder_type == "ctc": + task_decoder = CTCDecoder( + dictionary=tgt_dict, + in_dim=in_dim, + ) + else: + raise NotImplementedError( + "currently only support multitask decoder_type 'transformer', 'ctc'" + ) + + return task_decoder + + @classmethod + def build_model(cls, args, task): + encoder = cls.build_encoder(args) + decoder = ( + cls.build_decoder(args, task.target_dictionary) + if task.args.target_is_code + else cls.build_decoder(args) + ) + base_model = cls(encoder, decoder) + + # set up multitask decoders + base_model.multitask_decoders = {} + for task_name, task_obj in task.multitask_tasks.items(): + in_dim = ( + args.encoder_embed_dim + if task_obj.args.input_from == "encoder" + else args.decoder_embed_dim + ) + task_decoder = cls.build_multitask_decoder( + task_obj.args, task_obj.target_dictionary, in_dim + ) + + setattr(base_model, f"{task_name}_decoder", task_decoder) + decoder_model_cls = ( + FairseqEncoderModel + if task_obj.args.decoder_type == "ctc" + else FairseqLanguageModel + ) + base_model.multitask_decoders[task_name] = decoder_model_cls( + getattr(base_model, f"{task_name}_decoder") + ) + + return base_model + + def forward_encoder(self, src_tokens, src_lengths, speaker=None, **kwargs): + return self.encoder( + src_tokens, src_lengths=src_lengths, tgt_speaker=speaker, **kwargs + ) + + +@register_model("s2ut_transformer") +class S2UTTransformerModel(S2STransformerMultitaskModelBase): + """ + Direct speech-to-speech translation model with Transformer encoder + Transformer discrete unit decoder + https://arxiv.org/abs/2107.05604 + """ + + @staticmethod + def add_args(parser): + # input + parser.add_argument( + "--conv-kernel-sizes", + type=str, + metavar="STR", + help="kernel sizes of Conv1d (s2t_transformer) subsampling layers", + ) + parser.add_argument( + "--conv-channels", + type=int, + metavar="N", + help="# of channels in Conv1d (s2t_transformer) subsampling layers", + ) + parser.add_argument( + "--conv-out-channels", + type=int, + metavar="N", + help="# of channels in Conv2d (convtransformer) subsampling layers", + ) + parser.add_argument( + "--conv-version", + type=str, + default="s2t_transformer", + choices=["s2t_transformer", "convtransformer"], + help="version of frontend convolutional layers", + ) + # Transformer + parser.add_argument( + "--activation-fn", + type=str, + default="relu", + choices=utils.get_available_activation_fns(), + help="activation function to use", + ) + parser.add_argument( + "--dropout", type=float, metavar="D", help="dropout probability" + ) + parser.add_argument( + "--attention-dropout", + type=float, + metavar="D", + help="dropout probability for attention weights", + ) + parser.add_argument( + "--activation-dropout", + "--relu-dropout", + type=float, + metavar="D", + help="dropout probability after activation in FFN.", + ) + parser.add_argument( + "--encoder-embed-dim", + type=int, + metavar="N", + help="encoder embedding dimension", + ) + parser.add_argument( + "--encoder-ffn-embed-dim", + type=int, + metavar="N", + help="encoder embedding dimension for FFN", + ) + parser.add_argument( + "--encoder-layers", type=int, metavar="N", help="num encoder layers" + ) + parser.add_argument( + "--encoder-attention-heads", + type=int, + metavar="N", + help="num encoder attention heads", + ) + parser.add_argument( + "--encoder-normalize-before", + action="store_true", + help="apply layernorm before each encoder block", + ) + parser.add_argument( + "--decoder-embed-dim", + type=int, + metavar="N", + help="decoder embedding dimension", + ) + parser.add_argument( + "--decoder-ffn-embed-dim", + type=int, + metavar="N", + help="decoder embedding dimension for FFN", + ) + parser.add_argument( + "--decoder-layers", type=int, metavar="N", help="num decoder layers" + ) + parser.add_argument( + "--decoder-attention-heads", + type=int, + metavar="N", + help="num decoder attention heads", + ) + parser.add_argument( + "--decoder-normalize-before", + action="store_true", + help="apply layernorm before each decoder block", + ) + parser.add_argument( + "--share-decoder-input-output-embed", + action="store_true", + help="share decoder input and output embeddings", + ) + parser.add_argument( + "--layernorm-embedding", + action="store_true", + help="add layernorm to embedding", + ) + parser.add_argument( + "--no-scale-embedding", + action="store_true", + help="if True, dont scale embeddings", + ) + parser.add_argument( + "--load-pretrained-encoder-from", + type=str, + metavar="STR", + help="model to take encoder weights from (for initialization)", + ) + parser.add_argument( + "--encoder-freezing-updates", + type=int, + metavar="N", + help="freeze encoder for first N updates", + ) + # speaker + parser.add_argument( + "--speaker-embed-dim", + type=int, + metavar="N", + help="speaker embedding dimension", + ) + + @classmethod + def build_decoder(cls, args, tgt_dict): + num_embeddings = len(tgt_dict) + padding_idx = tgt_dict.pad() + embed_tokens = StackedEmbedding( + num_embeddings, + args.decoder_embed_dim, + padding_idx, + num_stacked=args.n_frames_per_step, + ) + + return TransformerUnitDecoder( + args, + tgt_dict, + embed_tokens, + ) + + def forward( + self, + src_tokens, + src_lengths, + prev_output_tokens, + tgt_speaker=None, + return_all_hiddens=False, + ): + encoder_out = self.encoder( + src_tokens, + src_lengths=src_lengths, + tgt_speaker=tgt_speaker, + return_all_hiddens=return_all_hiddens, + ) + decoder_out = self.decoder( + prev_output_tokens, + encoder_out=encoder_out, + ) + if return_all_hiddens: + decoder_out[-1]["encoder_states"] = encoder_out["encoder_states"] + decoder_out[-1]["encoder_padding_mask"] = encoder_out[ + "encoder_padding_mask" + ] + return decoder_out + + +@register_model("s2spect_transformer") +class S2SpecTTransformerModel(S2STransformerMultitaskModelBase): + """ + Speech-to-spectrogram model with S2T Transformer encoder + TTS Transformer decoder + """ + + @staticmethod + def add_args(parser): + # input + parser.add_argument( + "--conv-kernel-sizes", + type=str, + metavar="STR", + help="kernel sizes of Conv1d (s2t_transformer) subsampling layers", + ) + parser.add_argument( + "--conv-channels", + type=int, + metavar="N", + help="# of channels in Conv1d (s2t_transformer) subsampling layers", + ) + parser.add_argument( + "--conv-version", + type=str, + default="s2t_transformer", + choices=["s2t_transformer", "convtransformer"], + help="version of frontend convolutional layers", + ) + # Transformer + parser.add_argument( + "--activation-fn", + type=str, + default="relu", + choices=utils.get_available_activation_fns(), + help="activation function to use", + ) + parser.add_argument( + "--dropout", type=float, metavar="D", help="dropout probability" + ) + parser.add_argument( + "--attention-dropout", + type=float, + metavar="D", + help="dropout probability for attention weights", + ) + parser.add_argument( + "--activation-dropout", + "--relu-dropout", + type=float, + metavar="D", + help="dropout probability after activation in FFN.", + ) + parser.add_argument( + "--encoder-embed-dim", + type=int, + metavar="N", + help="encoder embedding dimension", + ) + parser.add_argument( + "--encoder-ffn-embed-dim", + type=int, + metavar="N", + help="encoder embedding dimension for FFN", + ) + parser.add_argument( + "--encoder-layers", type=int, metavar="N", help="num encoder layers" + ) + parser.add_argument( + "--encoder-attention-heads", + type=int, + metavar="N", + help="num encoder attention heads", + ) + parser.add_argument( + "--encoder-normalize-before", + action="store_true", + help="apply layernorm before each encoder block", + ) + parser.add_argument( + "--no-scale-embedding", + action="store_true", + help="if True, dont scale embeddings", + ) + parser.add_argument( + "--load-pretrained-encoder-from", + type=str, + metavar="STR", + help="model to take encoder weights from (for initialization)", + ) + parser.add_argument( + "--encoder-freezing-updates", + type=int, + metavar="N", + help="freeze encoder for first N updates", + ) + # speaker + parser.add_argument( + "--speaker-embed-dim", + type=int, + metavar="N", + help="speaker embedding dimension", + ) + # decoder + parser.add_argument("--output-frame-dim", type=int) + # decoder prenet + parser.add_argument("--prenet-dropout", type=float) + parser.add_argument("--prenet-layers", type=int) + parser.add_argument("--prenet-dim", type=int) + # decoder postnet + parser.add_argument("--postnet-dropout", type=float) + parser.add_argument("--postnet-layers", type=int) + parser.add_argument("--postnet-conv-dim", type=int) + parser.add_argument("--postnet-conv-kernel-size", type=int) + # decoder transformer layers + parser.add_argument("--decoder-transformer-layers", type=int) + parser.add_argument("--decoder-embed-dim", type=int) + parser.add_argument("--decoder-ffn-embed-dim", type=int) + parser.add_argument("--decoder-normalize-before", action="store_true") + parser.add_argument("--decoder-attention-heads", type=int) + + @classmethod + def build_decoder(cls, args): + return TTSTransformerDecoder(args, None, padding_idx=1) + + def forward( + self, + src_tokens, + src_lengths, + prev_output_tokens, + tgt_speaker=None, + incremental_state=None, + target_lengths=None, + speaker=None, + return_all_hiddens=False, + ): + encoder_out = self.encoder( + src_tokens, + src_lengths=src_lengths, + tgt_speaker=tgt_speaker, + return_all_hiddens=return_all_hiddens, + ) + decoder_out = self.decoder( + prev_output_tokens, + encoder_out=encoder_out, + incremental_state=incremental_state, + target_lengths=target_lengths, + speaker=speaker, + ) + if return_all_hiddens: + decoder_out[-1]["encoder_states"] = encoder_out["encoder_states"] + decoder_out[-1]["encoder_padding_mask"] = encoder_out[ + "encoder_padding_mask" + ] + return decoder_out + + +def base_multitask_text_transformer_decoder_arch(args): + args.dropout = getattr(args, "dropout", 0.3) + args.decoder_layerdrop = getattr(args, "decoder_layerdrop", 0.0) + args.share_decoder_input_output_embed = getattr( + args, "share_decoder_input_output_embed", True + ) + args.decoder_embed_dim = getattr(args, "decoder_embed_dim", 256) + args.decoder_output_dim = getattr( + args, "decoder_output_dim", args.decoder_embed_dim + ) + args.decoder_input_dim = getattr(args, "decoder_input_dim", args.decoder_embed_dim) + + args.max_target_positions = getattr(args, "max_target_positions", 1024) + args.no_scale_embedding = getattr(args, "no_scale_embedding", False) + + args.adaptive_input = getattr(args, "adaptive_input", False) + args.quant_noise_pq = getattr(args, "quant_noise_pq", 0) + + args.decoder_learned_pos = getattr(args, "decoder_learned_pos", False) + args.no_token_positional_embeddings = getattr( + args, "no_token_positional_embeddings", False + ) + + args.decoder_layers = getattr(args, "decoder_layers", 2) + + args.adaptive_softmax_cutoff = getattr(args, "adaptive_softmax_cutoff", None) + + # decoder layer + args.activation_dropout = getattr(args, "activation_dropout", args.dropout) + args.activation_fn = getattr(args, "activation_fn", "relu") + args.decoder_normalize_before = getattr(args, "decoder_normalize_before", True) + args.decoder_ffn_embed_dim = getattr(args, "decoder_ffn_embed_dim", 2048) + + args.attention_dropout = getattr(args, "attention_dropout", args.dropout) + args.decoder_attention_heads = getattr(args, "decoder_attention_heads", 4) + + +def base_s2st_transformer_encoder_architecture(args): + args.encoder_freezing_updates = getattr(args, "encoder_freezing_updates", 0) + + # Convolutional subsampler + args.input_channels = getattr(args, "input_channels", 1) + args.conv_kernel_sizes = getattr(args, "conv_kernel_sizes", "5,5") # for Conv1d + args.conv_channels = getattr(args, "conv_channels", 1024) # for Conv1d + args.conv_out_channels = getattr(args, "conv_out_channels", 256) # for Conv2d + args.conv_version = getattr(args, "conv_version", "s2t_transformer") + # Transformer + args.encoder_embed_dim = getattr(args, "encoder_embed_dim", 512) + args.encoder_ffn_embed_dim = getattr(args, "encoder_ffn_embed_dim", 2048) + args.encoder_layers = getattr(args, "encoder_layers", 12) + args.encoder_attention_heads = getattr(args, "encoder_attention_heads", 8) + args.encoder_normalize_before = getattr(args, "encoder_normalize_before", True) + args.no_scale_embedding = getattr(args, "no_scale_embedding", False) + + args.dropout = getattr(args, "dropout", 0.1) + args.attention_dropout = getattr(args, "attention_dropout", args.dropout) + args.activation_dropout = getattr(args, "activation_dropout", args.dropout) + args.activation_fn = getattr(args, "activation_fn", "relu") + + args.speaker_embed_dim = getattr(args, "speaker_embed_dim", 256) + + +@register_model_architecture( + model_name="s2ut_transformer", arch_name="s2ut_transformer" +) +def s2ut_architecture_base(args): + base_s2st_transformer_encoder_architecture(args) + + # decoder + args.decoder_embed_dim = getattr(args, "decoder_embed_dim", args.encoder_embed_dim) + args.decoder_ffn_embed_dim = getattr( + args, "decoder_ffn_embed_dim", args.encoder_ffn_embed_dim + ) + args.decoder_layers = getattr(args, "decoder_layers", 6) + args.decoder_attention_heads = getattr(args, "decoder_attention_heads", 8) + args.decoder_normalize_before = getattr(args, "decoder_normalize_before", True) + args.decoder_learned_pos = getattr(args, "decoder_learned_pos", False) + args.adaptive_softmax_cutoff = getattr(args, "adaptive_softmax_cutoff", None) + args.adaptive_softmax_dropout = getattr(args, "adaptive_softmax_dropout", 0) + args.share_decoder_input_output_embed = getattr( + args, "share_decoder_input_output_embed", False + ) + args.no_token_positional_embeddings = getattr( + args, "no_token_positional_embeddings", False + ) + args.adaptive_input = getattr(args, "adaptive_input", False) + args.decoder_layerdrop = getattr(args, "decoder_layerdrop", 0.0) + args.decoder_output_dim = getattr( + args, "decoder_output_dim", args.decoder_embed_dim + ) + args.decoder_input_dim = getattr(args, "decoder_input_dim", args.decoder_embed_dim) + args.quant_noise_pq = getattr(args, "quant_noise_pq", 0) + + +@register_model_architecture("s2ut_transformer", "s2ut_transformer_fisher") +def s2ut_architecture_fisher(args): + args.encoder_embed_dim = getattr(args, "encoder_embed_dim", 256) + args.encoder_attention_heads = getattr(args, "encoder_attention_heads", 4) + args.dropout = getattr(args, "dropout", 0.1) + + s2ut_architecture_base(args) + + +@register_model_architecture( + model_name="s2spect_transformer", arch_name="s2spect_transformer" +) +def s2spect_architecture_base(args): + base_s2st_transformer_encoder_architecture(args) + + # decoder + args.output_frame_dim = getattr(args, "output_frame_dim", 80) + # decoder prenet + args.prenet_dropout = getattr(args, "prenet_dropout", 0.5) + args.prenet_layers = getattr(args, "prenet_layers", 2) + args.prenet_dim = getattr(args, "prenet_dim", 256) + # decoder postnet + args.postnet_dropout = getattr(args, "postnet_dropout", 0.5) + args.postnet_layers = getattr(args, "postnet_layers", 5) + args.postnet_conv_dim = getattr(args, "postnet_conv_dim", 512) + args.postnet_conv_kernel_size = getattr(args, "postnet_conv_kernel_size", 5) + # decoder transformer layers + args.decoder_transformer_layers = getattr(args, "decoder_transformer_layers", 6) + args.decoder_embed_dim = getattr(args, "decoder_embed_dim", 512) + args.decoder_ffn_embed_dim = getattr( + args, "decoder_ffn_embed_dim", 4 * args.decoder_embed_dim + ) + args.decoder_normalize_before = getattr(args, "decoder_normalize_before", False) + args.decoder_attention_heads = getattr(args, "decoder_attention_heads", 4) + + +@register_model_architecture("s2spect_transformer", "s2spect_transformer_fisher") +def s2spect_architecture_fisher(args): + args.encoder_embed_dim = getattr(args, "encoder_embed_dim", 256) + args.encoder_ffn_embed_dim = getattr(args, "encoder_ffn_embed_dim", 256 * 8) + args.encoder_attention_heads = getattr(args, "encoder_attention_heads", 4) + args.dropout = getattr(args, "dropout", 0.1) + + # decoder + args.prenet_dim = getattr(args, "prenet_dim", 32) + + s2spect_architecture_base(args) diff --git a/data/fairseq/fairseq/models/speech_to_text/__init__.py b/data/fairseq/fairseq/models/speech_to_text/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..62ef663efba1b6d745169bddbc5ba98adadbc1da --- /dev/null +++ b/data/fairseq/fairseq/models/speech_to_text/__init__.py @@ -0,0 +1,13 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from .berard import * # noqa +from .convtransformer import * # noqa +from .multi_modality_model import * # noqa +from .s2t_conformer import * # noqa +from .s2t_transformer import * # noqa +from .s2t_wav_transformer import * # noqa +from .xm_transformer import * # noqa +from .xm_transformer_unity import * # noqa diff --git a/data/fairseq/fairseq/models/speech_to_text/__pycache__/__init__.cpython-38.pyc b/data/fairseq/fairseq/models/speech_to_text/__pycache__/__init__.cpython-38.pyc new file mode 100644 index 0000000000000000000000000000000000000000..a5ea710be67f236832d785c920897ee984867c7d Binary files /dev/null and b/data/fairseq/fairseq/models/speech_to_text/__pycache__/__init__.cpython-38.pyc differ diff --git a/data/fairseq/fairseq/models/speech_to_text/__pycache__/berard.cpython-38.pyc b/data/fairseq/fairseq/models/speech_to_text/__pycache__/berard.cpython-38.pyc new file mode 100644 index 0000000000000000000000000000000000000000..611f58fb17ff9880245f23770e55e8d24cf77d58 Binary files /dev/null and b/data/fairseq/fairseq/models/speech_to_text/__pycache__/berard.cpython-38.pyc differ diff --git a/data/fairseq/fairseq/models/speech_to_text/__pycache__/convtransformer.cpython-38.pyc b/data/fairseq/fairseq/models/speech_to_text/__pycache__/convtransformer.cpython-38.pyc new file mode 100644 index 0000000000000000000000000000000000000000..0948741a9f146ce97b1004c112cf2de05575e9e5 Binary files /dev/null and b/data/fairseq/fairseq/models/speech_to_text/__pycache__/convtransformer.cpython-38.pyc differ diff --git a/data/fairseq/fairseq/models/speech_to_text/__pycache__/hub_interface.cpython-38.pyc b/data/fairseq/fairseq/models/speech_to_text/__pycache__/hub_interface.cpython-38.pyc new file mode 100644 index 0000000000000000000000000000000000000000..3a2fc3878961b400caaebdb5a722cb4a95b81760 Binary files /dev/null and b/data/fairseq/fairseq/models/speech_to_text/__pycache__/hub_interface.cpython-38.pyc differ diff --git a/data/fairseq/fairseq/models/speech_to_text/__pycache__/multi_modality_model.cpython-38.pyc b/data/fairseq/fairseq/models/speech_to_text/__pycache__/multi_modality_model.cpython-38.pyc new file mode 100644 index 0000000000000000000000000000000000000000..f04b985014fd1c69882233b990bd072b473d9c8f Binary files /dev/null and b/data/fairseq/fairseq/models/speech_to_text/__pycache__/multi_modality_model.cpython-38.pyc differ diff --git a/data/fairseq/fairseq/models/speech_to_text/__pycache__/s2t_conformer.cpython-38.pyc b/data/fairseq/fairseq/models/speech_to_text/__pycache__/s2t_conformer.cpython-38.pyc new file mode 100644 index 0000000000000000000000000000000000000000..ef66c4558aaddc0cee19ce1221f0b7b89aafa280 Binary files /dev/null and b/data/fairseq/fairseq/models/speech_to_text/__pycache__/s2t_conformer.cpython-38.pyc differ diff --git a/data/fairseq/fairseq/models/speech_to_text/__pycache__/s2t_transformer.cpython-38.pyc b/data/fairseq/fairseq/models/speech_to_text/__pycache__/s2t_transformer.cpython-38.pyc new file mode 100644 index 0000000000000000000000000000000000000000..ba4322d77a6e0955cd7d97d23d10360bb39a77c2 Binary files /dev/null and b/data/fairseq/fairseq/models/speech_to_text/__pycache__/s2t_transformer.cpython-38.pyc differ diff --git a/data/fairseq/fairseq/models/speech_to_text/__pycache__/s2t_wav_transformer.cpython-38.pyc b/data/fairseq/fairseq/models/speech_to_text/__pycache__/s2t_wav_transformer.cpython-38.pyc new file mode 100644 index 0000000000000000000000000000000000000000..05b0c9d16d9b226472ece097d540bc2d2c8e93fd Binary files /dev/null and b/data/fairseq/fairseq/models/speech_to_text/__pycache__/s2t_wav_transformer.cpython-38.pyc differ diff --git a/data/fairseq/fairseq/models/speech_to_text/__pycache__/xm_transformer.cpython-38.pyc b/data/fairseq/fairseq/models/speech_to_text/__pycache__/xm_transformer.cpython-38.pyc new file mode 100644 index 0000000000000000000000000000000000000000..6e3d4ca532264ab9070631c8fee33de5b534ef23 Binary files /dev/null and b/data/fairseq/fairseq/models/speech_to_text/__pycache__/xm_transformer.cpython-38.pyc differ diff --git a/data/fairseq/fairseq/models/speech_to_text/__pycache__/xm_transformer_unity.cpython-38.pyc b/data/fairseq/fairseq/models/speech_to_text/__pycache__/xm_transformer_unity.cpython-38.pyc new file mode 100644 index 0000000000000000000000000000000000000000..4a15c4056a6ee11f6c7149c0ecd3ef68935c743e Binary files /dev/null and b/data/fairseq/fairseq/models/speech_to_text/__pycache__/xm_transformer_unity.cpython-38.pyc differ diff --git a/data/fairseq/fairseq/models/speech_to_text/berard.py b/data/fairseq/fairseq/models/speech_to_text/berard.py new file mode 100644 index 0000000000000000000000000000000000000000..107ac983c62d721eae0a00b633ee350c9f1673da --- /dev/null +++ b/data/fairseq/fairseq/models/speech_to_text/berard.py @@ -0,0 +1,607 @@ +#!/usr/bin/env python3 + +from ast import literal_eval +from typing import List, Tuple + +import torch +import torch.nn as nn +import torch.nn.functional as F + +from fairseq import checkpoint_utils, utils +from fairseq.data.data_utils import lengths_to_padding_mask +from fairseq.models import ( + FairseqEncoder, + FairseqEncoderDecoderModel, + FairseqIncrementalDecoder, + register_model, + register_model_architecture, +) + + +@register_model("s2t_berard") +class BerardModel(FairseqEncoderDecoderModel): + """Implementation of a model similar to https://arxiv.org/abs/1802.04200 + + Paper title: End-to-End Automatic Speech Translation of Audiobooks + An implementation is available in tensorflow at + https://github.com/eske/seq2seq + Relevant files in this implementation are the config + (https://github.com/eske/seq2seq/blob/master/config/LibriSpeech/AST.yaml) + and the model code + (https://github.com/eske/seq2seq/blob/master/translate/models.py). + The encoder and decoder try to be close to the original implementation. + The attention is an MLP as in Bahdanau et al. + (https://arxiv.org/abs/1409.0473). + There is no state initialization by averaging the encoder outputs. + """ + + def __init__(self, encoder, decoder): + super().__init__(encoder, decoder) + + @staticmethod + def add_args(parser): + parser.add_argument( + "--input-layers", + type=str, + metavar="EXPR", + help="List of linear layer dimensions. These " + "layers are applied to the input features and " + "are followed by tanh and possibly dropout.", + ) + parser.add_argument( + "--dropout", + type=float, + metavar="D", + help="Dropout probability to use in the encoder/decoder. " + "Note that this parameters control dropout in various places, " + "there is no fine-grained control for dropout for embeddings " + "vs LSTM layers for example.", + ) + parser.add_argument( + "--in-channels", + type=int, + metavar="N", + help="Number of encoder input channels. " "Typically value is 1.", + ) + parser.add_argument( + "--conv-layers", + type=str, + metavar="EXPR", + help="List of conv layers " "(format: (channels, kernel, stride)).", + ) + parser.add_argument( + "--num-blstm-layers", + type=int, + metavar="N", + help="Number of encoder bi-LSTM layers.", + ) + parser.add_argument( + "--lstm-size", type=int, metavar="N", help="LSTM hidden size." + ) + parser.add_argument( + "--decoder-embed-dim", + type=int, + metavar="N", + help="Embedding dimension of the decoder target tokens.", + ) + parser.add_argument( + "--decoder-hidden-dim", + type=int, + metavar="N", + help="Decoder LSTM hidden dimension.", + ) + parser.add_argument( + "--decoder-num-layers", + type=int, + metavar="N", + help="Number of decoder LSTM layers.", + ) + parser.add_argument( + "--attention-dim", + type=int, + metavar="N", + help="Hidden layer dimension in MLP attention.", + ) + parser.add_argument( + "--output-layer-dim", + type=int, + metavar="N", + help="Hidden layer dim for linear layer prior to output projection.", + ) + parser.add_argument( + "--load-pretrained-encoder-from", + type=str, + metavar="STR", + help="model to take encoder weights from (for initialization)", + ) + parser.add_argument( + "--load-pretrained-decoder-from", + type=str, + metavar="STR", + help="model to take decoder weights from (for initialization)", + ) + + @classmethod + def build_encoder(cls, args, task): + encoder = BerardEncoder( + input_layers=literal_eval(args.input_layers), + conv_layers=literal_eval(args.conv_layers), + in_channels=args.input_channels, + input_feat_per_channel=args.input_feat_per_channel, + num_blstm_layers=args.num_blstm_layers, + lstm_size=args.lstm_size, + dropout=args.dropout, + ) + if getattr(args, "load_pretrained_encoder_from", None) is not None: + encoder = checkpoint_utils.load_pretrained_component_from_model( + component=encoder, checkpoint=args.load_pretrained_encoder_from + ) + return encoder + + @classmethod + def build_decoder(cls, args, task): + decoder = LSTMDecoder( + dictionary=task.target_dictionary, + embed_dim=args.decoder_embed_dim, + num_layers=args.decoder_num_layers, + hidden_size=args.decoder_hidden_dim, + dropout=args.dropout, + encoder_output_dim=2 * args.lstm_size, # bidirectional + attention_dim=args.attention_dim, + output_layer_dim=args.output_layer_dim, + ) + if getattr(args, "load_pretrained_decoder_from", None) is not None: + decoder = checkpoint_utils.load_pretrained_component_from_model( + component=decoder, checkpoint=args.load_pretrained_decoder_from + ) + return decoder + + @classmethod + def build_model(cls, args, task): + """Build a new model instance.""" + encoder = cls.build_encoder(args, task) + decoder = cls.build_decoder(args, task) + + return cls(encoder, decoder) + + def get_normalized_probs(self, net_output, log_probs, sample=None): + # net_output['encoder_out'] is a (B, T, D) tensor + lprobs = super().get_normalized_probs(net_output, log_probs, sample) + # lprobs is a (B, T, D) tensor + lprobs.batch_first = True + return lprobs + + +class BerardEncoder(FairseqEncoder): + def __init__( + self, + input_layers: List[int], + conv_layers: List[Tuple[int]], + in_channels: int, + input_feat_per_channel: int, + num_blstm_layers: int, + lstm_size: int, + dropout: float, + ): + """ + Args: + input_layers: list of linear layer dimensions. These layers are + applied to the input features and are followed by tanh and + possibly dropout. + conv_layers: list of conv2d layer configurations. A configuration is + a tuple (out_channels, conv_kernel_size, stride). + in_channels: number of input channels. + input_feat_per_channel: number of input features per channel. These + are speech features, typically 40 or 80. + num_blstm_layers: number of bidirectional LSTM layers. + lstm_size: size of the LSTM hidden (and cell) size. + dropout: dropout probability. Dropout can be applied after the + linear layers and LSTM layers but not to the convolutional + layers. + """ + super().__init__(None) + + self.input_layers = nn.ModuleList() + in_features = input_feat_per_channel + for out_features in input_layers: + if dropout > 0: + self.input_layers.append( + nn.Sequential( + nn.Linear(in_features, out_features), nn.Dropout(p=dropout) + ) + ) + else: + self.input_layers.append(nn.Linear(in_features, out_features)) + in_features = out_features + + self.in_channels = in_channels + self.input_dim = input_feat_per_channel + self.conv_kernel_sizes_and_strides = [] + self.conv_layers = nn.ModuleList() + lstm_input_dim = input_layers[-1] + for conv_layer in conv_layers: + out_channels, conv_kernel_size, conv_stride = conv_layer + self.conv_layers.append( + nn.Conv2d( + in_channels, + out_channels, + conv_kernel_size, + stride=conv_stride, + padding=conv_kernel_size // 2, + ) + ) + self.conv_kernel_sizes_and_strides.append((conv_kernel_size, conv_stride)) + in_channels = out_channels + lstm_input_dim //= conv_stride + + lstm_input_dim *= conv_layers[-1][0] + self.lstm_size = lstm_size + self.num_blstm_layers = num_blstm_layers + self.lstm = nn.LSTM( + input_size=lstm_input_dim, + hidden_size=lstm_size, + num_layers=num_blstm_layers, + dropout=dropout, + bidirectional=True, + ) + self.output_dim = 2 * lstm_size # bidirectional + if dropout > 0: + self.dropout = nn.Dropout(p=dropout) + else: + self.dropout = None + + def forward(self, src_tokens, src_lengths=None, **kwargs): + """ + Args + src_tokens: padded tensor (B, T, C * feat) + src_lengths: tensor of original lengths of input utterances (B,) + """ + bsz, max_seq_len, _ = src_tokens.size() + # (B, C, T, feat) + x = ( + src_tokens.view(bsz, max_seq_len, self.in_channels, self.input_dim) + .transpose(1, 2) + .contiguous() + ) + + for input_layer in self.input_layers: + x = input_layer(x) + x = torch.tanh(x) + + for conv_layer in self.conv_layers: + x = conv_layer(x) + + bsz, _, output_seq_len, _ = x.size() + + # (B, C, T, feat) -> (B, T, C, feat) -> (T, B, C, feat) -> + # (T, B, C * feat) + x = x.transpose(1, 2).transpose(0, 1).contiguous().view(output_seq_len, bsz, -1) + + input_lengths = src_lengths.clone() + for k, s in self.conv_kernel_sizes_and_strides: + p = k // 2 + input_lengths = (input_lengths.float() + 2 * p - k) / s + 1 + input_lengths = input_lengths.floor().long() + + packed_x = nn.utils.rnn.pack_padded_sequence(x, input_lengths) + + h0 = x.new(2 * self.num_blstm_layers, bsz, self.lstm_size).zero_() + c0 = x.new(2 * self.num_blstm_layers, bsz, self.lstm_size).zero_() + packed_outs, _ = self.lstm(packed_x, (h0, c0)) + + # unpack outputs and apply dropout + x, output_lengths = nn.utils.rnn.pad_packed_sequence(packed_outs) + if self.dropout is not None: + x = self.dropout(x) + + encoder_padding_mask = ( + lengths_to_padding_mask(output_lengths).to(src_tokens.device).t() + ) + + return { + "encoder_out": x, # (T, B, C) + "encoder_padding_mask": encoder_padding_mask, # (T, B) + } + + def reorder_encoder_out(self, encoder_out, new_order): + encoder_out["encoder_out"] = encoder_out["encoder_out"].index_select( + 1, new_order + ) + encoder_out["encoder_padding_mask"] = encoder_out[ + "encoder_padding_mask" + ].index_select(1, new_order) + return encoder_out + + +class MLPAttention(nn.Module): + """The original attention from Badhanau et al. (2014) + + https://arxiv.org/abs/1409.0473, based on a Multi-Layer Perceptron. + The attention score between position i in the encoder and position j in the + decoder is: alpha_ij = V_a * tanh(W_ae * enc_i + W_ad * dec_j + b_a) + """ + + def __init__(self, decoder_hidden_state_dim, context_dim, attention_dim): + super().__init__() + + self.context_dim = context_dim + self.attention_dim = attention_dim + # W_ae and b_a + self.encoder_proj = nn.Linear(context_dim, self.attention_dim, bias=True) + # W_ad + self.decoder_proj = nn.Linear( + decoder_hidden_state_dim, self.attention_dim, bias=False + ) + # V_a + self.to_scores = nn.Linear(self.attention_dim, 1, bias=False) + + def forward(self, decoder_state, source_hids, encoder_padding_mask): + """The expected input dimensions are: + decoder_state: bsz x decoder_hidden_state_dim + source_hids: src_len x bsz x context_dim + encoder_padding_mask: src_len x bsz + """ + src_len, bsz, _ = source_hids.size() + # (src_len*bsz) x context_dim (to feed through linear) + flat_source_hids = source_hids.view(-1, self.context_dim) + # (src_len*bsz) x attention_dim + encoder_component = self.encoder_proj(flat_source_hids) + # src_len x bsz x attention_dim + encoder_component = encoder_component.view(src_len, bsz, self.attention_dim) + # 1 x bsz x attention_dim + decoder_component = self.decoder_proj(decoder_state).unsqueeze(0) + # Sum with broadcasting and apply the non linearity + # src_len x bsz x attention_dim + hidden_att = torch.tanh( + (decoder_component + encoder_component).view(-1, self.attention_dim) + ) + # Project onto the reals to get attentions scores (src_len x bsz) + attn_scores = self.to_scores(hidden_att).view(src_len, bsz) + + # Mask + softmax (src_len x bsz) + if encoder_padding_mask is not None: + attn_scores = ( + attn_scores.float() + .masked_fill_(encoder_padding_mask, float("-inf")) + .type_as(attn_scores) + ) # FP16 support: cast to float and back + # srclen x bsz + normalized_masked_attn_scores = F.softmax(attn_scores, dim=0) + + # Sum weighted sources (bsz x context_dim) + attn_weighted_context = ( + source_hids * normalized_masked_attn_scores.unsqueeze(2) + ).sum(dim=0) + + return attn_weighted_context, normalized_masked_attn_scores + + +class LSTMDecoder(FairseqIncrementalDecoder): + def __init__( + self, + dictionary, + embed_dim, + num_layers, + hidden_size, + dropout, + encoder_output_dim, + attention_dim, + output_layer_dim, + ): + """ + Args: + dictionary: target text dictionary. + embed_dim: embedding dimension for target tokens. + num_layers: number of LSTM layers. + hidden_size: hidden size for LSTM layers. + dropout: dropout probability. Dropout can be applied to the + embeddings, the LSTM layers, and the context vector. + encoder_output_dim: encoder output dimension (hidden size of + encoder LSTM). + attention_dim: attention dimension for MLP attention. + output_layer_dim: size of the linear layer prior to output + projection. + """ + super().__init__(dictionary) + self.num_layers = num_layers + self.hidden_size = hidden_size + num_embeddings = len(dictionary) + padding_idx = dictionary.pad() + self.embed_tokens = nn.Embedding(num_embeddings, embed_dim, padding_idx) + if dropout > 0: + self.dropout = nn.Dropout(p=dropout) + else: + self.dropout = None + + self.layers = nn.ModuleList() + for layer_id in range(num_layers): + input_size = embed_dim if layer_id == 0 else encoder_output_dim + self.layers.append( + nn.LSTMCell(input_size=input_size, hidden_size=hidden_size) + ) + + self.context_dim = encoder_output_dim + self.attention = MLPAttention( + decoder_hidden_state_dim=hidden_size, + context_dim=encoder_output_dim, + attention_dim=attention_dim, + ) + + self.deep_output_layer = nn.Linear( + hidden_size + encoder_output_dim + embed_dim, output_layer_dim + ) + self.output_projection = nn.Linear(output_layer_dim, num_embeddings) + + def forward( + self, prev_output_tokens, encoder_out=None, incremental_state=None, **kwargs + ): + encoder_padding_mask = encoder_out["encoder_padding_mask"] + encoder_outs = encoder_out["encoder_out"] + + if incremental_state is not None: + prev_output_tokens = prev_output_tokens[:, -1:] + bsz, seqlen = prev_output_tokens.size() + + srclen = encoder_outs.size(0) + + # embed tokens + embeddings = self.embed_tokens(prev_output_tokens) + x = embeddings + if self.dropout is not None: + x = self.dropout(x) + + # B x T x C -> T x B x C + x = x.transpose(0, 1) + + # initialize previous states (or get from cache during incremental + # generation) + cached_state = utils.get_incremental_state( + self, incremental_state, "cached_state" + ) + if cached_state is not None: + prev_hiddens, prev_cells = cached_state + else: + prev_hiddens = [encoder_out["encoder_out"].mean(dim=0)] * self.num_layers + prev_cells = [x.new_zeros(bsz, self.hidden_size)] * self.num_layers + + attn_scores = x.new_zeros(bsz, srclen) + attention_outs = [] + outs = [] + for j in range(seqlen): + input = x[j, :, :] + attention_out = None + for i, layer in enumerate(self.layers): + # the previous state is one layer below except for the bottom + # layer where the previous state is the state emitted by the + # top layer + hidden, cell = layer( + input, + ( + prev_hiddens[(i - 1) % self.num_layers], + prev_cells[(i - 1) % self.num_layers], + ), + ) + if self.dropout is not None: + hidden = self.dropout(hidden) + prev_hiddens[i] = hidden + prev_cells[i] = cell + if attention_out is None: + attention_out, attn_scores = self.attention( + hidden, encoder_outs, encoder_padding_mask + ) + if self.dropout is not None: + attention_out = self.dropout(attention_out) + attention_outs.append(attention_out) + input = attention_out + + # collect the output of the top layer + outs.append(hidden) + + # cache previous states (no-op except during incremental generation) + utils.set_incremental_state( + self, incremental_state, "cached_state", (prev_hiddens, prev_cells) + ) + + # collect outputs across time steps + x = torch.cat(outs, dim=0).view(seqlen, bsz, self.hidden_size) + attention_outs_concat = torch.cat(attention_outs, dim=0).view( + seqlen, bsz, self.context_dim + ) + + # T x B x C -> B x T x C + x = x.transpose(0, 1) + attention_outs_concat = attention_outs_concat.transpose(0, 1) + + # concat LSTM output, attention output and embedding + # before output projection + x = torch.cat((x, attention_outs_concat, embeddings), dim=2) + x = self.deep_output_layer(x) + x = torch.tanh(x) + if self.dropout is not None: + x = self.dropout(x) + # project back to size of vocabulary + x = self.output_projection(x) + + # to return the full attn_scores tensor, we need to fix the decoder + # to account for subsampling input frames + # return x, attn_scores + return x, None + + def reorder_incremental_state(self, incremental_state, new_order): + super().reorder_incremental_state(incremental_state, new_order) + cached_state = utils.get_incremental_state( + self, incremental_state, "cached_state" + ) + if cached_state is None: + return + + def reorder_state(state): + if isinstance(state, list): + return [reorder_state(state_i) for state_i in state] + return state.index_select(0, new_order) + + new_state = tuple(map(reorder_state, cached_state)) + utils.set_incremental_state(self, incremental_state, "cached_state", new_state) + + +@register_model_architecture(model_name="s2t_berard", arch_name="s2t_berard") +def berard(args): + """The original version: "End-to-End Automatic Speech Translation of + Audiobooks" (https://arxiv.org/abs/1802.04200) + """ + args.input_layers = getattr(args, "input_layers", "[256, 128]") + args.conv_layers = getattr(args, "conv_layers", "[(16, 3, 2), (16, 3, 2)]") + args.num_blstm_layers = getattr(args, "num_blstm_layers", 3) + args.lstm_size = getattr(args, "lstm_size", 256) + args.dropout = getattr(args, "dropout", 0.2) + args.decoder_embed_dim = getattr(args, "decoder_embed_dim", 128) + args.decoder_num_layers = getattr(args, "decoder_num_layers", 2) + args.decoder_hidden_dim = getattr(args, "decoder_hidden_dim", 512) + args.attention_dim = getattr(args, "attention_dim", 512) + args.output_layer_dim = getattr(args, "output_layer_dim", 128) + args.load_pretrained_encoder_from = getattr( + args, "load_pretrained_encoder_from", None + ) + args.load_pretrained_decoder_from = getattr( + args, "load_pretrained_decoder_from", None + ) + + +@register_model_architecture(model_name="s2t_berard", arch_name="s2t_berard_256_3_3") +def berard_256_3_3(args): + """Used in + * "Harnessing Indirect Training Data for End-to-End Automatic Speech + Translation: Tricks of the Trade" (https://arxiv.org/abs/1909.06515) + * "CoVoST: A Diverse Multilingual Speech-To-Text Translation Corpus" + (https://arxiv.org/pdf/2002.01320.pdf) + * "Self-Supervised Representations Improve End-to-End Speech Translation" + (https://arxiv.org/abs/2006.12124) + """ + args.decoder_num_layers = getattr(args, "decoder_num_layers", 3) + berard(args) + + +@register_model_architecture(model_name="s2t_berard", arch_name="s2t_berard_512_3_2") +def berard_512_3_2(args): + args.num_blstm_layers = getattr(args, "num_blstm_layers", 3) + args.lstm_size = getattr(args, "lstm_size", 512) + args.dropout = getattr(args, "dropout", 0.3) + args.decoder_embed_dim = getattr(args, "decoder_embed_dim", 256) + args.decoder_num_layers = getattr(args, "decoder_num_layers", 2) + args.decoder_hidden_dim = getattr(args, "decoder_hidden_dim", 1024) + args.attention_dim = getattr(args, "attention_dim", 512) + args.output_layer_dim = getattr(args, "output_layer_dim", 256) + berard(args) + + +@register_model_architecture(model_name="s2t_berard", arch_name="s2t_berard_512_5_3") +def berard_512_5_3(args): + args.num_blstm_layers = getattr(args, "num_blstm_layers", 5) + args.lstm_size = getattr(args, "lstm_size", 512) + args.dropout = getattr(args, "dropout", 0.3) + args.decoder_embed_dim = getattr(args, "decoder_embed_dim", 256) + args.decoder_num_layers = getattr(args, "decoder_num_layers", 3) + args.decoder_hidden_dim = getattr(args, "decoder_hidden_dim", 1024) + args.attention_dim = getattr(args, "attention_dim", 512) + args.output_layer_dim = getattr(args, "output_layer_dim", 256) + berard(args) diff --git a/data/fairseq/fairseq/models/speech_to_text/convtransformer.py b/data/fairseq/fairseq/models/speech_to_text/convtransformer.py new file mode 100644 index 0000000000000000000000000000000000000000..4d0fc02aee908ad50349946d7fbbfec563ec07e8 --- /dev/null +++ b/data/fairseq/fairseq/models/speech_to_text/convtransformer.py @@ -0,0 +1,443 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import logging +import math +from typing import Dict, List, Optional, Tuple + +import torch +import torch.nn as nn +import torch.nn.functional as F +from torch import Tensor + +from fairseq import checkpoint_utils, utils +from fairseq.data.data_utils import lengths_to_padding_mask +from fairseq.models import ( + FairseqEncoder, + FairseqEncoderDecoderModel, + register_model, + register_model_architecture, +) +from fairseq.models.speech_to_text.modules.convolution import infer_conv_output_dim +from fairseq.models.transformer import Embedding, TransformerDecoder +from fairseq.modules import LayerNorm, PositionalEmbedding, TransformerEncoderLayer + +logger = logging.getLogger(__name__) + + +@register_model("convtransformer") +class ConvTransformerModel(FairseqEncoderDecoderModel): + """ + Transformer-based Speech translation model from ESPNet-ST + https://arxiv.org/abs/2004.10234 + """ + + def __init__(self, encoder, decoder): + super().__init__(encoder, decoder) + + @staticmethod + def add_args(parser): + """Add model-specific arguments to the parser.""" + parser.add_argument( + "--input-feat-per-channel", + type=int, + metavar="N", + help="encoder input dimension per input channel", + ) + parser.add_argument( + "--activation-fn", + choices=utils.get_available_activation_fns(), + help="activation function to use", + ) + parser.add_argument( + "--dropout", type=float, metavar="D", help="dropout probability" + ) + parser.add_argument( + "--attention-dropout", + type=float, + metavar="D", + help="dropout probability for attention weights", + ) + parser.add_argument( + "--activation-dropout", + "--relu-dropout", + type=float, + metavar="D", + help="dropout probability after activation in FFN.", + ) + parser.add_argument( + "--encoder-embed-dim", + type=int, + metavar="N", + help="encoder embedding dimension", + ) + parser.add_argument( + "--encoder-ffn-embed-dim", + type=int, + metavar="N", + help="encoder embedding dimension for FFN", + ) + parser.add_argument( + "--encoder-layers", type=int, metavar="N", help="num encoder layers" + ) + parser.add_argument( + "--encoder-attention-heads", + type=int, + metavar="N", + help="num encoder attention heads", + ) + parser.add_argument( + "--encoder-normalize-before", + action="store_true", + help="apply layernorm before each encoder block", + ) + parser.add_argument( + "--decoder-embed-dim", + type=int, + metavar="N", + help="decoder embedding dimension", + ) + parser.add_argument( + "--decoder-ffn-embed-dim", + type=int, + metavar="N", + help="decoder embedding dimension for FFN", + ) + parser.add_argument( + "--decoder-layers", type=int, metavar="N", help="num decoder layers" + ) + parser.add_argument( + "--decoder-attention-heads", + type=int, + metavar="N", + help="num decoder attention heads", + ) + parser.add_argument( + "--decoder-normalize-before", + action="store_true", + help="apply layernorm before each decoder block", + ) + parser.add_argument( + "--decoder-output-dim", + type=int, + metavar="N", + help="decoder output dimension (extra linear layer if different from decoder embed dim)", + ) + parser.add_argument( + "--share-decoder-input-output-embed", + action="store_true", + help="share decoder input and output embeddings", + ) + parser.add_argument( + "--layernorm-embedding", + action="store_true", + help="add layernorm to embedding", + ) + parser.add_argument( + "--no-scale-embedding", + action="store_true", + help="if True, dont scale embeddings", + ) + parser.add_argument( + "--load-pretrained-encoder-from", + type=str, + metavar="STR", + help="model to take encoder weights from (for initialization)", + ) + parser.add_argument( + "--load-pretrained-decoder-from", + type=str, + metavar="STR", + help="model to take decoder weights from (for initialization)", + ) + parser.add_argument( + "--conv-out-channels", + type=int, + metavar="INT", + help="the number of output channels of conv layer", + ) + + @classmethod + def build_encoder(cls, args): + encoder = ConvTransformerEncoder(args) + if getattr(args, "load_pretrained_encoder_from", None) is not None: + encoder = checkpoint_utils.load_pretrained_component_from_model( + component=encoder, checkpoint=args.load_pretrained_encoder_from + ) + return encoder + + @classmethod + def build_decoder(cls, args, task, embed_tokens): + decoder = TransformerDecoderNoExtra(args, task.target_dictionary, embed_tokens) + if getattr(args, "load_pretrained_decoder_from", None) is not None: + decoder = checkpoint_utils.load_pretrained_component_from_model( + component=decoder, checkpoint=args.load_pretrained_decoder_from + ) + return decoder + + @classmethod + def build_model(cls, args, task): + """Build a new model instance.""" + + # make sure all arguments are present in older models + base_architecture(args) + + def build_embedding(dictionary, embed_dim): + num_embeddings = len(dictionary) + padding_idx = dictionary.pad() + return Embedding(num_embeddings, embed_dim, padding_idx) + + decoder_embed_tokens = build_embedding( + task.target_dictionary, args.decoder_embed_dim + ) + encoder = cls.build_encoder(args) + decoder = cls.build_decoder(args, task, decoder_embed_tokens) + return cls(encoder, decoder) + + @staticmethod + @torch.jit.unused + def set_batch_first(lprobs): + lprobs.batch_first = True + + def get_normalized_probs( + self, + net_output: Tuple[Tensor, Optional[Dict[str, List[Optional[Tensor]]]]], + log_probs: bool, + sample: Optional[Dict[str, Tensor]] = None, + ): + # net_output['encoder_out'] is a (B, T, D) tensor + lprobs = self.get_normalized_probs_scriptable(net_output, log_probs, sample) + if self.training: + self.set_batch_first(lprobs) + return lprobs + + def output_layout(self): + return "BTD" + + """ + The forward method inherited from the base class has a **kwargs argument in + its input, which is not supported in torchscript. This method overrites the forward + method definition without **kwargs. + """ + + def forward(self, src_tokens, src_lengths, prev_output_tokens): + encoder_out = self.encoder(src_tokens=src_tokens, src_lengths=src_lengths) + decoder_out = self.decoder( + prev_output_tokens=prev_output_tokens, encoder_out=encoder_out + ) + return decoder_out + + +class ConvTransformerEncoder(FairseqEncoder): + """Conv + Transformer encoder""" + + def __init__(self, args): + """Construct an Encoder object.""" + super().__init__(None) + + self.dropout = args.dropout + self.embed_scale = ( + 1.0 if args.no_scale_embedding else math.sqrt(args.encoder_embed_dim) + ) + self.padding_idx = 1 + self.in_channels = 1 + self.input_dim = args.input_feat_per_channel + self.conv = torch.nn.Sequential( + torch.nn.Conv2d(1, args.conv_out_channels, 3, stride=2, padding=3 // 2), + torch.nn.ReLU(), + torch.nn.Conv2d( + args.conv_out_channels, + args.conv_out_channels, + 3, + stride=2, + padding=3 // 2, + ), + torch.nn.ReLU(), + ) + transformer_input_dim = infer_conv_output_dim( + self.in_channels, self.input_dim, args.conv_out_channels + ) + self.out = torch.nn.Linear(transformer_input_dim, args.encoder_embed_dim) + self.embed_positions = PositionalEmbedding( + args.max_source_positions, + args.encoder_embed_dim, + self.padding_idx, + learned=False, + ) + + self.transformer_layers = nn.ModuleList([]) + self.transformer_layers.extend( + [TransformerEncoderLayer(args) for i in range(args.encoder_layers)] + ) + if args.encoder_normalize_before: + self.layer_norm = LayerNorm(args.encoder_embed_dim) + else: + self.layer_norm = None + + def pooling_ratio(self): + return 4 + + def forward(self, src_tokens, src_lengths): + """Encode input sequence. + :param torch.Tensor xs: input tensor + :param torch.Tensor masks: input mask + :return: position embedded tensor and mask + :rtype Tuple[torch.Tensor, torch.Tensor]: + """ + bsz, max_seq_len, _ = src_tokens.size() + x = ( + src_tokens.view(bsz, max_seq_len, self.in_channels, self.input_dim) + .transpose(1, 2) + .contiguous() + ) + x = self.conv(x) + bsz, _, output_seq_len, _ = x.size() + x = x.transpose(1, 2).transpose(0, 1).contiguous().view(output_seq_len, bsz, -1) + x = self.out(x) + x = self.embed_scale * x + + subsampling_factor = int(max_seq_len * 1.0 / output_seq_len + 0.5) + input_len_0 = (src_lengths.float() / subsampling_factor).ceil().long() + input_len_1 = x.size(0) * torch.ones([src_lengths.size(0)]).long().to( + input_len_0.device + ) + input_lengths = torch.min(input_len_0, input_len_1) + + encoder_padding_mask = lengths_to_padding_mask(input_lengths) + + positions = self.embed_positions(encoder_padding_mask).transpose(0, 1) + x += positions + x = F.dropout(x, p=self.dropout, training=self.training) + + for layer in self.transformer_layers: + x = layer(x, encoder_padding_mask) + + if not encoder_padding_mask.any(): + maybe_encoder_padding_mask = None + else: + maybe_encoder_padding_mask = encoder_padding_mask + + return { + "encoder_out": [x], + "encoder_padding_mask": [maybe_encoder_padding_mask] + if maybe_encoder_padding_mask is not None + else [], + "encoder_embedding": [], + "encoder_states": [], + "src_tokens": [], + "src_lengths": [], + } + + @torch.jit.export + def reorder_encoder_out(self, encoder_out: Dict[str, List[Tensor]], new_order): + """ + Reorder encoder output according to *new_order*. + + Args: + encoder_out: output from the ``forward()`` method + new_order (LongTensor): desired order + + Returns: + *encoder_out* rearranged according to *new_order* + """ + new_encoder_out = [encoder_out["encoder_out"][0].index_select(1, new_order)] + if len(encoder_out["encoder_padding_mask"]) == 0: + new_encoder_padding_mask = [] + else: + new_encoder_padding_mask = [ + (encoder_out["encoder_padding_mask"][0]).index_select(0, new_order) + ] + if len(encoder_out["encoder_embedding"]) == 0: + new_encoder_embedding = [] + else: + new_encoder_embedding = [ + (encoder_out["encoder_embedding"][0]).index_select(0, new_order) + ] + encoder_states = encoder_out["encoder_states"] + if len(encoder_states) > 0: + for idx, state in enumerate(encoder_states): + encoder_states[idx] = state.index_select(1, new_order) + + return { + "encoder_out": new_encoder_out, + "encoder_padding_mask": new_encoder_padding_mask, + "encoder_embedding": new_encoder_embedding, + "encoder_states": encoder_states, + "src_tokens": [], + "src_lengths": [], + } + + +class TransformerDecoderNoExtra(TransformerDecoder): + def extract_features( + self, + prev_output_tokens, + encoder_out: Optional[Dict[str, List[Tensor]]], + incremental_state: Optional[Dict[str, Dict[str, Optional[Tensor]]]] = None, + full_context_alignment: bool = False, + alignment_layer: Optional[int] = None, + alignment_heads: Optional[int] = None, + ): + # call scriptable method from parent class + x, _ = self.extract_features_scriptable( + prev_output_tokens, + encoder_out, + incremental_state, + full_context_alignment, + alignment_layer, + alignment_heads, + ) + return x, None + + +@register_model_architecture(model_name="convtransformer", arch_name="convtransformer") +def base_architecture(args): + args.input_feat_per_channel = getattr(args, "input_feat_per_channel", 80) + args.encoder_embed_dim = getattr(args, "encoder_embed_dim", 512) + args.encoder_ffn_embed_dim = getattr(args, "encoder_ffn_embed_dim", 2048) + args.encoder_layers = getattr(args, "encoder_layers", 6) + args.encoder_attention_heads = getattr(args, "encoder_attention_heads", 8) + args.encoder_normalize_before = getattr(args, "encoder_normalize_before", False) + args.decoder_embed_dim = getattr(args, "decoder_embed_dim", args.encoder_embed_dim) + args.decoder_ffn_embed_dim = getattr( + args, "decoder_ffn_embed_dim", args.encoder_ffn_embed_dim + ) + args.decoder_layers = getattr(args, "decoder_layers", 6) + args.decoder_attention_heads = getattr(args, "decoder_attention_heads", 8) + args.decoder_normalize_before = getattr(args, "decoder_normalize_before", False) + args.decoder_learned_pos = getattr(args, "decoder_learned_pos", False) + args.attention_dropout = getattr(args, "attention_dropout", 0.0) + args.activation_dropout = getattr(args, "activation_dropout", 0.0) + args.activation_fn = getattr(args, "activation_fn", "relu") + args.dropout = getattr(args, "dropout", 0.1) + args.adaptive_softmax_cutoff = getattr(args, "adaptive_softmax_cutoff", None) + args.adaptive_softmax_dropout = getattr(args, "adaptive_softmax_dropout", 0) + args.share_decoder_input_output_embed = getattr( + args, "share_decoder_input_output_embed", False + ) + args.no_token_positional_embeddings = getattr( + args, "no_token_positional_embeddings", False + ) + args.adaptive_input = getattr(args, "adaptive_input", False) + args.decoder_layerdrop = getattr(args, "decoder_layerdrop", 0.0) + + args.decoder_output_dim = getattr( + args, "decoder_output_dim", args.decoder_embed_dim + ) + args.decoder_input_dim = getattr(args, "decoder_input_dim", args.decoder_embed_dim) + args.no_scale_embedding = getattr(args, "no_scale_embedding", False) + args.quant_noise_pq = getattr(args, "quant_noise_pq", 0) + args.max_source_positions = getattr(args, "max_source_positions", 3000) + args.max_target_positions = getattr(args, "max_target_positions", 1024) + args.tie_adaptive_weights = getattr(args, "tie_adaptive_weights", False) + args.conv_out_channels = getattr(args, "conv_out_channels", args.encoder_embed_dim) + + +@register_model_architecture("convtransformer", "convtransformer_espnet") +def convtransformer_espnet(args): + args.encoder_embed_dim = getattr(args, "encoder_embed_dim", 256) + args.encoder_layers = getattr(args, "encoder_layers", 12) + args.encoder_attention_heads = getattr(args, "encoder_attention_heads", 4) + args.decoder_attention_heads = getattr(args, "decoder_attention_heads", 4) diff --git a/data/fairseq/fairseq/models/speech_to_text/hub_interface.py b/data/fairseq/fairseq/models/speech_to_text/hub_interface.py new file mode 100644 index 0000000000000000000000000000000000000000..d78427f68724c2262c61d48b31194d97e38773da --- /dev/null +++ b/data/fairseq/fairseq/models/speech_to_text/hub_interface.py @@ -0,0 +1,128 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import logging +from argparse import Namespace +from typing import Optional, Tuple, Union + +import torch +import torch.nn as nn +import torch.nn.functional as F + +import fairseq.data.audio.feature_transforms.utterance_cmvn as utt_cmvn +from fairseq.data import encoders +from fairseq.data.audio.audio_utils import convert_waveform as convert_wav +from fairseq.data.audio.audio_utils import get_fbank +from fairseq.data.audio.audio_utils import get_waveform as get_wav +from fairseq.data.audio.speech_to_text_dataset import SpeechToTextDataset + +logger = logging.getLogger(__name__) + + +class S2THubInterface(nn.Module): + def __init__(self, cfg, task, model): + super().__init__() + self.cfg = cfg + self.task = task + self.model = model + self.model.eval() + self.generator = self.task.build_generator([self.model], self.cfg.generation) + + @classmethod + def get_model_input(cls, task, audio: Union[str, torch.Tensor]): + input_type = task.data_cfg.hub.get("input_type", "fbank80") + if input_type == "fbank80_w_utt_cmvn": + if isinstance(audio, str): + feat = utt_cmvn.UtteranceCMVN()(get_fbank(audio)) + feat = feat.unsqueeze(0) # T x D -> 1 x T x D + else: + import torchaudio.compliance.kaldi as kaldi + + feat = kaldi.fbank(audio, num_mel_bins=80).numpy() # 1 x T x D + elif input_type in {"waveform", "standardized_waveform"}: + if isinstance(audio, str): + feat, sr = get_wav(audio) # C x T + feat, _ = convert_wav( + feat, sr, to_sample_rate=16_000, to_mono=True + ) # C x T -> 1 x T + else: + feat = audio.numpy() + else: + raise ValueError(f"Unknown value: input_type = {input_type}") + + src_lengths = torch.Tensor([feat.shape[1]]).long() + src_tokens = torch.from_numpy(feat) # 1 x T (x D) + if input_type == "standardized_waveform": + with torch.no_grad(): + src_tokens = F.layer_norm(src_tokens, src_tokens.shape) + + return { + "net_input": { + "src_tokens": src_tokens, + "src_lengths": src_lengths, + "prev_output_tokens": None, + }, + "target_lengths": None, + "speaker": None, + } + + @classmethod + def detokenize(cls, task, tokens): + text = task.tgt_dict.string(tokens) + tkn_cfg = task.data_cfg.bpe_tokenizer + tokenizer = encoders.build_bpe(Namespace(**tkn_cfg)) + return text if tokenizer is None else tokenizer.decode(text) + + @classmethod + def get_prefix_token(cls, task, lang): + prefix_size = int(task.data_cfg.prepend_tgt_lang_tag) + prefix_tokens = None + if prefix_size > 0: + assert lang is not None + lang_tag = SpeechToTextDataset.get_lang_tag_idx(lang, task.tgt_dict) + prefix_tokens = torch.Tensor([lang_tag]).long().unsqueeze(0) + return prefix_tokens + + @classmethod + def get_prediction( + cls, task, model, generator, sample, tgt_lang=None, synthesize_speech=False + ) -> Union[str, Tuple[str, Tuple[torch.Tensor, int]]]: + _tgt_lang = tgt_lang or task.data_cfg.hub.get("tgt_lang", None) + prefix = cls.get_prefix_token(task, _tgt_lang) + pred_tokens = generator.generate([model], sample, prefix_tokens=prefix) + pred = cls.detokenize(task, pred_tokens[0][0]["tokens"]) + eos_token = task.data_cfg.config.get("eos_token", None) + if eos_token: + pred = " ".join(pred.split(" ")[:-1]) + + if synthesize_speech: + pfx = f"{_tgt_lang}_" if task.data_cfg.prepend_tgt_lang_tag else "" + tts_model_id = task.data_cfg.hub.get(f"{pfx}tts_model_id", None) + speaker = task.data_cfg.hub.get(f"{pfx}speaker", None) + if tts_model_id is None: + logger.warning("TTS model configuration not found") + else: + _repo, _id = tts_model_id.split(":") + tts_model = torch.hub.load(_repo, _id, verbose=False) + pred = (pred, tts_model.predict(pred, speaker=speaker)) + return pred + + def predict( + self, + audio: Union[str, torch.Tensor], + tgt_lang: Optional[str] = None, + synthesize_speech: bool = False, + ) -> Union[str, Tuple[str, Tuple[torch.Tensor, int]]]: + # `audio` is either a file path or a 1xT Tensor + # return either text or (text, synthetic speech) + sample = self.get_model_input(self.task, audio) + return self.get_prediction( + self.task, + self.model, + self.generator, + sample, + tgt_lang=tgt_lang, + synthesize_speech=synthesize_speech, + ) diff --git a/data/fairseq/fairseq/models/speech_to_text/modules/__init__.py b/data/fairseq/fairseq/models/speech_to_text/modules/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391 diff --git a/data/fairseq/fairseq/models/speech_to_text/modules/__pycache__/__init__.cpython-38.pyc b/data/fairseq/fairseq/models/speech_to_text/modules/__pycache__/__init__.cpython-38.pyc new file mode 100644 index 0000000000000000000000000000000000000000..763e098591a2a53271c4340ffaa7c93bfe3eac67 Binary files /dev/null and b/data/fairseq/fairseq/models/speech_to_text/modules/__pycache__/__init__.cpython-38.pyc differ diff --git a/data/fairseq/fairseq/models/speech_to_text/modules/__pycache__/convolution.cpython-38.pyc b/data/fairseq/fairseq/models/speech_to_text/modules/__pycache__/convolution.cpython-38.pyc new file mode 100644 index 0000000000000000000000000000000000000000..4bdd67863bbd54db153674e81559052cbd619d16 Binary files /dev/null and b/data/fairseq/fairseq/models/speech_to_text/modules/__pycache__/convolution.cpython-38.pyc differ diff --git a/data/fairseq/fairseq/models/speech_to_text/modules/augmented_memory_attention.py b/data/fairseq/fairseq/models/speech_to_text/modules/augmented_memory_attention.py new file mode 100644 index 0000000000000000000000000000000000000000..2d330f96f689864a51a0bd6d952d0ec3070186ac --- /dev/null +++ b/data/fairseq/fairseq/models/speech_to_text/modules/augmented_memory_attention.py @@ -0,0 +1,487 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from typing import List, Tuple + +import torch +import torch.nn.functional as F +from torch import Tensor, nn + +from fairseq.models import FairseqEncoder +from fairseq.models.speech_to_text import ConvTransformerEncoder +from fairseq.models.speech_to_text.utils import ( + attention_suppression, + lengths_to_encoder_padding_mask, + segments_to_sequence, + sequence_to_segments, +) +from fairseq.modules import MultiheadAttention, TransformerEncoderLayer + +# ------------------------------------------------------------------------------ +# AugmentedMemoryConvTransformerEncoder +# ------------------------------------------------------------------------------ + + +class AugmentedMemoryConvTransformerEncoder(ConvTransformerEncoder): + def __init__(self, args): + super().__init__(args) + + args.encoder_stride = self.stride() + + self.left_context = args.left_context // args.encoder_stride + + self.right_context = args.right_context // args.encoder_stride + + self.left_context_after_stride = args.left_context // args.encoder_stride + self.right_context_after_stride = args.right_context // args.encoder_stride + + self.transformer_layers = nn.ModuleList([]) + self.transformer_layers.extend( + [ + AugmentedMemoryTransformerEncoderLayer(args) + for i in range(args.encoder_layers) + ] + ) + + def stride(self): + # Hard coded here. Should infer from convs in future + stride = 4 + return stride + + def forward(self, src_tokens, src_lengths, states=None): + """Encode input sequence. + :param torch.Tensor xs: input tensor + :param torch.Tensor masks: input mask + :return: position embedded tensor and mask + :rtype Tuple[torch.Tensor, torch.Tensor]: + """ + bsz, max_seq_len, _ = src_tokens.size() + x = ( + src_tokens.view(bsz, max_seq_len, self.in_channels, self.input_dim) + .transpose(1, 2) + .contiguous() + ) + x = self.conv(x) + bsz, _, output_seq_len, _ = x.size() + x = x.transpose(1, 2).transpose(0, 1).contiguous().view(output_seq_len, bsz, -1) + x = self.out(x) + x = self.embed_scale * x + + subsampling_factor = 1.0 * max_seq_len / output_seq_len + input_lengths = torch.max( + (src_lengths.float() / subsampling_factor).ceil().long(), + x.size(0) * src_lengths.new_ones([src_lengths.size(0)]).long(), + ) + + encoder_padding_mask, _ = lengths_to_encoder_padding_mask( + input_lengths, batch_first=True + ) + + # TODO: fix positional embedding + positions = self.embed_positions(encoder_padding_mask).transpose(0, 1) + + x += positions + x = F.dropout(x, p=self.dropout, training=self.training) + + # State to store memory banks etc. + if states is None: + states = [ + {"memory_banks": None, "encoder_states": None} + for i in range(len(self.transformer_layers)) + ] + + for i, layer in enumerate(self.transformer_layers): + # x size: + # (self.left_size + self.segment_size + self.right_size) + # / self.stride, num_heads, dim + # TODO: Consider mask here + x = layer(x, states[i]) + states[i]["encoder_states"] = x[ + self.left_context_after_stride : -self.right_context_after_stride + ] + + lengths = ( + ( + ~encoder_padding_mask[ + :, self.left_context_after_stride : -self.right_context_after_stride + ] + ) + .sum(dim=1, keepdim=True) + .long() + ) + + return states[-1]["encoder_states"], lengths, states + + +# ------------------------------------------------------------------------------ +# AugmentedMemoryTransformerEncoderLayer +# ------------------------------------------------------------------------------ +class AugmentedMemoryTransformerEncoderLayer(TransformerEncoderLayer): + def __init__(self, args): + super().__init__(args) + + self.left_context = args.left_context // args.encoder_stride + self.right_context = args.right_context // args.encoder_stride + + def forward(self, x, state): + + length, batch_size, x_dim = x.size() + + residual = x + + if self.normalize_before: + x = self.self_attn_layer_norm(x) + + # init_state + if state.get("memory_banks", None) is None: + state["memory_banks"] = [] + + # TODO reseach new sum_query method + seg_start = self.left_context + seg_end = length - self.right_context + if seg_start < seg_end: + summarization_query = torch.mean(x[seg_start:seg_end], keepdim=True, dim=0) + else: + summarization_query = x.new_zeros(1, batch_size, x_dim) + + x = torch.cat([x, summarization_query], dim=0) + + x = self.self_attn(input_and_summary=x, state=state) + + x = self.dropout_module(x) + x = residual + x + + if not self.normalize_before: + x = self.self_attn_layer_norm(x) + + residual = x + if self.normalize_before: + x = self.final_layer_norm(x) + + x = self.activation_fn(self.fc1(x)) + x = self.activation_dropout_module(x) + x = self.fc2(x) + x = self.dropout_module(x) + x = residual + x + if not self.normalize_before: + x = self.final_layer_norm(x) + + return x + + def build_self_attention(self, embed_dim, args): + return AugmentedMemoryMultiheadAttention( + embed_dim=embed_dim, + num_heads=args.encoder_attention_heads, + dropout=args.attention_dropout, + self_attention=True, + q_noise=self.quant_noise, + qn_block_size=self.quant_noise_block_size, + tanh_on_mem=True, + max_memory_size=args.max_memory_size, + ) + + +# ------------------------------------------------------------------------------ +# AugmentedMemoryMultiheadAttention +# ------------------------------------------------------------------------------ +class AugmentedMemoryMultiheadAttention(MultiheadAttention): + """ + Augmented Memory Attention from + Streaming Transformer-based Acoustic Models + Using Self-attention with Augmented Memory + https://arxiv.org/abs/2005.08042 + """ + + def __init__( + self, + embed_dim, + num_heads, + kdim=None, + vdim=None, + dropout=0.0, + bias=True, + add_bias_kv=False, + add_zero_attn=False, + self_attention=False, + encoder_decoder_attention=False, + q_noise=0.0, + qn_block_size=8, + tanh_on_mem=False, + memory_dim=None, + std_scale=0.5, # 0.5 based on https://arxiv.org/abs/2005.09137 + max_memory_size=-1, + disable_mem_on_mem_attn=True, + ): + super().__init__( + embed_dim, + num_heads, + kdim, + vdim, + dropout, + bias, + add_bias_kv, + add_zero_attn, + self_attention, + encoder_decoder_attention, + q_noise, + qn_block_size, + ) + + self.memory_dim = memory_dim if memory_dim is not None else embed_dim + self.std_scale = std_scale + self.disable_mem_on_mem_attn = disable_mem_on_mem_attn + + # This Operator was used for factorization in PySpeech + self.v2e = lambda x: x + + if tanh_on_mem: + self.squash_mem = torch.tanh + self.nonlinear_squash_mem = True + else: + self.squash_mem = lambda x: x + self.nonlinear_squash_mem = False + + self.max_memory_size = max_memory_size + + def forward(self, input_and_summary, state): + """ + input: Encoder states of current segment with left or right context, + plus one summarization query + + """ + + length, batch_size, _ = input_and_summary.shape + length = length - 1 # not include sum_query, last index + + memory = state["memory_banks"] + # TODO: positional embedding on memory + + if self.max_memory_size > -1 and len(memory) > self.max_memory_size: + # TODO: need to fix here + if self.max_memory_size == 0: + memory = memory.new_zeros(1, memory.size(1), self.memory_dim) + else: + memory = memory[-self.max_memory_size :] + + memory_and_input = torch.cat(memory + [input_and_summary[:-1]], dim=0) + input_and_sum_query = input_and_summary + + q = self.q_proj(self.v2e(input_and_sum_query)) + k = self.k_proj(self.v2e(memory_and_input)) + v = self.v_proj(self.v2e(memory_and_input)) + + q = ( + q.contiguous() + .view(-1, batch_size * self.num_heads, self.head_dim) + .transpose(0, 1) + * self.scaling + ) + k = ( + k.contiguous() + .view(-1, batch_size * self.num_heads, self.head_dim) + .transpose(0, 1) + ) + + v = ( + v.contiguous() + .view(-1, batch_size * self.num_heads, self.head_dim) + .transpose(0, 1) + ) + + attention_weights = torch.bmm(q, k.transpose(1, 2)) + + if self.disable_mem_on_mem_attn: + attention_weights = self.suppress_mem_on_mem_attention( + batch_size, self.num_heads, len(memory), attention_weights + ) + + if self.std_scale is not None: + attention_weights = attention_suppression(attention_weights, self.std_scale) + + assert list(attention_weights.shape) == [ + batch_size * self.num_heads, + length + 1, + length + len(memory), + ] + + attention_weights = torch.nn.functional.softmax( + attention_weights.float(), dim=-1 + ).type_as(attention_weights) + + attention_probs = self.dropout_module(attention_weights) + + # [T, T, B, n_head] + [T, B, n_head, d_head] -> [T, B, n_head, d_head] + attention = torch.bmm(attention_probs, v) + + assert list(attention.shape) == [ + batch_size * self.num_heads, + length + 1, + self.head_dim, + ] + + attention = ( + attention.transpose(0, 1) + .contiguous() + .view(length + 1, batch_size, self.embed_dim) + ) + + output_and_memory = self.out_proj(attention) + + next_m = output_and_memory[-1:] + next_m = self.squash_mem(next_m) + output = output_and_memory[:-1] + + state["memory_banks"].append(next_m) + + return output + + def suppress_mem_on_mem_attention( + self, B: int, num_heads: int, mem_size: int, attention_weight: Tensor + ): + """ + Arguments: + - B: batch size + - num_heads: number of attention heads + - mem_size: size of memory bank + - attention_weight: a [B*num_heads, T + 1, T + mem_size] vector + + Return: + modified attention_weight with [B*num_heads, -1, :mem_size] = -inf + """ + attention_weight[:, -1, :mem_size] = float("-inf") + return attention_weight + + +# ------------------------------------------------------------------------------ +# SequenceEncoder +# ------------------------------------------------------------------------------ +class SequenceEncoder(FairseqEncoder): + """ + SequenceEncoder encodes sequences. + + More specifically, `src_tokens` and `src_lengths` in `forward()` should + describe a batch of "complete" sequences rather than segments. + + Segment-by-segment inference can be triggered by `segment_size`: + 1) `segment_size` is None: + SequenceEncoder treats the input sequence as one single segment. + 2) `segment_size` is not None (some int instead): + SequenceEncoder does the following: + 1. breaks the input sequence into several segments + 2. inference on each segment and collect the outputs + 3. concatanete segment outputs into the output sequence. + Note that `segment_size` here shouldn't include additional left/right + contexts needed, for example if we wish to infer with LC-BLSTM where the + middle chunk size is 100 and right context is 20, `segment_size` should be + 100. + """ + + def __init__(self, args, module): + super().__init__(None) + + self.module = module + self.input_time_axis = 1 + self.output_time_axis = 0 + self.segment_size = args.segment_size + self.left_context = args.left_context + self.right_context = args.right_context + + def forward( + self, + src_tokens: Tensor, + src_lengths: Tensor, + states=None, + ): + + seg_src_tokens_lengths = sequence_to_segments( + sequence=src_tokens, + time_axis=self.input_time_axis, + lengths=src_lengths, + segment_size=self.segment_size, + extra_left_context=self.left_context, + extra_right_context=self.right_context, + ) + + seg_encoder_states_lengths: List[Tuple[Tensor, Tensor]] = [] + + for seg_src_tokens, seg_src_lengths in seg_src_tokens_lengths: + (seg_encoder_states, seg_enc_lengths, states) = self.module( + seg_src_tokens, + seg_src_lengths, + states=states, + ) + + seg_encoder_states_lengths.append((seg_encoder_states, seg_enc_lengths)) + + encoder_out, enc_lengths = segments_to_sequence( + segments=seg_encoder_states_lengths, time_axis=self.output_time_axis + ) + + encoder_padding_mask, _ = lengths_to_encoder_padding_mask( + enc_lengths, batch_first=True + ) + + if not encoder_padding_mask.any(): + encoder_padding_mask = None + + return { + "encoder_out": [encoder_out], + "encoder_padding_mask": [encoder_padding_mask], + "encoder_embedding": [], + "encoder_states": [states], + "src_tokens": [], + "src_lengths": [], + } + + def incremental_encode( + self, + seg_src_tokens: Tensor, + seg_src_lengths: Tensor, + states=None, + ): + """ + Different from forward function, this function takes segmented speech + as input, and append encoder states to previous states + """ + (seg_encoder_states, seg_enc_lengths, states) = self.module( + seg_src_tokens, + seg_src_lengths, + states=states, + ) + return seg_encoder_states, seg_enc_lengths, states + + +# ------------------------------------------------------------------------------ +# Augmented memory model decorator +# ------------------------------------------------------------------------------ +def augmented_memory(klass): + class StreamSeq2SeqModel(klass): + @staticmethod + def add_args(parser): + super(StreamSeq2SeqModel, StreamSeq2SeqModel).add_args(parser) + parser.add_argument( + "--segment-size", type=int, required=True, help="Length of the segment." + ) + parser.add_argument( + "--left-context", + type=int, + default=0, + help="Left context for the segment.", + ) + parser.add_argument( + "--right-context", + type=int, + default=0, + help="Right context for the segment.", + ) + parser.add_argument( + "--max-memory-size", + type=int, + default=-1, + help="Right context for the segment.", + ) + + StreamSeq2SeqModel.__name__ = klass.__name__ + return StreamSeq2SeqModel diff --git a/data/fairseq/fairseq/models/speech_to_text/modules/convolution.py b/data/fairseq/fairseq/models/speech_to_text/modules/convolution.py new file mode 100644 index 0000000000000000000000000000000000000000..526d7540c59bd46a5f49f462e4c52c8b16d76aca --- /dev/null +++ b/data/fairseq/fairseq/models/speech_to_text/modules/convolution.py @@ -0,0 +1,126 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + + +from typing import List + +import torch +import torch.nn as nn + + +class Conv1dSubsampler(nn.Module): + """Convolutional subsampler: a stack of 1D convolution (along temporal + dimension) followed by non-linear activation via gated linear units + (https://arxiv.org/abs/1911.08460) + + Args: + in_channels (int): the number of input channels + mid_channels (int): the number of intermediate channels + out_channels (int): the number of output channels + kernel_sizes (List[int]): the kernel size for each convolutional layer + """ + + def __init__( + self, + in_channels: int, + mid_channels: int, + out_channels: int, + kernel_sizes: List[int] = (3, 3), + ): + super(Conv1dSubsampler, self).__init__() + self.n_layers = len(kernel_sizes) + self.conv_layers = nn.ModuleList( + nn.Conv1d( + in_channels if i == 0 else mid_channels // 2, + mid_channels if i < self.n_layers - 1 else out_channels * 2, + k, + stride=2, + padding=k // 2, + ) + for i, k in enumerate(kernel_sizes) + ) + + def get_out_seq_lens_tensor(self, in_seq_lens_tensor): + out = in_seq_lens_tensor.clone() + for _ in range(self.n_layers): + out = ((out.float() - 1) / 2 + 1).floor().long() + return out + + def forward(self, src_tokens, src_lengths): + bsz, in_seq_len, _ = src_tokens.size() # B x T x (C x D) + x = src_tokens.transpose(1, 2).contiguous() # -> B x (C x D) x T + for conv in self.conv_layers: + x = conv(x) + x = nn.functional.glu(x, dim=1) + _, _, out_seq_len = x.size() + x = x.transpose(1, 2).transpose(0, 1).contiguous() # -> T x B x (C x D) + return x, self.get_out_seq_lens_tensor(src_lengths) + + +def infer_conv_output_dim(in_channels, input_dim, out_channels): + sample_seq_len = 200 + sample_bsz = 10 + x = torch.randn(sample_bsz, in_channels, sample_seq_len, input_dim) + x = torch.nn.Conv2d(in_channels, out_channels, 3, stride=2, padding=3 // 2)(x) + x = torch.nn.Conv2d(out_channels, out_channels, 3, stride=2, padding=3 // 2)(x) + x = x.transpose(1, 2) + mb, seq = x.size()[:2] + return x.contiguous().view(mb, seq, -1).size(-1) + + +class Conv2dSubsampler(nn.Module): + """Convolutional subsampler: a stack of 2D convolution based on ESPnet implementation + (https://github.com/espnet/espnet) + + Args: + input_channels (int): the number of input channels + input_feat_per_channel (int): encoder input dimension per input channel + conv_out_channels (int): the number of output channels of conv layer + encoder_embed_dim (int): encoder dimentions + """ + + def __init__( + self, + input_channels: int, + input_feat_per_channel: int, + conv_out_channels: int, + encoder_embed_dim: int, + ): + super().__init__() + assert input_channels == 1, input_channels + self.conv = torch.nn.Sequential( + torch.nn.Conv2d( + input_channels, conv_out_channels, 3, stride=2, padding=3 // 2 + ), + torch.nn.ReLU(), + torch.nn.Conv2d( + conv_out_channels, + conv_out_channels, + 3, + stride=2, + padding=3 // 2, + ), + torch.nn.ReLU(), + ) + transformer_input_dim = infer_conv_output_dim( + input_channels, input_feat_per_channel, conv_out_channels + ) + self.out = torch.nn.Linear(transformer_input_dim, encoder_embed_dim) + + def forward(self, src_tokens, src_lengths): + B, T_i, C = src_tokens.size() + x = src_tokens.view(B, T_i, 1, C).transpose(1, 2).contiguous() + x = self.conv(x) + B, _, T_o, _ = x.size() + x = x.transpose(1, 2).transpose(0, 1).contiguous().view(T_o, B, -1) + x = self.out(x) + + subsampling_factor = int(T_i * 1.0 / T_o + 0.5) + input_len_0 = (src_lengths.float() / subsampling_factor).ceil().long() + input_len_1 = x.size(0) * torch.ones([src_lengths.size(0)]).long().to( + input_len_0.device + ) + input_lengths = torch.min(input_len_0, input_len_1) + return x, input_lengths diff --git a/data/fairseq/fairseq/models/speech_to_text/modules/emformer.py b/data/fairseq/fairseq/models/speech_to_text/modules/emformer.py new file mode 100644 index 0000000000000000000000000000000000000000..935d5930787f3fd3d3dbe227e233c140a10f65b2 --- /dev/null +++ b/data/fairseq/fairseq/models/speech_to_text/modules/emformer.py @@ -0,0 +1,1844 @@ +#!/usr/bin/env python3 +# Copyright (c) 2017-present, Facebook, Inc. +# All rights reserved. +# +# This source code is licensed under the license found in the LICENSE file in +# the root directory of this source tree. An additional grant of patent rights +# can be found in the PATENTS file in the same directory. + + +import math +import re +from functools import partial +from typing import List, Optional, Tuple + +import torch +import torch.nn as nn +from torch import Tensor +from torch import device as Device + +from fairseq.models import FairseqEncoder +from fairseq.models.speech_to_text.utils import ( + NoOp, + attention_suppression, + layer_norm_backward_hook, + lengths_to_padding_mask, + segments_to_sequence, +) + +try: + import torch.ao.quantization as quantization + from torch.ao.quantization.qconfig import ( + default_dynamic_qconfig, + per_channel_dynamic_qconfig, + ) +except ImportError: + import torch.quantization as quantization + from torch.quantization.qconfig import ( + default_dynamic_qconfig, + per_channel_dynamic_qconfig, + ) + + +class RelativePositionEmbedding(nn.Module): + """ + Implementation according to https://arxiv.org/abs/1803.02155 + """ + + def __init__(self, head_dim, max_position, norm_init=True): + super().__init__() + self.head_dim = head_dim + self.max_position = max_position + self.embeddings = nn.Parameter(torch.Tensor(max_position * 2 + 1, head_dim)) + if norm_init: + nn.init.xavier_normal_(self.embeddings) + else: + nn.init.xavier_uniform_(self.embeddings) + + def forward(self, input: Tensor): + output = nn.functional.embedding(input.long(), self.embeddings) + return output + + +class Fp32LayerNorm(nn.Module): + def __init__( + self, + input_dim, + clamp_grad=True, + max_grad_value=256, + eps=1e-5, + elementwise_affine=True, + ): + super().__init__() + self.torch_module = torch.nn.LayerNorm( + input_dim, eps=eps, elementwise_affine=elementwise_affine + ) + if clamp_grad: + hook = partial(layer_norm_backward_hook, clamp_value=max_grad_value) + self.torch_module.register_backward_hook(hook) + + def forward(self, input): + output = torch.nn.functional.layer_norm( + input.float(), + self.torch_module.normalized_shape, + self.torch_module.weight.float() + if self.torch_module.weight is not None + else None, + self.torch_module.bias.float() + if self.torch_module.bias is not None + else None, + self.torch_module.eps, + ).type_as(input) + return output + + +# ------------------------------------------------------------------------------ +# PositionwiseFF +# ------------------------------------------------------------------------------ + + +class PositionwiseFF(nn.Module): + """ + FFN layer in transformer. + + Args: + input_dim: input embedding dimension + ffn_dim: FFN layer inner dimension + dropout_on_fc1: dropout for first linear layer + dropout_on_fc2: dropout fr second linear layer + activation_fn: activation function used after first linear layer. \ + Only relu or gelu is supported. + + """ + + def __init__( + self, input_dim, ffn_dim, dropout_on_fc1, dropout_on_fc2, activation_fn + ): + super(PositionwiseFF, self).__init__() + + self.input_dim = input_dim + self.ffn_dim = ffn_dim + if activation_fn == "relu": + ac = nn.ReLU() + elif activation_fn == "gelu": + ac = nn.GELU() + else: + raise ValueError("Unsupported activation_fn = ({})".format(activation_fn)) + + # fc1 -> ac -> dropout -> fc2 -> dropout + self.module = nn.Sequential( + nn.Linear(input_dim, ffn_dim), + ac, + nn.Dropout(dropout_on_fc1), + nn.Linear(ffn_dim, input_dim), + nn.Dropout(dropout_on_fc2), + ) + + self.layer_norm = Fp32LayerNorm(input_dim) + + def forward(self, input): + module_out = self.module(self.layer_norm(input)) + output = module_out + input + + return output + + def quantize_(self, params=None): + if params and "per_channel" in params and params["per_channel"]: + qconfig = per_channel_dynamic_qconfig + else: + qconfig = default_dynamic_qconfig + quantization.quantize_dynamic( + self, {torch.nn.Linear: qconfig}, dtype=torch.qint8, inplace=True + ) + return self + + +# ------------------------------------------------------------------------------ +# SummarizationLayer +# ------------------------------------------------------------------------------ + + +class SummarizationLayer(nn.Module): + def __init__(self, method, segment_size, embedding_dim): + super(SummarizationLayer, self).__init__() + self.segment_size = segment_size + self.embedding_dim = embedding_dim + nonlin_match = re.match(r"nonlinear\((?P[a-z]+),(?P[0-9]+)\)", method) + self.method = method + if method == "mean": + self.module = nn.AvgPool1d( + kernel_size=segment_size, + stride=segment_size, + ceil_mode=True, + ) + elif method == "max": + self.module = nn.MaxPool1d( + kernel_size=segment_size, + stride=segment_size, + ceil_mode=True, + ) + elif method == "linear": + self.module = nn.Linear(segment_size, 1) + elif nonlin_match: + nonlin_args = nonlin_match.groupdict() + act_type = nonlin_args["act"] + hid_dim = int(nonlin_args["dim"]) + if act_type == "relu": + act = nn.ReLU() + elif act_type == "gelu": + act = nn.GELU() + else: + raise ValueError("Unsupported activation_fn = ({})".format(act_type)) + self.module = nn.Sequential( + nn.Linear(segment_size, hid_dim), + act, + nn.Linear(hid_dim, 1), + ) + else: + raise ValueError("Unsupported summarization method = ({})".format(method)) + + def forward(self, input): + # T, B, D -> B, D, T + input = input.permute(1, 2, 0) + + if self.method == "mean" or self.method == "max": + output = self.module(input) + output = output.permute(2, 0, 1) + return output + + full_seg_length = input.size(2) // self.segment_size * self.segment_size + if full_seg_length > 0: + # at least one seg is full + B = input.size(0) + D = input.size(1) + input_todo = ( + input[:, :, :full_seg_length] + .contiguous() + .view(B, -1, self.segment_size) + ) + output = self.module(input_todo) + output = output.view(B, D, -1) + else: + output = input.new_zeros(input.size(0), input.size(1), 0) + left = input.size(2) - full_seg_length + if left > 0: + # when last seg is not full, use zeros as last memory placeholder + zeros = input.new_zeros(input.size(0), input.size(1), 1) + output = torch.cat([output, zeros], dim=2) + output = output.permute(2, 0, 1) + return output + + +# ------------------------------------------------------------------------------ +# NoSegAugmentedMemoryMultiheadAttentionBmm +# ------------------------------------------------------------------------------ + + +class NoSegAugmentedMemoryMultiheadAttentionBmm(nn.Module): + """ + Whole utterance augmented memory multihead attention using BMM. + + Different with previous augmented memory multihead attention where + the utterance is chunked into segments. Here we use attention mask + achieve so. The input embedding [right_context, utterance, summary] + is a concatenation of right context, utterance and summary. + + Right context block is the concatenation of all the right context for + each segments. [right_context_0, right_context_1, ..., right_context_n] + For example, if we have utterance = [v0, v1, v2, ...., v20]. segment + size 8, right_context size 4. Then the right context blocks = + [v8, v9, v10, v11, v16, v17, v18, v19, 0, 0, 0, 0], where v8, v9, v10, + and v11 are the right context for first segment. v16, v17, v18 and v19 + are the right context for second segment. 0, 0, 0 and 0 are right context + for the last segment. + + utterance is corresponding to input embedding sequence + + summary is concatenation of average of each segments. [summary_0, + summary_1, ..., ]. + + In augmented memory multihead attention, the query is [right_context, + utterance, summary], key is [memory, right_context, utterance]. Different + with AugmentedMemoryMultiheadAttentionBmm, memory here is passed from + previous attention layer. For the first attention layer, memory is average + of each segment. + + Memory is a concatenation of memory from each segments in previous attention + layer. For example, current layer is i, then memory is [m_0, m_1, ..., m_n]. + Each m_k is the output from seg_k in layer i-1. + + args: + input_dim: input embedding dimension + num_heads: number of heads in multihead self-attention + dropout: attention dropout + std_scale: if std_scale is not None. The weak attention suppression is + turned on. For std_scale = 0.5, all the attention smaller than + mean + 0.5 * std will be suppressed. + scaled_init: whether to use scaled init for linear weight + tanh_on_mem: whether to use tanh on memory output + use_mem: whether to use memory or not. When max_memory_size is 0, then + we don't have memory anymore. + layer_index: current self-attention layer index that is used in depth + initialization + max_relative_position: max relative position used in relative position + embedding + rpe_old_option: To be compatible with previous model. The previous model + was trained with attention += attention + rpe. The correct equation + should be attention = attention + rpe + + """ + + def __init__( + self, + input_dim, + num_heads, + dropout=0.0, + std_scale=None, + scaled_init=False, + tanh_on_mem=False, + use_mem=True, + mini_batches=False, + negative_inf="-inf", + layer_index=-1, + max_relative_position=0, + rpe_old_option=True, + ): + if input_dim % num_heads: + raise ValueError( + "input_dim ({}) must be divisible by num_heads ({})".format( + input_dim, num_heads + ) + ) + + super().__init__() + + embed_dim = input_dim + self.e2h_kv = torch.nn.Linear(input_dim, 2 * input_dim, bias=True) + self.e2h_q = torch.nn.Linear(input_dim, input_dim, bias=True) + self.rpe_old_option = rpe_old_option + if max_relative_position > 0: + self.use_rpe = True + self.rpe_k = RelativePositionEmbedding( + head_dim=input_dim // num_heads, + max_position=max_relative_position, + ) + self.rpe_v = RelativePositionEmbedding( + head_dim=input_dim // num_heads, + max_position=max_relative_position, + ) + else: + self.use_rpe = False + self.rpe_k = None + self.rpe_v = None + if scaled_init: + if layer_index == -1: + gain = 1.0 / math.sqrt(2) + else: + # https://arxiv.org/abs/2005.09684 depthwise initialization + # stablize the training greatly. Use depthwise initialization to + # replace incremental loss. + gain = 1.0 / math.sqrt(layer_index + 1) + torch.nn.init.xavier_uniform_(self.e2h_kv.weight, gain=gain) + torch.nn.init.xavier_uniform_(self.e2h_q.weight, gain=gain) + + self.out_proj = torch.nn.Linear(embed_dim, embed_dim, bias=True) + + self.embed_dim = embed_dim + self.num_heads = num_heads + self.dropout = dropout + + self.head_dim = embed_dim // num_heads + self.scaling = self.head_dim**-0.5 + + self.std_scale = std_scale + self.use_mem = use_mem + self.mini_batches = mini_batches + self.negative_inf = negative_inf + + if tanh_on_mem: + self.squash_mem = torch.tanh + self.nonlinear_squash_mem = True + else: + self.squash_mem = NoOp() + self.nonlinear_squash_mem = False + + def prepare_qkv( + self, + input: Tensor, + mems: Tensor, + lengths: Tensor, + summary_length: int, + lc_length: int, + ): + # T: right_context length + utterance_length + summary_length + T, B, D = input.shape + mem_length = mems.size(0) + utterance_length = torch.max(lengths) + + right_context_blocks_length = T - utterance_length - summary_length + rc_block = input[:right_context_blocks_length, :, :] + utterance_block = input[right_context_blocks_length : T - summary_length, :, :] + + if B == 1: + padding_mask = None + else: + klengths = lengths + mem_length + right_context_blocks_length + lc_length + padding_mask = lengths_to_padding_mask(lengths=klengths) + + mem_rc_input = torch.cat([mems, rc_block, utterance_block], dim=0) + + # In training lc_length = 0 + key_length = mem_rc_input.size(0) + lc_length + rc_input_sum = input + q = self.e2h_q(rc_input_sum) + kv = self.e2h_kv(mem_rc_input) + k, v = kv.chunk(chunks=2, dim=2) + result_qkv = (q, k, v) + input_shape = (T, B, D) + result_lengths_info = ( + mem_length, + utterance_length, + right_context_blocks_length, + key_length, + ) + if padding_mask is not None: + assert padding_mask.size(0) == B + assert padding_mask.size(1) == key_length + + return result_qkv, input_shape, result_lengths_info, padding_mask + + def prepare_attention_weights( + self, + q: Tensor, + new_k: Tensor, + new_v: Tensor, + input_shape: Tuple[int, int, int], + rpe: Optional[Tensor], + ) -> Tuple[Tensor, Tensor, Tensor]: + T, B, D = input_shape + q = ( + q.contiguous().view(-1, B * self.num_heads, self.head_dim).transpose(0, 1) + * self.scaling + ) + + k = ( + new_k.contiguous() + .view(-1, B * self.num_heads, self.head_dim) + .transpose(0, 1) + ) + + v = ( + new_v.contiguous() + .view(-1, B * self.num_heads, self.head_dim) + .transpose(0, 1) + ) + + attention_weights = torch.bmm(q, k.transpose(1, 2)) + if self.use_rpe and rpe is not None and self.rpe_v is not None: + r_k = self.rpe_k(rpe) + # [q, B*h, d] * [q, k, d] -> [B*h, q, k] + attention_weights_rpe = torch.matmul( + q.transpose(0, 1), r_k.transpose(1, 2) + ).transpose(0, 1) + attention_weights = attention_weights + attention_weights_rpe + attention_weights_float = attention_weights.float() + + return attention_weights, attention_weights_float, v + + def prepare_attention_output( + self, + attention_weights: Tensor, + attention_weights_float: Tensor, + v: Tensor, + input_shape: Tuple[int, int, int], + key_length: int, + padding_mask: Optional[Tensor], + rpe: Optional[Tensor], + ) -> Tensor: + T, B, D = input_shape + if padding_mask is not None: + attention_weights_float = attention_weights_float.view( + B, self.num_heads, T, key_length + ) + attention_weights_float = attention_weights_float.masked_fill( + padding_mask.unsqueeze(1).unsqueeze(2).to(torch.bool), float("-inf") + ) + attention_weights_float = attention_weights_float.view( + B * self.num_heads, T, key_length + ) + + if self.std_scale is not None: + attention_weights_float = attention_suppression( + attention_weights_float, self.std_scale + ) + + attention_weights_float = torch.nn.functional.softmax( + attention_weights_float, dim=-1 + ) + attention_weights = attention_weights_float.type_as(attention_weights) + + attention_probs = torch.nn.functional.dropout( + attention_weights, p=self.dropout, training=self.training + ) + + # [T, key_length, B, n_head]+ [key_length, B, n_head, d_head] + # -> [T, B, n_head, d_head] + attention = torch.bmm(attention_probs, v) + if self.use_rpe and rpe is not None and self.rpe_v is not None: + r_v = self.rpe_v(rpe) + attention_rpe = torch.matmul( + attention_probs.transpose(0, 1), r_v + ).transpose(0, 1) + + if self.rpe_old_option: + attention += attention + attention_rpe + else: + attention = attention + attention_rpe + + assert list(attention.shape) == [B * self.num_heads, T, self.head_dim] + + attention = attention.transpose(0, 1).contiguous().view(T, B, self.embed_dim) + + rc_output_memory = self.out_proj(attention) + return rc_output_memory + + @torch.jit.unused + def forward( + self, + input: Tensor, + lengths: Tensor, + mems: Tensor, + attention_mask: Tensor, + pre_mems: Optional[Tensor] = None, + left_context_key: Optional[Tensor] = None, + left_context_val: Optional[Tensor] = None, + rpe: Optional[Tensor] = None, + ) -> Tuple[Tensor, Tensor, Tensor, Tensor]: + """ + forward function for NoSegAugmentedMemoryMultiheadAttentionBmm in training. + + args: + input: formed in the following way + [right_context_0, right_contex_1, ..., seg_0, seg_1, + ..., summary_0, summary_1,..] + lengths: the length of query which is [seg_0, seg_1, ....] + mems: [mem_0, mem_1, ...]. + attention_mask: attention mask for query = [right_context, query, summary] + key = [mem, right_context, query]. This is only used for traing. + + """ + if self.use_mem: + mem_length = mems.size(0) + summary_length = mem_length + 1 + if pre_mems is not None: + mems = torch.cat([pre_mems, mems], dim=0) + else: + mem_length = 0 + summary_length = 0 + + # In training, lc_length = 0 + if left_context_key is not None: + lc_length = left_context_key.size(0) + else: + lc_length = 0 + results = self.prepare_qkv( + input=input, + mems=mems, + lengths=lengths, + summary_length=summary_length, + lc_length=lc_length, + ) + result_qkv, input_shape, result_lengths_info, padding_mask = results + q, k, v = result_qkv + ( + mem_length, + utterance_length, + right_context_blocks_length, + key_length, + ) = result_lengths_info + + if left_context_key is not None: + # add the cache key and value + new_k = torch.cat( + [ + k[: mem_length + right_context_blocks_length, :, :], + left_context_key, + k[-utterance_length:, :, :], + ], + dim=0, + ) + new_v = torch.cat( + [ + v[: mem_length + right_context_blocks_length, :, :], + left_context_val, + v[-utterance_length:, :, :], + ], + dim=0, + ) + next_k = new_k[mem_length + right_context_blocks_length :, :, :] + next_v = new_v[mem_length + right_context_blocks_length :, :, :] + else: + new_k = k + new_v = v + next_k = None + next_v = None + + attention_weights, attention_weights_float, v = self.prepare_attention_weights( + q=q, + new_k=new_k, + new_v=new_v, + input_shape=input_shape, + rpe=rpe, + ) + + # mask attention + attention_mask = attention_mask.unsqueeze(0) + attention_weights_float = attention_weights_float.masked_fill( + attention_mask, float(self.negative_inf) + ) + + rc_output_memory = self.prepare_attention_output( + attention_weights=attention_weights, + attention_weights_float=attention_weights_float, + v=v, + input_shape=input_shape, + key_length=key_length, + padding_mask=padding_mask, + rpe=rpe, + ) + + if self.use_mem: + # next_m length equals to summary length - 1 + # last memory is ignored + if self.mini_batches: + next_m = rc_output_memory[-summary_length:] + else: + next_m = rc_output_memory[-summary_length:-1] + + next_m = self.squash_mem(next_m) + # rc and output + rc_output = rc_output_memory[:-summary_length] + if not self.nonlinear_squash_mem: + next_m = torch.clamp(next_m, min=-10, max=10) + else: + next_m = mems + rc_output = rc_output_memory + + return rc_output, next_m, next_k, next_v + + @torch.jit.export + def forward_jit( + self, + input: Tensor, + lengths: Tensor, + mems: Tensor, + left_context_key: Tensor, + left_context_val: Tensor, + rpe: Optional[Tensor], + ) -> Tuple[Tensor, Tensor, Tensor, Tensor]: + """ + forward function for NoSegAugmentedMemoryMultiheadAttentionBmm in decoding. + + args: + input: formed in the following way + [right_context_0, right_contex_1, ..., seg_0, seg_1, + ..., summary_0, summary_1,..] + lengths: the length of query which is [seg_0, seg_1, ....] + mems: [mem_0, mem_1, ...]. + left_context_key: left_context for key part. This is only used for online + decoding. In training, this is empty tensor + left_context_val: left_context for value part. This is only used for online + decoding. In training, this is empty tensor + + """ + lc_length = left_context_key.size(0) + + # In decoding, summary_length = 1 or 0 + if self.use_mem: + summary_length = 1 + else: + summary_length = 0 + + results = self.prepare_qkv( + input=input, + mems=mems, + lengths=lengths, + summary_length=summary_length, + lc_length=lc_length, + ) + result_qkv, input_shape, result_lengths_info, padding_mask = results + q, k, v = result_qkv + ( + mem_length, + utterance_length, + right_context_blocks_length, + key_length, + ) = result_lengths_info + + # add the cache key and value + new_k = torch.cat( + [ + k[: mem_length + right_context_blocks_length, :, :], + left_context_key, + k[-utterance_length:, :, :], + ], + dim=0, + ) + new_v = torch.cat( + [ + v[: mem_length + right_context_blocks_length, :, :], + left_context_val, + v[-utterance_length:, :, :], + ], + dim=0, + ) + next_k = new_k[mem_length + right_context_blocks_length :, :, :] + next_v = new_v[mem_length + right_context_blocks_length :, :, :] + + attention_weights, attention_weights_float, v = self.prepare_attention_weights( + q=q, + new_k=new_k, + new_v=new_v, + input_shape=input_shape, + rpe=rpe, + ) + # In online decoding, we don't have attention mask. But we still need + # to disable the attention from summary query to memory + attention_weights_float[:, -1, :mem_length] = float(self.negative_inf) + rc_output_memory = self.prepare_attention_output( + attention_weights=attention_weights, + attention_weights_float=attention_weights_float, + v=v, + input_shape=input_shape, + key_length=key_length, + padding_mask=padding_mask, + rpe=rpe, + ) + + # In decoding, summary length is 1 + if self.use_mem: + next_m = rc_output_memory[-1:] + next_m = self.squash_mem(next_m) + # rc and output + rc_output = rc_output_memory[:-1] + if not self.nonlinear_squash_mem: + next_m = torch.clamp(next_m, min=-10, max=10) + else: + rc_output = rc_output_memory + # empty tensor as input mems + next_m = mems + + return rc_output, next_m, next_k, next_v + + def quantize_(self, params=None): + if params and "per_channel" in params and params["per_channel"]: + qconfig = per_channel_dynamic_qconfig + else: + qconfig = default_dynamic_qconfig + quantization.quantize_dynamic( + self, {torch.nn.Linear: qconfig}, dtype=torch.qint8, inplace=True + ) + return self + + +class NoSegAugmentedMemoryTransformer(nn.Module): + """ + Whole utterance augmented memory transformer. + + This is not pyspeech nn layer. It is used as a module in a master layer where + multiple transformers is used. + """ + + def __init__( + self, + input_dim, + num_heads, + ffn_dim, + dropout_in_attn=0.0, + dropout_on_attn=None, + dropout_on_fc1=None, + dropout_on_fc2=None, + activation_fn="relu", + tanh_on_mem=False, + std_scale=None, + scaled_init=False, + segment_size=128, + use_mem=True, + mini_batches=False, + negative_inf="-inf", + layer_index=-1, + summarization_method="mean", + max_relative_position=0, + rpe_old_option=True, + ): + super(NoSegAugmentedMemoryTransformer, self).__init__() + + self.attention = NoSegAugmentedMemoryMultiheadAttentionBmm( + input_dim=input_dim, + num_heads=num_heads, + dropout=dropout_in_attn, + scaled_init=scaled_init, + tanh_on_mem=tanh_on_mem, + std_scale=std_scale, + use_mem=use_mem, + mini_batches=mini_batches, + negative_inf=negative_inf, + layer_index=layer_index, + max_relative_position=max_relative_position, + ) + self.dropout = nn.Dropout(dropout_on_attn) + self.pos_ff = PositionwiseFF( + input_dim=input_dim, + ffn_dim=ffn_dim, + dropout_on_fc1=dropout_on_fc1, + dropout_on_fc2=dropout_on_fc2, + activation_fn=activation_fn, + ) + self.layer_norm_pre = Fp32LayerNorm(input_dim) + self.layer_norm = Fp32LayerNorm(input_dim) + self.segment_size = segment_size + self.use_mem = use_mem + + self.memory_op = SummarizationLayer( + summarization_method, segment_size, input_dim + ) + + def set_mini_batches(self, mini_batches): + self.attention.mini_batches = mini_batches + + def gen_summary_queries(self, input): + sum_input = self.memory_op(input) + return sum_input + + def pre_attention_ops(self, input, right_context_blocks): + rc_length = right_context_blocks.size(0) + input_length = input.size(0) + + rc_and_input = torch.cat([right_context_blocks, input], dim=0) + residual_input = rc_and_input + rc_and_input = self.layer_norm_pre(rc_and_input) + + query_input = rc_and_input[-input_length:, :, :] + return rc_length, input_length, residual_input, query_input, rc_and_input + + def after_attention_ops(self, attention_output, residual_input): + output = self.dropout(attention_output) + output = output + residual_input + output = self.pos_ff(output) + output = self.layer_norm(output) + return output + + @torch.jit.export + def forward_jit( + self, + input: Tensor, + lengths: Tensor, + mems: Tensor, + left_context_key: Tensor, + left_context_val: Tensor, + right_context_blocks: Tensor, + rpe: Optional[Tensor], + ) -> Tuple[Tensor, Tensor, Tensor, Tensor, Tensor]: + + results = self.pre_attention_ops(input, right_context_blocks) + rc_length, input_length, residual_input, query_input, rc_and_input = results + + # In online decoding, the summary query size is always 1 or 0 + if self.use_mem: + summary_query = self.gen_summary_queries(query_input) + summary_query = summary_query[0:1, :, :] + rc_qu_su = torch.cat([rc_and_input, summary_query], dim=0) + else: + rc_qu_su = rc_and_input + + rc_output, next_m, next_k, next_v = self.attention.forward_jit( + input=rc_qu_su, + lengths=lengths, + mems=mems, + left_context_key=left_context_key, + left_context_val=left_context_val, + rpe=rpe, + ) + rc_output = self.after_attention_ops(rc_output, residual_input) + results = ( + rc_output[-input_length:, :, :], + next_m, + rc_output[0:rc_length, :, :], + next_k, + next_v, + ) + return results + + @torch.jit.unused + def forward( + self, + input, + lengths, + mems, + right_context_blocks, + attention_mask, + pre_mems, + left_context_key, + left_context_val, + rpe, + ): + + results = self.pre_attention_ops(input, right_context_blocks) + rc_length, input_length, residual_input, query_input, rc_and_input = results + if self.use_mem: + summary_query = self.gen_summary_queries(query_input) + rc_qu_su = torch.cat([rc_and_input, summary_query], dim=0) + else: + rc_qu_su = rc_and_input + + rc_output, next_m, next_k, next_v = self.attention( + input=rc_qu_su, + lengths=lengths, + mems=mems, + attention_mask=attention_mask, + pre_mems=pre_mems, + left_context_key=left_context_key, + left_context_val=left_context_val, + rpe=rpe, + ) + + # [TODO] Note memory did not go through pos_ff. What happen if we pass + # memory through the pos_ff as well? + rc_output = self.after_attention_ops(rc_output, residual_input) + results = ( + rc_output[-input_length:, :, :], + next_m, + rc_output[0:rc_length, :, :], + next_k, + next_v, + ) + + return results + + +class NoSegAugmentedMemoryTransformerEncoderLayer(FairseqEncoder): + """ + Whole utterance augmented memory transformer encoder layer. This is a master layer + where we can define multiple augmented memory transformers. There are two reasons + to setup the master layer. + 1. We only need to define once about the attention mask. All the layers in the master + layer share the same mask. + 2. pyspeech nn layer has special input and output format. Defining one master layer is + easier to passing memory between different layes inside the master layer + + args: + input_dim: input embedding dimension + num_heads: number of heads in multihead self-attention + ffn_dim: ffn dimension in FFN layer + num_layers: number of augmented memory transformer layers + dropout_in_attn: dropout used in multi-head self-attention + dropout_on_attn: dropout used for output from te multihead self-attention + dropout_on_fc1: dropout used in FFN layer for the first linear layer + dropout_on_fc2: dropout used in FFN layer for the second linear layer + segment_size: segment size for each segment + context_config: (left_context_size, right_context_size) defines the surround context size + for each segment + max_memory_size: maximum memory size used for each segment + scaled_init: whether use scaled init for weight initialization in attention layer + std_scale: if std_scale is not None. The weak attention suppression is + turned on. For std_scale = 0.5, all the attention smaller than + mean + 0.5 * std will be suppressed. + activation_fn: activation function used in FFN layer. [ReLU, GELU] supported + tanh_on_mem: whether use tanh on memory + mini_batches: use mini-btach training + negative_inf: the negative infinity value used in attention masking. default is "-inf". + For some situation, e.g. LM. it is better to use "-1e8" to avoid nan issue. + summarization_method: method to generate segment summrization embedding + max_relative_position: max relatie position for relative position embedding + rpe_old_option: To be compatible with previous model. The previous model + was trained with attention += attention + rpe. The correct equation + should be attention = attention + rpe + [TODO]: remove the rpe_old_option by the end of 2021 Q1. + + """ + + def __init__( + self, + input_dim, + num_heads, + ffn_dim, + num_layers=1, + dropout_in_attn=0.0, + dropout_on_attn=0.0, + dropout_on_fc1=0.0, + dropout_on_fc2=0.0, + segment_size=128, + context_config=(0, 0), + max_memory_size=0, + scaled_init=True, + std_scale=None, + activation_fn="relu", + tanh_on_mem=False, + mini_batches=False, + negative_inf="-inf", + deep_init=True, + summarization_method="mean", + max_relative_position=0, + rpe_old_option=True, + ): + super().__init__(None) + if input_dim % num_heads: + raise ValueError( + "input_dim ({}) must be divisible by num_heads ({})".format( + input_dim, num_heads + ) + ) + + # we used to support growing memory size. However, it will cause + # cross stream batching failure. Now we need to have exact max memory size + if max_memory_size < 0: + raise ValueError("max_memory_size must be >= 0") + + # Only assign right_context. In decoding, left context will be cached. + # No need to let the online decoder to re-assign the left context + self.left_context, self.right_context = context_config + self.segment_size = segment_size + self.memory_dim = input_dim + self.max_memory_size = max_memory_size + self.mini_batches = mini_batches + if self.max_memory_size != 0: + self.use_mem = True + else: + self.use_mem = False + + self.memory_op = SummarizationLayer( + summarization_method, segment_size, input_dim + ) + + self.layers = torch.nn.ModuleList() + self.num_layers = num_layers + self.max_relative_position = max_relative_position + if self.max_relative_position > 0: + self.use_rpe = True + else: + self.use_rpe = False + for i in range(self.num_layers): + if deep_init: + layer_index = i + else: + layer_index = -1 + + self.layers.append( + NoSegAugmentedMemoryTransformer( + num_heads=num_heads, + input_dim=input_dim, + ffn_dim=ffn_dim, + dropout_in_attn=dropout_in_attn, + dropout_on_attn=dropout_on_attn, + dropout_on_fc1=dropout_on_fc1, + dropout_on_fc2=dropout_on_fc2, + segment_size=segment_size, + std_scale=std_scale, + activation_fn=activation_fn, + tanh_on_mem=tanh_on_mem, + scaled_init=scaled_init, + use_mem=self.use_mem, + mini_batches=mini_batches, + negative_inf=negative_inf, + layer_index=layer_index, + summarization_method=summarization_method, + max_relative_position=max_relative_position, + rpe_old_option=rpe_old_option, + ) + ) + + def set_mini_batches(self, mini_batches): + # handy function only used for unit test + self.mini_batches = mini_batches + for layer in self.layers: + layer.set_mini_batches(mini_batches) + + def _get_relative_position( + self, + input: Tensor, + max_relative_position: int, + left_context_length: int, + past_length: int, + is_decoding: bool, + ): + # For training, we copy the right context to the start of the utterance + # First dimension in distance is corresponding to query. + # [right context, utterance, summary vector] + # Second dimension in distance is corresponding to key. + # [Memory bank, right context, utterance] + # For summary vector in query part, the distance with + # all other position is 2*max_position. For memory bank in key, + # the distance with all other positions is 0. + + T, B, D = input.shape + num_segs = math.ceil((T - self.right_context) / self.segment_size) + + # utterance + u_st = past_length * self.segment_size + u_ed = u_st + T + utterance_ranges = torch.arange(u_st, u_ed - self.right_context) + + # left context. Only in minibatch or decoding + left_context_ranges = torch.arange(u_st - left_context_length, u_st) + + # Right context block + # right context + utterance + right_context_blocks = [] + for i in range(0, num_segs - 1): + st = (i + 1) * self.segment_size + u_st + ed = st + self.right_context + assert ed < u_ed + temp = torch.arange(st, ed) + right_context_blocks.append(temp) + right_context_blocks.append(torch.arange(u_ed - self.right_context, u_ed)) + right_context_ranges = torch.cat(right_context_blocks) + + if self.use_mem: + # Memory bank + # The position for memory -n, .., -1 + if is_decoding: + memory_size = min(past_length, self.max_memory_size) + else: + memory_size = num_segs + past_length - 1 + memory_bank_ranges = torch.arange( + -max_relative_position - 1, -max_relative_position - 1 - memory_size, -1 + ) + + # summary vector + # The position for summary vector as the T+max_relative_position+1. + # After the clamping, the relative position is max_relative_position + summary_pos_st = u_ed + max_relative_position + 1 + summary_vector_ranges = torch.arange( + summary_pos_st, summary_pos_st + num_segs + ) + + key_ranges = torch.cat( + [ + memory_bank_ranges, + right_context_ranges, + left_context_ranges, + utterance_ranges, + ] + ) + + query_ranges = torch.cat( + [right_context_ranges, utterance_ranges, summary_vector_ranges] + ) + else: + key_ranges = torch.cat( + [right_context_ranges, left_context_ranges, utterance_ranges] + ) + + query_ranges = torch.cat([right_context_ranges, utterance_ranges]) + + distance = key_ranges[None, :] - query_ranges[:, None] + distance_clamp = ( + torch.clamp(distance, -max_relative_position, max_relative_position) + + max_relative_position + ) + distance_clamp = distance_clamp.to(input.device).long().detach() + return distance_clamp + + def _get_attention_mask(self, input, past_length=0, left_context_cache=0): + # attention mask for each query contains three parts: + # 1. memory part + # 2. left_context + segment + # 3. right_context_block + # so for each segment and its correspoinding right context block, + # the attention matrix is formed by 9 parts: + # [0, m, 0, 0, right_context, 0, 0, seg, 0] + # [before memory, memory, after memory, before right context, right_context, + # after right context, before seg, seg, after seg] + # + # Query is formed in the way as [right_context_blocks, utterance, summary] + # + # Note: put m and right_context before segment is convenient + # for padding_mask operation. + # Key lengths = m_length + right_context_block_length + lengths + utterance_length, batch_size, _ = input.shape + summary_length = math.ceil(utterance_length / self.segment_size) + num_segs = summary_length + rc_length = self.right_context * num_segs + rc = self.right_context + lc = self.left_context + + # using mini-batches, there is left context cache available for current + # sequence. + lcc = left_context_cache + + # max_memory_size is 0 then we don't have memory and summary + # past_length is the memory carry from previous sequence + if self.use_mem: + mem_length = num_segs - 1 + past_length + else: + mem_length = 0 + rc_mask = [] + query_mask = [] + summary_mask = [] + for j in range(0, num_segs): + ssize = min(self.segment_size, utterance_length - j * self.segment_size) + + rc_size = rc + rc_mat = [] + q_mat = [] + s_mat = [] + m_start = max(j + past_length - self.max_memory_size, 0) + + # max_memory_size is 0, then we don't use memory + if self.use_mem: + # part 0: before memory + rc_mat.append(input.new_zeros(rc_size, m_start)) + q_mat.append(input.new_zeros(ssize, m_start)) + s_mat.append(input.new_zeros(1, m_start)) + + # part 1: memory + col_1 = j + past_length - m_start + rc_mat.append(torch.ones(rc_size, col_1, device=input.device)) + q_mat.append(torch.ones(ssize, col_1, device=input.device)) + # based on D22875746, disable summary query attention + # on memeory is better for long form utterance + s_mat.append(input.new_zeros(1, col_1)) + + # part 2: after memory + col_2 = mem_length - (j + past_length) + rc_mat.append(input.new_zeros(rc_size, col_2)) + q_mat.append(input.new_zeros(ssize, col_2)) + s_mat.append(input.new_zeros(1, col_2)) + + # part 3: before right context + rc_start = j * rc + rc_mat.append(input.new_zeros(rc_size, rc_start)) + q_mat.append(input.new_zeros(ssize, rc_start)) + s_mat.append(input.new_zeros(1, rc_start)) + + # part 4: right context + rc_end = rc_start + rc + col_4 = rc + rc_mat.append(torch.ones(rc_size, col_4, device=input.device)) + q_mat.append(torch.ones(ssize, col_4, device=input.device)) + s_mat.append(torch.ones(1, col_4, device=input.device)) + + # part 5: after right context + col_5 = rc_length - rc_end + rc_mat.append(input.new_zeros(rc_size, col_5)) + q_mat.append(input.new_zeros(ssize, col_5)) + s_mat.append(input.new_zeros(1, col_5)) + + # part 6: before query segment + seg_start = max(j * self.segment_size + lcc - lc, 0) + rc_mat.append(input.new_zeros(rc_size, seg_start)) + q_mat.append(input.new_zeros(ssize, seg_start)) + s_mat.append(input.new_zeros(1, seg_start)) + + # part 7: query segment + # note: right context is put in right context block + # here we only need to consider about left context + seg_end = min((j + 1) * self.segment_size + lcc, utterance_length + lcc) + col_7 = seg_end - seg_start + rc_mat.append(torch.ones(rc_size, col_7, device=input.device)) + q_mat.append(torch.ones(ssize, col_7, device=input.device)) + s_mat.append(torch.ones(1, col_7, device=input.device)) + + # part 8: after query segment + col_8 = utterance_length + lcc - seg_end + rc_mat.append(input.new_zeros(rc_size, col_8)) + q_mat.append(input.new_zeros(ssize, col_8)) + s_mat.append(input.new_zeros(1, col_8)) + + rc_mask.append(torch.cat(rc_mat, dim=1)) + query_mask.append(torch.cat(q_mat, dim=1)) + summary_mask.append(torch.cat(s_mat, dim=1)) + + # no memory, then we don't need summary either + if self.use_mem: + attention_mask = ( + 1 + - torch.cat( + [ + torch.cat(rc_mask, dim=0), + torch.cat(query_mask, dim=0), + torch.cat(summary_mask, dim=0), + ], + dim=0, + ) + ).to(torch.bool) + else: + attention_mask = ( + 1 + - torch.cat( + [torch.cat(rc_mask, dim=0), torch.cat(query_mask, dim=0)], dim=0 + ) + ).to(torch.bool) + + return attention_mask + + @torch.jit.export + def init_state( + self, batch_size: int, device: Optional[Device] = None + ) -> List[Tensor]: + empty_memory = torch.zeros( + self.num_layers, + self.max_memory_size, + batch_size, + self.memory_dim, + device=device, + ) + left_context_key = torch.zeros( + self.num_layers, + self.left_context, + batch_size, + self.memory_dim, + device=device, + ) + left_context_val = torch.zeros( + self.num_layers, + self.left_context, + batch_size, + self.memory_dim, + device=device, + ) + past_length = torch.zeros(1, batch_size, dtype=torch.int32, device=device) + + return [empty_memory, left_context_key, left_context_val, past_length] + + @torch.jit.export + def batch_state(self, states: List[List[Tensor]]) -> List[Tensor]: + if len(states) == 0: + return [] + batched_m = [] + batched_lc_key = [] + batched_lc_val = [] + batched_past_length = [] + for state in states: + if len(state) == 0: + continue + m, lc_key, lc_val, past_length = state + batched_m.append(m) + batched_lc_key.append(lc_key) + batched_lc_val.append(lc_val) + batched_past_length.append(past_length) + + if ( + (len(batched_m) == 0) + or (len(batched_lc_key) == 0) + or (len(batched_lc_val) == 0) + or (len(batched_past_length) == 0) + ): + return [ + torch.tensor([]), + torch.tensor([]), + torch.tensor([]), + torch.tensor([]), + ] + + batched_m = torch.cat(batched_m, dim=2) + batched_lc_key = torch.cat(batched_lc_key, dim=2) + batched_lc_val = torch.cat(batched_lc_val, dim=2) + batched_past_length = torch.cat(batched_past_length, dim=1) + return [batched_m, batched_lc_key, batched_lc_val, batched_past_length] + + @torch.jit.export + def reorder_state(self, state: List[Tensor], indices: Tensor) -> List[Tensor]: + if len(state) == 0: + return [] + m, lc_key, lc_val, past_length = state + indices = indices.to(device=m.device) + reord_m = torch.index_select(m, 2, indices) + reord_lc_key = torch.index_select(lc_key, 2, indices) + reord_lc_val = torch.index_select(lc_val, 2, indices) + reord_past_length = torch.index_select(past_length, 1, indices) + return [reord_m, reord_lc_key, reord_lc_val, reord_past_length] + + @torch.jit.export + def reset_state(self, state: List[Tensor], indices: Tensor) -> List[Tensor]: + m, lc_key, lc_val, past_length = state + m = m.index_fill(dim=2, index=indices, value=0.0) + lc_key = lc_key.index_fill(dim=2, index=indices, value=0.0) + lc_val = lc_val.index_fill(dim=2, index=indices, value=0.0) + past_length = past_length.index_fill(dim=1, index=indices, value=0) + + return [m, lc_key, lc_val, past_length] + + @torch.jit.export + def state_size(self) -> int: + return 4 + + @torch.jit.export + def batch_size_in_state( + self, state: Optional[List[Tensor]], sloppy: bool = True + ) -> Optional[int]: + if state is None: + return None + return state[0].size(2) + + def gen_summary_queries(self, input): + sum_input = self.memory_op(input) + return sum_input + + def _gen_right_context_padded_input(self, input): + # This function deals with input that is already + # padded with right context (e.g. minibatch training) + right_context_blocks = [] + T, B, D = input.shape + num_segs = math.ceil((T - self.right_context) / self.segment_size) + for i in range(0, num_segs - 1): + st = (i + 1) * self.segment_size + ed = st + self.right_context + assert ed < T + temp = input[st:ed, :, :] + right_context_blocks.append(temp) + + # last segment right context is already available + right_context_blocks.append(input[T - self.right_context :, :, :]) + return torch.cat(right_context_blocks, dim=0) + + def _gen_segs_right_context(self, input, lengths): + segments = [] + T, B, D = input.size() + nT = T - self.right_context + + # assume input is right context padded + num_segs = math.ceil(nT / self.segment_size) + # pad zeros to the utterance to make sure each + # segment has the same right context. For the + for i in range(0, num_segs - 1): + st = i * self.segment_size + ed = min(T, st + self.segment_size + self.right_context) + temp = input[st:ed, :, :] + rest_lengths = torch.clamp( + lengths - self.segment_size, min=0, max=nT - (i + 1) * self.segment_size + ) + segments.append((temp, lengths - rest_lengths + self.right_context)) + lengths = rest_lengths + + last_seg = input[st + self.segment_size :, :, :] + segments.append((last_seg, rest_lengths + self.right_context)) + + return segments + + @torch.jit.unused + def forward( + self, input: Tensor, padding_masks: Tensor, state: Optional[List[Tensor]] = None + ) -> Tuple[Tensor, Tensor, List[Tensor], List[Tensor]]: + # Xutai: originally the second argument is lengths. + lengths = (~padding_masks).sum(dim=1).long() + # mini batch training. + if self.mini_batches: + return self.forward_mini_batches(input, lengths, state) + + # regular full sequence training. Note, assume the right context in provided + # in the input. + T, B, D = input.size() + right_context_blocks = self._gen_right_context_padded_input(input) + + # generate the relative positional embedding + if self.use_rpe: + rpe = self._get_relative_position( + input=input, + max_relative_position=self.max_relative_position, + left_context_length=0, + past_length=0, + is_decoding=False, + ) + else: + rpe = None + input = input[: T - self.right_context, :, :] + + attention_mask = self._get_attention_mask(input) + + # firt layer use each segment mean as memory + # ignore the last one seg average + if self.use_mem: + mems = self.gen_summary_queries(input)[:-1, :, :] + else: + mems = torch.zeros(0, input.size(1), input.size(2), device=input.device) + mems = mems.type_as(input) + + output = input + all_outputs = [] + + for layer in self.layers: + output, mems, right_context_blocks, _, _ = layer( + input=output, + lengths=lengths, + attention_mask=attention_mask, + mems=mems, + right_context_blocks=right_context_blocks, + pre_mems=None, + left_context_key=None, + left_context_val=None, + rpe=rpe, + ) + all_outputs.append(output) + return output, padding_masks, [], all_outputs + + def forward_jit_mini_batch_init( + self, + seg: Tensor, + state: Optional[List[Tensor]] = None, + is_decoding: bool = False, + ): + # Prepare state. In whole sequence training, state is ignored. + # For minibatch training, we need to prepare state + if state is None: + state = self.init_state(batch_size=seg.size(1), device=seg.device) + if seg.dtype == torch.half: + state = [state[0].half(), state[1].half(), state[2].half(), state[3]] + + if self.use_mem: + # note input average only on seg, not on right context + # first layer use each segmetn mean as memory. the last + # one segment average is used in state + full_mems = self.gen_summary_queries(seg) + if is_decoding: + mems = full_mems[0:1, :, :] + state_mems = torch.cat([state[0][0], mems], dim=0) + else: + mems = full_mems[:-1, :, :] + state_mems = torch.cat([state[0][0], full_mems], dim=0) + else: + mems = state[0][0] + state_mems = mems + + # track processed segment number or memory number + # the same batch as the same bumber of past length + past_length = state[3][0][0].item() + past_left_context = min(past_length * self.segment_size, self.left_context) + past_length = min(self.max_memory_size, past_length) + + return state, mems, state_mems, past_length, past_left_context + + def state_update_before( + self, layer: int, state: List[Tensor], past_length: int, past_left_context: int + ): + pre_mems = state[0][layer][self.max_memory_size - past_length :, :, :] + lc_key = state[1][layer][self.left_context - past_left_context :, :, :] + lc_val = state[2][layer][self.left_context - past_left_context :, :, :] + return pre_mems, lc_key, lc_val + + def state_update_after( + self, + layer: int, + state: List[Tensor], + mems: Tensor, + next_key: Tensor, + next_val: Tensor, + mems_list: List[Tensor], + lc_key_list: List[Tensor], + lc_val_list: List[Tensor], + ): + # mems is used for next layer + if layer < self.num_layers - 1: + state_mems = torch.cat([state[0][layer + 1], mems], dim=0) + mems_list.append(state_mems[-self.max_memory_size :, :, :]) + + # when mems pass to next sequence, we need the last memory. when mems + # use for the next layer, we can ignore the last memory + mems = mems[:-1, :, :] + + # note state[1][i] and state[2][i] original length equals to self.left_context + new_k = torch.cat([state[1][layer], next_key], dim=0) + new_v = torch.cat([state[2][layer], next_val], dim=0) + lc_key_list.append(new_k[-self.left_context :, :, :]) + lc_val_list.append(new_v[-self.left_context :, :, :]) + return mems_list, lc_key_list, lc_val_list, mems + + def state_update_after_loop( + self, + state: List[Tensor], + mems_list: List[Tensor], + lc_key_list: List[Tensor], + lc_val_list: List[Tensor], + update_length: int, + ): + state[0] = torch.stack(mems_list, dim=0) + state[1] = torch.stack(lc_key_list, dim=0) + state[2] = torch.stack(lc_val_list, dim=0) + state[3] = state[3] + update_length + return state + + @torch.jit.unused + def forward_mini_batches( + self, input: Tensor, lengths: Tensor, state: Optional[List[Tensor]] = None + ) -> Tuple[Tensor, Tensor, List[Tensor], List[Tensor]]: + T, B, D = input.size() + + # input without right context + seg = input[: T - self.right_context, :, :] + + # get right context blocks + right_context_blocks = self._gen_right_context_padded_input(input) + + mems_list = [] + lc_key_list = [] + lc_val_list = [] + results = self.forward_jit_mini_batch_init(seg, state, False) + state, mems, state_mems, past_length, past_left_context = results + + # relative position embedding + if self.use_rpe: + rpe = self._get_relative_position( + input=input, + max_relative_position=self.max_relative_position, + left_context_length=past_left_context, + past_length=past_length, + is_decoding=False, + ) + else: + rpe = None + + # get attention mask based on seg (not include right context) and available + # left context + attention_mask = self._get_attention_mask(seg, past_length, past_left_context) + mems_list.append(state_mems[-self.max_memory_size :, :, :]) + output = seg + i = 0 + all_outputs = [] + for layer in self.layers: + # In order to make cross stream batching work, mem, left context key + # and left context value in the state should always be the same shape. + # We use the past length to track the processed segment number. In this + # way, we take out the essential memory, left context key and left + # context val from the state. After finish the forward for current segment + # we add the new memory, left context key and left context value into the + # staate and trim out the oldest part to keep the shape consistent. + pre_mems, lc_key, lc_val = self.state_update_before( + i, state, past_length, past_left_context + ) + + output, mems, right_context_blocks, next_key, next_val = layer.forward( + input=output, + lengths=lengths, + attention_mask=attention_mask, + mems=mems, + right_context_blocks=right_context_blocks, + pre_mems=pre_mems, + left_context_key=lc_key, + left_context_val=lc_val, + rpe=rpe, + ) + all_outputs.append(output) + mems_list, lc_key_list, lc_val_list, mems = self.state_update_after( + layer=i, + state=state, + mems=mems, + next_key=next_key, + next_val=next_val, + mems_list=mems_list, + lc_key_list=lc_key_list, + lc_val_list=lc_val_list, + ) + + i += 1 + + # update state + update_length = math.ceil((T - self.right_context) / self.segment_size) + state = self.state_update_after_loop( + state=state, + mems_list=mems_list, + lc_key_list=lc_key_list, + lc_val_list=lc_val_list, + update_length=update_length, + ) + + return output, lengths, state, all_outputs + + def forward_jit_test( + self, input: Tensor, lengths: Tensor, state: Optional[List[Tensor]] = None + ) -> Tuple[Tensor, Tensor, List[Tensor]]: + """ + This one simulate sequence encoder forward jit. This is for unit test purpose. + It is not used in training or decoding. Note, extra_right_context is set in + the model. In unit test, input = [utterance, right_context], lengths = + [utterance_length]. + args: + input: input utterance + lengths: utterance input length + state: None here. input is whole utterance + """ + # [TODO] sequence_to_segment has bug in lengths. + seg_src_tokens_lengths = self._gen_segs_right_context(input, lengths) + + seg_enc_tokens_lengths: List[Tuple[Tensor, Tensor]] = [] + state: Optional[List[Tensor]] = None + for seg_src_tokens, seg_src_lengths in seg_src_tokens_lengths: + seg_enc_tokens, seg_enc_lengths, state = self.forward_jit( + input=seg_src_tokens, lengths=seg_src_lengths, state=state + ) + seg_enc_tokens_lengths.append((seg_enc_tokens, seg_enc_lengths)) + + enc_tokens, enc_lengths = segments_to_sequence( + segments=seg_enc_tokens_lengths, time_axis=0 + ) + + state = [] # returns trivial state + + return enc_tokens, enc_lengths, state + + @torch.jit.export + def forward_jit( + self, input: Tensor, lengths: Tensor, state: Optional[List[Tensor]] = None + ) -> Tuple[Tensor, Tensor, List[Tensor]]: + """ + Forward helper for online decoding. + + args: + input: [seg, right_context]. We assume in online we + always padding the right context to the preset right context size. + For the last segment, we may have short segment size, but right + context size is the same as other segments + lengths: utterance input length is the utterance segment length and + right context size + state: [memory, left_context_key, left_context_val]. To improve throughput, + in addition to memory, we also cache key and value for left_context in + multihead self-attention + """ + # In online decoding, input = [segment, right_context] + # Lengths = [segment_length, right_context_length] + # so we need strip right context in output + T, B, D = input.size() + rc_str = T - self.right_context + rc_end = T + right_context_blocks = input[rc_str:rc_end, :, :] + seg = input[:rc_str, :, :] + lengths = torch.clamp(lengths - self.right_context, min=0) + mems_list = [] + lc_key_list = [] + lc_val_list = [] + + results = self.forward_jit_mini_batch_init(seg, state, True) + state, mems, state_mems, past_length, past_left_context = results + + # relative position embedding + if self.use_rpe: + rpe = self._get_relative_position( + input=input, + max_relative_position=self.max_relative_position, + left_context_length=past_left_context, + past_length=past_length, + is_decoding=True, + ) + else: + rpe = None + + # memory for first layer. + mems_list.append(state_mems[-self.max_memory_size :, :, :]) + output = seg + i = 0 + for layer in self.layers: + # In order to make cross stream batching work, mem, left context key + # and left context value in the state should always be the same shape. + # We use the past length to track the processed segment number. In this + # way, we take out the essential memory, left context key and left + # context val from the state. After finish the forward for current segment + # we add the new memory, left context key and left context value into the + # staate and trim out the oldest part to keep the shape consistent. + true_mems, lc_key, lc_val = self.state_update_before( + layer=i, + state=state, + past_length=past_length, + past_left_context=past_left_context, + ) + + output, mems, right_context_blocks, next_key, next_val = layer.forward_jit( + input=output, + lengths=lengths, + mems=true_mems, + right_context_blocks=right_context_blocks, + left_context_key=lc_key, + left_context_val=lc_val, + rpe=rpe, + ) + # mems is used for next layer + mems_list, lc_key_list, lc_val_list, _ = self.state_update_after( + layer=i, + state=state, + mems_list=mems_list, + mems=mems, + next_key=next_key, + next_val=next_val, + lc_key_list=lc_key_list, + lc_val_list=lc_val_list, + ) + i += 1 + + # update state + state = self.state_update_after_loop( + state=state, + mems_list=mems_list, + lc_key_list=lc_key_list, + lc_val_list=lc_val_list, + update_length=1, + ) + + return output, lengths, state + + def quantize_(self, params=None): + if params and "per_channel" in params and params["per_channel"]: + qconfig = per_channel_dynamic_qconfig + else: + qconfig = default_dynamic_qconfig + quantization.quantize_dynamic( + self, {torch.nn.Linear: qconfig}, dtype=torch.qint8, inplace=True + ) + return self + + +# ------------------------------------------------------------------------------ +# Emformer encoder for seq2seq model +# This is a wrapper over the original emformer +# ------------------------------------------------------------------------------ +def emformer_encoder(klass): + class SpeechEncoder(klass): + def __init__(self, args): + super().__init__(args) + stride = SpeechEncoder.conv_layer_stride(args) + trf_left_context = args.segment_left_context // stride + trf_right_context = args.segment_right_context // stride + context_config = [trf_left_context, trf_right_context] + self.transformer_layers = nn.ModuleList( + [ + NoSegAugmentedMemoryTransformerEncoderLayer( + input_dim=args.encoder_embed_dim, + num_heads=args.encoder_attention_heads, + ffn_dim=args.encoder_ffn_embed_dim, + num_layers=args.encoder_layers, + dropout_in_attn=args.dropout, + dropout_on_attn=args.dropout, + dropout_on_fc1=args.dropout, + dropout_on_fc2=args.dropout, + activation_fn=args.activation_fn, + context_config=context_config, + segment_size=args.segment_length, + max_memory_size=args.max_memory_size, + scaled_init=True, # TODO: use constant for now. + tanh_on_mem=args.amtrf_tanh_on_mem, + ) + ] + ) + + def forward(self, src_tokens, src_lengths): + encoder_out = super().forward(src_tokens, src_lengths) + output = encoder_out["encoder_out"][0] + encoder_padding_masks = encoder_out["encoder_padding_mask"][0] + + # This is because that in the original implementation + # the output didn't consider the last segment as right context. + encoder_padding_masks = encoder_padding_masks[:, : output.size(0)] + + return { + "encoder_out": [output], + "encoder_padding_mask": [encoder_padding_masks], + "encoder_embedding": [], + "encoder_states": [], + "src_tokens": [], + "src_lengths": [], + } + + @staticmethod + def conv_layer_stride(args): + # TODO: make it configurable from the args + return 4 + + SpeechEncoder.__name__ = klass.__name__ + return SpeechEncoder diff --git a/data/fairseq/fairseq/models/speech_to_text/multi_modality_model.py b/data/fairseq/fairseq/models/speech_to_text/multi_modality_model.py new file mode 100644 index 0000000000000000000000000000000000000000..046421620ae9e78a6138632c26e8a25ce3b64478 --- /dev/null +++ b/data/fairseq/fairseq/models/speech_to_text/multi_modality_model.py @@ -0,0 +1,49 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from fairseq.models import FairseqDecoder, FairseqEncoder + + +# a container for different encoders with training samples from different modality +# each time, only one encoder is selected +class MultiModalityEncoder(FairseqEncoder): + def __init__(self, dictionary): + super().__init__(dictionary) + + def select_encoder(self, mode, **kwargs): + raise NotImplementedError("Model must implement the select_encoder method") + return None, kwargs + + # def post_encoder(self, encoder_out, src_tokens, src_lengths, mode, **kwargs): + # # Default do nothing + # return encoder_out + + # get sample data from JointSpeechTextDataset + def forward(self, src_tokens, src_lengths=None, mode="", **kwargs): + encoder, kwargs = self.select_encoder(mode, **kwargs) + # return self.post_encoder(encoder(src_tokens, src_lengths, **kwargs), src_tokens, src_lengths, mode, **kwargs) + return encoder(src_tokens, src_lengths, **kwargs) + + +# a container for different decoders with training samples from different modality +# each time, only one decoder is selected +class MultiInputDecoder(FairseqDecoder): + def __init__(self, dictionary): + super().__init__(dictionary) + + def select_decoder(self, mode, **kwargs): + raise NotImplementedError("Model must implement the select_decoder method") + return None, kwargs + + def forward( + self, prev_output_tokens, encoder_out, incremental_state=None, mode="", **kwargs + ): + decoder, kwargs = self.select_decoder(mode, **kwargs) + return decoder( + prev_output_tokens, + encoder_out, + incremental_state=incremental_state, + **kwargs + ) diff --git a/data/fairseq/fairseq/models/speech_to_text/s2t_conformer.py b/data/fairseq/fairseq/models/speech_to_text/s2t_conformer.py new file mode 100644 index 0000000000000000000000000000000000000000..79dbbec1b90a05ab101d5b086cd18eec858cdc8c --- /dev/null +++ b/data/fairseq/fairseq/models/speech_to_text/s2t_conformer.py @@ -0,0 +1,234 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import logging +import math +from pathlib import Path + +import torch + +from fairseq import checkpoint_utils +from fairseq.data.data_utils import lengths_to_padding_mask +from fairseq.models import FairseqEncoder, register_model, register_model_architecture +from fairseq.models.speech_to_text.modules.convolution import ( + Conv1dSubsampler, + Conv2dSubsampler, +) +from fairseq.models.speech_to_text.s2t_transformer import ( + S2TTransformerEncoder, + S2TTransformerModel, +) +from fairseq.models.speech_to_text.s2t_transformer import ( + base_architecture as transformer_base_architecture, +) +from fairseq.modules import PositionalEmbedding, RelPositionalEncoding +from fairseq.modules.conformer_layer import ConformerEncoderLayer + +logger = logging.getLogger(__name__) + + +class S2TConformerEncoder(FairseqEncoder): + """Conformer Encoder for speech translation based on https://arxiv.org/abs/2005.08100""" + + def __init__(self, args): + super().__init__(None) + + self.encoder_freezing_updates = args.encoder_freezing_updates + self.num_updates = 0 + + self.embed_scale = math.sqrt(args.encoder_embed_dim) + if args.no_scale_embedding: + self.embed_scale = 1.0 + self.padding_idx = 1 + self.conv_version = args.conv_version + if self.conv_version == "s2t_transformer": + self.subsample = Conv1dSubsampler( + args.input_feat_per_channel * args.input_channels, + args.conv_channels, + args.encoder_embed_dim, + [int(k) for k in args.conv_kernel_sizes.split(",")], + ) + elif self.conv_version == "convtransformer": + self.subsample = Conv2dSubsampler( + args.input_channels, + args.input_feat_per_channel, + args.conv_out_channels, + args.encoder_embed_dim, + ) + self.pos_enc_type = args.pos_enc_type + if self.pos_enc_type == "rel_pos": + self.embed_positions = RelPositionalEncoding( + args.max_source_positions, args.encoder_embed_dim + ) + elif self.pos_enc_type == "rope": + self.embed_positions = None + else: # Use absolute positional embedding + self.pos_enc_type = "abs" + self.embed_positions = PositionalEmbedding( + args.max_source_positions, args.encoder_embed_dim, self.padding_idx + ) + + self.linear = torch.nn.Linear(args.encoder_embed_dim, args.encoder_embed_dim) + self.dropout = torch.nn.Dropout(args.dropout) + self.conformer_layers = torch.nn.ModuleList( + [ + ConformerEncoderLayer( + embed_dim=args.encoder_embed_dim, + ffn_embed_dim=args.encoder_ffn_embed_dim, + attention_heads=args.encoder_attention_heads, + dropout=args.dropout, + depthwise_conv_kernel_size=args.depthwise_conv_kernel_size, + attn_type=args.attn_type, + pos_enc_type=self.pos_enc_type, + use_fp16=args.fp16, + ) + for _ in range(args.encoder_layers) + ] + ) + + def _forward(self, src_tokens, src_lengths, return_all_hiddens=False): + """ + Args: + src_tokens: Input source tokens Tensor of shape B X T X C + src_lengths: Lengths Tensor corresponding to input source tokens + return_all_hiddens: If true will append the self attention states to the encoder states + Returns: + encoder_out: Tensor of shape B X T X C + encoder_padding_mask: Optional Tensor with mask + encoder_embedding: Optional Tensor. Always empty here + encoder_states: List of Optional Tensors wih self attention states + src_tokens: Optional Tensor. Always empty here + src_lengths: Optional Tensor. Always empty here + """ + x, input_lengths = self.subsample(src_tokens, src_lengths) # returns T X B X C + encoder_padding_mask = lengths_to_padding_mask(input_lengths) + x = self.embed_scale * x + if self.pos_enc_type == "rel_pos": + positions = self.embed_positions(x) + + elif self.pos_enc_type == "rope": + positions = None + + else: + positions = self.embed_positions(encoder_padding_mask).transpose(0, 1) + x += positions + positions = None + + x = self.linear(x) + x = self.dropout(x) + encoder_states = [] + + # x is T X B X C + for layer in self.conformer_layers: + x, _ = layer(x, encoder_padding_mask, positions) + if return_all_hiddens: + encoder_states.append(x) + + return { + "encoder_out": [x], # T x B x C + "encoder_padding_mask": [encoder_padding_mask] + if encoder_padding_mask.any() + else [], # B x T + "encoder_embedding": [], # B x T x C + "encoder_states": encoder_states, # List[T x B x C] + "src_tokens": [], + "src_lengths": [], + } + + def forward(self, src_tokens, src_lengths, return_all_hiddens=False): + if self.num_updates < self.encoder_freezing_updates: + with torch.no_grad(): + x = self._forward( + src_tokens, + src_lengths, + return_all_hiddens=return_all_hiddens, + ) + else: + x = self._forward( + src_tokens, + src_lengths, + return_all_hiddens=return_all_hiddens, + ) + return x + + def reorder_encoder_out(self, encoder_out, new_order): + """Required method for a FairseqEncoder. Calls the method from the parent class""" + return S2TTransformerEncoder.reorder_encoder_out(self, encoder_out, new_order) + + def set_num_updates(self, num_updates): + super().set_num_updates(num_updates) + self.num_updates = num_updates + + +@register_model("s2t_conformer") +class S2TConformerModel(S2TTransformerModel): + def __init__(self, encoder, decoder): + super().__init__(encoder, decoder) + + @staticmethod + def add_args(parser): + S2TTransformerModel.add_args(parser) + parser.add_argument( + "--input-feat-per-channel", + type=int, + metavar="N", + help="dimension of input features per channel", + ) + parser.add_argument( + "--input-channels", + type=int, + metavar="N", + help="number of chennels of input features", + ) + parser.add_argument( + "--depthwise-conv-kernel-size", + type=int, + metavar="N", + help="kernel size of depthwise convolution layers", + ) + parser.add_argument( + "--attn-type", + type=str, + metavar="STR", + help="If not specified uses fairseq MHA. Other valid option is espnet", + ) + parser.add_argument( + "--pos-enc-type", + type=str, + metavar="STR", + help="Must be specified in addition to attn-type=espnet for rel_pos and rope", + ) + + @classmethod + def build_encoder(cls, args): + encoder = S2TConformerEncoder(args) + pretraining_path = getattr(args, "load_pretrained_encoder_from", None) + if pretraining_path is not None: + if not Path(pretraining_path).exists(): + logger.warning( + f"skipped pretraining because {pretraining_path} does not exist" + ) + else: + encoder = checkpoint_utils.load_pretrained_component_from_model( + component=encoder, checkpoint=pretraining_path + ) + logger.info(f"loaded pretrained encoder from: {pretraining_path}") + return encoder + + +@register_model_architecture("s2t_conformer", "s2t_conformer") +def conformer_base_architecture(args): + args.attn_type = getattr(args, "attn_type", None) + args.pos_enc_type = getattr(args, "pos_enc_type", "abs") + args.input_feat_per_channel = getattr(args, "input_feat_per_channel", 80) + args.input_channels = getattr(args, "input_channels", 1) + args.max_source_positions = getattr(args, "max_source_positions", 6000) + args.encoder_embed_dim = getattr(args, "encoder_embed_dim", 256) + args.encoder_ffn_embed_dim = getattr(args, "encoder_ffn_embed_dim", 2048) + args.encoder_attention_heads = getattr(args, "encoder_attention_heads", 4) + args.dropout = getattr(args, "dropout", 0.1) + args.encoder_layers = getattr(args, "encoder_layers", 16) + args.depthwise_conv_kernel_size = getattr(args, "depthwise_conv_kernel_size", 31) + transformer_base_architecture(args) diff --git a/data/fairseq/fairseq/models/speech_to_text/s2t_transformer.py b/data/fairseq/fairseq/models/speech_to_text/s2t_transformer.py new file mode 100644 index 0000000000000000000000000000000000000000..50fae2ffa276368e9fb1c339fda904f27363c45d --- /dev/null +++ b/data/fairseq/fairseq/models/speech_to_text/s2t_transformer.py @@ -0,0 +1,552 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import logging +import math +from pathlib import Path +from typing import Dict, List, Optional, Tuple + +import torch +import torch.nn as nn +from torch import Tensor + +from fairseq import checkpoint_utils, utils +from fairseq.data.data_utils import lengths_to_padding_mask +from fairseq.models import ( + FairseqEncoder, + FairseqEncoderDecoderModel, + register_model, + register_model_architecture, +) +from fairseq.models.speech_to_text.hub_interface import S2THubInterface +from fairseq.models.speech_to_text.modules.convolution import ( + Conv1dSubsampler, + Conv2dSubsampler, +) +from fairseq.models.transformer import Embedding, TransformerDecoder +from fairseq.modules import ( + FairseqDropout, + LayerNorm, + PositionalEmbedding, + TransformerEncoderLayer, +) + +logger = logging.getLogger(__name__) + + +@register_model("s2t_transformer") +class S2TTransformerModel(FairseqEncoderDecoderModel): + """Adapted Transformer model (https://arxiv.org/abs/1706.03762) for + speech-to-text tasks. The Transformer encoder/decoder remains the same. + A trainable input subsampler is prepended to the Transformer encoder to + project inputs into the encoder dimension as well as downsample input + sequence for computational efficiency.""" + + @classmethod + def hub_models(cls): + base_url = "http://dl.fbaipublicfiles.com/fairseq/s2t" + model_ids = [ + "s2t_transformer_s-en-asr-librispeech", + "s2t_transformer_m-en-asr-librispeech", + "s2t_transformer_l-en-asr-librispeech", + ] + return {i: f"{base_url}/{i}.tar.gz" for i in model_ids} + + @classmethod + def from_pretrained( + cls, + model_name_or_path, + checkpoint_file="model.pt", + data_name_or_path=".", + config_yaml="config.yaml", + **kwargs, + ): + from fairseq import hub_utils + + x = hub_utils.from_pretrained( + model_name_or_path, + checkpoint_file, + data_name_or_path, + archive_map=cls.hub_models(), + config_yaml=config_yaml, + **kwargs, + ) + return S2THubInterface(x["args"], x["task"], x["models"][0]) + + def __init__(self, encoder, decoder): + super().__init__(encoder, decoder) + + @staticmethod + def add_args(parser): + """Add model-specific arguments to the parser.""" + # input + parser.add_argument( + "--conv-kernel-sizes", + type=str, + metavar="STR", + help="kernel sizes of Conv1d (s2t_transformer) subsampling layers", + ) + parser.add_argument( + "--conv-channels", + type=int, + metavar="N", + help="# of channels in Conv1d (s2t_transformer) subsampling layers", + ) + parser.add_argument( + "--conv-out-channels", + type=int, + metavar="N", + help="# of channels in Conv2d (convtransformer) subsampling layers", + ) + parser.add_argument( + "--conv-version", + type=str, + default="s2t_transformer", + choices=["s2t_transformer", "convtransformer"], + help="version of frontend convolutional layers", + ) + # Transformer + parser.add_argument( + "--activation-fn", + type=str, + default="relu", + choices=utils.get_available_activation_fns(), + help="activation function to use", + ) + parser.add_argument( + "--dropout", type=float, metavar="D", help="dropout probability" + ) + parser.add_argument( + "--attention-dropout", + type=float, + metavar="D", + help="dropout probability for attention weights", + ) + parser.add_argument( + "--activation-dropout", + "--relu-dropout", + type=float, + metavar="D", + help="dropout probability after activation in FFN.", + ) + parser.add_argument( + "--encoder-embed-dim", + type=int, + metavar="N", + help="encoder embedding dimension", + ) + parser.add_argument( + "--encoder-ffn-embed-dim", + type=int, + metavar="N", + help="encoder embedding dimension for FFN", + ) + parser.add_argument( + "--encoder-layers", type=int, metavar="N", help="num encoder layers" + ) + parser.add_argument( + "--encoder-attention-heads", + type=int, + metavar="N", + help="num encoder attention heads", + ) + parser.add_argument( + "--encoder-normalize-before", + action="store_true", + help="apply layernorm before each encoder block", + ) + parser.add_argument( + "--decoder-embed-dim", + type=int, + metavar="N", + help="decoder embedding dimension", + ) + parser.add_argument( + "--decoder-ffn-embed-dim", + type=int, + metavar="N", + help="decoder embedding dimension for FFN", + ) + parser.add_argument( + "--decoder-layers", type=int, metavar="N", help="num decoder layers" + ) + parser.add_argument( + "--decoder-attention-heads", + type=int, + metavar="N", + help="num decoder attention heads", + ) + parser.add_argument( + "--decoder-normalize-before", + action="store_true", + help="apply layernorm before each decoder block", + ) + parser.add_argument( + "--share-decoder-input-output-embed", + action="store_true", + help="share decoder input and output embeddings", + ) + parser.add_argument( + "--layernorm-embedding", + action="store_true", + help="add layernorm to embedding", + ) + parser.add_argument( + "--no-scale-embedding", + action="store_true", + help="if True, dont scale embeddings", + ) + parser.add_argument( + "--load-pretrained-encoder-from", + type=str, + metavar="STR", + help="model to take encoder weights from (for initialization)", + ) + parser.add_argument( + "--encoder-freezing-updates", + type=int, + metavar="N", + help="freeze encoder for first N updates", + ) + + @classmethod + def build_encoder(cls, args): + encoder = S2TTransformerEncoder(args) + pretraining_path = getattr(args, "load_pretrained_encoder_from", None) + if pretraining_path is not None: + if not Path(pretraining_path).exists(): + logger.warning( + f"skipped pretraining because {pretraining_path} does not exist" + ) + else: + encoder = checkpoint_utils.load_pretrained_component_from_model( + component=encoder, checkpoint=pretraining_path + ) + logger.info(f"loaded pretrained encoder from: {pretraining_path}") + return encoder + + @classmethod + def build_decoder(cls, args, task, embed_tokens): + return TransformerDecoderScriptable(args, task.target_dictionary, embed_tokens) + + @classmethod + def build_model(cls, args, task): + """Build a new model instance.""" + + # make sure all arguments are present in older models + base_architecture(args) + + def build_embedding(dictionary, embed_dim): + num_embeddings = len(dictionary) + padding_idx = dictionary.pad() + return Embedding(num_embeddings, embed_dim, padding_idx) + + decoder_embed_tokens = build_embedding( + task.target_dictionary, args.decoder_embed_dim + ) + args.tgt_dict_size = len(task.target_dictionary) + encoder = cls.build_encoder(args) + decoder = cls.build_decoder(args, task, decoder_embed_tokens) + return cls(encoder, decoder) + + def get_normalized_probs( + self, + net_output: Tuple[Tensor, Optional[Dict[str, List[Optional[Tensor]]]]], + log_probs: bool, + sample: Optional[Dict[str, Tensor]] = None, + ): + # net_output['encoder_out'] is a (B, T, D) tensor + lprobs = self.get_normalized_probs_scriptable(net_output, log_probs, sample) + lprobs.batch_first = True + return lprobs + + def get_ctc_target(self, sample: Optional[Dict[str, Tensor]]): + return sample["target"], sample["target_lengths"] + + def get_ctc_output( + self, + net_output: Tuple[Tensor, Optional[Dict[str, List[Optional[Tensor]]]]], + sample: Optional[Dict[str, Tensor]], + ): + encoder_out = net_output[1]["encoder_out"]["encoder_out"][0] + logits = self.encoder.ctc_proj(encoder_out) # T x B x C + out = utils.log_softmax(logits.float(), dim=-1) + padding_mask = net_output[1]["encoder_out"]["encoder_padding_mask"] + lens = out.new_full((out.shape[1],), out.shape[0]).long() + if len(padding_mask) > 0: + lens -= padding_mask[0].sum(dim=-1) + return out, lens + + def forward(self, src_tokens, src_lengths, prev_output_tokens): + """ + The forward method inherited from the base class has a **kwargs + argument in its input, which is not supported in torchscript. This + method overwrites the forward method definition without **kwargs. + """ + encoder_out = self.encoder(src_tokens=src_tokens, src_lengths=src_lengths) + decoder_out = self.decoder( + prev_output_tokens=prev_output_tokens, encoder_out=encoder_out + ) + return decoder_out + + +class S2TTransformerEncoder(FairseqEncoder): + """Speech-to-text Transformer encoder that consists of input subsampler and + Transformer encoder.""" + + def __init__(self, args): + super().__init__(None) + + self.encoder_freezing_updates = args.encoder_freezing_updates + self.num_updates = 0 + + self.dropout_module = FairseqDropout( + p=args.dropout, module_name=self.__class__.__name__ + ) + self.embed_scale = math.sqrt(args.encoder_embed_dim) + if args.no_scale_embedding: + self.embed_scale = 1.0 + self.padding_idx = 1 + + self.conv_version = args.conv_version + if self.conv_version == "s2t_transformer": + self.subsample = Conv1dSubsampler( + args.input_feat_per_channel * args.input_channels, + args.conv_channels, + args.encoder_embed_dim, + [int(k) for k in args.conv_kernel_sizes.split(",")], + ) + elif self.conv_version == "convtransformer": + self.subsample = Conv2dSubsampler( + args.input_channels, + args.input_feat_per_channel, + args.conv_out_channels, + args.encoder_embed_dim, + ) + + self.embed_positions = PositionalEmbedding( + args.max_source_positions, args.encoder_embed_dim, self.padding_idx + ) + + self.transformer_layers = nn.ModuleList( + [TransformerEncoderLayer(args) for _ in range(args.encoder_layers)] + ) + if args.encoder_normalize_before: + self.layer_norm = LayerNorm(args.encoder_embed_dim) + else: + self.layer_norm = None + + self.ctc_proj = None + if getattr(args, "ctc_weight", 0.0) > 0.0: + self.ctc_proj = nn.Linear(args.encoder_embed_dim, args.tgt_dict_size) + + def _forward(self, src_tokens, src_lengths, return_all_hiddens=False): + x, input_lengths = self.subsample(src_tokens, src_lengths) + x = self.embed_scale * x + + encoder_padding_mask = lengths_to_padding_mask(input_lengths) + positions = self.embed_positions(encoder_padding_mask).transpose(0, 1) + x += positions + x = self.dropout_module(x) + + encoder_states = [] + + for layer in self.transformer_layers: + x = layer(x, encoder_padding_mask) + if return_all_hiddens: + encoder_states.append(x) + + if self.layer_norm is not None: + x = self.layer_norm(x) + + return { + "encoder_out": [x], # T x B x C + "encoder_padding_mask": [encoder_padding_mask] + if encoder_padding_mask.any() + else [], # B x T + "encoder_embedding": [], # B x T x C + "encoder_states": encoder_states, # List[T x B x C] + "src_tokens": [], + "src_lengths": [], + } + + def forward(self, src_tokens, src_lengths, return_all_hiddens=False): + if self.num_updates < self.encoder_freezing_updates: + with torch.no_grad(): + x = self._forward( + src_tokens, src_lengths, return_all_hiddens=return_all_hiddens + ) + else: + x = self._forward( + src_tokens, src_lengths, return_all_hiddens=return_all_hiddens + ) + return x + + def reorder_encoder_out(self, encoder_out, new_order): + new_encoder_out = ( + [] + if len(encoder_out["encoder_out"]) == 0 + else [x.index_select(1, new_order) for x in encoder_out["encoder_out"]] + ) + + new_encoder_padding_mask = ( + [] + if len(encoder_out["encoder_padding_mask"]) == 0 + else [ + x.index_select(0, new_order) + for x in encoder_out["encoder_padding_mask"] + ] + ) + + new_encoder_embedding = ( + [] + if len(encoder_out["encoder_embedding"]) == 0 + else [ + x.index_select(0, new_order) for x in encoder_out["encoder_embedding"] + ] + ) + + encoder_states = encoder_out["encoder_states"] + if len(encoder_states) > 0: + for idx, state in enumerate(encoder_states): + encoder_states[idx] = state.index_select(1, new_order) + + return { + "encoder_out": new_encoder_out, # T x B x C + "encoder_padding_mask": new_encoder_padding_mask, # B x T + "encoder_embedding": new_encoder_embedding, # B x T x C + "encoder_states": encoder_states, # List[T x B x C] + "src_tokens": [], # B x T + "src_lengths": [], # B x 1 + } + + def set_num_updates(self, num_updates): + super().set_num_updates(num_updates) + self.num_updates = num_updates + + +class TransformerDecoderScriptable(TransformerDecoder): + def extract_features( + self, + prev_output_tokens, + encoder_out: Optional[Dict[str, List[Tensor]]] = None, + incremental_state: Optional[Dict[str, Dict[str, Optional[Tensor]]]] = None, + full_context_alignment: bool = False, + alignment_layer: Optional[int] = None, + alignment_heads: Optional[int] = None, + ): + # call scriptable method from parent class + x, _ = self.extract_features_scriptable( + prev_output_tokens, + encoder_out, + incremental_state, + full_context_alignment, + alignment_layer, + alignment_heads, + ) + extra = {"encoder_out": encoder_out} if incremental_state is None else None + return x, extra + + +@register_model_architecture(model_name="s2t_transformer", arch_name="s2t_transformer") +def base_architecture(args): + args.encoder_freezing_updates = getattr(args, "encoder_freezing_updates", 0) + # Convolutional subsampler + args.input_channels = getattr(args, "input_channels", 1) + args.conv_kernel_sizes = getattr(args, "conv_kernel_sizes", "5,5") # for Conv1d + args.conv_channels = getattr(args, "conv_channels", 1024) # for Conv1d + args.conv_out_channels = getattr(args, "conv_out_channels", 256) # for Conv2d + args.conv_version = getattr(args, "conv_version", "s2t_transformer") + # Transformer + args.encoder_embed_dim = getattr(args, "encoder_embed_dim", 512) + args.encoder_ffn_embed_dim = getattr(args, "encoder_ffn_embed_dim", 2048) + args.encoder_layers = getattr(args, "encoder_layers", 12) + args.encoder_attention_heads = getattr(args, "encoder_attention_heads", 8) + args.encoder_normalize_before = getattr(args, "encoder_normalize_before", True) + args.decoder_embed_dim = getattr(args, "decoder_embed_dim", args.encoder_embed_dim) + args.decoder_ffn_embed_dim = getattr( + args, "decoder_ffn_embed_dim", args.encoder_ffn_embed_dim + ) + args.decoder_layers = getattr(args, "decoder_layers", 6) + args.decoder_attention_heads = getattr(args, "decoder_attention_heads", 8) + args.decoder_normalize_before = getattr(args, "decoder_normalize_before", True) + args.decoder_learned_pos = getattr(args, "decoder_learned_pos", False) + args.dropout = getattr(args, "dropout", 0.1) + args.attention_dropout = getattr(args, "attention_dropout", args.dropout) + args.activation_dropout = getattr(args, "activation_dropout", args.dropout) + args.activation_fn = getattr(args, "activation_fn", "relu") + args.adaptive_softmax_cutoff = getattr(args, "adaptive_softmax_cutoff", None) + args.adaptive_softmax_dropout = getattr(args, "adaptive_softmax_dropout", 0) + args.share_decoder_input_output_embed = getattr( + args, "share_decoder_input_output_embed", False + ) + args.no_token_positional_embeddings = getattr( + args, "no_token_positional_embeddings", False + ) + args.adaptive_input = getattr(args, "adaptive_input", False) + args.decoder_layerdrop = getattr(args, "decoder_layerdrop", 0.0) + args.decoder_output_dim = getattr( + args, "decoder_output_dim", args.decoder_embed_dim + ) + args.decoder_input_dim = getattr(args, "decoder_input_dim", args.decoder_embed_dim) + args.no_scale_embedding = getattr(args, "no_scale_embedding", False) + args.quant_noise_pq = getattr(args, "quant_noise_pq", 0) + + +@register_model_architecture("s2t_transformer", "s2t_transformer_s") +def s2t_transformer_s(args): + args.encoder_embed_dim = getattr(args, "encoder_embed_dim", 256) + args.encoder_ffn_embed_dim = getattr(args, "encoder_ffn_embed_dim", 256 * 8) + args.encoder_attention_heads = getattr(args, "encoder_attention_heads", 4) + args.decoder_attention_heads = getattr(args, "decoder_attention_heads", 4) + args.dropout = getattr(args, "dropout", 0.1) + base_architecture(args) + + +@register_model_architecture("s2t_transformer", "s2t_transformer_xs") +def s2t_transformer_xs(args): + args.encoder_layers = getattr(args, "encoder_layers", 6) + args.decoder_layers = getattr(args, "decoder_layers", 3) + args.encoder_ffn_embed_dim = getattr(args, "encoder_ffn_embed_dim", 256 * 4) + args.dropout = getattr(args, "dropout", 0.3) + s2t_transformer_s(args) + + +@register_model_architecture("s2t_transformer", "s2t_transformer_sp") +def s2t_transformer_sp(args): + args.encoder_layers = getattr(args, "encoder_layers", 16) + s2t_transformer_s(args) + + +@register_model_architecture("s2t_transformer", "s2t_transformer_m") +def s2t_transformer_m(args): + args.encoder_embed_dim = getattr(args, "encoder_embed_dim", 512) + args.encoder_ffn_embed_dim = getattr(args, "encoder_ffn_embed_dim", 512 * 4) + args.encoder_attention_heads = getattr(args, "encoder_attention_heads", 8) + args.decoder_attention_heads = getattr(args, "decoder_attention_heads", 8) + args.dropout = getattr(args, "dropout", 0.15) + base_architecture(args) + + +@register_model_architecture("s2t_transformer", "s2t_transformer_mp") +def s2t_transformer_mp(args): + args.encoder_layers = getattr(args, "encoder_layers", 16) + s2t_transformer_m(args) + + +@register_model_architecture("s2t_transformer", "s2t_transformer_l") +def s2t_transformer_l(args): + args.encoder_embed_dim = getattr(args, "encoder_embed_dim", 1024) + args.encoder_ffn_embed_dim = getattr(args, "encoder_ffn_embed_dim", 1024 * 4) + args.encoder_attention_heads = getattr(args, "encoder_attention_heads", 16) + args.decoder_attention_heads = getattr(args, "decoder_attention_heads", 16) + args.dropout = getattr(args, "dropout", 0.2) + base_architecture(args) + + +@register_model_architecture("s2t_transformer", "s2t_transformer_lp") +def s2t_transformer_lp(args): + args.encoder_layers = getattr(args, "encoder_layers", 16) + s2t_transformer_l(args) diff --git a/data/fairseq/fairseq/models/speech_to_text/s2t_wav_transformer.py b/data/fairseq/fairseq/models/speech_to_text/s2t_wav_transformer.py new file mode 100644 index 0000000000000000000000000000000000000000..ad21aeeb1a6f2c47a0ed380478853a189e3577a4 --- /dev/null +++ b/data/fairseq/fairseq/models/speech_to_text/s2t_wav_transformer.py @@ -0,0 +1,504 @@ +#!/usr/bin/env python3 + +import math + +import torch +import torch.nn as nn + +from fairseq.data.data_utils import compute_mask_indices +from fairseq.models import FairseqEncoder +from fairseq.models.wav2vec import ConvFeatureExtractionModel +from fairseq.modules import GradMultiply, LayerNorm, SamePad, TransformerEncoderLayer + + +# Transformer encoder with wave input, it is adopted from wav2vec 2.0 Encoder. +# use wav input +# use trained position embedding so it is easier to match with text input +class SpeechWavTransformerEncoder(FairseqEncoder): + + # extra parameters for speech encoder besides those defined in transformermodel + @staticmethod + def add_args(parser): + parser.add_argument( + "--dropout-input", + type=float, + metavar="D", + help="dropout to apply to the input (after feat extr)", + ) + parser.add_argument( + "--dropout-features", + type=float, + metavar="D", + help="dropout to apply to the unmasked features (after feat extr)", + ) + parser.add_argument( + "--speech-extractor-mode", + type=str, + default="layer_norm", + choices=["default", "layer_norm"], + help="feature extractor norm", + ) + + parser.add_argument( + "--speech-conv-bias", + action="store_true", + help="include bias in speech conv encoder", + ) + + parser.add_argument( + "--conv-feature-layers", + default="[(512, 10, 5)] + [(512, 3, 2)] * 4 + [(512,2,2)] + [(512,2,2)]", + help="string describing convolutional feature extraction layers in form of a python list that contains [(dim, kernel_size, stride), ...]", + ) + + parser.add_argument( + "--speech-mask-length", + type=int, + help="repeat the mask indices multiple times", + ) + + parser.add_argument( + "--speech-mask-prob", + type=float, + help="probability of replacing a token with mask", + ) + + parser.add_argument( + "--speech-mask-selection", + type=str, + choices=["static", "uniform", "normal", "poisson"], + help="how to choose masks", + ) + + parser.add_argument( + "--speech-mask-other", + type=float, + help="stdev of the mask length in case of 'normal' selection strategy", + ) + + parser.add_argument( + "--speech-no-mask-overlap", + action="store_true", + help="whether to allow masks to overlap", + ) + + parser.add_argument( + "--speech-mask-min-space", + type=int, + help="min space between spans (if no overlap is enabled)", + ) + + parser.add_argument( + "--speech-mask-channel-length", + type=int, + help="repeat the mask indices multiple times", + ) + + parser.add_argument( + "--speech-mask-channel-prob", + type=float, + help="probability of replacing a token with mask", + ) + + parser.add_argument( + "--speech-mask-channel-selection", + type=str, + choices=["static", "uniform", "normal", "poisson"], + help="how to choose masks", + ) + + parser.add_argument( + "--speech-mask-channel-other", + type=float, + help="stdev of the mask length in case of 'normal' selection strategy", + ) + + parser.add_argument( + "--speech-no-mask-channel-overlap", + action="store_true", + help="whether to allow masks to overlap", + ) + + parser.add_argument( + "--no-scale-feature", + action="store_true", + help="no scale for the calculated features", + ) + + parser.add_argument( + "--speech-mask-channel-min-space", + type=int, + help="min space between spans (if no overlap is enabled)", + ) + + parser.add_argument( + "--feature-grad-mult", + type=float, + help="reset feature grad mult in wav2vec 2.0 to this", + ) + + # positional embeddings + parser.add_argument( + "--conv-pos", + type=int, + default=128, + help="number of filters for convolutional positional embeddings", + ) + + parser.add_argument( + "--conv-pos-groups", + type=int, + default=16, + help="number of groups for convolutional positional embedding", + ) + # model configures + parser.add_argument( + "--speech-encoder-layers", + type=int, + help="number of speech encoder layers", + ) + parser.add_argument( + "--text-encoder-layers", + type=int, + help="number of text encoder layers", + ) + + def __init__(self, args, alway_mask=False): + super().__init__(args) + self.args = args + self.dropout = args.dropout + self.embedding_dim = args.encoder_embed_dim + self.feat_scale = math.sqrt(args.encoder_embed_dim) + if args.no_scale_feature: + self.feat_scale = 1.0 + + subsample = ConvFeatureExtractionModel( + conv_layers=eval(args.conv_feature_layers), + dropout=0.0, + mode=args.speech_extractor_mode, # default, layer_norm + conv_bias=args.speech_conv_bias, + ) + self.feature_enc_layers = eval(args.conv_feature_layers) + self.subsample = subsample + self.feat_proj = ( + nn.Linear(self.feature_enc_layers[-1][0], self.embedding_dim) + if self.feature_enc_layers[-1][0] != self.embedding_dim + else None + ) + + self.feat_layer_norm = LayerNorm(self.feature_enc_layers[-1][0]) + + self.embed_positions = nn.Conv1d( + self.embedding_dim, + self.embedding_dim, + kernel_size=args.conv_pos, + padding=args.conv_pos // 2, + groups=args.conv_pos_groups, + ) + std = math.sqrt(4 / (args.conv_pos * self.embedding_dim)) + nn.init.normal_(self.embed_positions.weight, mean=0, std=std) + nn.init.constant_(self.embed_positions.bias, 0) + + self.embed_positions = nn.utils.weight_norm( + self.embed_positions, name="weight", dim=2 + ) + self.embed_positions = nn.Sequential( + self.embed_positions, SamePad(args.conv_pos), nn.GELU() + ) + + self.mask_prob = args.speech_mask_prob + self.mask_selection = args.speech_mask_selection + self.mask_other = args.speech_mask_other + self.mask_length = args.speech_mask_length + self.no_mask_overlap = args.speech_no_mask_overlap + self.mask_min_space = args.speech_mask_min_space + + self.mask_channel_prob = args.speech_mask_channel_prob + self.mask_channel_selection = args.speech_mask_channel_selection + self.mask_channel_other = args.speech_mask_channel_other + self.mask_channel_length = args.speech_mask_channel_length + self.no_mask_channel_overlap = args.speech_no_mask_channel_overlap + self.mask_channel_min_space = args.speech_mask_channel_min_space + + self.dropout_input = nn.Dropout(args.dropout_input) + self.dropout_features = nn.Dropout(args.dropout_features) + + self.feature_grad_mult = args.feature_grad_mult + + self.mask_emb = nn.Parameter( + torch.FloatTensor(args.encoder_embed_dim).uniform_() + ) + + self.layers = nn.ModuleList( + [TransformerEncoderLayer(args) for _ in range(args.encoder_layers)] + ) + self.layer_norm = LayerNorm(args.encoder_embed_dim) + self.normalize_before = args.encoder_normalize_before + self.alway_mask = alway_mask + + def _get_feat_extract_output_lengths(self, input_lengths: torch.LongTensor): + """ + Computes the output length of the convolutional layers + """ + + def _conv_out_length(input_length, kernel_size, stride): + return torch.floor((input_length - kernel_size) / stride + 1) + + for i in range(len(self.feature_enc_layers)): + input_lengths = _conv_out_length( + input_lengths, + self.feature_enc_layers[i][1], + self.feature_enc_layers[i][2], + ) + + return input_lengths.to(torch.long) + + def apply_mask(self, x, padding_mask): + B, T, C = x.shape + if self.mask_prob > 0: + mask_indices = compute_mask_indices( + (B, T), + padding_mask, + self.mask_prob, + self.mask_length, + self.mask_selection, + self.mask_other, + min_masks=2, + no_overlap=self.no_mask_overlap, + min_space=self.mask_min_space, + ) + mask_indices = torch.from_numpy(mask_indices).to(x.device) + x[mask_indices] = self.mask_emb + else: + mask_indices = None + + if self.mask_channel_prob > 0: + mask_channel_indices = compute_mask_indices( + (B, C), + None, + self.mask_channel_prob, + self.mask_channel_length, + self.mask_channel_selection, + self.mask_channel_other, + no_overlap=self.no_mask_channel_overlap, + min_space=self.mask_channel_min_space, + ) + mask_channel_indices = ( + torch.from_numpy(mask_channel_indices) + .to(x.device) + .unsqueeze(1) + .expand(-1, T, -1) + ) + x[mask_channel_indices] = 0 + + return x, mask_indices + + def forward( + self, + src_tokens, + src_lengths, + return_all_hiddens=False, + padding_mask=None, + features_only=True, + ): + mask = self.training or self.alway_mask + if self.feature_grad_mult > 0 and self.training: + features = self.subsample(src_tokens) + if self.feature_grad_mult != 1.0: + features = GradMultiply.apply(features, self.feature_grad_mult) + else: + with torch.no_grad(): + features = self.subsample(src_tokens) + features = features.transpose(1, 2) + features = self.feat_layer_norm(features) + if self.feat_proj is not None: + features = self.feat_proj(features) + + if padding_mask is not None: + input_lengths = (1 - padding_mask.long()).sum(-1) + else: + input_lengths = src_lengths + # apply conv formula to get real output_lengths + output_lengths = self._get_feat_extract_output_lengths(input_lengths) + + padding_mask = torch.zeros( + features.shape[:2], dtype=features.dtype, device=features.device + ) + + # these two operations makes sure that all values + # before the output lengths indices are attended to + padding_mask[ + ( + torch.arange(padding_mask.shape[0], device=padding_mask.device), + output_lengths - 1, + ) + ] = 1 + padding_mask = (1 - padding_mask.flip([-1]).cumsum(-1).flip([-1])).bool() + + features = self.feat_scale * features if self.feat_scale != 1.0 else features + unmasked_features = features.clone() + + features = self.dropout_input(features) + unmasked_features = self.dropout_features(unmasked_features) + if mask: + x, mask_indices = self.apply_mask(features, padding_mask) + else: + x = features + mask_indices = None + + def cal_transformer_layers(x, encoder_padding_mask, return_all_hiddens=False): + # x: B x T x C + positions = self.embed_positions(x.transpose(1, 2)).transpose(1, 2) + x = x + positions + if not self.normalize_before: + x = self.layer_norm(x) + + # B x T x C -> T x B x C + x = x.transpose(0, 1) + encoder_states = [] + for layer in self.layers: + x = layer(x, encoder_padding_mask) + if return_all_hiddens: + encoder_states.append(x) + if self.normalize_before: + x = self.layer_norm(x) + return x, encoder_states + + x, encoder_states = cal_transformer_layers(x, padding_mask, return_all_hiddens) + if features_only: + return { + "encoder_out": [x], # [T x B x C] + "encoder_padding_mask": [padding_mask] + if padding_mask is not None + else [], # B x T + "encoder_embedding": [], # + "encoder_states": encoder_states, # List[T x B x C] + "src_tokens": [], + "src_lengths": [], + "mask_indices": [mask_indices], + } + + x_unmasked = x + if self.mask_prob > 0 or self.mask_channel_prob > 0: + x_unmasked, _ = cal_transformer_layers(unmasked_features, padding_mask) + return { + "encoder_out": [x], # [T x B x C] + "encoder_unmasked_out": [x_unmasked], # [T x B x C] + "encoder_padding_mask": [padding_mask] + if padding_mask is not None + else [], # B x T + "encoder_embedding": [], # + "encoder_states": encoder_states, # List[T x B x C] + "src_tokens": [], + "src_lengths": [], + "mask_indices": [mask_indices] if mask_indices is not None else [], # B X T + } + + def reorder_encoder_out(self, encoder_out, new_order): + new_encoder_out = ( + [] + if len(encoder_out["encoder_out"]) == 0 + else [x.index_select(1, new_order) for x in encoder_out["encoder_out"]] + ) + + new_encoder_padding_mask = ( + [] + if len(encoder_out["encoder_padding_mask"]) == 0 + else [ + x.index_select(0, new_order) + for x in encoder_out["encoder_padding_mask"] + ] + ) + + new_encoder_embedding = ( + [] + if len(encoder_out["encoder_embedding"]) == 0 + else [ + x.index_select(0, new_order) for x in encoder_out["encoder_embedding"] + ] + ) + + encoder_states = encoder_out["encoder_states"] + if len(encoder_states) > 0: + for idx, state in enumerate(encoder_states): + encoder_states[idx] = state.index_select(1, new_order) + + return { + "encoder_out": new_encoder_out, # T x B x C + "encoder_padding_mask": new_encoder_padding_mask, # B x T + "encoder_embedding": new_encoder_embedding, # B x T x C + "encoder_states": encoder_states, # List[T x B x C] + "src_tokens": [], # B x T + "src_lengths": [], # B x 1 + } + + +class StackedSpeechWavTransformerEncoder(FairseqEncoder): + def __init__(self, speech_enc, text_enc_layers, text_layer_norm): + super().__init__(None) + self.speech_encoder = speech_enc + self.text_encoder_layers = text_enc_layers + self.final_layer_norm = text_layer_norm + + def forward( + self, + src_tokens, + src_lengths=None, + return_all_hiddens=False, + padding_mask=None, + features_only=True, + ): + + out = self.speech_encoder.forward( + src_tokens, + src_lengths, + return_all_hiddens, + padding_mask=padding_mask, + features_only=features_only, + ) + x = out["encoder_out"][0] + encoder_padding_mask = None + if len(out["encoder_padding_mask"]) > 0: + encoder_padding_mask = out["encoder_padding_mask"][0] + + def cal_text_layers(x, padding_mask, return_all_hiddens=False): + encoder_states = [] + for layer in self.text_encoder_layers: + x = layer(x, padding_mask) + if return_all_hiddens: + encoder_states.append(x) + if self.final_layer_norm is not None: + x = self.final_layer_norm(x) + return x, encoder_states + + x, encoder_states = cal_text_layers(x, encoder_padding_mask, return_all_hiddens) + if features_only: + return { + "encoder_out": [x], # T x B x C + "encoder_padding_mask": [encoder_padding_mask] + if encoder_padding_mask is not None + else [], # B x T + "encoder_embedding": [], # B x T x C + "encoder_states": encoder_states, # List[T x B x C] + "src_tokens": [], + "src_lengths": [], + } + + x_u = out["encoder_unmasked_out"][0] + x_u, _ = cal_text_layers(x_u, encoder_padding_mask) + + return { + "encoder_out": [x], # [T x B x C] + "encoder_unmasked_out": [x_u], # [T x B x C] + "encoder_padding_mask": [encoder_padding_mask] + if encoder_padding_mask is not None + else [], # B x T + "encoder_embedding": [], # + "encoder_states": encoder_states, # List[T x B x C] + "src_tokens": [], + "src_lengths": [], + "mask_indices": out["mask_indices"], # B X T + } + + def reorder_encoder_out(self, encoder_out, new_order): + return self.speech_encoder.reorder_encoder_out(encoder_out, new_order) diff --git a/data/fairseq/fairseq/models/speech_to_text/utils.py b/data/fairseq/fairseq/models/speech_to_text/utils.py new file mode 100644 index 0000000000000000000000000000000000000000..33117446a5e2f3b71c64f0a7b6b8122a1ac7c182 --- /dev/null +++ b/data/fairseq/fairseq/models/speech_to_text/utils.py @@ -0,0 +1,562 @@ +# Copyright (c) 2017-present, Facebook, Inc. +# All rights reserved. +# +# This source code is licensed under the license found in the LICENSE file in +# the root directory of this source tree. An additional grant of patent rights +# can be found in the PATENTS file in the same directory. + + +import logging +from collections.abc import Iterable +from itertools import repeat +from typing import List, Optional, Tuple + +import torch +from torch import Tensor + +# ------------------------------------------------------------------------------ +# assert_equal() +# ------------------------------------------------------------------------------ + + +def assert_equal(value1, value2, name1=None, name2=None): + """Asserts two values are equal otherwise raise an error.""" + + str_name1 = "" if name1 is None else "{} ".format(name1) + str_name2 = "" if name2 is None else "{} ".format(name2) + if value1 != value2: + str_value1 = "{}" if name1 is None else "({})" + str_value1 = str_value1.format(value1) + str_value2 = "{}" if name2 is None else "({})" + str_value2 = str_value2.format(value2) + raise ValueError( + "Expected {}{} == {}{}".format(str_name1, str_value1, str_name2, str_value2) + ) + + +def fill_config(config, key, value): + if value is not None: + if key not in config or config[key] is None: + config[key] = value + assert_equal(value, config[key], "value", f'config["{key}"]') + + +# ------------------------------------------------------------------------------ +# check_and_return_expected() +# ------------------------------------------------------------------------------ + + +def check_and_return_expected(value, undefined_value, expected_value, name=None): + """ + Return the expected value while checking if the given value is undefined or + equal to the expected value. + """ + if (undefined_value is None and value is None) or (undefined_value == value): + return expected_value + if value != expected_value: + str_name = "" if name is None else "{} ".format(name) + str_value = "{}" if name is None else "({})" + str_value = str_value.format(value) + raise ValueError( + "Expected {}{} == {}".format(str_name, str_value, expected_value) + ) + return expected_value + + +# ------------------------------------------------------------------------------ +# get_time_axis() +# ------------------------------------------------------------------------------ + + +def get_time_axis(layout): + """ + Extract the time axis from the layout, for example for breaking sequence into + segments. + """ + if layout in ["TB", "TBD"]: + return 0 + if layout in ["BT", "BTD"]: + return 1 + if layout in ["BCTD"]: + return 2 + raise ValueError("Unsupported layout = {}".format(layout)) + + +# ------------------------------------------------------------------------------ +# get_batch_axis() +# ------------------------------------------------------------------------------ + + +def get_batch_axis(layout): + """ + Extract the batch axis from the layout + """ + if layout in ["TB", "TBD"]: + return 1 + if layout in ["BT", "BTD", "BCTD"]: + return 0 + raise ValueError("Unsupported layout = {}".format(layout)) + + +# ------------------------------------------------------------------------------ +# monotonically_increasing_and_bounded() +# ------------------------------------------------------------------------------ + + +def monotonically_increasing_and_bounded(iterable, min=None, max=None): + """ + Check if the elements in the given iterable are monotonically increasing and + bounded by upper/lower bounds. + """ + if not isinstance(iterable, Iterable): + raise TypeError( + "Expected iterable to be of type Iterable, got ({})".format( + iterable.__class__.__name__ + ) + ) + for i in range(len(iterable)): + if min is not None and iterable[i] < min: + return False + if max is not None and iterable[i] > max: + return False + if i > 0 and iterable[i] <= iterable[i - 1]: + return False + return True + + +# ------------------------------------------------------------------------------ +# to_pair() +# ------------------------------------------------------------------------------ + + +def to_pair(value, name): + """Make a pair (of type tuple) of given value.""" + if isinstance(value, Iterable): + if len(value) != 2: + raise ValueError( + "Expected `{}` to have exactly 2 elements, got: ({})".format( + name, value + ) + ) + return value + return tuple(repeat(value, 2)) + + +# ------------------------------------------------------------------------------ +# infer_conv_output_attrs() +# ------------------------------------------------------------------------------ + + +# TODO(cfyeh): figure out if we can get `output_dim` without calling the module. +def infer_conv_output_attrs( + module, input_channels, input_dim, batch_size=1, max_length=8 +): + """Get output attributes of a module with input.""" + input = torch.randn(batch_size, input_channels, max_length, input_dim) + output = module(input) + output_channels = output.shape[1] + output_dim = output.shape[-1] + return output_channels, output_dim + + +# ------------------------------------------------------------------------------ +# NoOp +# ------------------------------------------------------------------------------ + + +class NoOp(torch.nn.Module): + """ + NoOp simply passes the input as the output. + """ + + def __init__(self): + super().__init__() + + def forward(self, input: Tensor) -> Tensor: + return input + + +# ------------------------------------------------------------------------------ +# Permute: a torch.nn.Module applies permutation on the input tensor. +# ------------------------------------------------------------------------------ + + +class Permute(torch.nn.Module): + def __init__(self, dims): + super().__init__() + self.dims = dims + + def forward(self, input: Tensor) -> Tensor: + return input.permute(self.dims).contiguous() + + +# ------------------------------------------------------------------------------ +# lengths_to_padding_mask() +# ------------------------------------------------------------------------------ + + +def lengths_to_padding_mask(lengths: Tensor) -> Tensor: + """Convert lengths of shape (B, ) to padding mask.""" + batch_size = lengths.shape[0] + max_length = int(torch.max(lengths).item()) + padding_mask = torch.arange( # [0, ..., T-1] + max_length, device=lengths.device, dtype=lengths.dtype + ).expand(batch_size, max_length) >= lengths.unsqueeze(1) + + return padding_mask + + +# ------------------------------------------------------------------------------ +# lengths_to_attention_mask() +# ------------------------------------------------------------------------------ + + +def lengths_to_attention_mask( + lengths: Tensor, + left_context: Optional[int] = None, + right_context: Optional[int] = None, +) -> Optional[Tensor]: + """ + Generate attention mask based on (lengths, left_context, right_context). + left_context is None means unlimited left context. + right_context is None means unlimited right context. + """ + + if left_context is None and right_context is None: + return None + + max_length = int(torch.max(lengths).item()) + + # For example, with `max_length` == 5, + # indices = tensor([ + # [ 0, 1, 2, 3, 4, 5], + # [-1, 0, 1, 2, 3, 4], + # [-2, -1, 0, 1, 2, 3], + # [-3, -2, -1, 0, 1, 2], + # [-4, -3, -2, -1, 0, 1], + # [-5, -4, -3, -2, -1, 0], + # ]) + + # In some cases the second torch.arange is created on cpu which causes a + # failure. Adding the device option to guard against it. + indices = torch.arange( + max_length, device=lengths.device, dtype=lengths.dtype + ).expand(max_length, max_length) - torch.arange( + max_length, device=lengths.device + ).view( + max_length, -1 + ) + + # For example, with `max_length` == 5, + # bool_mask = tensor([ + # [True, True, True, True, True], + # [True, True, True, True, True], + # [True, True, True, True, True], + # [True, True, True, True, True], + # [True, True, True, True, True], + # ]) + bool_mask = ( + torch.tensor([True]).to(device=lengths.device).expand(max_length, max_length) + ) + + # For example, with `max_length` == 5, left_context == 2 + # left_mask = tensor([ + # [ True, True, True, True, True], + # [ True, True, True, True, True], + # [ True, True, True, True, True], + # [False, True, True, True, True], + # [False, False, True, True, True], + # ]) + if left_context is not None: + left_mask = indices >= -left_context + bool_mask = bool_mask & left_mask + + # For example, with `max_length` == 5, right_context == 1 + # right_mask = tensor([ + # [True, True, False, False, False], + # [True, True, True, False, False], + # [True, True, True, True, False], + # [True, True, True, True, True], + # [True, True, True, True, True], + # ]) + if right_context is not None: + right_mask = indices <= right_context + bool_mask = bool_mask & right_mask + + bool_mask = (~bool_mask).to(device=lengths.device) + return bool_mask + + +# ------------------------------------------------------------------------------ +# infer_output_norm() +# ------------------------------------------------------------------------------ + + +def infer_output_norm(module, output_norm=None): + """ + Infer the output norm (string and module) needed on the module gvien desired + output normalization. + """ + if output_norm == module.output_norm(): + # output_norm already matches module.output_norm(). + return (None, NoOp()) + + if output_norm is None and module.output_norm() is not None: + logger = logging.getLogger("infer_output_norm()") + logger.warning( + "trying to set output_norm ({}) ".format(output_norm) + + "but got module.output_norm() ({}), ".format(module.output_norm()) + + "the combined output_norm() will be ({})".format(module.output_norm()) + ) + return (None, NoOp()) + + if output_norm == "log_softmax": + if module.output_norm() is not None: + raise ValueError( + "incompatible output_norm ({}) ".format(output_norm) + + "and module.output_norm() ({})".format(module.output_norm()) + ) + else: + return ("log_softmax", torch.nn.LogSoftmax(dim=-1)) + + if output_norm == "softmax": + if module.output_norm() is not None: + raise ValueError( + "incompatible output_norm ({}) ".format(output_norm) + + "and module.output_norm() ({})".format(module.output_norm()) + ) + else: + return ("softmax", torch.nn.Softmax(dim=-1)) + + raise ValueError( + "output_norm ({}) not in ".format(output_norm) + + "supported list = [None, softmax, log_softmax]" + ) + + +# ------------------------------------------------------------------------------ +# infer_channels_from_layout() +# ------------------------------------------------------------------------------ + + +def infer_channels_from_layout(layout, channels): + """Extract the number of channels from the layout.""" + if layout in ("TBD", "BTD"): + if channels is not None and channels != 1: + raise ValueError( + "Expected channels ({}) to be 1 for layout = {}".format( + channels, layout + ) + ) + if channels is None: + return 1 + return channels + + +# ------------------------------------------------------------------------------ +# pad_sequence() +# ------------------------------------------------------------------------------ + + +@torch.jit.export +def pad_sequence( + sequence: Tensor, + time_axis: int, + extra_left_context: int = 0, + extra_right_context: int = 0, +) -> Tensor: + """Pad extra left/right contexts to the sequence.""" + + if extra_left_context == 0 and extra_right_context == 0: + return sequence + + tensors_to_concat = [] + + if extra_left_context: + size = (extra_left_context,) + fill_value = 0 + indices = torch.full( + size=size, + fill_value=fill_value, + dtype=torch.long, + device=sequence.device, + ) + left_padding = torch.index_select(sequence, time_axis, indices) + tensors_to_concat.append(left_padding) + + tensors_to_concat.append(sequence) + + # NOTE(cfyeh): for efficiency reason we pad 0 instead of the last frame for + # extra right contexts. + if extra_right_context: + size = list(sequence.shape) + size[time_axis] = extra_right_context + right_padding = torch.zeros(size, dtype=sequence.dtype, device=sequence.device) + tensors_to_concat.append(right_padding) + + padded_sequence = torch.cat(tensors_to_concat, dim=time_axis) + return padded_sequence + + +# ------------------------------------------------------------------------------ +# sequence_to_segments() +# ------------------------------------------------------------------------------ + + +@torch.jit.export +def sequence_to_segments( + sequence: Tensor, + time_axis: int, + lengths: Tensor, + segment_size: Optional[int] = None, + extra_left_context: int = 0, + extra_right_context: int = 0, +) -> List[Tuple[Tensor, Tensor]]: + """Breaks sequence into segments.""" + + sequence = pad_sequence( + sequence=sequence, + time_axis=time_axis, + extra_left_context=extra_left_context, + extra_right_context=extra_right_context, + ) + + lengths = lengths + extra_left_context + extra_right_context + + segments: List[Tuple[Tensor, Tensor]] = [] + + if segment_size is None: + segments.append((sequence, lengths)) + return segments + + offset = 0 + end = sequence.shape[time_axis] + step = segment_size + size = extra_left_context + segment_size + extra_right_context + + while offset + extra_left_context + extra_right_context < end: + clamped_size = min(size, end - offset) + segment_lengths = torch.clamp(lengths - offset, min=0, max=clamped_size) + indices = torch.arange( + start=offset, + end=(offset + clamped_size), + step=1, + dtype=torch.long, + device=sequence.device, + ) + segment_tensor = torch.index_select(sequence, time_axis, indices) + segments.append((segment_tensor, segment_lengths)) + offset = offset + step + + return segments + + +# ------------------------------------------------------------------------------ +# segments_to_sequence() +# ------------------------------------------------------------------------------ + + +@torch.jit.export +def segments_to_sequence( + segments: List[Tuple[Tensor, Tensor]], time_axis: int +) -> Tuple[Tensor, Tensor]: + """Concatenate segments into a full sequence.""" + if len(segments) == 1: + return segments[0] + + tensors_to_concat: List[Tensor] = [] + lengths_to_stack: List[Tensor] = [] + + for tensor, lengths in segments: + tensors_to_concat.append(tensor) + lengths_to_stack.append(lengths) + + sequence = torch.cat(tensors_to_concat, dim=time_axis) + lengths = torch.stack(lengths_to_stack, dim=0) + lengths = torch.sum(lengths, dim=0) + + return sequence, lengths + + +def lengths_to_encoder_padding_mask(lengths, batch_first: bool = False): + """ + convert lengths (a 1-D Long/Int tensor) to 2-D binary tensor + + Args: + lengths: a (B, )-shaped tensor + batch_first: whether to return a (B, T) tensor + + Return: + max_length: maximum length of B sequences + encoder_padding_mask: a (max_length, B) binary mask, where + [t, b] = False for t < lengths[b] and True otherwise + + TODO: + kernelize this function if benchmarking shows this function is slow + """ + max_lengths = torch.max(lengths).item() + bsz = lengths.size(0) + encoder_padding_mask = torch.arange( + max_lengths + ).to( # a (T, ) tensor with [0, ..., T-1] + lengths.device + ).view( # move to the right device + 1, max_lengths + ).expand( # reshape to (1, T)-shaped tensor + bsz, -1 + ) > lengths.view( # expand to (B, T)-shaped tensor + bsz, 1 + ).expand( + -1, max_lengths + ) + if not batch_first: + return encoder_padding_mask.t(), max_lengths + else: + return encoder_padding_mask, max_lengths + + +# ------------------------------------------------------------------------------ +# attention suppression +# ------------------------------------------------------------------------------ + + +def attention_suppression(attention_weights: Tensor, scale: float): + # B, H, qlen, klen -> B, H, qlen, 1 + attention_prob = torch.nn.functional.softmax(attention_weights.float(), dim=-1) + attention_nozeros = attention_prob.to(torch.bool) + nozeros_sum = torch.sum(attention_nozeros.to(torch.float), dim=-1, keepdim=True) + + # For very sparse situation, we need get round about 0s + key_sum = torch.sum(attention_prob, dim=-1, keepdim=True) + + # nozeros_sum should > 1 + key_mean = key_sum / (nozeros_sum + 1e-8) + + # std calculation + dis = (attention_prob - key_mean) * (attention_prob - key_mean) + + # if attention_prob[i] < threshold, then dis_masked[i] = 0; for all i + dis_masked = torch.where( + attention_nozeros, dis, attention_prob.new_zeros(attention_prob.size()) + ) + + key_var = torch.sum(dis_masked, dim=-1, keepdim=True) + key_var = key_var / (nozeros_sum - 1.0 + 1e-8) + key_std = torch.sqrt(key_var) + key_thread = key_mean - scale * key_std + + # if attention_prob[i] >= key_thread, then attention_prob[i] + # , otherwise "-inf" + inf_tensor = attention_prob.new_zeros(attention_prob.size()).detach() + inf_tensor[:] = float("-inf") + attention_weights_float = torch.where( + attention_prob < key_thread, + inf_tensor, + attention_weights.float(), + ) + + return attention_weights_float.type_as(attention_weights) + + +def layer_norm_backward_hook(module, grad_input, grad_output, clamp_value): + return tuple(torch.clamp(v, min=-clamp_value, max=clamp_value) for v in grad_input) diff --git a/data/fairseq/fairseq/models/speech_to_text/xm_transformer.py b/data/fairseq/fairseq/models/speech_to_text/xm_transformer.py new file mode 100644 index 0000000000000000000000000000000000000000..7b4b23464182b233ef178be2a3babfb13872259a --- /dev/null +++ b/data/fairseq/fairseq/models/speech_to_text/xm_transformer.py @@ -0,0 +1,855 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import copy +import logging +from typing import Dict, List, Optional, Tuple + +import numpy as np +import torch +import torch.nn as nn +from torch import Tensor + +from fairseq import checkpoint_utils, utils +from fairseq.data.data_utils import lengths_to_padding_mask +from fairseq.models import ( + FairseqEncoder, + FairseqEncoderDecoderModel, + FairseqEncoderModel, + FairseqLanguageModel, + register_model, + register_model_architecture, +) +from fairseq.models.speech_to_speech.modules.ctc_decoder import CTCDecoder +from fairseq.models.speech_to_text.hub_interface import S2THubInterface +from fairseq.models.transformer import ( + Embedding, + TransformerDecoder, + TransformerModelBase, +) +from fairseq.models.wav2vec import Wav2VecEncoder +from fairseq.modules.layer_norm import LayerNorm + +logger = logging.getLogger(__name__) + + +def build_embedding(dictionary, embed_dim): + num_embeddings = len(dictionary) + padding_idx = dictionary.pad() + return Embedding(num_embeddings, embed_dim, padding_idx) + + +class Conv1dAdaptor(nn.Module): + def __init__( + self, + in_dim, + out_dim, + n_layers=3, + kernel_size=3, + stride=2, + layerdrop=0.0, + layernorm=False, + proj=False, + ): + super().__init__() + self.proj, self.proj_ln = None, None + self.post_proj, self.post_proj_ln = None, None + if proj: + self.proj = nn.Sequential( + nn.Linear(in_dim, in_dim * 4), nn.ReLU(), nn.Linear(in_dim * 4, in_dim) + ) + self.proj_ln = LayerNorm(in_dim) + self.post_proj = nn.Sequential( + nn.Linear(out_dim, out_dim * 4), + nn.ReLU(), + nn.Linear(out_dim * 4, out_dim), + ) + self.post_proj_ln = LayerNorm(out_dim) + + self.layers = nn.ModuleList( + nn.Conv1d( + in_dim if i == 0 else out_dim, + out_dim * 2, + kernel_size, + stride=stride, + padding=kernel_size // 2, + ) + for i in range(n_layers) + ) + self.stride = stride + self.layerdrop = layerdrop + self.layernorm = LayerNorm(in_dim) if layernorm else None + + @classmethod + def add_args(cls, parser): + parser.add_argument("--adaptor-n-layers", type=int) + parser.add_argument("--adaptor-kernel-size", type=int) + parser.add_argument("--adaptor-stride", type=int) + parser.add_argument("--adaptor-layerdrop", type=float) + parser.add_argument("--adaptor-layernorm", action="store_true") + parser.add_argument("--adaptor-proj", action="store_true") + + def forward(self, x, padding_mask: Optional[torch.Tensor]): + if self.layernorm is not None: + x = self.layernorm(x) + + if self.proj is not None: + x = x + 0.5 * self.proj(x) + x = self.proj_ln(x) + + if padding_mask is not None: + x = utils.index_put(x, padding_mask.T, 0) + + # T x B x C -> B x C x T + x = x.transpose(0, 1).transpose(1, 2) + out_lens = None + if padding_mask is not None: + out_lens = (~padding_mask).sum(1).float() + + for layer in self.layers: + layerdrop_prob = np.random.random() + if not self.training or (layerdrop_prob > self.layerdrop): + x = nn.functional.glu(layer(x), dim=1) + if padding_mask is not None: + out_lens = ((out_lens - 1) / self.stride + 1).floor() + # B x C x T -> T x B x C + x = x.transpose(1, 2).transpose(0, 1) + + if self.post_proj is not None: + x = x + 0.5 * self.post_proj(x) + x = self.post_proj_ln(x) + + out_padding_mask = None + if padding_mask is not None: + out_padding_mask = lengths_to_padding_mask(out_lens.long()) + x = utils.index_put(x, out_padding_mask.T, 0) + return x, out_padding_mask + + +def add_wav2vec_asr_args(parser): + parser.add_argument("--w2v-path", help="path to wav2vec 2.0 model") + parser.add_argument( + "--no-pretrained-weights", + action="store_true", + help="if true, does not load pretrained weights", + ) + parser.add_argument( + "--dropout-input", + type=float, + metavar="D", + help="dropout to apply to the input (after feat extr)", + ) + parser.add_argument( + "--final-dropout", + type=float, + metavar="D", + help="dropout after transformer and before final projection", + ) + parser.add_argument( + "--apply-mask", action="store_true", help="apply masking during fine-tuning" + ) + parser.add_argument( + "--dropout", + type=float, + metavar="D", + help="dropout probability inside wav2vec 2.0 model", + ) + parser.add_argument( + "--attention-dropout", + type=float, + metavar="D", + help="dropout probability for attention weights inside wav2vec 2.0 model", + ) + parser.add_argument( + "--activation-dropout", + "--relu-dropout", + type=float, + metavar="D", + help="dropout probability after activation in FFN inside wav2vec 2.0 model", + ) + parser.add_argument( + "--mask-length", type=int, help="repeat the mask indices multiple times" + ) + parser.add_argument( + "--mask-prob", type=float, help="probability of replacing a token with mask" + ) + parser.add_argument( + "--mask-selection", + type=str, + choices=["static", "uniform", "normal", "poisson"], + help="how to choose masks", + ) + parser.add_argument( + "--mask-other", + type=float, + help="stdev of the mask length in case of 'normal' selection strategy", + ) + parser.add_argument( + "--no-mask-overlap", + action="store_true", + help="whether to allow masks to overlap", + ) + parser.add_argument( + "--mask-channel-length", type=int, help="repeat the mask indices multiple times" + ) + parser.add_argument( + "--mask-channel-prob", + type=float, + help="probability of replacing a token with mask", + ) + parser.add_argument( + "--mask-channel-selection", + type=str, + choices=["static", "uniform", "normal", "poisson"], + help="how to choose masks", + ) + parser.add_argument( + "--mask-channel-other", + type=float, + help="stdev of the mask length in case of 'normal' selection strategy", + ) + parser.add_argument( + "--no-mask-channel-overlap", + action="store_true", + help="whether to allow masks to overlap", + ) + parser.add_argument( + "--freeze-finetune-updates", + type=int, + metavar="N", + help="dont finetune wav2vec for this many updates", + ) + parser.add_argument( + "--feature-grad-mult", + type=float, + metavar="D", + help="reset feature grad mult in wav2vec 2.0 to this", + ) + parser.add_argument( + "--layerdrop", + type=float, + metavar="D", + help="probability of dropping a layer in wav2vec 2.0", + ) + parser.add_argument( + "--max-positions", + type=int, + metavar="N", + help="Max input positions to be used in the conformer encoder in wav2vec 2.0", + ) + parser.add_argument("--encoder-proj", action="store_true") + parser.add_argument("--w2v-args", default=None) + parser.add_argument( + "--remove-weight-norm", + action="store_true", + help="if set, then the weight-norm (in one pos_conv layer) is removed from the model", + ) + parser.add_argument( + "--encoder-embed-dim", + type=int, + metavar="N", + help="encoder embedding dimension to be used when w2v_path is None and no encoder_proj is set", + ) + + +def need_finetuning(ft_params, param_name): + if ft_params == "all": + return True + ft_params_list = ft_params.split(",") + for ft_param in ft_params_list: + if ft_param in param_name: + return True + return False + + +class Wav2VecEncoderWithAdaptor(FairseqEncoder): + def build_adaptor(self, args): + adaptor = None + if args.adaptor_n_layers > 0: + adaptor = Conv1dAdaptor( + args.decoder_embed_dim, + args.decoder_embed_dim, + n_layers=args.adaptor_n_layers, + kernel_size=args.adaptor_kernel_size, + stride=args.adaptor_stride, + layerdrop=args.adaptor_layerdrop, + layernorm=args.adaptor_layernorm, + proj=args.adaptor_proj, + ) + return adaptor + + def __init__(self, args): + super().__init__(None) + self.w2v_encoder = Wav2VecEncoder(args) + self.is_v0_arch = not args.adaptor_proj + self.w2v_proj_ln = None + if not self.is_v0_arch and self.w2v_encoder.proj is not None: + self.w2v_proj_ln = LayerNorm(args.decoder_embed_dim) + self.adaptor = self.build_adaptor(args) + + self.num_updates = 0 + self.freezing_updates = args.w2v_freezing_updates + self.finetuning_params = args.finetune_w2v_params + for k, p in self.w2v_encoder.w2v_model.named_parameters(): + p.requires_grad = need_finetuning(self.finetuning_params, k) + + @classmethod + def add_args(cls, parser): + """Add model-specific arguments to the parser.""" + add_wav2vec_asr_args(parser) + parser.add_argument( + "--normalize", + action="store_true", + help="if set, normalizes input to have 0 mean and unit variance", + ) + parser.add_argument( + "--finetune-w2v-params", + type=str, + metavar="STR", + help="comma-separated param strings to finetune.", + ) + parser.add_argument("--w2v-freezing-updates", type=int) + parser.add_argument("--load-pretrained-encoder-from", type=str, metavar="STR") + Conv1dAdaptor.add_args(parser) + + def set_num_updates(self, num_updates): + super().set_num_updates(num_updates) + self.num_updates = num_updates + + def forward(self, src_tokens, src_lengths=None, **kwargs): + if ( + self.freezing_updates is not None + and self.num_updates > self.freezing_updates + ): + for p in self.w2v_encoder.w2v_model.parameters(): + p.requires_grad = True + + padding_mask = lengths_to_padding_mask(src_lengths) + out = self.w2v_encoder.forward(src_tokens, padding_mask, tbc=True) + x, padding_mask = out["encoder_out"], out["padding_mask"] + if self.w2v_proj_ln is not None: + x = self.w2v_proj_ln(x) + + if self.adaptor is not None: + x, padding_mask = self.adaptor(x, padding_mask) + + return { + "encoder_out": [x], # T x B x C + "encoder_padding_mask": [] + if padding_mask is None + else [padding_mask], # B x T + "encoder_embedding": [], # B x T x C + "encoder_states": [], # List[T x B x C] + "src_tokens": [], + "src_lengths": [], + } + + def reorder_encoder_out(self, encoder_out, new_order): + new_encoder_out = ( + [] + if len(encoder_out["encoder_out"]) == 0 + else [x.index_select(1, new_order) for x in encoder_out["encoder_out"]] + ) + + new_encoder_padding_mask = ( + [] + if len(encoder_out["encoder_padding_mask"]) == 0 + else [ + x.index_select(0, new_order) + for x in encoder_out["encoder_padding_mask"] + ] + ) + + new_encoder_embedding = ( + [] + if len(encoder_out["encoder_embedding"]) == 0 + else [ + x.index_select(0, new_order) for x in encoder_out["encoder_embedding"] + ] + ) + + encoder_states = encoder_out["encoder_states"] + if len(encoder_states) > 0: + for idx, state in enumerate(encoder_states): + encoder_states[idx] = state.index_select(1, new_order) + + return { + "encoder_out": new_encoder_out, # T x B x C + "encoder_padding_mask": new_encoder_padding_mask, # B x T + "encoder_embedding": new_encoder_embedding, # B x T x C + "encoder_states": encoder_states, # List[T x B x C] + "src_tokens": [], # B x T + "src_lengths": [], # B x 1 + } + + +def add_decoder_args(parser): + parser.add_argument( + "--activation-fn", + type=str, + default="relu", + choices=utils.get_available_activation_fns(), + help="activation function to use", + ) + parser.add_argument( + "--decoder-dropout", type=float, metavar="D", help="dropout probability" + ) + parser.add_argument( + "--decoder-attention-dropout", + type=float, + metavar="D", + help="dropout probability for attention weights", + ) + parser.add_argument( + "--decoder-activation-dropout", + type=float, + metavar="D", + help="dropout probability after activation in FFN.", + ) + parser.add_argument( + "--decoder-embed-dim", type=int, metavar="N", help="decoder embedding dimension" + ) + parser.add_argument( + "--decoder-ffn-embed-dim", + type=int, + metavar="N", + help="decoder embedding dimension for FFN", + ) + parser.add_argument( + "--decoder-layers", type=int, metavar="N", help="num decoder layers" + ) + parser.add_argument( + "--decoder-attention-heads", + type=int, + metavar="N", + help="num decoder attention heads", + ) + parser.add_argument( + "--decoder-normalize-before", + action="store_true", + help="apply layernorm before each decoder block", + ) + parser.add_argument( + "--layernorm-embedding", action="store_true", help="add layernorm to embedding" + ) + parser.add_argument( + "--decoder-layerdrop", + type=float, + metavar="D", + help="layerdrop probability for decoder", + ) + parser.add_argument( + "--decoder-learned-pos", + action="store_true", + help="learn positional embedding in decoder", + ) + parser.add_argument( + "--share-decoder-input-output-embed", + action="store_true", + help="share decoder input and output embeddings", + ) + parser.add_argument( + "--no-scale-embedding", + action="store_true", + help="if True, dont scale embeddings", + ) + parser.add_argument( + "--load-pretrained-decoder-from", + type=str, + metavar="STR", + help="model to take decoder weights from (for initialization)", + ) + parser.add_argument( + "--finetune-decoder-params", + type=str, + metavar="STR", + help="comma-separated param strings to finetune.", + ) + + +def remove_weight_norm_from_model(model): + from functools import reduce + + layers_with_wn = [] + for param_name, _ in model.named_parameters(): + if param_name.endswith("_g"): + # retrieve the module with this param_name + module_names = param_name.split(".")[ + :-1 + ] # exclude the actual parameter name + wn_module = reduce(getattr, module_names, model) + layers_with_wn.append(wn_module) + for wn_module in layers_with_wn: + torch.nn.utils.remove_weight_norm(wn_module) + logger.warning(f"Weight norm removed from module with {wn_module}\n") + + +@register_model("xm_transformer") +class XMTransformerModel(FairseqEncoderDecoderModel): + @classmethod + def hub_models(cls): + base_url = "http://dl.fbaipublicfiles.com/fairseq/s2t" + model_ids = [ + "xm_transformer_600m-es_en-multi_domain", + "xm_transformer_600m-ru_en-multi_domain", + "xm_transformer_600m-fr_en-multi_domain", + "xm_transformer_600m-en_es-multi_domain", + "xm_transformer_600m-en_ru-multi_domain", + "xm_transformer_600m-en_fr-multi_domain", + "xm_transformer_600m-en_zh-multi_domain", + "xm_transformer_600m-en_ar-multi_domain", + "xm_transformer_600m-en_tr-multi_domain", + "xm_transformer_600m-en_vi-multi_domain", + "xm_transformer-21_en-xls_r_300m", + "xm_transformer-en_15-xls_r_300m", + "xm_transformer-21_en-xls_r_1b", + "xm_transformer-en_15-xls_r_1b", + "xm_transformer-21_en-xls_r_2b", + "xm_transformer-en_15-xls_r_2b", + "xm_transformer-22_16-xls_r_2b", + "xm_transformer_s2ut_800m-es-en-st-asr-bt_h1_2022", + "xm_transformer_s2ut_800m-en-es-st_plus_asr", + "xm_transformer_s2ut_800m-hk-en-h1_2022", + "xm_transformer_s2ut_800m-en-hk-h1_2022", + ] + return {i: f"{base_url}/{i}.tar.gz" for i in model_ids} + + @classmethod + def from_pretrained( + cls, + model_name_or_path, + checkpoint_file="model.pt", + data_name_or_path=".", + config_yaml="config.yaml", + task="speech_to_text", + generation_args=None, + **kwargs, + ): + from fairseq import hub_utils + + x = hub_utils.from_pretrained( + model_name_or_path, + checkpoint_file, + data_name_or_path, + archive_map=cls.hub_models(), + config_yaml=config_yaml, + task=task, + generation_args=generation_args, + **kwargs, + ) + return S2THubInterface(x["args"], x["task"], x["models"][0]) + + def __init__(self, encoder, decoder): + super().__init__(encoder, decoder) + + @classmethod + def add_args(cls, parser): + """Add model-specific arguments to the parser.""" + Wav2VecEncoderWithAdaptor.add_args(parser) + add_decoder_args(parser) + parser.add_argument("--checkpoint-activations", action="store_true") + parser.add_argument("--offload-activations", action="store_true") + parser.add_argument("--min-params-to-wrap", type=int, metavar="N") + + @classmethod + def maybe_load_pretrained(cls, component, checkpoint: Optional[str] = None): + if checkpoint is None: + return component + + _load = checkpoint_utils.load_pretrained_component_from_model + try: + return _load(component, checkpoint) + except RuntimeError as e: + logger.warning(e) + return _load(component, checkpoint, strict=False) + + @classmethod + def build_encoder(cls, args): + _args = copy.deepcopy(args) + if not args.adaptor_proj and not args.encoder_proj: # V0 arch + if args.w2v_path: + state = checkpoint_utils.load_checkpoint_to_cpu(args.w2v_path) + if state.get("cfg") is not None: + encoder_embed_dim = state["cfg"]._content["model"][ + "encoder_embed_dim" + ] + elif state.get("args") is not None: + encoder_embed_dim = state["args"].encoder_embed_dim + else: + raise ValueError(f"Invalid config in {args.w2v_path}") + _args.decoder_embed_dim = encoder_embed_dim + del state + else: + _args.decoder_embed_dim = args.encoder_embed_dim + + encoder = Wav2VecEncoderWithAdaptor(_args) + encoder = cls.maybe_load_pretrained( + encoder, getattr(args, "load_pretrained_encoder_from", None) + ) + if args.remove_weight_norm: + # remove the wn for EMA usage + logger.warning("Removing weight norm from wav2vec encoder") + remove_weight_norm_from_model(encoder) + + return encoder + + @classmethod + def get_decoder_args_from_checkpoint(cls, ckpt_args): + assert "model" in ckpt_args, "Model args not found in checkpoint cfg!" + decoder_args = {} + for k, v in ckpt_args["model"].__dict__.items(): + if "decoder" in k: + decoder_args[k] = v + + return decoder_args + + @classmethod + def override_decoder_args(cls, cli_args, decoder_args_dict): + for k, v in decoder_args_dict.items(): + if v != getattr(cli_args, k, None): + logger.warning( + f"Overriding decoder arg {k}: from {getattr(cli_args, k, None)} to {v}" + ) + setattr(cli_args, k, v) + + return cli_args + + @classmethod + def build_decoder(cls, args, task, embed_tokens): + _args = copy.deepcopy(args) + if args.adaptor_proj or args.encoder_proj: # not V0 arch + _args.encoder_embed_dim = _args.decoder_embed_dim + _args.dropout = args.decoder_dropout + _args.attention_dropout = args.decoder_attention_dropout + _args.activation_dropout = args.decoder_activation_dropout + _args.layerdrop = _args.decoder_layerdrop + + decoder = TransformerDecoder(_args, task.target_dictionary, embed_tokens) + decoder = cls.maybe_load_pretrained( + decoder, getattr(args, "load_pretrained_decoder_from", None) + ) + + for k, p in decoder.named_parameters(): + p.requires_grad = need_finetuning(args.finetune_decoder_params, k) + return decoder + + @classmethod + def build_model(cls, args, task): + """Build a new model instance.""" + + # make sure all arguments are present in older models + base_architecture(args) + if getattr(args, "load_pretrained_decoder_from", None) is not None: + ckpt = torch.load(getattr(args, "load_pretrained_decoder_from", None)) + decoder_args_dict = cls.get_decoder_args_from_checkpoint(ckpt["cfg"]) + args = cls.override_decoder_args(args, decoder_args_dict) + + decoder_embed_tokens = build_embedding( + task.target_dictionary, args.decoder_embed_dim + ) + + encoder = cls.build_encoder(args) + decoder = cls.build_decoder(args, task, decoder_embed_tokens) + base_model = cls(encoder, decoder) + + # set up multitask decoders + base_model.multitask_decoders = {} + for i, (task_name, task_obj) in enumerate(task.multitask_tasks.items()): + # dummy auxiliary decoder + if task_obj.args.get_loss_weight(0) == 0: + continue + + task_decoder = cls.build_multitask_decoder( + args, task_obj.args, task_obj.target_dictionary, args.decoder_embed_dim + ) + + setattr(base_model, f"{task_name}_decoder", task_decoder) + decoder_model_cls = ( + FairseqEncoderModel + if task_obj.args.decoder_type == "ctc" + else FairseqLanguageModel + ) + base_model.multitask_decoders[task_name] = decoder_model_cls( + getattr(base_model, f"{task_name}_decoder") + ) + return base_model + + @classmethod + def build_multitask_decoder( + cls, + args, + mtl_args, + tgt_dict, + in_dim, + is_first_pass_decoder=False, + ): + decoder_args = mtl_args.decoder_args + decoder_args.encoder_embed_dim = in_dim + if mtl_args.decoder_type == "transformer": + if is_first_pass_decoder: + task_decoder = cls.build_text_decoder(args, tgt_dict) + else: + from fairseq.models.speech_to_speech import ( + base_multitask_text_transformer_decoder_arch, + ) + + base_multitask_text_transformer_decoder_arch(decoder_args) # 2L + task_decoder = TransformerDecoder( + decoder_args, + tgt_dict, + embed_tokens=TransformerModelBase.build_embedding( + decoder_args, + tgt_dict, + decoder_args.decoder_embed_dim, + ), + ) + elif mtl_args.decoder_type == "ctc": + task_decoder = CTCDecoder( + dictionary=tgt_dict, + in_dim=in_dim, + ) + else: + raise NotImplementedError( + "currently only support multitask decoder_type 'transformer', 'ctc'" + ) + + return task_decoder + + def get_normalized_probs( + self, + net_output: Tuple[Tensor, Optional[Dict[str, List[Optional[Tensor]]]]], + log_probs: bool, + sample: Optional[Dict[str, Tensor]] = None, + ): + return self.get_normalized_probs_scriptable(net_output, log_probs, sample) + + def forward( + self, + src_tokens, + src_lengths, + prev_output_tokens, + return_all_hiddens=False, + **kwargs, + ): + """ + The forward method inherited from the base class has a **kwargs + argument in its input, which is not supported in torchscript. This + method overwrites the forward method definition without **kwargs. + """ + encoder_out = self.encoder( + src_tokens=src_tokens, src_lengths=src_lengths, **kwargs + ) + decoder_out = self.decoder( + prev_output_tokens=prev_output_tokens, encoder_out=encoder_out + ) + if return_all_hiddens: + decoder_out[-1]["encoder_states"] = encoder_out["encoder_out"] + # NOTE: from the top layer + decoder_out[-1]["encoder_padding_mask"] = encoder_out[ + "encoder_padding_mask" + ] + return decoder_out + + def upgrade_state_dict(self, state_dict): + for k, _ in state_dict.items(): + if "adaptor.layers" in state_dict: + new = k.replace("adaptor.layers", "adaptor_layers") + state_dict[new] = state_dict[k] + del state_dict[k] + + +def set_default_w2v_encoder_args(args): + args.no_pretrained_weights = getattr(args, "no_pretrained_weights", False) + args.dropout_input = getattr(args, "dropout_input", 0) + args.final_dropout = getattr(args, "final_dropout", 0) + args.apply_mask = getattr(args, "apply_mask", False) + args.dropout = getattr(args, "dropout", 0) + args.attention_dropout = getattr(args, "attention_dropout", 0) + args.activation_dropout = getattr(args, "activation_dropout", 0) + args.encoder_proj = getattr(args, "encoder_proj", False) + args.remove_weight_norm = getattr(args, "remove_weight_norm", False) + + args.mask_length = getattr(args, "mask_length", 10) + args.mask_prob = getattr(args, "mask_prob", 0.5) + args.mask_selection = getattr(args, "mask_selection", "static") + args.mask_other = getattr(args, "mask_other", 0) + args.no_mask_overlap = getattr(args, "no_mask_overlap", False) + args.mask_channel_length = getattr(args, "mask_channel_length", 10) + args.mask_channel_prob = getattr(args, "mask_channel_prob", 0.5) + args.mask_channel_before = getattr(args, "mask_channel_before", False) + args.mask_channel_selection = getattr(args, "mask_channel_selection", "static") + args.mask_channel_other = getattr(args, "mask_channel_other", 0) + args.no_mask_channel_overlap = getattr(args, "no_mask_channel_overlap", False) + + args.freeze_finetune_updates = getattr(args, "freeze_finetune_updates", 0) + args.feature_grad_mult = 0.1 + args.layerdrop = getattr(args, "layerdrop", 0.0) + + args.normalize = getattr(args, "normalize", False) + args.finetune_w2v_params = getattr(args, "finetune_w2v_params", "all") + args.w2v_freezing_updates = getattr(args, "w2v_freezing_updates", None) + args.encoder_embed_dim = getattr(args, "encoder_embed_dim", 1024) + + +def set_default_adaptor_args(args): + args.adaptor_n_layers = getattr(args, "adaptor_n_layers", 3) + args.adaptor_kernel_size = getattr(args, "adaptor_kernel_size", 3) + args.adaptor_stride = getattr(args, "adaptor_stride", 2) + args.adaptor_layerdrop = getattr(args, "adaptor_layerdrop", 0.0) + args.adaptor_layernorm = getattr(args, "adaptor_layernorm", False) + args.adaptor_proj = getattr(args, "adaptor_proj", False) + + +def set_default_transformer_decoder_args(args): + args.decoder_embed_path = getattr(args, "decoder_embed_path", None) + args.decoder_embed_dim = getattr(args, "decoder_embed_dim", 1024) + args.decoder_ffn_embed_dim = getattr(args, "decoder_ffn_embed_dim", 4 * 1024) + args.decoder_layers = getattr(args, "decoder_layers", 12) + args.decoder_attention_heads = getattr(args, "decoder_attention_heads", 16) + args.decoder_normalize_before = getattr(args, "decoder_normalize_before", False) + args.decoder_learned_pos = getattr(args, "decoder_learned_pos", False) + args.decoder_layerdrop = getattr(args, "decoder_layerdrop", 0.0) + args.adaptive_input = getattr(args, "adaptive_input", False) + args.decoder_attention_dropout = getattr(args, "decoder_attention_dropout", 0.0) + args.decoder_activation_dropout = getattr(args, "decoder_activation_dropout", 0.0) + args.decoder_dropout = getattr(args, "decoder_dropout", 0.1) + args.adaptive_softmax_cutoff = getattr(args, "adaptive_softmax_cutoff", None) + args.adaptive_softmax_dropout = getattr(args, "adaptive_softmax_dropout", 0) + args.share_decoder_input_output_embed = getattr( + args, "share_decoder_input_output_embed", False + ) + args.no_token_positional_embeddings = getattr( + args, "no_token_positional_embeddings", False + ) + + args.decoder_output_dim = getattr( + args, "decoder_output_dim", args.decoder_embed_dim + ) + args.decoder_input_dim = getattr(args, "decoder_input_dim", args.decoder_embed_dim) + + args.no_scale_embedding = getattr(args, "no_scale_embedding", False) + args.quant_noise_pq = getattr(args, "quant_noise_pq", 0) + args.layernorm_embedding = getattr(args, "layernorm_embedding", False) + + args.activation_fn = getattr(args, "activation_fn", "gelu") + args.pooler_activation_fn = getattr(args, "pooler_activation_fn", "tanh") + args.pooler_dropout = getattr(args, "pooler_dropout", 0.0) + + args.finetune_decoder_params = getattr(args, "finetune_decoder_params", "all") + + +def set_default_general_args(args): + args.checkpoint_activations = getattr(args, "checkpoint_activations", False) + args.offload_activations = getattr(args, "offload_activations", False) + args.min_params_to_wrap = getattr(args, "min_params_to_wrap", int(1e8)) + args.max_positions = getattr(args, "max_positions", 3000) + + +@register_model_architecture(model_name="xm_transformer", arch_name="xm_transformer") +def base_architecture(args): + set_default_general_args(args) + set_default_w2v_encoder_args(args) + set_default_adaptor_args(args) + set_default_transformer_decoder_args(args) diff --git a/data/fairseq/fairseq/models/speech_to_text/xm_transformer_unity.py b/data/fairseq/fairseq/models/speech_to_text/xm_transformer_unity.py new file mode 100644 index 0000000000000000000000000000000000000000..f77ef4e5707fdce1bd93b46e9bbaad5890f8e7bf --- /dev/null +++ b/data/fairseq/fairseq/models/speech_to_text/xm_transformer_unity.py @@ -0,0 +1,315 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import copy +import logging + +from fairseq.models import ( + FairseqEncoderModel, + FairseqLanguageModel, + register_model, + register_model_architecture, +) +from fairseq.models.speech_to_speech.modules.ctc_decoder import CTCDecoder +from fairseq.models.speech_to_speech.modules.transformer_encoder import ( + TransformerEncoderNoEmb, +) +from fairseq.models.speech_to_text.xm_transformer import XMTransformerModel +from fairseq.models.speech_to_text.xm_transformer import ( + base_architecture as xm_t_base_architecture, +) +from fairseq.models.speech_to_text.xm_transformer import ( + build_embedding, + need_finetuning, + set_default_adaptor_args, + set_default_general_args, + set_default_transformer_decoder_args, + set_default_w2v_encoder_args, +) +from fairseq.models.transformer import Linear, TransformerDecoder, TransformerModelBase +from fairseq.models.transformer.transformer_decoder_aug import AugTransformerDecoder + +logger = logging.getLogger(__name__) + + +def unit_transformer_decoder_arch_base( + args, decoder_layers=6, decoder_embed_dim=768, decoder_attention_heads=12 +): + args.encoder_layers = decoder_layers + args.decoder_layers = decoder_layers + args.decoder_embed_dim = decoder_embed_dim + args.decoder_ffn_embed_dim = decoder_embed_dim * 4 + args.decoder_attention_heads = decoder_attention_heads + args.encoder_embed_dim = args.decoder_embed_dim + args.decoder_output_dim = decoder_embed_dim + args.decoder_input_dim = decoder_embed_dim + + +def unit_transformer_decoder_arch_large( + args, decoder_layers=12, decoder_embed_dim=1024, decoder_attention_heads=16 +): + args.encoder_layers = decoder_layers + args.decoder_layers = decoder_layers + args.decoder_embed_dim = decoder_embed_dim + args.decoder_ffn_embed_dim = decoder_embed_dim * 4 + args.decoder_attention_heads = decoder_attention_heads + args.encoder_embed_dim = args.decoder_embed_dim + args.decoder_output_dim = decoder_embed_dim + args.decoder_input_dim = decoder_embed_dim + + +@register_model("unity_xm_transformer") +class XMTransformerModelUnitY(XMTransformerModel): + @classmethod + def hub_models(cls): + base_url = "http://dl.fbaipublicfiles.com/fairseq/s2t" + model_ids = [] + return {i: f"{base_url}/{i}.tar.gz" for i in model_ids} + + def __init__(self, encoder, decoder): + super().__init__(encoder, decoder) + + @classmethod + def add_args(cls, parser): + """Add model-specific arguments to the parser.""" + XMTransformerModel.add_args(parser) + parser.add_argument( + "--translation-decoder-layers", + type=int, + default=4, + metavar="N", + help="num decoder layers in the first-pass translation module", + ) + parser.add_argument( + "--synthesizer-encoder-layers", + type=int, + default=0, + metavar="N", + help="num encoder layers in the second-pass synthesizer module", + ) + parser.add_argument( + "--synthesizer-augmented-cross-attention", + action="store_true", + default=False, + help="augmented cross-attention over speech encoder output", + ) + parser.add_argument( + "--load-pretrained-aux-decoder-from", + type=str, + metavar="STR", + help="model to take decoder weights from (for initialization)", + ) + + @classmethod + def build_text_decoder(cls, args, tgt_dict): + _args = copy.deepcopy(args) + + if args.adaptor_proj or args.encoder_proj: # not V0 arch + _args.encoder_embed_dim = _args.decoder_embed_dim + _args.dropout = args.decoder_dropout + _args.attention_dropout = args.decoder_attention_dropout + _args.activation_dropout = args.decoder_activation_dropout + _args.layerdrop = _args.decoder_layerdrop + _args.decoder_layers = _args.translation_decoder_layers + + embed_tokens = build_embedding(tgt_dict, _args.decoder_embed_dim) + decoder = TransformerDecoder(_args, tgt_dict, embed_tokens) + + if getattr(args, "load_pretrained_aux_decoder_from", None) is not None: + decoder = cls.maybe_load_pretrained( + decoder, getattr(args, "load_pretrained_aux_decoder_from", None) + ) + + for k, p in decoder.named_parameters(): + p.requires_grad = need_finetuning(args.finetune_decoder_params, k) + return decoder + + @classmethod + def build_decoder(cls, args, task, aug_attn=False): + _args = copy.deepcopy(args) + _args.layerdrop = 0.0 # turn off layerdrop for shallow layers + + _args.encoder_embed_dim = args.decoder_embed_dim + + proj = None + if args.decoder_embed_dim != _args.decoder_embed_dim: + proj = Linear(args.decoder_embed_dim, _args.decoder_embed_dim) + + embed_tokens = build_embedding(task.target_dictionary, _args.decoder_embed_dim) + decoder_cls = AugTransformerDecoder if aug_attn else TransformerDecoder + decoder = decoder_cls(_args, task.target_dictionary, embed_tokens) + + if getattr(args, "load_pretrained_decoder_from", None) is not None: + # load all layers first and then discard the bottom layers + embed_tokens = build_embedding( + task.target_dictionary, _args.decoder_embed_dim + ) + decoder_tmp = decoder_cls(_args, task.target_dictionary, embed_tokens) + decoder_tmp = cls.maybe_load_pretrained( + decoder_tmp, getattr(_args, "load_pretrained_decoder_from", None) + ) + state_dict = decoder_tmp.state_dict() + for k, p in decoder.named_parameters(): + p.data = state_dict[k].data + p.requires_grad = need_finetuning(_args.finetune_decoder_params, k) + decoder.layers = decoder.layers[-_args.decoder_layers :] + + return decoder, proj, _args + + @classmethod + def build_model(cls, args, task): + """Build a new model instance.""" + + # make sure all arguments are present in older models + xm_t_base_architecture(args) + + encoder = cls.build_encoder(args) + decoder, proj, unit_args = cls.build_decoder( + args, + task, + aug_attn=getattr(args, "synthesizer_augmented_cross_attention", False), + ) + base_model = cls(encoder, decoder) + setattr(base_model, "proj", proj) + + base_model.t2u_augmented_cross_attn = getattr( + args, "synthesizer_augmented_cross_attention", False + ) + + # set up multitask decoders + base_model.mt_task_name = None + base_model.multitask_decoders = {} + has_first_pass_decoder = False + for task_name, task_obj in task.multitask_tasks.items(): + if task_obj.is_first_pass_decoder: + has_first_pass_decoder = True + base_model.mt_task_name = task_name + + task_decoder = cls.build_multitask_decoder( + args, + task_obj.args, + task_obj.target_dictionary, + args.decoder_embed_dim, + task_obj.is_first_pass_decoder, + ) + + setattr(base_model, f"{task_name}_decoder", task_decoder) + decoder_model_cls = ( + FairseqEncoderModel + if task_obj.args.decoder_type == "ctc" + else FairseqLanguageModel + ) + base_model.multitask_decoders[task_name] = decoder_model_cls( + getattr(base_model, f"{task_name}_decoder") + ) + + assert has_first_pass_decoder, "set at least one intermediate non-CTC decoder" + + # set up encoder on top of the auxiliary MT decoder + if getattr(args, "synthesizer_encoder_layers", 0) > 0: + base_model.synthesizer_encoder = cls.build_t2u_encoder(unit_args) + else: + base_model.synthesizer_encoder = None + + return base_model + + @classmethod + def build_t2u_encoder(cls, args): + _args = copy.deepcopy(args) + _args.encoder_layers = _args.synthesizer_encoder_layers + _args.encoder_embed_dim = args.decoder_embed_dim + _args.encoder_ffn_embed_dim = args.decoder_ffn_embed_dim + _args.encoder_attention_heads = args.decoder_attention_heads + _args.encoder_normalize_before = True + return TransformerEncoderNoEmb(_args) + + def forward( + self, + src_tokens, + src_lengths, + prev_output_tokens, + prev_output_tokens_mt, + return_all_hiddens=False, + tgt_speaker=None, + **kwargs, + ): + """ + The forward method inherited from the base class has a **kwargs + argument in its input, which is not supported in torchscript. This + method overwrites the forward method definition without **kwargs. + """ + encoder_out = self.encoder( + src_tokens=src_tokens, src_lengths=src_lengths, **kwargs + ) + + # 1. MT decoder + mt_decoder = getattr(self, f"{self.mt_task_name}_decoder") + mt_decoder_out = mt_decoder( + prev_output_tokens_mt, + encoder_out=encoder_out, + ) + x = mt_decoder_out[1]["inner_states"][-1] + if mt_decoder.layer_norm is not None: + x = mt_decoder.layer_norm(x) + if self.proj is not None: + x = self.proj(x) + + mt_decoder_padding_mask = None + if prev_output_tokens_mt.eq(mt_decoder.padding_idx).any(): + mt_decoder_padding_mask = prev_output_tokens_mt.eq(mt_decoder.padding_idx) + + # 2. T2U encoder + if self.synthesizer_encoder is not None: + t2u_encoder_out = self.synthesizer_encoder( + x, + mt_decoder_padding_mask, + ) + else: + t2u_encoder_out = { + "encoder_out": [x], # T x B x C + "encoder_padding_mask": [mt_decoder_padding_mask], # B x T + } + + # 3. T2U decoder + if self.t2u_augmented_cross_attn: + decoder_out = self.decoder( + prev_output_tokens, + encoder_out=encoder_out, + encoder_out_aug=t2u_encoder_out, + ) + else: + decoder_out = self.decoder( + prev_output_tokens, + encoder_out=t2u_encoder_out, + ) + if return_all_hiddens: + decoder_out[-1]["encoder_states"] = encoder_out["encoder_out"] + # NOTE: from the top layer + decoder_out[-1]["encoder_padding_mask"] = encoder_out[ + "encoder_padding_mask" + ] + decoder_out[-1]["mt_decoder_out"] = mt_decoder_out + return decoder_out + + +@register_model_architecture( + model_name="unity_xm_transformer", arch_name="unity_xm_transformer" +) +def base_architecture_unity(args): + set_default_general_args(args) + set_default_w2v_encoder_args(args) + set_default_adaptor_args(args) + set_default_transformer_decoder_args(args) + + args.layernorm_embedding = False + args.decoder_learned_pos = False + + +# for old models +@register_model_architecture( + model_name="unity_xm_transformer", arch_name="xm_transformer_t2" +) +def base_architecture_unity_legacy(args): + base_architecture_unity(args) diff --git a/data/fairseq/fairseq/models/text_to_speech/__pycache__/vocoder.cpython-38.pyc b/data/fairseq/fairseq/models/text_to_speech/__pycache__/vocoder.cpython-38.pyc new file mode 100644 index 0000000000000000000000000000000000000000..7325dd42f6fc9799333a320b98f367baba282732 Binary files /dev/null and b/data/fairseq/fairseq/models/text_to_speech/__pycache__/vocoder.cpython-38.pyc differ diff --git a/data/fairseq/fairseq/models/transformer/__init__.py b/data/fairseq/fairseq/models/transformer/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..681fca3d4553f6832a65f61fc186793bc4ee0679 --- /dev/null +++ b/data/fairseq/fairseq/models/transformer/__init__.py @@ -0,0 +1,50 @@ +# Copyright (c) Facebook Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. +"""isort:skip_file""" + +from .transformer_config import ( + TransformerConfig, + DEFAULT_MAX_SOURCE_POSITIONS, + DEFAULT_MAX_TARGET_POSITIONS, + DEFAULT_MIN_PARAMS_TO_WRAP, +) +from .transformer_decoder import TransformerDecoder, TransformerDecoderBase, Linear +from .transformer_encoder import TransformerEncoder, TransformerEncoderBase +from .transformer_legacy import ( + TransformerModel, + base_architecture, + tiny_architecture, + transformer_iwslt_de_en, + transformer_wmt_en_de, + transformer_vaswani_wmt_en_de_big, + transformer_vaswani_wmt_en_fr_big, + transformer_wmt_en_de_big, + transformer_wmt_en_de_big_t2t, +) +from .transformer_base import TransformerModelBase, Embedding + + +__all__ = [ + "TransformerModelBase", + "TransformerConfig", + "TransformerDecoder", + "TransformerDecoderBase", + "TransformerEncoder", + "TransformerEncoderBase", + "TransformerModel", + "Embedding", + "Linear", + "base_architecture", + "tiny_architecture", + "transformer_iwslt_de_en", + "transformer_wmt_en_de", + "transformer_vaswani_wmt_en_de_big", + "transformer_vaswani_wmt_en_fr_big", + "transformer_wmt_en_de_big", + "transformer_wmt_en_de_big_t2t", + "DEFAULT_MAX_SOURCE_POSITIONS", + "DEFAULT_MAX_TARGET_POSITIONS", + "DEFAULT_MIN_PARAMS_TO_WRAP", +] diff --git a/data/fairseq/fairseq/models/transformer/__pycache__/__init__.cpython-38.pyc b/data/fairseq/fairseq/models/transformer/__pycache__/__init__.cpython-38.pyc new file mode 100644 index 0000000000000000000000000000000000000000..7d5f6c4f386eabe005e2115f8dbf8362a609ce8c Binary files /dev/null and b/data/fairseq/fairseq/models/transformer/__pycache__/__init__.cpython-38.pyc differ diff --git a/data/fairseq/fairseq/models/transformer/__pycache__/transformer_base.cpython-38.pyc b/data/fairseq/fairseq/models/transformer/__pycache__/transformer_base.cpython-38.pyc new file mode 100644 index 0000000000000000000000000000000000000000..4ac21cae13882de21b52978cd2fdf49c5cfda187 Binary files /dev/null and b/data/fairseq/fairseq/models/transformer/__pycache__/transformer_base.cpython-38.pyc differ diff --git a/data/fairseq/fairseq/models/transformer/__pycache__/transformer_config.cpython-38.pyc b/data/fairseq/fairseq/models/transformer/__pycache__/transformer_config.cpython-38.pyc new file mode 100644 index 0000000000000000000000000000000000000000..ad9f9f38f9422f7b43159b6f6abf6483fb97b2fe Binary files /dev/null and b/data/fairseq/fairseq/models/transformer/__pycache__/transformer_config.cpython-38.pyc differ diff --git a/data/fairseq/fairseq/models/transformer/__pycache__/transformer_decoder.cpython-38.pyc b/data/fairseq/fairseq/models/transformer/__pycache__/transformer_decoder.cpython-38.pyc new file mode 100644 index 0000000000000000000000000000000000000000..e80267e726de39a5ed3de853d3134d1fcc0587c7 Binary files /dev/null and b/data/fairseq/fairseq/models/transformer/__pycache__/transformer_decoder.cpython-38.pyc differ diff --git a/data/fairseq/fairseq/models/transformer/__pycache__/transformer_decoder_aug.cpython-38.pyc b/data/fairseq/fairseq/models/transformer/__pycache__/transformer_decoder_aug.cpython-38.pyc new file mode 100644 index 0000000000000000000000000000000000000000..14eb714c8799976e51f9ba5f91b73ab0695713c1 Binary files /dev/null and b/data/fairseq/fairseq/models/transformer/__pycache__/transformer_decoder_aug.cpython-38.pyc differ diff --git a/data/fairseq/fairseq/models/transformer/__pycache__/transformer_encoder.cpython-38.pyc b/data/fairseq/fairseq/models/transformer/__pycache__/transformer_encoder.cpython-38.pyc new file mode 100644 index 0000000000000000000000000000000000000000..eb9d84605463e59ff862d94ca389be36e211991d Binary files /dev/null and b/data/fairseq/fairseq/models/transformer/__pycache__/transformer_encoder.cpython-38.pyc differ diff --git a/data/fairseq/fairseq/models/transformer/__pycache__/transformer_legacy.cpython-38.pyc b/data/fairseq/fairseq/models/transformer/__pycache__/transformer_legacy.cpython-38.pyc new file mode 100644 index 0000000000000000000000000000000000000000..e53562d964c1c39b9869c59aa6903d0f8feacf4e Binary files /dev/null and b/data/fairseq/fairseq/models/transformer/__pycache__/transformer_legacy.cpython-38.pyc differ diff --git a/data/fairseq/fairseq/models/transformer/transformer_base.py b/data/fairseq/fairseq/models/transformer/transformer_base.py new file mode 100644 index 0000000000000000000000000000000000000000..f9f097f04b15ff9d87a7d95c2bbfd9e81a58b813 --- /dev/null +++ b/data/fairseq/fairseq/models/transformer/transformer_base.py @@ -0,0 +1,193 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from typing import Dict, List, Optional, Tuple + +import torch +import torch.nn as nn +from torch import Tensor + +import logging + +from fairseq import utils +from fairseq.dataclass.utils import gen_parser_from_dataclass +from fairseq.distributed import fsdp_wrap +from fairseq.models import FairseqEncoderDecoderModel +from fairseq.models.transformer import ( + TransformerConfig, + TransformerDecoderBase, + TransformerEncoderBase, +) + + +logger = logging.getLogger(__name__) + + +class TransformerModelBase(FairseqEncoderDecoderModel): + """ + Transformer model from `"Attention Is All You Need" (Vaswani, et al, 2017) + `_. + + Args: + encoder (TransformerEncoder): the encoder + decoder (TransformerDecoder): the decoder + + The Transformer model provides the following named architectures and + command-line arguments: + + .. argparse:: + :ref: fairseq.models.transformer_parser + :prog: + """ + + def __init__(self, cfg, encoder, decoder): + super().__init__(encoder, decoder) + self.cfg = cfg + self.supports_align_args = True + + @classmethod + def add_args(cls, parser): + """Add model-specific arguments to the parser.""" + # we want to build the args recursively in this case. + gen_parser_from_dataclass( + parser, TransformerConfig(), delete_default=False, with_prefix="" + ) + + @classmethod + def build_model(cls, cfg, task): + """Build a new model instance.""" + + # -- TODO T96535332 + # bug caused by interaction between OmegaConf II and argparsing + cfg.decoder.input_dim = int(cfg.decoder.input_dim) + cfg.decoder.output_dim = int(cfg.decoder.output_dim) + # -- + + if cfg.encoder.layers_to_keep: + cfg.encoder.layers = len(cfg.encoder.layers_to_keep.split(",")) + if cfg.decoder.layers_to_keep: + cfg.decoder.layers = len(cfg.decoder.layers_to_keep.split(",")) + + src_dict, tgt_dict = task.source_dictionary, task.target_dictionary + + if cfg.share_all_embeddings: + if src_dict != tgt_dict: + raise ValueError("--share-all-embeddings requires a joined dictionary") + if cfg.encoder.embed_dim != cfg.decoder.embed_dim: + raise ValueError( + "--share-all-embeddings requires --encoder-embed-dim to match --decoder-embed-dim" + ) + if cfg.decoder.embed_path and ( + cfg.decoder.embed_path != cfg.encoder.embed_path + ): + raise ValueError( + "--share-all-embeddings not compatible with --decoder-embed-path" + ) + encoder_embed_tokens = cls.build_embedding( + cfg, src_dict, cfg.encoder.embed_dim, cfg.encoder.embed_path + ) + decoder_embed_tokens = encoder_embed_tokens + cfg.share_decoder_input_output_embed = True + elif cfg.merge_src_tgt_embed: + logger.info(f"source dict size: {len(src_dict)}") + logger.info(f"target dict size: {len(tgt_dict)}") + src_dict.update(tgt_dict) + task.src_dict = src_dict + task.tgt_dict = src_dict + logger.info(f"merged dict size: {len(src_dict)}") + encoder_embed_tokens = cls.build_embedding( + cfg, src_dict, cfg.encoder.embed_dim + ) + decoder_embed_tokens = encoder_embed_tokens + cfg.share_decoder_input_output_embed = True + else: + encoder_embed_tokens = cls.build_embedding( + cfg, src_dict, cfg.encoder.embed_dim, cfg.encoder.embed_path + ) + decoder_embed_tokens = cls.build_embedding( + cfg, tgt_dict, cfg.decoder.embed_dim, cfg.decoder.embed_path + ) + if cfg.offload_activations: + cfg.checkpoint_activations = True # offloading implies checkpointing + encoder = cls.build_encoder(cfg, src_dict, encoder_embed_tokens) + decoder = cls.build_decoder(cfg, tgt_dict, decoder_embed_tokens) + return cls(cfg, encoder, decoder) + + @classmethod + def build_embedding(cls, cfg, dictionary, embed_dim, path=None): + num_embeddings = len(dictionary) + padding_idx = dictionary.pad() + + emb = Embedding(num_embeddings, embed_dim, padding_idx) + # if provided, load from preloaded dictionaries + if path: + embed_dict = utils.parse_embedding(path) + utils.load_embedding(embed_dict, dictionary, emb) + return emb + + @classmethod + def build_encoder(cls, cfg, src_dict, embed_tokens): + return TransformerEncoderBase(cfg, src_dict, embed_tokens) + + @classmethod + def build_decoder(cls, cfg, tgt_dict, embed_tokens): + return TransformerDecoderBase( + cfg, + tgt_dict, + embed_tokens, + no_encoder_attn=cfg.no_cross_attention, + ) + + # TorchScript doesn't support optional arguments with variable length (**kwargs). + # Current workaround is to add union of all arguments in child classes. + def forward( + self, + src_tokens, + src_lengths, + prev_output_tokens, + return_all_hiddens: bool = True, + features_only: bool = False, + alignment_layer: Optional[int] = None, + alignment_heads: Optional[int] = None, + ): + """ + Run the forward pass for an encoder-decoder model. + + Copied from the base class, but without ``**kwargs``, + which are not supported by TorchScript. + """ + encoder_out = self.encoder( + src_tokens, src_lengths=src_lengths, return_all_hiddens=return_all_hiddens + ) + decoder_out = self.decoder( + prev_output_tokens, + encoder_out=encoder_out, + features_only=features_only, + alignment_layer=alignment_layer, + alignment_heads=alignment_heads, + src_lengths=src_lengths, + return_all_hiddens=return_all_hiddens, + ) + return decoder_out + + # Since get_normalized_probs is in the Fairseq Model which is not scriptable, + # I rewrite the get_normalized_probs from Base Class to call the + # helper function in the Base Class. + @torch.jit.export + def get_normalized_probs( + self, + net_output: Tuple[Tensor, Optional[Dict[str, List[Optional[Tensor]]]]], + log_probs: bool, + sample: Optional[Dict[str, Tensor]] = None, + ): + """Get normalized probabilities (or log probs) from a net's output.""" + return self.get_normalized_probs_scriptable(net_output, log_probs, sample) + + +def Embedding(num_embeddings, embedding_dim, padding_idx): + m = nn.Embedding(num_embeddings, embedding_dim, padding_idx=padding_idx) + nn.init.normal_(m.weight, mean=0, std=embedding_dim**-0.5) + nn.init.constant_(m.weight[padding_idx], 0) + return m diff --git a/data/fairseq/fairseq/models/transformer/transformer_config.py b/data/fairseq/fairseq/models/transformer/transformer_config.py new file mode 100644 index 0000000000000000000000000000000000000000..4650de2e1710a7bb790431dea855a82ac9b49def --- /dev/null +++ b/data/fairseq/fairseq/models/transformer/transformer_config.py @@ -0,0 +1,341 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + + +import re +from dataclasses import dataclass, field, fields +from typing import List, Optional + +from omegaconf import II + +from fairseq import utils +from fairseq.dataclass import ChoiceEnum, FairseqDataclass +from fairseq.utils import safe_getattr, safe_hasattr + +DEFAULT_MAX_SOURCE_POSITIONS = 1024 +DEFAULT_MAX_TARGET_POSITIONS = 1024 + +DEFAULT_MIN_PARAMS_TO_WRAP = int(1e8) + +_NAME_PARSER = r"(decoder|encoder|quant_noise)_(.*)" + + +@dataclass +class EncDecBaseConfig(FairseqDataclass): + embed_path: Optional[str] = field( + default=None, metadata={"help": "path to pre-trained embedding"} + ) + embed_dim: Optional[int] = field( + default=512, metadata={"help": "embedding dimension"} + ) + ffn_embed_dim: int = field( + default=2048, metadata={"help": "embedding dimension for FFN"} + ) + layers: int = field(default=6, metadata={"help": "number of layers"}) + attention_heads: int = field( + default=8, metadata={"help": "number of attention heads"} + ) + normalize_before: bool = field( + default=False, metadata={"help": "apply layernorm before each block"} + ) + learned_pos: bool = field( + default=False, metadata={"help": "use learned positional embeddings"} + ) + # args for "Reducing Transformer Depth on Demand with Structured Dropout" (Fan et al., 2019) + layerdrop: float = field(default=0, metadata={"help": "LayerDrop probability"}) + layers_to_keep: Optional[List[int]] = field( + default=None, metadata={"help": "which layers to *keep* when pruning"} + ) + + xformers_att_config: Optional[str] = field( + default=None, + metadata={ + "help": "config for xFormers attention, defined in xformers.components.attention.AttentionConfig" + }, + ) + + +@dataclass +class DecoderConfig(EncDecBaseConfig): + input_dim: int = II("model.decoder.embed_dim") + output_dim: int = field( + default=II("model.decoder.embed_dim"), + metadata={ + "help": "decoder output dimension (extra linear layer if different from decoder embed dim)" + }, + ) + + def __post_init__(self): + # II doesn't work if we are just creating the object outside of hydra so fix that + if self.input_dim == II("model.decoder.embed_dim"): + self.input_dim = self.embed_dim + if self.output_dim == II("model.decoder.embed_dim"): + self.output_dim = self.embed_dim + + +@dataclass +class QuantNoiseConfig(FairseqDataclass): + pq: float = field( + default=0.0, + metadata={"help": "iterative PQ quantization noise at training time"}, + ) + pq_block_size: int = field( + default=8, + metadata={"help": "block size of quantization noise at training time"}, + ) + scalar: float = field( + default=0.0, + metadata={ + "help": "scalar quantization noise and scalar quantization at training time" + }, + ) + + +@dataclass +class TransformerConfig(FairseqDataclass): + activation_fn: ChoiceEnum(utils.get_available_activation_fns()) = field( + default="relu", + metadata={"help": "activation function to use"}, + ) + dropout: float = field(default=0.1, metadata={"help": "dropout probability"}) + attention_dropout: float = field( + default=0.0, metadata={"help": "dropout probability for attention weights"} + ) + activation_dropout: float = field( + default=0.0, + metadata={ + "help": "dropout probability after activation in FFN.", + "alias": "--relu-dropout", + }, + ) + adaptive_input: bool = False + encoder: EncDecBaseConfig = EncDecBaseConfig() + # TODO should really be in the encoder config + max_source_positions: int = field( + default=DEFAULT_MAX_SOURCE_POSITIONS, + metadata={"help": "Maximum input length supported by the encoder"}, + ) + decoder: DecoderConfig = DecoderConfig() + # TODO should really be in the decoder config + max_target_positions: int = field( + default=DEFAULT_MAX_TARGET_POSITIONS, + metadata={"help": "Maximum output length supported by the decoder"}, + ) + share_decoder_input_output_embed: bool = field( + default=False, metadata={"help": "share decoder input and output embeddings"} + ) + share_all_embeddings: bool = field( + default=False, + metadata={ + "help": "share encoder, decoder and output embeddings (requires shared dictionary and embed dim)" + }, + ) + merge_src_tgt_embed: bool = field( + default=False, + metadata={ + "help": "if true then the source and target embedding table is " + "merged into one table. This is going to make the model smaller but " + "it might hurt performance." + }, + ) + no_token_positional_embeddings: bool = field( + default=False, + metadata={ + "help": "if True, disables positional embeddings (outside self attention)" + }, + ) + adaptive_softmax_cutoff: Optional[List[int]] = field( + default=None, + metadata={ + "help": "list of adaptive softmax cutoff points. Must be used with adaptive_loss criterion" + }, + ) + adaptive_softmax_dropout: float = field( + default=0.0, + metadata={"help": "sets adaptive softmax dropout for the tail projections"}, + ) + adaptive_softmax_factor: float = field( + default=4, metadata={"help": "adaptive input factor"} + ) + layernorm_embedding: bool = field( + default=False, metadata={"help": "add layernorm to embedding"} + ) + tie_adaptive_weights: bool = field( + default=False, + metadata={ + "help": "if set, ties the weights of adaptive softmax and adaptive input" + }, + ) + tie_adaptive_proj: bool = field( + default=False, + metadata={ + "help": "if set, ties the projection weights of adaptive softmax and adaptive input" + }, + ) + no_scale_embedding: bool = field( + default=False, metadata={"help": "if True, dont scale embeddings"} + ) + checkpoint_activations: bool = field( + default=False, + metadata={ + "help": "checkpoint activations at each layer, which saves GPU memory usage at the cost of some additional compute" + }, + ) + offload_activations: bool = field( + default=False, + metadata={ + "help": "checkpoint activations at each layer, then save to gpu. Sets --checkpoint-activations." + }, + ) + # args for "Cross+Self-Attention for Transformer Models" (Peitz et al., 2019) + no_cross_attention: bool = field( + default=False, metadata={"help": "do not perform cross-attention"} + ) + cross_self_attention: bool = field( + default=False, metadata={"help": "perform cross+self-attention"} + ) + # args for Training with Quantization Noise for Extreme Model Compression ({Fan*, Stock*} et al., 2020) + quant_noise: QuantNoiseConfig = field(default=QuantNoiseConfig()) + min_params_to_wrap: int = field( + default=DEFAULT_MIN_PARAMS_TO_WRAP, + metadata={ + "help": "minimum number of params for a layer to be wrapped with FSDP() when " + "training with --ddp-backend=fully_sharded. Smaller values will " + "improve memory efficiency, but may make torch.distributed " + "communication less efficient due to smaller input sizes. This option " + "is set to 0 (i.e., always wrap) when --checkpoint-activations or " + "--offload-activations are passed." + }, + ) + # DEPRECATED field, but some old checkpoints might have it + char_inputs: bool = field( + default=False, metadata={"help": "if set, model takes character ids as input"} + ) + relu_dropout: float = 0.0 + # config for "BASE Layers: Simplifying Training of Large, Sparse Models" + base_layers: Optional[int] = field( + default=0, metadata={"help": "number of BASE layers in total"} + ) + base_sublayers: Optional[int] = field( + default=1, metadata={"help": "number of sublayers in each BASE layer"} + ) + base_shuffle: Optional[int] = field( + default=1, + metadata={"help": "shuffle tokens between workers before computing assignment"}, + ) + + export: bool = field( + default=False, + metadata={"help": "make the layernorm exportable with torchscript."}, + ) + + # copied from transformer_lm but expected in transformer_decoder: + no_decoder_final_norm: bool = field( + default=False, + metadata={"help": "don't add an extra layernorm after the last decoder block"}, + ) + + # We need to make this hierarchical dataclass like the flat namespace + # __getattr__ and __setattr__ here allow backward compatibility + # for subclasses of Transformer(Legacy) that depend on read/write on + # the flat namespace. + + def __getattr__(self, name): + match = re.match(_NAME_PARSER, name) + if match: + sub = safe_getattr(self, match[1]) + return safe_getattr(sub, match[2]) + raise AttributeError(f"invalid argument {name}.") + + def __setattr__(self, name, value): + match = re.match(_NAME_PARSER, name) + if match: + sub = safe_getattr(self, match[1]) + setattr(sub, match[2], value) + else: + super().__setattr__(name, value) + + @staticmethod + def _copy_keys(args, cls, prefix, seen): + """ + copy the prefixed keys (decoder_embed_dim) to the DC fields: decoder.embed_dim + """ + cfg = cls() + for fld in fields(cls): + # for all the fields in the DC, find the fields (e.g. embed_dim) + # in the namespace with the prefix (e.g. decoder) + # and set it on the dc. + args_key = f"{prefix}_{fld.name}" + if safe_hasattr(args, args_key): + seen.add(args_key) + setattr(cfg, fld.name, safe_getattr(args, args_key)) + if safe_hasattr(args, fld.name): + seen.add(fld.name) + setattr(cfg, fld.name, safe_getattr(args, fld.name)) + return cfg + + @classmethod + def from_namespace(cls, args): + if args is None: + return None + if not isinstance(args, cls): + seen = set() + config = cls() + # currently, we can go generically from DC fields to args hierarchically + # but we can't easily deconstruct a flat namespace to a hierarchical + # DC. Mostly because we could have a sub-dc called `decoder-foo` that should not + # go to the sub struct called `decoder`. There are ways to go around this, but let's keep it simple + # for now. + for fld in fields(cls): + # concretelly, the transformer_config know what sub-dc it has, so we go through all the dc fields + # and if it's one that has a sub-dc, we build that sub-dc with `copy_keys()` + if fld.name == "decoder": + if safe_hasattr(args, "decoder"): + # in some cases, the args we receive is already structured (as DictConfigs), so let's just build the correct DC + seen.add("decoder") + config.decoder = DecoderConfig(**args.decoder) + else: + config.decoder = cls._copy_keys( + args, DecoderConfig, "decoder", seen + ) + elif fld.name == "encoder": + # same but for encoder + if safe_hasattr(args, "encoder"): + seen.add("encoder") + config.encoder = EncDecBaseConfig(**args.encoder) + else: + config.encoder = cls._copy_keys( + args, EncDecBaseConfig, "encoder", seen + ) + elif fld.name == "quant_noise": + # same but for quant_noise + if safe_hasattr(args, "quant_noise"): + seen.add("quant_noise") + config.quant_noise = QuantNoiseConfig(**args.quant_noise) + else: + config.quant_noise = cls._copy_keys( + args, QuantNoiseConfig, "quant_noise", seen + ) + elif safe_hasattr(args, fld.name): + # if it's not a structure field, it's just a normal field, copy it over + seen.add(fld.name) + setattr(config, fld.name, safe_getattr(args, fld.name)) + # we got all the fields defined in the dataclass, but + # the argparse namespace might have extra args for two reasons: + # - we are in a legacy class so all the args are not declared in the dataclass. Ideally once everyone has defined a dataclass for their model, we won't need this + # - some places expect args to be there but never define them + args_dict = ( + args._asdict() + if safe_hasattr(args, "_asdict") + else vars(args) + if safe_hasattr(args, "__dict__") + else {} + ) # namedtupled doesn't have __dict__ :-/ + for key, value in args_dict.items(): + if key not in seen: + setattr(config, key, value) + return config + else: + return args diff --git a/data/fairseq/fairseq/models/transformer/transformer_decoder.py b/data/fairseq/fairseq/models/transformer/transformer_decoder.py new file mode 100644 index 0000000000000000000000000000000000000000..744c73f4f8cda6f0e9d6899d146ca1343d514b45 --- /dev/null +++ b/data/fairseq/fairseq/models/transformer/transformer_decoder.py @@ -0,0 +1,474 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import math +from typing import Any, Dict, List, Optional + +import torch +import torch.nn as nn +from torch import Tensor + +from fairseq import utils +from fairseq.distributed import fsdp_wrap +from fairseq.models import FairseqIncrementalDecoder +from fairseq.models.transformer import TransformerConfig +from fairseq.modules import ( + AdaptiveSoftmax, + BaseLayer, + FairseqDropout, + LayerDropModuleList, + LayerNorm, + PositionalEmbedding, + SinusoidalPositionalEmbedding, + transformer_layer, +) +from fairseq.modules.checkpoint_activations import checkpoint_wrapper +from fairseq.modules.quant_noise import quant_noise as apply_quant_noise_ + + +# rewrite name for backward compatibility in `make_generation_fast_` +def module_name_fordropout(module_name: str) -> str: + if module_name == "TransformerDecoderBase": + return "TransformerDecoder" + else: + return module_name + + +class TransformerDecoderBase(FairseqIncrementalDecoder): + """ + Transformer decoder consisting of *cfg.decoder.layers* layers. Each layer + is a :class:`TransformerDecoderLayer`. + + Args: + cfg (argparse.Namespace): parsed command-line arguments + dictionary (~fairseq.data.Dictionary): decoding dictionary + embed_tokens (torch.nn.Embedding): output embedding + no_encoder_attn (bool, optional): whether to attend to encoder outputs + (default: False). + """ + + def __init__( + self, + cfg, + dictionary, + embed_tokens, + no_encoder_attn=False, + output_projection=None, + ): + self.cfg = cfg + super().__init__(dictionary) + self.register_buffer("version", torch.Tensor([3])) + self._future_mask = torch.empty(0) + + self.dropout_module = FairseqDropout( + cfg.dropout, module_name=module_name_fordropout(self.__class__.__name__) + ) + self.decoder_layerdrop = cfg.decoder.layerdrop + self.share_input_output_embed = cfg.share_decoder_input_output_embed + + input_embed_dim = embed_tokens.embedding_dim + embed_dim = cfg.decoder.embed_dim + self.embed_dim = embed_dim + self.output_embed_dim = cfg.decoder.output_dim + + self.padding_idx = embed_tokens.padding_idx + self.max_target_positions = cfg.max_target_positions + + self.embed_tokens = embed_tokens + + self.embed_scale = 1.0 if cfg.no_scale_embedding else math.sqrt(embed_dim) + + if not cfg.adaptive_input and cfg.quant_noise.pq > 0: + self.quant_noise = apply_quant_noise_( + nn.Linear(embed_dim, embed_dim, bias=False), + cfg.quant_noise.pq, + cfg.quant_noise.pq_block_size, + ) + else: + self.quant_noise = None + + self.project_in_dim = ( + Linear(input_embed_dim, embed_dim, bias=False) + if embed_dim != input_embed_dim + else None + ) + self.embed_positions = ( + PositionalEmbedding( + self.max_target_positions, + embed_dim, + self.padding_idx, + learned=cfg.decoder.learned_pos, + ) + if not cfg.no_token_positional_embeddings + else None + ) + if cfg.layernorm_embedding: + self.layernorm_embedding = LayerNorm(embed_dim, export=cfg.export) + else: + self.layernorm_embedding = None + + self.cross_self_attention = cfg.cross_self_attention + + if self.decoder_layerdrop > 0.0: + self.layers = LayerDropModuleList(p=self.decoder_layerdrop) + else: + self.layers = nn.ModuleList([]) + self.layers.extend( + [ + self.build_decoder_layer(cfg, no_encoder_attn) + for _ in range(cfg.decoder.layers) + ] + ) + self.num_layers = len(self.layers) + + if cfg.decoder.normalize_before and not cfg.no_decoder_final_norm: + self.layer_norm = LayerNorm(embed_dim, export=cfg.export) + else: + self.layer_norm = None + + self.project_out_dim = ( + Linear(embed_dim, self.output_embed_dim, bias=False) + if embed_dim != self.output_embed_dim and not cfg.tie_adaptive_weights + else None + ) + + self.adaptive_softmax = None + self.output_projection = output_projection + if self.output_projection is None: + self.build_output_projection(cfg, dictionary, embed_tokens) + + def build_output_projection(self, cfg, dictionary, embed_tokens): + if cfg.adaptive_softmax_cutoff is not None: + self.adaptive_softmax = AdaptiveSoftmax( + len(dictionary), + self.output_embed_dim, + utils.eval_str_list(cfg.adaptive_softmax_cutoff, type=int), + dropout=cfg.adaptive_softmax_dropout, + adaptive_inputs=embed_tokens if cfg.tie_adaptive_weights else None, + factor=cfg.adaptive_softmax_factor, + tie_proj=cfg.tie_adaptive_proj, + ) + elif self.share_input_output_embed: + self.output_projection = nn.Linear( + self.embed_tokens.weight.shape[1], + self.embed_tokens.weight.shape[0], + bias=False, + ) + self.output_projection.weight = self.embed_tokens.weight + else: + self.output_projection = nn.Linear( + self.output_embed_dim, len(dictionary), bias=False + ) + nn.init.normal_( + self.output_projection.weight, mean=0, std=self.output_embed_dim**-0.5 + ) + num_base_layers = cfg.base_layers + for i in range(num_base_layers): + self.layers.insert( + ((i + 1) * cfg.decoder.layers) // (num_base_layers + 1), + BaseLayer(cfg), + ) + + def build_decoder_layer(self, cfg, no_encoder_attn=False): + layer = transformer_layer.TransformerDecoderLayerBase(cfg, no_encoder_attn) + checkpoint = cfg.checkpoint_activations + if checkpoint: + offload_to_cpu = cfg.offload_activations + layer = checkpoint_wrapper(layer, offload_to_cpu=offload_to_cpu) + # if we are checkpointing, enforce that FSDP always wraps the + # checkpointed layer, regardless of layer size + min_params_to_wrap = cfg.min_params_to_wrap if not checkpoint else 0 + layer = fsdp_wrap(layer, min_num_params=min_params_to_wrap) + return layer + + def forward( + self, + prev_output_tokens, + encoder_out: Optional[Dict[str, List[Tensor]]] = None, + incremental_state: Optional[Dict[str, Dict[str, Optional[Tensor]]]] = None, + features_only: bool = False, + full_context_alignment: bool = False, + alignment_layer: Optional[int] = None, + alignment_heads: Optional[int] = None, + src_lengths: Optional[Any] = None, + return_all_hiddens: bool = False, + ): + """ + Args: + prev_output_tokens (LongTensor): previous decoder outputs of shape + `(batch, tgt_len)`, for teacher forcing + encoder_out (optional): output from the encoder, used for + encoder-side attention, should be of size T x B x C + incremental_state (dict): dictionary used for storing state during + :ref:`Incremental decoding` + features_only (bool, optional): only return features without + applying output layer (default: False). + full_context_alignment (bool, optional): don't apply + auto-regressive mask to self-attention (default: False). + + Returns: + tuple: + - the decoder's output of shape `(batch, tgt_len, vocab)` + - a dictionary with any model-specific outputs + """ + + x, extra = self.extract_features( + prev_output_tokens, + encoder_out=encoder_out, + incremental_state=incremental_state, + full_context_alignment=full_context_alignment, + alignment_layer=alignment_layer, + alignment_heads=alignment_heads, + ) + + if not features_only: + x = self.output_layer(x) + return x, extra + + def extract_features( + self, + prev_output_tokens, + encoder_out: Optional[Dict[str, List[Tensor]]], + incremental_state: Optional[Dict[str, Dict[str, Optional[Tensor]]]] = None, + full_context_alignment: bool = False, + alignment_layer: Optional[int] = None, + alignment_heads: Optional[int] = None, + ): + return self.extract_features_scriptable( + prev_output_tokens, + encoder_out, + incremental_state, + full_context_alignment, + alignment_layer, + alignment_heads, + ) + + """ + A scriptable subclass of this class has an extract_features method and calls + super().extract_features, but super() is not supported in torchscript. A copy of + this function is made to be used in the subclass instead. + """ + + def extract_features_scriptable( + self, + prev_output_tokens, + encoder_out: Optional[Dict[str, List[Tensor]]], + incremental_state: Optional[Dict[str, Dict[str, Optional[Tensor]]]] = None, + full_context_alignment: bool = False, + alignment_layer: Optional[int] = None, + alignment_heads: Optional[int] = None, + ): + """ + Similar to *forward* but only return features. + + Includes several features from "Jointly Learning to Align and + Translate with Transformer Models" (Garg et al., EMNLP 2019). + + Args: + full_context_alignment (bool, optional): don't apply + auto-regressive mask to self-attention (default: False). + alignment_layer (int, optional): return mean alignment over + heads at this layer (default: last layer). + alignment_heads (int, optional): only average alignment over + this many heads (default: all heads). + + Returns: + tuple: + - the decoder's features of shape `(batch, tgt_len, embed_dim)` + - a dictionary with any model-specific outputs + """ + bs, slen = prev_output_tokens.size() + if alignment_layer is None: + alignment_layer = self.num_layers - 1 + + enc: Optional[Tensor] = None + padding_mask: Optional[Tensor] = None + if encoder_out is not None and len(encoder_out["encoder_out"]) > 0: + enc = encoder_out["encoder_out"][0] + if encoder_out is not None and len(encoder_out["encoder_padding_mask"]) > 0: + padding_mask = encoder_out["encoder_padding_mask"][0] + + # embed positions + positions = None + if self.embed_positions is not None: + positions = self.embed_positions( + prev_output_tokens, incremental_state=incremental_state + ) + + if incremental_state is not None: + prev_output_tokens = prev_output_tokens[:, -1:] + if positions is not None: + positions = positions[:, -1:] + + # Prevent torchscript exporting issue for dynamic quant embedding + prev_output_tokens = prev_output_tokens.contiguous() + # embed tokens and positions + x = self.embed_scale * self.embed_tokens(prev_output_tokens) + + if self.quant_noise is not None: + x = self.quant_noise(x) + + if self.project_in_dim is not None: + x = self.project_in_dim(x) + + if positions is not None: + x += positions + + if self.layernorm_embedding is not None: + x = self.layernorm_embedding(x) + + x = self.dropout_module(x) + + # B x T x C -> T x B x C + x = x.transpose(0, 1) + + self_attn_padding_mask: Optional[Tensor] = None + if self.cross_self_attention or prev_output_tokens.eq(self.padding_idx).any(): + self_attn_padding_mask = prev_output_tokens.eq(self.padding_idx) + + # decoder layers + attn: Optional[Tensor] = None + inner_states: List[Optional[Tensor]] = [x] + for idx, layer in enumerate(self.layers): + if incremental_state is None and not full_context_alignment: + self_attn_mask = self.buffered_future_mask(x) + else: + self_attn_mask = None + + x, layer_attn, _ = layer( + x, + enc, + padding_mask, + incremental_state, + self_attn_mask=self_attn_mask, + self_attn_padding_mask=self_attn_padding_mask, + need_attn=bool((idx == alignment_layer)), + need_head_weights=bool((idx == alignment_layer)), + ) + inner_states.append(x) + if layer_attn is not None and idx == alignment_layer: + attn = layer_attn.float().to(x) + + if attn is not None: + if alignment_heads is not None: + attn = attn[:alignment_heads] + + # average probabilities over heads + attn = attn.mean(dim=0) + + if self.layer_norm is not None: + x = self.layer_norm(x) + + # T x B x C -> B x T x C + x = x.transpose(0, 1) + + if self.project_out_dim is not None: + x = self.project_out_dim(x) + + return x, {"attn": [attn], "inner_states": inner_states} + + def output_layer(self, features): + """Project features to the vocabulary size.""" + if self.adaptive_softmax is None: + # project back to size of vocabulary + return self.output_projection(features) + else: + return features + + def max_positions(self): + """Maximum output length supported by the decoder.""" + if self.embed_positions is None: + return self.max_target_positions + return min(self.max_target_positions, self.embed_positions.max_positions) + + def buffered_future_mask(self, tensor): + dim = tensor.size(0) + # self._future_mask.device != tensor.device is not working in TorchScript. This is a workaround. + if ( + self._future_mask.size(0) == 0 + or (not self._future_mask.device == tensor.device) + or self._future_mask.size(0) < dim + ): + self._future_mask = torch.triu( + utils.fill_with_neg_inf(torch.zeros([dim, dim])), 1 + ) + self._future_mask = self._future_mask.to(tensor) + return self._future_mask[:dim, :dim] + + def upgrade_state_dict_named(self, state_dict, name): + """Upgrade a (possibly old) state dict for new versions of fairseq.""" + if f"{name}.output_projection.weight" not in state_dict: + if self.share_input_output_embed: + embed_out_key = f"{name}.embed_tokens.weight" + else: + embed_out_key = f"{name}.embed_out" + if embed_out_key in state_dict: + state_dict[f"{name}.output_projection.weight"] = state_dict[ + embed_out_key + ] + if not self.share_input_output_embed: + del state_dict[embed_out_key] + + for i in range(self.num_layers): + # update layer norms + layer_norm_map = { + "0": "self_attn_layer_norm", + "1": "encoder_attn_layer_norm", + "2": "final_layer_norm", + } + for old, new in layer_norm_map.items(): + for m in ("weight", "bias"): + k = "{}.layers.{}.layer_norms.{}.{}".format(name, i, old, m) + if k in state_dict: + state_dict[ + "{}.layers.{}.{}.{}".format(name, i, new, m) + ] = state_dict[k] + del state_dict[k] + + version_key = "{}.version".format(name) + if utils.item(state_dict.get(version_key, torch.Tensor([1]))[0]) <= 2: + # earlier checkpoints did not normalize after the stack of layers + self.layer_norm = None + self.normalize = False + state_dict[version_key] = torch.Tensor([1]) + + return state_dict + + +def Linear(in_features, out_features, bias=True): + m = nn.Linear(in_features, out_features, bias) + nn.init.xavier_uniform_(m.weight) + if bias: + nn.init.constant_(m.bias, 0.0) + return m + + +class TransformerDecoder(TransformerDecoderBase): + def __init__( + self, + args, + dictionary, + embed_tokens, + no_encoder_attn=False, + output_projection=None, + ): + self.args = args + super().__init__( + TransformerConfig.from_namespace(args), + dictionary, + embed_tokens, + no_encoder_attn=no_encoder_attn, + output_projection=output_projection, + ) + + def build_output_projection(self, args, dictionary, embed_tokens): + super().build_output_projection( + TransformerConfig.from_namespace(args), dictionary, embed_tokens + ) + + def build_decoder_layer(self, args, no_encoder_attn=False): + return super().build_decoder_layer( + TransformerConfig.from_namespace(args), no_encoder_attn=no_encoder_attn + ) diff --git a/data/fairseq/fairseq/models/transformer/transformer_decoder_aug.py b/data/fairseq/fairseq/models/transformer/transformer_decoder_aug.py new file mode 100644 index 0000000000000000000000000000000000000000..b73c06e02a0a9749120e50ce6e9f65c35a0462dc --- /dev/null +++ b/data/fairseq/fairseq/models/transformer/transformer_decoder_aug.py @@ -0,0 +1,384 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from typing import Any, Dict, List, Optional + +import torch +import torch.nn as nn +from torch import Tensor + +from fairseq import utils +from fairseq.distributed import fsdp_wrap +from fairseq.models.transformer import TransformerConfig +from fairseq.models.transformer.transformer_decoder import TransformerDecoderBase +from fairseq.modules import ( + LayerDropModuleList, + SinusoidalPositionalEmbedding, + transformer_layer_aug, +) +from fairseq.modules.checkpoint_activations import checkpoint_wrapper + + +class AugTransformerDecoderBase(TransformerDecoderBase): + """ + Transformer decoder augmented with an additional cross-attention. Each layer + is a :class:`AugTransformerDecoderLayerBase`. + + Args: + cfg (argparse.Namespace): parsed command-line arguments + dictionary (~fairseq.data.Dictionary): decoding dictionary + embed_tokens (torch.nn.Embedding): output embedding + encoder_attn_merge_type (str, optional): the way to combine outputs from + two cross-attention modules. If "sequential" is set, two cross-attention + modules are stacked sequentially. If "parallel" is set, they are processed + in parallel and combined before feeding it to FFN (default: sequential). + dropnet_ratio (float, optional): a probability to drop each cross-attention + module during training (default: 0.0). + """ + + def __init__( + self, + cfg, + dictionary, + embed_tokens, + output_projection=None, + encoder_attn_merge_type="sequential", + dropnet_ratio=0.0, + ): + super().__init__( + cfg, + dictionary, + embed_tokens, + no_encoder_attn=False, + output_projection=output_projection, + ) + # assert cfg.cross_self_attention + self.cross_self_attention = cfg.cross_self_attention + + if self.decoder_layerdrop > 0.0: + self.layers = LayerDropModuleList(p=self.decoder_layerdrop) + else: + self.layers = nn.ModuleList([]) + self.layers.extend( + [ + self.build_decoder_layer(cfg, encoder_attn_merge_type, dropnet_ratio) + for _ in range(cfg.decoder.layers) + ] + ) + + def build_decoder_layer( + self, + cfg, + encoder_attn_merge_type="sequential", + dropnet_ratio=0, + ): + layer = transformer_layer_aug.AugTransformerDecoderLayerBase( + cfg, + no_encoder_attn=False, + encoder_attn_merge_type=encoder_attn_merge_type, + dropnet_ratio=dropnet_ratio, + ) + checkpoint = cfg.checkpoint_activations + if checkpoint: + offload_to_cpu = cfg.offload_activations + layer = checkpoint_wrapper(layer, offload_to_cpu=offload_to_cpu) + # if we are checkpointing, enforce that FSDP always wraps the + # checkpointed layer, regardless of layer size + min_params_to_wrap = cfg.min_params_to_wrap if not checkpoint else 0 + layer = fsdp_wrap(layer, min_num_params=min_params_to_wrap) + return layer + + def forward( + self, + prev_output_tokens, + encoder_out: Optional[Dict[str, List[Tensor]]] = None, + encoder_out_aug: Optional[Dict[str, List[Tensor]]] = None, + incremental_state: Optional[Dict[str, Dict[str, Optional[Tensor]]]] = None, + features_only: bool = False, + full_context_alignment: bool = False, + alignment_layer: Optional[int] = None, + alignment_heads: Optional[int] = None, + src_lengths: Optional[Any] = None, + return_all_hiddens: bool = False, + ): + """ + Args: + prev_output_tokens (LongTensor): previous decoder outputs of shape + `(batch, tgt_len)`, for teacher forcing + encoder_out (optional): output from the encoder, used for + encoder-side attention, should be of size T x B x C + incremental_state (dict): dictionary used for storing state during + :ref:`Incremental decoding` + features_only (bool, optional): only return features without + applying output layer (default: False). + full_context_alignment (bool, optional): don't apply + auto-regressive mask to self-attention (default: False). + + Returns: + tuple: + - the decoder's output of shape `(batch, tgt_len, vocab)` + - a dictionary with any model-specific outputs + """ + + x, extra = self.extract_features( + prev_output_tokens, + encoder_out=encoder_out, + encoder_out_aug=encoder_out_aug, + incremental_state=incremental_state, + full_context_alignment=full_context_alignment, + alignment_layer=alignment_layer, + alignment_heads=alignment_heads, + ) + + if not features_only: + x = self.output_layer(x) + return x, extra + + def extract_features( + self, + prev_output_tokens, + encoder_out: Optional[Dict[str, List[Tensor]]], + encoder_out_aug: Optional[Dict[str, List[Tensor]]], + incremental_state: Optional[Dict[str, Dict[str, Optional[Tensor]]]] = None, + full_context_alignment: bool = False, + alignment_layer: Optional[int] = None, + alignment_heads: Optional[int] = None, + ): + return self.extract_features_scriptable( + prev_output_tokens, + encoder_out, + encoder_out_aug, + incremental_state, + full_context_alignment, + alignment_layer, + alignment_heads, + ) + + """ + A scriptable subclass of this class has an extract_features method and calls + super().extract_features, but super() is not supported in torchscript. A copy of + this function is made to be used in the subclass instead. + """ + + def extract_features_scriptable( + self, + prev_output_tokens, + encoder_out: Optional[Dict[str, List[Tensor]]], + encoder_out_aug: Optional[Dict[str, List[Tensor]]], + incremental_state: Optional[Dict[str, Dict[str, Optional[Tensor]]]] = None, + full_context_alignment: bool = False, + alignment_layer: Optional[int] = None, + alignment_heads: Optional[int] = None, + ): + """ + Similar to *forward* but only return features. + + Includes several features from "Jointly Learning to Align and + Translate with Transformer Models" (Garg et al., EMNLP 2019). + + Args: + full_context_alignment (bool, optional): don't apply + auto-regressive mask to self-attention (default: False). + alignment_layer (int, optional): return mean alignment over + heads at this layer (default: last layer). + alignment_heads (int, optional): only average alignment over + this many heads (default: all heads). + + Returns: + tuple: + - the decoder's features of shape `(batch, tgt_len, embed_dim)` + - a dictionary with any model-specific outputs + """ + bs, slen = prev_output_tokens.size() + if alignment_layer is None: + alignment_layer = self.num_layers - 1 + + enc: Optional[Tensor] = None + padding_mask: Optional[Tensor] = None + if encoder_out is not None and len(encoder_out["encoder_out"]) > 0: + enc = encoder_out["encoder_out"][0] + if encoder_out is not None and len(encoder_out["encoder_padding_mask"]) > 0: + padding_mask = encoder_out["encoder_padding_mask"][0] + + enc_aug: Optional[Tensor] = None + padding_mask_aug: Optional[Tensor] = None + if encoder_out_aug is not None and len(encoder_out_aug["encoder_out"]) > 0: + enc_aug = encoder_out_aug["encoder_out"][0] + if ( + encoder_out_aug is not None + and len(encoder_out_aug["encoder_padding_mask"]) > 0 + ): + padding_mask_aug = encoder_out_aug["encoder_padding_mask"][0] + + # embed positions + positions = None + if self.embed_positions is not None: + positions = self.embed_positions( + prev_output_tokens, incremental_state=incremental_state + ) + + if incremental_state is not None: + prev_output_tokens = prev_output_tokens[:, -1:] + if positions is not None: + positions = positions[:, -1:] + + # Prevent torchscript exporting issue for dynamic quant embedding + prev_output_tokens = prev_output_tokens.contiguous() + # embed tokens and positions + x = self.embed_scale * self.embed_tokens(prev_output_tokens) + + if self.quant_noise is not None: + x = self.quant_noise(x) + + if self.project_in_dim is not None: + x = self.project_in_dim(x) + + if positions is not None: + x += positions + + if self.layernorm_embedding is not None: + x = self.layernorm_embedding(x) + + x = self.dropout_module(x) + + # B x T x C -> T x B x C + x = x.transpose(0, 1) + + self_attn_padding_mask: Optional[Tensor] = None + if self.cross_self_attention or prev_output_tokens.eq(self.padding_idx).any(): + self_attn_padding_mask = prev_output_tokens.eq(self.padding_idx) + + # decoder layers + attn: Optional[Tensor] = None + attn_aug: Optional[Tensor] = None + inner_states: List[Optional[Tensor]] = [x] + for idx, layer in enumerate(self.layers): + if incremental_state is None and not full_context_alignment: + self_attn_mask = self.buffered_future_mask(x) + else: + self_attn_mask = None + + x, layer_attn, layer_attn_aug, _ = layer( + x, + enc, + padding_mask, + enc_aug, + padding_mask_aug, + incremental_state, + self_attn_mask=self_attn_mask, + self_attn_padding_mask=self_attn_padding_mask, + need_attn=bool((idx == alignment_layer)), + need_head_weights=bool((idx == alignment_layer)), + ) + inner_states.append(x) + if layer_attn is not None and idx == alignment_layer: + attn = layer_attn.float().to(x) + if layer_attn_aug is not None and idx == alignment_layer: + attn_aug = layer_attn_aug.float().to(x) + + if attn is not None: + if alignment_heads is not None: + attn = attn[:alignment_heads] + + # average probabilities over heads + attn = attn.mean(dim=0) + + if attn_aug is not None: + if alignment_heads is not None: + attn_aug = attn_aug[:alignment_heads] + + # average probabilities over heads + attn_aug = attn_aug.mean(dim=0) + + if self.layer_norm is not None: + x = self.layer_norm(x) + + # T x B x C -> B x T x C + x = x.transpose(0, 1) + + if self.project_out_dim is not None: + x = self.project_out_dim(x) + + return x, {"attn": [attn], "attn_aug": [attn_aug], "inner_states": inner_states} + + def upgrade_state_dict_named(self, state_dict, name): + """Upgrade a (possibly old) state dict for new versions of fairseq.""" + if f"{name}.output_projection.weight" not in state_dict: + if self.share_input_output_embed: + embed_out_key = f"{name}.embed_tokens.weight" + else: + embed_out_key = f"{name}.embed_out" + if embed_out_key in state_dict: + state_dict[f"{name}.output_projection.weight"] = state_dict[ + embed_out_key + ] + if not self.share_input_output_embed: + del state_dict[embed_out_key] + + for i in range(self.num_layers): + # update layer norms + layer_norm_map = { + "0": "self_attn_layer_norm", + "1": "encoder_attn_layer_norm", + "2": "encoder_attn_layer_norm2", + "3": "final_layer_norm", + } + for old, new in layer_norm_map.items(): + for m in ("weight", "bias"): + k = "{}.layers.{}.layer_norms.{}.{}".format(name, i, old, m) + if k in state_dict: + state_dict[ + "{}.layers.{}.{}.{}".format(name, i, new, m) + ] = state_dict[k] + del state_dict[k] + + version_key = "{}.version".format(name) + if utils.item(state_dict.get(version_key, torch.Tensor([1]))[0]) <= 2: + # earlier checkpoints did not normalize after the stack of layers + self.layer_norm = None + self.normalize = False + state_dict[version_key] = torch.Tensor([1]) + + return state_dict + + +class AugTransformerDecoder(AugTransformerDecoderBase): + def __init__( + self, + args, + dictionary, + embed_tokens, + output_projection=None, + ): + self.args = args + super().__init__( + TransformerConfig.from_namespace(args), + dictionary, + embed_tokens, + no_encoder_attn=False, + output_projection=output_projection, + encoder_attn_merge_type=getattr( + args, "synthesizer_augmented_cross_attention_merge_type", "sequential" + ), + dropnet_ratio=getattr(args, "dropnet_ratio", 0), + ) + + def build_output_projection(self, args, dictionary, embed_tokens): + super().build_output_projection( + TransformerConfig.from_namespace(args), dictionary, embed_tokens + ) + + def build_decoder_layer( + self, + args, + encoder_attn_merge_type="sequential", + dropnet_ratio=0, + ): + return super().build_decoder_layer( + TransformerConfig.from_namespace(args), + no_encoder_attn=False, + encoder_attn_merge_type=encoder_attn_merge_type, + dropnet_ratio=dropnet_ratio, + ) diff --git a/data/fairseq/fairseq/models/transformer/transformer_encoder.py b/data/fairseq/fairseq/models/transformer/transformer_encoder.py new file mode 100644 index 0000000000000000000000000000000000000000..a684fcb448e9ac6a7a6440c825bbe3212789cf88 --- /dev/null +++ b/data/fairseq/fairseq/models/transformer/transformer_encoder.py @@ -0,0 +1,362 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import math +from typing import Dict, List, Optional + +import torch +import torch.nn as nn +from torch import Tensor + +from fairseq import utils +from fairseq.distributed import fsdp_wrap +from fairseq.models import FairseqEncoder +from fairseq.models.transformer import TransformerConfig +from fairseq.modules import ( + FairseqDropout, + LayerDropModuleList, + LayerNorm, + PositionalEmbedding, + SinusoidalPositionalEmbedding, + transformer_layer, +) +from fairseq.modules.checkpoint_activations import checkpoint_wrapper +from fairseq.modules.quant_noise import quant_noise as apply_quant_noise_ + + +# rewrite name for backward compatibility in `make_generation_fast_` +def module_name_fordropout(module_name: str) -> str: + if module_name == "TransformerEncoderBase": + return "TransformerEncoder" + else: + return module_name + + +class TransformerEncoderBase(FairseqEncoder): + """ + Transformer encoder consisting of *cfg.encoder.layers* layers. Each layer + is a :class:`TransformerEncoderLayer`. + + Args: + args (argparse.Namespace): parsed command-line arguments + dictionary (~fairseq.data.Dictionary): encoding dictionary + embed_tokens (torch.nn.Embedding): input embedding + """ + + def __init__(self, cfg, dictionary, embed_tokens, return_fc=False): + self.cfg = cfg + super().__init__(dictionary) + self.register_buffer("version", torch.Tensor([3])) + + self.dropout_module = FairseqDropout( + cfg.dropout, module_name=module_name_fordropout(self.__class__.__name__) + ) + self.encoder_layerdrop = cfg.encoder.layerdrop + self.return_fc = return_fc + + embed_dim = embed_tokens.embedding_dim + self.padding_idx = embed_tokens.padding_idx + self.max_source_positions = cfg.max_source_positions + + self.embed_tokens = embed_tokens + + self.embed_scale = 1.0 if cfg.no_scale_embedding else math.sqrt(embed_dim) + + self.embed_positions = ( + PositionalEmbedding( + cfg.max_source_positions, + embed_dim, + self.padding_idx, + learned=cfg.encoder.learned_pos, + ) + if not cfg.no_token_positional_embeddings + else None + ) + if cfg.layernorm_embedding: + self.layernorm_embedding = LayerNorm(embed_dim, export=cfg.export) + else: + self.layernorm_embedding = None + + if not cfg.adaptive_input and cfg.quant_noise.pq > 0: + self.quant_noise = apply_quant_noise_( + nn.Linear(embed_dim, embed_dim, bias=False), + cfg.quant_noise.pq, + cfg.quant_noise.pq_block_size, + ) + else: + self.quant_noise = None + + if self.encoder_layerdrop > 0.0: + self.layers = LayerDropModuleList(p=self.encoder_layerdrop) + else: + self.layers = nn.ModuleList([]) + self.layers.extend( + [self.build_encoder_layer(cfg) for i in range(cfg.encoder.layers)] + ) + self.num_layers = len(self.layers) + + if cfg.encoder.normalize_before: + self.layer_norm = LayerNorm(embed_dim, export=cfg.export) + else: + self.layer_norm = None + + def build_encoder_layer(self, cfg): + layer = transformer_layer.TransformerEncoderLayerBase( + cfg, return_fc=self.return_fc + ) + checkpoint = cfg.checkpoint_activations + if checkpoint: + offload_to_cpu = cfg.offload_activations + layer = checkpoint_wrapper(layer, offload_to_cpu=offload_to_cpu) + # if we are checkpointing, enforce that FSDP always wraps the + # checkpointed layer, regardless of layer size + min_params_to_wrap = cfg.min_params_to_wrap if not checkpoint else 0 + layer = fsdp_wrap(layer, min_num_params=min_params_to_wrap) + return layer + + def forward_embedding( + self, src_tokens, token_embedding: Optional[torch.Tensor] = None + ): + # embed tokens and positions + if token_embedding is None: + token_embedding = self.embed_tokens(src_tokens) + x = embed = self.embed_scale * token_embedding + if self.embed_positions is not None: + x = embed + self.embed_positions(src_tokens) + if self.layernorm_embedding is not None: + x = self.layernorm_embedding(x) + x = self.dropout_module(x) + if self.quant_noise is not None: + x = self.quant_noise(x) + return x, embed + + def forward( + self, + src_tokens, + src_lengths: Optional[torch.Tensor] = None, + return_all_hiddens: bool = False, + token_embeddings: Optional[torch.Tensor] = None, + ): + """ + Args: + src_tokens (LongTensor): tokens in the source language of shape + `(batch, src_len)` + src_lengths (torch.LongTensor): lengths of each source sentence of + shape `(batch)` + return_all_hiddens (bool, optional): also return all of the + intermediate hidden states (default: False). + token_embeddings (torch.Tensor, optional): precomputed embeddings + default `None` will recompute embeddings + + Returns: + dict: + - **encoder_out** (Tensor): the last encoder layer's output of + shape `(src_len, batch, embed_dim)` + - **encoder_padding_mask** (ByteTensor): the positions of + padding elements of shape `(batch, src_len)` + - **encoder_embedding** (Tensor): the (scaled) embedding lookup + of shape `(batch, src_len, embed_dim)` + - **encoder_states** (List[Tensor]): all intermediate + hidden states of shape `(src_len, batch, embed_dim)`. + Only populated if *return_all_hiddens* is True. + """ + return self.forward_scriptable( + src_tokens, src_lengths, return_all_hiddens, token_embeddings + ) + + # TorchScript doesn't support super() method so that the scriptable Subclass + # can't access the base class model in Torchscript. + # Current workaround is to add a helper function with different name and + # call the helper function from scriptable Subclass. + def forward_scriptable( + self, + src_tokens, + src_lengths: Optional[torch.Tensor] = None, + return_all_hiddens: bool = False, + token_embeddings: Optional[torch.Tensor] = None, + ): + """ + Args: + src_tokens (LongTensor): tokens in the source language of shape + `(batch, src_len)` + src_lengths (torch.LongTensor): lengths of each source sentence of + shape `(batch)` + return_all_hiddens (bool, optional): also return all of the + intermediate hidden states (default: False). + token_embeddings (torch.Tensor, optional): precomputed embeddings + default `None` will recompute embeddings + + Returns: + dict: + - **encoder_out** (Tensor): the last encoder layer's output of + shape `(src_len, batch, embed_dim)` + - **encoder_padding_mask** (ByteTensor): the positions of + padding elements of shape `(batch, src_len)` + - **encoder_embedding** (Tensor): the (scaled) embedding lookup + of shape `(batch, src_len, embed_dim)` + - **encoder_states** (List[Tensor]): all intermediate + hidden states of shape `(src_len, batch, embed_dim)`. + Only populated if *return_all_hiddens* is True. + """ + # compute padding mask + encoder_padding_mask = src_tokens.eq(self.padding_idx) + has_pads = ( + torch.tensor(src_tokens.device.type == "xla") or encoder_padding_mask.any() + ) + # Torchscript doesn't handle bool Tensor correctly, so we need to work around. + if torch.jit.is_scripting(): + has_pads = torch.tensor(1) if has_pads else torch.tensor(0) + + x, encoder_embedding = self.forward_embedding(src_tokens, token_embeddings) + + # account for padding while computing the representation + x = x * ( + 1 - encoder_padding_mask.unsqueeze(-1).type_as(x) * has_pads.type_as(x) + ) + + # B x T x C -> T x B x C + x = x.transpose(0, 1) + + encoder_states = [] + fc_results = [] + + if return_all_hiddens: + encoder_states.append(x) + + # encoder layers + for layer in self.layers: + lr = layer( + x, encoder_padding_mask=encoder_padding_mask if has_pads else None + ) + + if isinstance(lr, tuple) and len(lr) == 2: + x, fc_result = lr + else: + x = lr + fc_result = None + + if return_all_hiddens and not torch.jit.is_scripting(): + assert encoder_states is not None + encoder_states.append(x) + fc_results.append(fc_result) + + if self.layer_norm is not None: + x = self.layer_norm(x) + + # The Pytorch Mobile lite interpreter does not supports returning NamedTuple in + # `forward` so we use a dictionary instead. + # TorchScript does not support mixed values so the values are all lists. + # The empty list is equivalent to None. + src_lengths = ( + src_tokens.ne(self.padding_idx) + .sum(dim=1, dtype=torch.int32) + .reshape(-1, 1) + .contiguous() + ) + return { + "encoder_out": [x], # T x B x C + "encoder_padding_mask": [encoder_padding_mask], # B x T + "encoder_embedding": [encoder_embedding], # B x T x C + "encoder_states": encoder_states, # List[T x B x C] + "fc_results": fc_results, # List[T x B x C] + "src_tokens": [], + "src_lengths": [src_lengths], + } + + @torch.jit.export + def reorder_encoder_out(self, encoder_out: Dict[str, List[Tensor]], new_order): + """ + Reorder encoder output according to *new_order*. + + Args: + encoder_out: output from the ``forward()`` method + new_order (LongTensor): desired order + + Returns: + *encoder_out* rearranged according to *new_order* + """ + if len(encoder_out["encoder_out"]) == 0: + new_encoder_out = [] + else: + new_encoder_out = [encoder_out["encoder_out"][0].index_select(1, new_order)] + if len(encoder_out["encoder_padding_mask"]) == 0: + new_encoder_padding_mask = [] + else: + new_encoder_padding_mask = [ + encoder_out["encoder_padding_mask"][0].index_select(0, new_order) + ] + if len(encoder_out["encoder_embedding"]) == 0: + new_encoder_embedding = [] + else: + new_encoder_embedding = [ + encoder_out["encoder_embedding"][0].index_select(0, new_order) + ] + + if len(encoder_out["src_tokens"]) == 0: + src_tokens = [] + else: + src_tokens = [(encoder_out["src_tokens"][0]).index_select(0, new_order)] + + if len(encoder_out["src_lengths"]) == 0: + src_lengths = [] + else: + src_lengths = [(encoder_out["src_lengths"][0]).index_select(0, new_order)] + + encoder_states = encoder_out["encoder_states"] + if len(encoder_states) > 0: + for idx, state in enumerate(encoder_states): + encoder_states[idx] = state.index_select(1, new_order) + + return { + "encoder_out": new_encoder_out, # T x B x C + "encoder_padding_mask": new_encoder_padding_mask, # B x T + "encoder_embedding": new_encoder_embedding, # B x T x C + "encoder_states": encoder_states, # List[T x B x C] + "src_tokens": src_tokens, # B x T + "src_lengths": src_lengths, # B x 1 + } + + @torch.jit.export + def _reorder_encoder_out(self, encoder_out: Dict[str, List[Tensor]], new_order): + """Dummy re-order function for beamable enc-dec attention""" + return encoder_out + + def max_positions(self): + """Maximum input length supported by the encoder.""" + if self.embed_positions is None: + return self.max_source_positions + return min(self.max_source_positions, self.embed_positions.max_positions) + + def upgrade_state_dict_named(self, state_dict, name): + """Upgrade a (possibly old) state dict for new versions of fairseq.""" + for i in range(self.num_layers): + # update layer norms + self.layers[i].upgrade_state_dict_named( + state_dict, "{}.layers.{}".format(name, i) + ) + + version_key = "{}.version".format(name) + if utils.item(state_dict.get(version_key, torch.Tensor([1]))[0]) < 2: + # earlier checkpoints did not normalize after the stack of layers + self.layer_norm = None + self.normalize = False + state_dict[version_key] = torch.Tensor([1]) + return state_dict + + +class TransformerEncoder(TransformerEncoderBase): + def __init__(self, args, dictionary, embed_tokens, return_fc=False): + self.args = args + super().__init__( + TransformerConfig.from_namespace(args), + dictionary, + embed_tokens, + return_fc=return_fc, + ) + + def build_encoder_layer(self, args): + return super().build_encoder_layer( + TransformerConfig.from_namespace(args), + ) diff --git a/data/fairseq/fairseq/models/transformer/transformer_legacy.py b/data/fairseq/fairseq/models/transformer/transformer_legacy.py new file mode 100644 index 0000000000000000000000000000000000000000..00d14a7dde778a4b3b2faecde50eecdd5115a864 --- /dev/null +++ b/data/fairseq/fairseq/models/transformer/transformer_legacy.py @@ -0,0 +1,277 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from fairseq.dataclass.utils import gen_parser_from_dataclass +from fairseq.models import ( + register_model, + register_model_architecture, +) +from fairseq.models.transformer.transformer_config import ( + TransformerConfig, + DEFAULT_MAX_SOURCE_POSITIONS, + DEFAULT_MAX_TARGET_POSITIONS, + DEFAULT_MIN_PARAMS_TO_WRAP, +) +from fairseq.models.transformer.transformer_base import ( + TransformerModelBase, +) + + +@register_model("transformer") +class TransformerModel(TransformerModelBase): + """ + This is the legacy implementation of the transformer model that + uses argparse for configuration. + """ + + @classmethod + def hub_models(cls): + # fmt: off + + def moses_subword(path): + return { + 'path': path, + 'tokenizer': 'moses', + 'bpe': 'subword_nmt', + } + + def moses_fastbpe(path): + return { + 'path': path, + 'tokenizer': 'moses', + 'bpe': 'fastbpe', + } + + def spm(path): + return { + 'path': path, + 'bpe': 'sentencepiece', + 'tokenizer': 'space', + } + + return { + 'transformer.wmt14.en-fr': moses_subword('https://dl.fbaipublicfiles.com/fairseq/models/wmt14.en-fr.joined-dict.transformer.tar.bz2'), + 'transformer.wmt16.en-de': 'https://dl.fbaipublicfiles.com/fairseq/models/wmt16.en-de.joined-dict.transformer.tar.bz2', + 'transformer.wmt18.en-de': moses_subword('https://dl.fbaipublicfiles.com/fairseq/models/wmt18.en-de.ensemble.tar.gz'), + 'transformer.wmt19.en-de': moses_fastbpe('https://dl.fbaipublicfiles.com/fairseq/models/wmt19.en-de.joined-dict.ensemble.tar.gz'), + 'transformer.wmt19.en-ru': moses_fastbpe('https://dl.fbaipublicfiles.com/fairseq/models/wmt19.en-ru.ensemble.tar.gz'), + 'transformer.wmt19.de-en': moses_fastbpe('https://dl.fbaipublicfiles.com/fairseq/models/wmt19.de-en.joined-dict.ensemble.tar.gz'), + 'transformer.wmt19.ru-en': moses_fastbpe('https://dl.fbaipublicfiles.com/fairseq/models/wmt19.ru-en.ensemble.tar.gz'), + 'transformer.wmt19.en-de.single_model': moses_fastbpe('https://dl.fbaipublicfiles.com/fairseq/models/wmt19.en-de.joined-dict.single_model.tar.gz'), + 'transformer.wmt19.en-ru.single_model': moses_fastbpe('https://dl.fbaipublicfiles.com/fairseq/models/wmt19.en-ru.single_model.tar.gz'), + 'transformer.wmt19.de-en.single_model': moses_fastbpe('https://dl.fbaipublicfiles.com/fairseq/models/wmt19.de-en.joined-dict.single_model.tar.gz'), + 'transformer.wmt19.ru-en.single_model': moses_fastbpe('https://dl.fbaipublicfiles.com/fairseq/models/wmt19.ru-en.single_model.tar.gz'), + 'transformer.wmt20.en-ta': spm('https://dl.fbaipublicfiles.com/fairseq/models/wmt20.en-ta.single.tar.gz'), + 'transformer.wmt20.en-iu.news': spm('https://dl.fbaipublicfiles.com/fairseq/models/wmt20.en-iu.news.single.tar.gz'), + 'transformer.wmt20.en-iu.nh': spm('https://dl.fbaipublicfiles.com/fairseq/models/wmt20.en-iu.nh.single.tar.gz'), + 'transformer.wmt20.ta-en': spm('https://dl.fbaipublicfiles.com/fairseq/models/wmt20.ta-en.single.tar.gz'), + 'transformer.wmt20.iu-en.news': spm('https://dl.fbaipublicfiles.com/fairseq/models/wmt20.iu-en.news.single.tar.gz'), + 'transformer.wmt20.iu-en.nh': spm('https://dl.fbaipublicfiles.com/fairseq/models/wmt20.iu-en.nh.single.tar.gz'), + 'transformer.flores101.mm100.615M': spm('https://dl.fbaipublicfiles.com/flores101/pretrained_models/flores101_mm100_615M.tar.gz'), + 'transformer.flores101.mm100.175M': spm('https://dl.fbaipublicfiles.com/flores101/pretrained_models/flores101_mm100_175M.tar.gz'), + } + # fmt: on + + def __init__(self, args, encoder, decoder): + cfg = TransformerConfig.from_namespace(args) + super().__init__(cfg, encoder, decoder) + self.args = args + + @classmethod + def add_args(cls, parser): + """Add model-specific arguments to the parser.""" + # we want to build the args recursively in this case. + # do not set defaults so that settings defaults from various architectures still works + gen_parser_from_dataclass( + parser, TransformerConfig(), delete_default=True, with_prefix="" + ) + + @classmethod + def build_model(cls, args, task): + """Build a new model instance.""" + + # make sure all arguments are present in older models + base_architecture(args) + + if args.encoder_layers_to_keep: + args.encoder_layers = len(args.encoder_layers_to_keep.split(",")) + if args.decoder_layers_to_keep: + args.decoder_layers = len(args.decoder_layers_to_keep.split(",")) + + if getattr(args, "max_source_positions", None) is None: + args.max_source_positions = DEFAULT_MAX_SOURCE_POSITIONS + if getattr(args, "max_target_positions", None) is None: + args.max_target_positions = DEFAULT_MAX_TARGET_POSITIONS + + src_dict, tgt_dict = task.source_dictionary, task.target_dictionary + + if args.share_all_embeddings: + if src_dict != tgt_dict: + raise ValueError("--share-all-embeddings requires a joined dictionary") + if args.encoder_embed_dim != args.decoder_embed_dim: + raise ValueError( + "--share-all-embeddings requires --encoder-embed-dim to match --decoder-embed-dim" + ) + if args.decoder_embed_path and ( + args.decoder_embed_path != args.encoder_embed_path + ): + raise ValueError( + "--share-all-embeddings not compatible with --decoder-embed-path" + ) + args.share_decoder_input_output_embed = True + + if getattr(args, "offload_activations", False): + args.checkpoint_activations = True # offloading implies checkpointing + + if not args.share_all_embeddings: + args.min_params_to_wrap = getattr( + args, "min_params_to_wrap", DEFAULT_MIN_PARAMS_TO_WRAP + ) + cfg = TransformerConfig.from_namespace(args) + return super().build_model(cfg, task) + + @classmethod + def build_embedding(cls, args, dictionary, embed_dim, path=None): + return super().build_embedding( + TransformerConfig.from_namespace(args), dictionary, embed_dim, path + ) + + @classmethod + def build_encoder(cls, args, src_dict, embed_tokens): + return super().build_encoder( + TransformerConfig.from_namespace(args), src_dict, embed_tokens + ) + + @classmethod + def build_decoder(cls, args, tgt_dict, embed_tokens): + return super().build_decoder( + TransformerConfig.from_namespace(args), tgt_dict, embed_tokens + ) + + +# architectures + + +@register_model_architecture("transformer", "transformer_tiny") +def tiny_architecture(args): + args.encoder_embed_dim = getattr(args, "encoder_embed_dim", 64) + args.encoder_ffn_embed_dim = getattr(args, "encoder_ffn_embed_dim", 64) + args.encoder_layers = getattr(args, "encoder_layers", 2) + args.encoder_attention_heads = getattr(args, "encoder_attention_heads", 2) + args.decoder_layers = getattr(args, "decoder_layers", 2) + args.decoder_attention_heads = getattr(args, "decoder_attention_heads", 2) + return base_architecture(args) + + +@register_model_architecture("transformer", "transformer") +def base_architecture(args): + args.encoder_embed_path = getattr(args, "encoder_embed_path", None) + args.encoder_embed_dim = getattr(args, "encoder_embed_dim", 512) + args.encoder_ffn_embed_dim = getattr(args, "encoder_ffn_embed_dim", 2048) + args.encoder_layers = getattr(args, "encoder_layers", 6) + args.encoder_attention_heads = getattr(args, "encoder_attention_heads", 8) + args.encoder_normalize_before = getattr(args, "encoder_normalize_before", False) + args.encoder_learned_pos = getattr(args, "encoder_learned_pos", False) + + args.decoder_embed_path = getattr(args, "decoder_embed_path", None) + args.decoder_embed_dim = getattr(args, "decoder_embed_dim", args.encoder_embed_dim) + args.decoder_ffn_embed_dim = getattr( + args, "decoder_ffn_embed_dim", args.encoder_ffn_embed_dim + ) + args.decoder_layers = getattr(args, "decoder_layers", 6) + args.decoder_attention_heads = getattr(args, "decoder_attention_heads", 8) + args.decoder_normalize_before = getattr(args, "decoder_normalize_before", False) + args.decoder_learned_pos = getattr(args, "decoder_learned_pos", False) + args.attention_dropout = getattr(args, "attention_dropout", 0.0) + args.activation_dropout = getattr(args, "activation_dropout", 0.0) + args.activation_fn = getattr(args, "activation_fn", "relu") + args.dropout = getattr(args, "dropout", 0.1) + args.adaptive_softmax_cutoff = getattr(args, "adaptive_softmax_cutoff", None) + args.adaptive_softmax_dropout = getattr(args, "adaptive_softmax_dropout", 0) + args.share_decoder_input_output_embed = getattr( + args, "share_decoder_input_output_embed", False + ) + args.share_all_embeddings = getattr(args, "share_all_embeddings", False) + args.merge_src_tgt_embed = getattr(args, "merge_src_tgt_embed", False) + args.no_token_positional_embeddings = getattr( + args, "no_token_positional_embeddings", False + ) + args.adaptive_input = getattr(args, "adaptive_input", False) + args.no_cross_attention = getattr(args, "no_cross_attention", False) + args.cross_self_attention = getattr(args, "cross_self_attention", False) + + args.decoder_output_dim = getattr( + args, "decoder_output_dim", args.decoder_embed_dim + ) + args.decoder_input_dim = getattr(args, "decoder_input_dim", args.decoder_embed_dim) + + args.no_scale_embedding = getattr(args, "no_scale_embedding", False) + args.layernorm_embedding = getattr(args, "layernorm_embedding", False) + args.tie_adaptive_weights = getattr(args, "tie_adaptive_weights", False) + args.checkpoint_activations = getattr(args, "checkpoint_activations", False) + args.offload_activations = getattr(args, "offload_activations", False) + if args.offload_activations: + args.checkpoint_activations = True + args.encoder_layers_to_keep = getattr(args, "encoder_layers_to_keep", None) + args.decoder_layers_to_keep = getattr(args, "decoder_layers_to_keep", None) + args.encoder_layerdrop = getattr(args, "encoder_layerdrop", 0) + args.decoder_layerdrop = getattr(args, "decoder_layerdrop", 0) + args.quant_noise_pq = getattr(args, "quant_noise_pq", 0) + args.quant_noise_pq_block_size = getattr(args, "quant_noise_pq_block_size", 8) + args.quant_noise_scalar = getattr(args, "quant_noise_scalar", 0) + + +@register_model_architecture("transformer", "transformer_iwslt_de_en") +def transformer_iwslt_de_en(args): + args.encoder_embed_dim = getattr(args, "encoder_embed_dim", 512) + args.encoder_ffn_embed_dim = getattr(args, "encoder_ffn_embed_dim", 1024) + args.encoder_attention_heads = getattr(args, "encoder_attention_heads", 4) + args.encoder_layers = getattr(args, "encoder_layers", 6) + args.decoder_embed_dim = getattr(args, "decoder_embed_dim", 512) + args.decoder_ffn_embed_dim = getattr(args, "decoder_ffn_embed_dim", 1024) + args.decoder_attention_heads = getattr(args, "decoder_attention_heads", 4) + args.decoder_layers = getattr(args, "decoder_layers", 6) + base_architecture(args) + + +@register_model_architecture("transformer", "transformer_wmt_en_de") +def transformer_wmt_en_de(args): + base_architecture(args) + + +# parameters used in the "Attention Is All You Need" paper (Vaswani et al., 2017) +@register_model_architecture("transformer", "transformer_vaswani_wmt_en_de_big") +def transformer_vaswani_wmt_en_de_big(args): + args.encoder_embed_dim = getattr(args, "encoder_embed_dim", 1024) + args.encoder_ffn_embed_dim = getattr(args, "encoder_ffn_embed_dim", 4096) + args.encoder_attention_heads = getattr(args, "encoder_attention_heads", 16) + args.encoder_normalize_before = getattr(args, "encoder_normalize_before", False) + args.decoder_embed_dim = getattr(args, "decoder_embed_dim", 1024) + args.decoder_ffn_embed_dim = getattr(args, "decoder_ffn_embed_dim", 4096) + args.decoder_attention_heads = getattr(args, "decoder_attention_heads", 16) + args.dropout = getattr(args, "dropout", 0.3) + base_architecture(args) + + +@register_model_architecture("transformer", "transformer_vaswani_wmt_en_fr_big") +def transformer_vaswani_wmt_en_fr_big(args): + args.dropout = getattr(args, "dropout", 0.1) + transformer_vaswani_wmt_en_de_big(args) + + +@register_model_architecture("transformer", "transformer_wmt_en_de_big") +def transformer_wmt_en_de_big(args): + args.attention_dropout = getattr(args, "attention_dropout", 0.1) + transformer_vaswani_wmt_en_de_big(args) + + +# default parameters used in tensor2tensor implementation +@register_model_architecture("transformer", "transformer_wmt_en_de_big_t2t") +def transformer_wmt_en_de_big_t2t(args): + args.encoder_normalize_before = getattr(args, "encoder_normalize_before", True) + args.decoder_normalize_before = getattr(args, "decoder_normalize_before", True) + args.attention_dropout = getattr(args, "attention_dropout", 0.1) + args.activation_dropout = getattr(args, "activation_dropout", 0.1) + transformer_vaswani_wmt_en_de_big(args) diff --git a/data/fairseq/fairseq/models/wav2vec/__init__.py b/data/fairseq/fairseq/models/wav2vec/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..b756e4580b82b42c735d9e097c62f63cb7798f5e --- /dev/null +++ b/data/fairseq/fairseq/models/wav2vec/__init__.py @@ -0,0 +1,10 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from .wav2vec import * # noqa +from .wav2vec2 import * # noqa +from .wav2vec2_asr import * # noqa +from .wav2vec2_laser import * # noqa +from .wav2vec2_classification import * # noqa diff --git a/data/fairseq/fairseq/models/wav2vec/__pycache__/__init__.cpython-38.pyc b/data/fairseq/fairseq/models/wav2vec/__pycache__/__init__.cpython-38.pyc new file mode 100644 index 0000000000000000000000000000000000000000..964427dc632760ecc31079dbf45700e07b95927d Binary files /dev/null and b/data/fairseq/fairseq/models/wav2vec/__pycache__/__init__.cpython-38.pyc differ diff --git a/data/fairseq/fairseq/models/wav2vec/__pycache__/utils.cpython-38.pyc b/data/fairseq/fairseq/models/wav2vec/__pycache__/utils.cpython-38.pyc new file mode 100644 index 0000000000000000000000000000000000000000..e914c2c4f0d2ef7c13eb0ef909de4346f1a7bc2a Binary files /dev/null and b/data/fairseq/fairseq/models/wav2vec/__pycache__/utils.cpython-38.pyc differ diff --git a/data/fairseq/fairseq/models/wav2vec/__pycache__/wav2vec.cpython-38.pyc b/data/fairseq/fairseq/models/wav2vec/__pycache__/wav2vec.cpython-38.pyc new file mode 100644 index 0000000000000000000000000000000000000000..a5655abcb558372d4e47fa1fa0c4447e9eae64d2 Binary files /dev/null and b/data/fairseq/fairseq/models/wav2vec/__pycache__/wav2vec.cpython-38.pyc differ diff --git a/data/fairseq/fairseq/models/wav2vec/__pycache__/wav2vec2.cpython-38.pyc b/data/fairseq/fairseq/models/wav2vec/__pycache__/wav2vec2.cpython-38.pyc new file mode 100644 index 0000000000000000000000000000000000000000..ae5fc6cfd3c4de9a1af510043f2278e8c487a0e0 Binary files /dev/null and b/data/fairseq/fairseq/models/wav2vec/__pycache__/wav2vec2.cpython-38.pyc differ diff --git a/data/fairseq/fairseq/models/wav2vec/__pycache__/wav2vec2_asr.cpython-38.pyc b/data/fairseq/fairseq/models/wav2vec/__pycache__/wav2vec2_asr.cpython-38.pyc new file mode 100644 index 0000000000000000000000000000000000000000..3a4b6283c249cd389261b4a425fb7d44b42b3f58 Binary files /dev/null and b/data/fairseq/fairseq/models/wav2vec/__pycache__/wav2vec2_asr.cpython-38.pyc differ diff --git a/data/fairseq/fairseq/models/wav2vec/__pycache__/wav2vec2_classification.cpython-38.pyc b/data/fairseq/fairseq/models/wav2vec/__pycache__/wav2vec2_classification.cpython-38.pyc new file mode 100644 index 0000000000000000000000000000000000000000..217bb05ecc461ac588b9bc25f3765aac5398e0c0 Binary files /dev/null and b/data/fairseq/fairseq/models/wav2vec/__pycache__/wav2vec2_classification.cpython-38.pyc differ diff --git a/data/fairseq/fairseq/models/wav2vec/__pycache__/wav2vec2_laser.cpython-38.pyc b/data/fairseq/fairseq/models/wav2vec/__pycache__/wav2vec2_laser.cpython-38.pyc new file mode 100644 index 0000000000000000000000000000000000000000..abf8aab8cd65b160c56a5a46f09f854344406f88 Binary files /dev/null and b/data/fairseq/fairseq/models/wav2vec/__pycache__/wav2vec2_laser.cpython-38.pyc differ diff --git a/data/fairseq/fairseq/models/wav2vec/utils.py b/data/fairseq/fairseq/models/wav2vec/utils.py new file mode 100644 index 0000000000000000000000000000000000000000..dd52d86242ec5289b480cb085b50a68a713cc306 --- /dev/null +++ b/data/fairseq/fairseq/models/wav2vec/utils.py @@ -0,0 +1,21 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import math +import torch.nn.functional as F + + +def pad_to_multiple(x, multiple, dim=-1, value=0): + # Inspired from https://github.com/lucidrains/local-attention/blob/master/local_attention/local_attention.py#L41 + if x is None: + return None, 0 + tsz = x.size(dim) + m = tsz / multiple + remainder = math.ceil(m) * multiple - tsz + if m.is_integer(): + return x, 0 + pad_offset = (0,) * (-1 - dim) * 2 + + return F.pad(x, (*pad_offset, 0, remainder), value=value), remainder diff --git a/data/fairseq/fairseq/models/wav2vec/wav2vec.py b/data/fairseq/fairseq/models/wav2vec/wav2vec.py new file mode 100644 index 0000000000000000000000000000000000000000..af6604da10f504baabff50bf14a6eb2214bffef3 --- /dev/null +++ b/data/fairseq/fairseq/models/wav2vec/wav2vec.py @@ -0,0 +1,630 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from dataclasses import dataclass, field +import logging +import math +from typing import Optional, Tuple +from omegaconf import II +import sys + +import torch +import torch.nn as nn +import torch.nn.functional as F +from fairseq.dataclass import ChoiceEnum, FairseqDataclass +from fairseq.models import BaseFairseqModel, register_model +from fairseq.modules import ( + Fp32GroupNorm, + Fp32LayerNorm, + GumbelVectorQuantizer, + KmeansVectorQuantizer, + TransposeLast, +) +from fairseq.tasks import FairseqTask +from fairseq.utils import buffered_arange + + +logger = logging.getLogger(__name__) + + +AGGREGATOR_CHOICES = ChoiceEnum(["cnn", "gru"]) +PROJECT_FEATURES_CHOICES = ChoiceEnum(["none", "same", "new"]) +ACTIVATION_CHOICES = ChoiceEnum(["relu", "gelu"]) +VQ_TYPE_CHOICES = ChoiceEnum(["none", "gumbel", "kmeans"]) + + +@dataclass +class Wav2VecConfig(FairseqDataclass): + prediction_steps: int = field( + default=12, metadata={"help": "number of steps ahead to predict"} + ) + sample_distance: Optional[int] = field( + default=None, + metadata={ + "help": "sample distance from target. does not work properly with cross-sampling" + }, + ) + cross_sample_negatives: int = field( + default=0, metadata={"help": "num of cross sampled negatives"} + ) + num_negatives: int = field( + default=10, metadata={"help": "num of sampled negatives"} + ) + conv_feature_layers: str = field( + default="[(512, 10, 5), (512, 8, 4), (512, 4, 2), (512, 4, 2), (512, 4, 2), (512, 1, 1), (512, 1, 1), (512, 1, 1)]", + metadata={ + "help": "convolutional feature extraction layers [(dim, kernel_size, stride), ...]" + }, + ) + conv_aggregator_layers: str = field( + default="[(512, 2, 1), (512, 3, 1), (512, 4, 1), (512, 5, 1), (512, 6, 1), (512, 7, 1), (512, 8, 1), (512, 9, 1), (512, 10, 1), (512, 11, 1), (512, 12, 1), (512, 13, 1)]", + metadata={ + "help": "convolutional aggregator layers [(dim, kernel_size, stride), ...]" + }, + ) + dropout: float = field( + default=0.0, metadata={"help": "dropout to apply within the model"} + ) + dropout_features: float = field( + default=0.0, metadata={"help": "dropout to apply to the features"} + ) + dropout_agg: float = field( + default=0.0, metadata={"help": "dropout to apply after aggregation step"} + ) + aggregator: AGGREGATOR_CHOICES = field( + default="cnn", metadata={"help": "type of aggregator to use"} + ) + gru_dim: int = field(default=512, metadata={"help": "GRU dimensionality"}) + no_conv_bias: bool = field( + default=False, metadata={"help": "if set, does not learn bias for conv layers"} + ) + agg_zero_pad: bool = field( + default=False, + metadata={"help": "if set, zero pads in aggregator instead of repl pad"}, + ) + skip_connections_feat: bool = field( + default=False, + metadata={"help": "if set, adds skip connections to the feature extractor"}, + ) + skip_connections_agg: bool = field( + default=True, + metadata={"help": "if set, adds skip connections to the aggregator"}, + ) + residual_scale: float = field( + default=0.5, metadata={"help": "scales residual by sqrt(value)"} + ) + log_compression: bool = field( + default=True, + metadata={"help": "if set, adds a log compression to feature extractor"}, + ) + balanced_classes: bool = field( + default=False, + metadata={"help": "if set, loss is scaled to balance for number of negatives"}, + ) + project_features: PROJECT_FEATURES_CHOICES = field( + default="none", + metadata={ + "help": "if not none, features are projected using the (same or new) aggregator" + }, + ) + non_affine_group_norm: bool = field( + default=False, metadata={"help": "if set, group norm is not affine"} + ) + offset: str = field( + default="auto", + metadata={ + "help": "if set to 'auto', it is computed automatically from the receptive field, else set to int value" + }, + ) + activation: ACTIVATION_CHOICES = field( + default="relu", + metadata={ + "help": "if set to 'auto', it is computed automatically from the receptive field, else set to int value" + }, + ) + vq_type: VQ_TYPE_CHOICES = field( + default="none", metadata={"help": "which type of quantizer to use"} + ) + vq_vars: int = field( + default=320, + metadata={"help": "project to this many vector quantized variables per group"}, + ) + vq_groups: int = field( + default=2, metadata={"help": "number of groups of latent variables"} + ) + vq_dim: int = field( + default=0, + metadata={ + "help": "uses this dimensionality for quantized vectors. 0 to use model dim // groups" + }, + ) + vq_depth: int = field( + default=1, metadata={"help": "number of layers for vq weight projection"} + ) + combine_groups: bool = field( + default=False, metadata={"help": "if set, variables are shared among groups"} + ) + vq_temp: Tuple[float, float, float] = field( + default=(2.0, 0.5, 0.999995), + metadata={ + "help": "temperature for latent variable sampling with gumbel softmax. should be a tuple of 3 values (start, end, decay)" + }, + ) + vq_gamma: float = field( + default=0.25, + metadata={"help": "gamma parameter for kmeans style vector quantization"}, + ) + infonce: bool = II("criterion.infonce") + + +@register_model("wav2vec", dataclass=Wav2VecConfig) +class Wav2VecModel(BaseFairseqModel): + @classmethod + def build_model(cls, cfg: Wav2VecConfig, task: FairseqTask): + """Build a new model instance.""" + + model = Wav2VecModel(cfg) + logger.info(model) + return model + + def __init__(self, cfg: Wav2VecConfig): + super().__init__() + + self.prediction_steps = cfg.prediction_steps + offset = cfg.offset + + if cfg.activation == "relu": + activation = nn.ReLU() + elif cfg.activation == "gelu": + activation = nn.GELU() + else: + raise Exception("unknown activation " + cfg.activation) + + feature_enc_layers = eval(cfg.conv_feature_layers) + self.feature_extractor = ConvFeatureExtractionModel( + conv_layers=feature_enc_layers, + dropout=0.0, + log_compression=cfg.log_compression, + skip_connections=cfg.skip_connections_feat, + residual_scale=cfg.residual_scale, + non_affine_group_norm=cfg.non_affine_group_norm, + activation=activation, + ) + embed = feature_enc_layers[-1][0] + + self.vector_quantizer = None + if cfg.vq_type == "gumbel": + self.vector_quantizer = GumbelVectorQuantizer( + dim=embed, + num_vars=cfg.vq_vars, + temp=cfg.vq_temp, + groups=cfg.vq_groups, + combine_groups=cfg.combine_groups, + vq_dim=cfg.vq_dim if cfg.vq_dim > 0 else embed, + time_first=False, + activation=activation, + weight_proj_depth=cfg.vq_depth, + weight_proj_factor=2, + ) + elif cfg.vq_type == "kmeans": + self.vector_quantizer = KmeansVectorQuantizer( + dim=embed, + num_vars=cfg.vq_vars, + groups=cfg.vq_groups, + combine_groups=cfg.combine_groups, + vq_dim=cfg.vq_dim if cfg.vq_dim > 0 else embed, + time_first=False, + gamma=cfg.vq_gamma, + ) + else: + assert ( + cfg.vq_type == "none" or cfg.vq_type is None + ), "Unknown quantizer type" + + if cfg.offset == "auto": + jin = 0 + rin = 0 + for _, k, stride in feature_enc_layers: + if rin == 0: + rin = k + rin = rin + (k - 1) * jin + if jin == 0: + jin = stride + else: + jin *= stride + offset = math.ceil(rin / jin) + + offset = int(offset) + + def make_aggregator(): + if cfg.aggregator == "cnn": + agg_layers = eval(cfg.conv_aggregator_layers) + agg_dim = agg_layers[-1][0] + feature_aggregator = ConvAggegator( + conv_layers=agg_layers, + embed=embed, + dropout=cfg.dropout, + skip_connections=cfg.skip_connections_agg, + residual_scale=cfg.residual_scale, + non_affine_group_norm=cfg.non_affine_group_norm, + conv_bias=not cfg.no_conv_bias, + zero_pad=cfg.agg_zero_pad, + activation=activation, + ) + elif cfg.aggregator == "gru": + agg_dim = cfg.gru_dim + feature_aggregator = nn.Sequential( + TransposeLast(), + nn.GRU( + input_size=embed, + hidden_size=agg_dim, + num_layers=1, + dropout=cfg.dropout, + ), + TransposeLast(deconstruct_idx=0), + ) + else: + raise Exception("unknown aggregator type " + cfg.aggregator) + + return feature_aggregator, agg_dim + + self.feature_aggregator, agg_dim = make_aggregator() + + self.wav2vec_predictions = Wav2VecPredictionsModel( + in_dim=agg_dim, + out_dim=embed, + prediction_steps=cfg.prediction_steps, + n_negatives=cfg.num_negatives, + cross_sample_negatives=cfg.cross_sample_negatives, + sample_distance=cfg.sample_distance, + dropout=cfg.dropout, + offset=offset, + balanced_classes=cfg.balanced_classes, + infonce=cfg.infonce, + ) + + self.dropout_feats = nn.Dropout(p=cfg.dropout_features) + self.dropout_agg = nn.Dropout(p=cfg.dropout_agg) + + if cfg.project_features == "none": + self.project_features = None + elif cfg.project_features == "same": + self.project_features = self.feature_aggregator + elif cfg.project_features == "new": + self.project_features, _ = make_aggregator() + + def forward(self, source): + result = {} + + features = self.feature_extractor(source) + if self.vector_quantizer: + q_res = self.vector_quantizer(features) + features = q_res["x"] + for k in q_res.keys(): + if k != "x": + result[k] = q_res[k] + + x = self.dropout_feats(features) + x = self.feature_aggregator(x) + x = self.dropout_agg(x) + + if self.project_features is not None: + features = self.project_features(features) + x, targets = self.wav2vec_predictions(x, features) + result["cpc_logits"] = x + result["cpc_targets"] = targets + + return result + + def upgrade_state_dict_named(self, state_dict, name): + super().upgrade_state_dict_named(state_dict, name) + + def max_positions(self): + """Maximum length supported by the model.""" + return sys.maxsize + + def get_logits(self, net_output): + logits = net_output["cpc_logits"] + return logits + + def get_targets(self, sample, net_output): + t = net_output["cpc_targets"] + if isinstance(t, tuple): + t = t[0] + return t.contiguous() + + def get_target_weights(self, targets, net_output): + targets = net_output["cpc_targets"] + if isinstance(targets, tuple) and targets[-1] is not None: + return targets[-1] + return None + + def get_extra_losses(self, net_output): + loss = None + if "prob_perplexity" in net_output: + loss = net_output["num_vars"] - net_output["prob_perplexity"] + elif "kmeans_loss" in net_output: + loss = net_output["kmeans_loss"] + + return loss + + +def norm_block(is_layer_norm, dim, affine=True): + if is_layer_norm: + mod = nn.Sequential( + TransposeLast(), + Fp32LayerNorm(dim, elementwise_affine=affine), + TransposeLast(), + ) + else: + mod = Fp32GroupNorm(1, dim, affine=affine) + + return mod + + +class ConvFeatureExtractionModel(nn.Module): + def __init__( + self, + conv_layers, + dropout, + log_compression, + skip_connections, + residual_scale, + non_affine_group_norm, + activation, + ): + super().__init__() + + def block(n_in, n_out, k, stride): + return nn.Sequential( + nn.Conv1d(n_in, n_out, k, stride=stride, bias=False), + nn.Dropout(p=dropout), + norm_block( + is_layer_norm=False, dim=n_out, affine=not non_affine_group_norm + ), + activation, + ) + + in_d = 1 + self.conv_layers = nn.ModuleList() + for dim, k, stride in conv_layers: + self.conv_layers.append(block(in_d, dim, k, stride)) + in_d = dim + + self.log_compression = log_compression + self.skip_connections = skip_connections + self.residual_scale = math.sqrt(residual_scale) + + def forward(self, x): + # BxT -> BxCxT + x = x.unsqueeze(1) + + for conv in self.conv_layers: + residual = x + x = conv(x) + if self.skip_connections and x.size(1) == residual.size(1): + tsz = x.size(2) + r_tsz = residual.size(2) + residual = residual[..., :: r_tsz // tsz][..., :tsz] + x = (x + residual) * self.residual_scale + + if self.log_compression: + x = x.abs() + x = x + 1 + x = x.log() + + return x + + +class ZeroPad1d(nn.Module): + def __init__(self, pad_left, pad_right): + super().__init__() + self.pad_left = pad_left + self.pad_right = pad_right + + def forward(self, x): + return F.pad(x, (self.pad_left, self.pad_right)) + + +class ConvAggegator(nn.Module): + def __init__( + self, + conv_layers, + embed, + dropout, + skip_connections, + residual_scale, + non_affine_group_norm, + conv_bias, + zero_pad, + activation, + ): + super().__init__() + + def block(n_in, n_out, k, stride): + # padding dims only really make sense for stride = 1 + ka = k // 2 + kb = ka - 1 if k % 2 == 0 else ka + + pad = ( + ZeroPad1d(ka + kb, 0) if zero_pad else nn.ReplicationPad1d((ka + kb, 0)) + ) + + return nn.Sequential( + pad, + nn.Conv1d(n_in, n_out, k, stride=stride, bias=conv_bias), + nn.Dropout(p=dropout), + norm_block(False, n_out, affine=not non_affine_group_norm), + activation, + ) + + in_d = embed + self.conv_layers = nn.ModuleList() + self.residual_proj = nn.ModuleList() + for dim, k, stride in conv_layers: + if in_d != dim and skip_connections: + self.residual_proj.append(nn.Conv1d(in_d, dim, 1, bias=False)) + else: + self.residual_proj.append(None) + + self.conv_layers.append(block(in_d, dim, k, stride)) + in_d = dim + self.conv_layers = nn.Sequential(*self.conv_layers) + self.skip_connections = skip_connections + self.residual_scale = math.sqrt(residual_scale) + + def forward(self, x): + for rproj, conv in zip(self.residual_proj, self.conv_layers): + residual = x + x = conv(x) + if self.skip_connections: + if rproj is not None: + residual = rproj(residual) + x = (x + residual) * self.residual_scale + return x + + +class Wav2VecPredictionsModel(nn.Module): + def __init__( + self, + in_dim, + out_dim, + prediction_steps, + n_negatives, + cross_sample_negatives, + sample_distance, + dropout, + offset, + balanced_classes, + infonce, + ): + super().__init__() + + self.n_negatives = n_negatives + self.cross_sample_negatives = cross_sample_negatives + self.sample_distance = sample_distance + self.project_to_steps = nn.ConvTranspose2d( + in_dim, out_dim, (1, prediction_steps) + ) + self.dropout = nn.Dropout(p=dropout) + self.offset = offset + self.balanced_classes = balanced_classes + self.infonce = infonce + + def sample_negatives(self, y): + bsz, fsz, tsz = y.shape + + y = y.transpose(0, 1) # BCT -> CBT + y = y.contiguous().view(fsz, -1) # CBT => C(BxT) + + cross_high = tsz * bsz + high = tsz if self.sample_distance is None else min(tsz, self.sample_distance) + assert high > 1 + + neg_idxs = torch.randint(low=0, high=high, size=(bsz, self.n_negatives * tsz)) + + with torch.no_grad(): + if self.n_negatives > 0: + tszs = ( + buffered_arange(tsz) + .unsqueeze(-1) + .expand(-1, self.n_negatives) + .flatten() + ) + + neg_idxs = torch.randint( + low=0, high=high - 1, size=(bsz, self.n_negatives * tsz) + ) + neg_idxs[neg_idxs >= tszs] += 1 + + if self.cross_sample_negatives > 0: + tszs = ( + buffered_arange(tsz) + .unsqueeze(-1) + .expand(-1, self.cross_sample_negatives) + .flatten() + ) + + cross_neg_idxs = torch.randint( + low=0, + high=cross_high - 1, + size=(bsz, self.cross_sample_negatives * tsz), + ) + cross_neg_idxs[cross_neg_idxs >= tszs] += 1 + + if self.n_negatives > 0: + for i in range(1, bsz): + neg_idxs[i] += i * high + else: + neg_idxs = cross_neg_idxs + + if self.cross_sample_negatives > 0 and self.n_negatives > 0: + neg_idxs = torch.cat([neg_idxs, cross_neg_idxs], dim=1) + + negs = y[..., neg_idxs.view(-1)] + negs = negs.view( + fsz, bsz, self.n_negatives + self.cross_sample_negatives, tsz + ).permute( + 2, 1, 0, 3 + ) # to NxBxCxT + + return negs + + def forward(self, x, y): + + x = x.unsqueeze(-1) + x = self.project_to_steps(x) # BxCxTxS + x = self.dropout(x) + + negatives = self.sample_negatives(y) + y = y.unsqueeze(0) + targets = torch.cat([y, negatives], dim=0) # Copies x B x C x T + + copies = targets.size(0) + bsz, dim, tsz, steps = x.shape + steps = min(steps, tsz - self.offset) + + predictions = x.new( + bsz * copies * (tsz - self.offset + 1) * steps + - ((steps + 1) * steps // 2) * copies * bsz + ) + if self.infonce: + labels = predictions.new_full( + (predictions.shape[0] // copies,), 0, dtype=torch.long + ) + else: + labels = torch.zeros_like(predictions) + weights = ( + torch.full_like(labels, 1 / self.n_negatives) + if self.balanced_classes and not self.infonce + else None + ) + + start = end = 0 + for i in range(steps): + offset = i + self.offset + end = start + (tsz - offset) * bsz * copies + if self.infonce: + predictions[start:end] = torch.einsum( + "bct,nbct->tbn", x[..., :-offset, i], targets[..., offset:] + ).flatten() + else: + pos_num = (end - start) // copies + predictions[start:end] = torch.einsum( + "bct,nbct->nbt", x[..., :-offset, i], targets[..., offset:] + ).flatten() + labels[start : start + pos_num] = 1.0 + if weights is not None: + weights[start : start + pos_num] = 1.0 + start = end + assert end == predictions.numel(), "{} != {}".format(end, predictions.numel()) + + if self.infonce: + predictions = predictions.view(-1, copies) + else: + if weights is not None: + labels = (labels, weights) + + return predictions, labels diff --git a/data/fairseq/fairseq/models/wav2vec/wav2vec2.py b/data/fairseq/fairseq/models/wav2vec/wav2vec2.py new file mode 100644 index 0000000000000000000000000000000000000000..0faba77f8be8e5acfc9358972606714ee2da5da6 --- /dev/null +++ b/data/fairseq/fairseq/models/wav2vec/wav2vec2.py @@ -0,0 +1,1499 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import math +from dataclasses import dataclass, field +from typing import List, Tuple + +import numpy as np +import torch +import torch.nn as nn +import torch.nn.functional as F + +from fairseq import utils +from fairseq.data.data_utils import compute_mask_indices +from fairseq.dataclass import ChoiceEnum, FairseqDataclass +from fairseq.distributed import fsdp_wrap +from fairseq.models import BaseFairseqModel, register_model +from fairseq.distributed.fully_sharded_data_parallel import FullyShardedDataParallel +from fairseq.modules import ( + Fp32GroupNorm, + Fp32LayerNorm, + GradMultiply, + GumbelVectorQuantizer, + LayerNorm, + MultiheadAttention, + RelPositionalEncoding, + SamePad, + TransposeLast, +) +from fairseq.modules.checkpoint_activations import checkpoint_wrapper +from fairseq.modules.conformer_layer import ConformerWav2Vec2EncoderLayer +from fairseq.modules.transformer_sentence_encoder import init_bert_params +from fairseq.utils import buffered_arange, index_put, is_xla_tensor + +from .utils import pad_to_multiple + +EXTRACTOR_MODE_CHOICES = ChoiceEnum(["default", "layer_norm"]) +MASKING_DISTRIBUTION_CHOICES = ChoiceEnum(["static", "uniform", "normal", "poisson"]) +LAYER_TYPE_CHOICES = ChoiceEnum(["transformer", "conformer", "trf_adp"]) + + +@dataclass +class Wav2Vec2Config(FairseqDataclass): + extractor_mode: EXTRACTOR_MODE_CHOICES = field( + default="default", + metadata={ + "help": "mode for feature extractor. default has a single group norm with d " + "groups in the first conv block, whereas layer_norm has layer norms in " + "every block (meant to use with normalize=True)" + }, + ) + encoder_layers: int = field( + default=12, metadata={"help": "num encoder layers in the transformer"} + ) + encoder_embed_dim: int = field( + default=768, metadata={"help": "encoder embedding dimension"} + ) + encoder_ffn_embed_dim: int = field( + default=3072, metadata={"help": "encoder embedding dimension for FFN"} + ) + encoder_attention_heads: int = field( + default=12, metadata={"help": "num encoder attention heads"} + ) + activation_fn: ChoiceEnum(utils.get_available_activation_fns()) = field( + default="gelu", metadata={"help": "activation function to use"} + ) + layer_type: LAYER_TYPE_CHOICES = field( + default="transformer", metadata={"help": "layer type in encoder"} + ) + # dropouts + dropout: float = field( + default=0.1, metadata={"help": "dropout probability for the transformer"} + ) + attention_dropout: float = field( + default=0.1, metadata={"help": "dropout probability for attention weights"} + ) + activation_dropout: float = field( + default=0.0, metadata={"help": "dropout probability after activation in FFN"} + ) + encoder_layerdrop: float = field( + default=0.0, metadata={"help": "probability of dropping a tarnsformer layer"} + ) + dropout_input: float = field( + default=0.0, + metadata={"help": "dropout to apply to the input (after feat extr)"}, + ) + dropout_features: float = field( + default=0.0, + metadata={"help": "dropout to apply to the features (after feat extr)"}, + ) + + final_dim: int = field( + default=0, + metadata={ + "help": "project final representations and targets to this many dimensions." + "set to encoder_embed_dim is <= 0" + }, + ) + layer_norm_first: bool = field( + default=False, metadata={"help": "apply layernorm first in the transformer"} + ) + conv_feature_layers: str = field( + default="[(512, 10, 5)] + [(512, 3, 2)] * 4 + [(512,2,2)] + [(512,2,2)]", + metadata={ + "help": "string describing convolutional feature extraction layers in form of a python list that contains " + "[(dim, kernel_size, stride), ...]" + }, + ) + conv_bias: bool = field( + default=False, metadata={"help": "include bias in conv encoder"} + ) + logit_temp: float = field( + default=0.1, metadata={"help": "temperature to divide logits by"} + ) + quantize_targets: bool = field( + default=False, metadata={"help": "use quantized targets"} + ) + quantize_input: bool = field( + default=False, metadata={"help": "use quantized inputs"} + ) + same_quantizer: bool = field( + default=False, metadata={"help": "use same quantizer for inputs and targets"} + ) + target_glu: bool = field( + default=False, metadata={"help": "adds projection + glu to targets"} + ) + feature_grad_mult: float = field( + default=1.0, metadata={"help": "multiply feature extractor var grads by this"} + ) + quantizer_depth: int = field( + default=1, + metadata={"help": "number of quantizer layers"}, + ) + quantizer_factor: int = field( + default=3, + metadata={ + "help": "dimensionality increase for inner quantizer layers (if depth > 1)" + }, + ) + latent_vars: int = field( + default=320, + metadata={"help": "number of latent variables V in each group of the codebook"}, + ) + latent_groups: int = field( + default=2, + metadata={"help": "number of groups G of latent variables in the codebook"}, + ) + latent_dim: int = field( + default=0, + metadata={ + "help": "if > 0, uses this dimensionality for latent variables. " + "otherwise uses final_dim / latent_groups" + }, + ) + + # masking + mask_length: int = field(default=10, metadata={"help": "mask length"}) + mask_prob: float = field( + default=0.65, metadata={"help": "probability of replacing a token with mask"} + ) + mask_selection: MASKING_DISTRIBUTION_CHOICES = field( + default="static", metadata={"help": "how to choose mask length"} + ) + mask_other: float = field( + default=0, + metadata={ + "help": "secondary mask argument (used for more complex distributions), " + "see help in compute_mask_indices" + }, + ) + no_mask_overlap: bool = field( + default=False, metadata={"help": "whether to allow masks to overlap"} + ) + mask_min_space: int = field( + default=1, + metadata={"help": "min space between spans (if no overlap is enabled)"}, + ) + require_same_masks: bool = field( + default=True, + metadata={ + "help": "whether to number of masked timesteps must be the same across all " + "examples in a batch" + }, + ) + mask_dropout: float = field( + default=0.0, + metadata={"help": "percent of masks to unmask for each sample"}, + ) + + # channel masking + mask_channel_length: int = field( + default=10, metadata={"help": "length of the mask for features (channels)"} + ) + mask_channel_prob: float = field( + default=0.0, metadata={"help": "probability of replacing a feature with 0"} + ) + mask_channel_before: bool = False + mask_channel_selection: MASKING_DISTRIBUTION_CHOICES = field( + default="static", + metadata={"help": "how to choose mask length for channel masking"}, + ) + mask_channel_other: float = field( + default=0, + metadata={ + "help": "secondary mask argument (used for more complex distributions), " + "see help in compute_mask_indicesh" + }, + ) + no_mask_channel_overlap: bool = field( + default=False, metadata={"help": "whether to allow channel masks to overlap"} + ) + mask_channel_min_space: int = field( + default=1, + metadata={"help": "min space between spans (if no overlap is enabled)"}, + ) + + # negative selection + num_negatives: int = field( + default=100, + metadata={"help": "number of negative examples from the same sample"}, + ) + negatives_from_everywhere: bool = field( + default=False, + metadata={"help": "sample negatives from everywhere, not just masked states"}, + ) + cross_sample_negatives: int = field( + default=0, metadata={"help": "number of negative examples from the any sample"} + ) + codebook_negatives: int = field( + default=0, metadata={"help": "number of negative examples codebook"} + ) + + # positional embeddings + conv_pos: int = field( + default=128, + metadata={"help": "number of filters for convolutional positional embeddings"}, + ) + conv_pos_groups: int = field( + default=16, + metadata={"help": "number of groups for convolutional positional embedding"}, + ) + pos_conv_depth: int = field( + default=1, + metadata={"help": "depth of positional encoder network"}, + ) + + latent_temp: Tuple[float, float, float] = field( + default=(2, 0.5, 0.999995), + metadata={ + "help": "temperature for latent variable sampling. " + "can be tuple of 3 values (start, end, decay)" + }, + ) + max_positions: int = field(default=100000, metadata={"help": "Max positions"}) + checkpoint_activations: bool = field( + default=False, + metadata={"help": "recompute activations and save memory for extra compute"}, + ) + + # FP16 optimization + required_seq_len_multiple: int = field( + default=2, + metadata={ + "help": "pad the input to encoder such that the sequence length is divisible by multiple" + }, + ) + crop_seq_to_multiple: int = field( + default=1, + metadata={ + "help": "crop convolutional feature extractor output such that the sequence length is divisible by multiple" + }, + ) + + # Conformer + depthwise_conv_kernel_size: int = field( + default=31, + metadata={ + "help": "depthwise-conv-kernel-size for convolution in conformer layer" + }, + ) + attn_type: str = field( + default="", + metadata={"help": "if espnet use ESPNET MHA"}, + ) + pos_enc_type: str = field( + default="abs", + metadata={"help": "Positional encoding type to use in conformer"}, + ) + fp16: bool = field(default=False, metadata={"help": "If fp16 is being used"}) + + # Adapter num + adp_num: int = field( + default=-1 + ) + adp_dim: int = field( + default=64 + ) + adp_act_fn: str = field( + default="relu" + ) + adp_trf_idx: str = field( + default="all", + ) + + +@register_model("wav2vec2", dataclass=Wav2Vec2Config) +class Wav2Vec2Model(BaseFairseqModel): + def __init__(self, cfg: Wav2Vec2Config): + super().__init__() + self.cfg = cfg + + feature_enc_layers = eval(cfg.conv_feature_layers) + self.embed = feature_enc_layers[-1][0] + + self.feature_extractor = ConvFeatureExtractionModel( + conv_layers=feature_enc_layers, + dropout=0.0, + mode=cfg.extractor_mode, + conv_bias=cfg.conv_bias, + ) + + self.post_extract_proj = ( + nn.Linear(self.embed, cfg.encoder_embed_dim) + if self.embed != cfg.encoder_embed_dim and not cfg.quantize_input + else None + ) + + self.crop_seq_to_multiple = cfg.crop_seq_to_multiple + + self.mask_prob = cfg.mask_prob + self.mask_selection = cfg.mask_selection + self.mask_other = cfg.mask_other + self.mask_length = cfg.mask_length + self.no_mask_overlap = cfg.no_mask_overlap + self.mask_min_space = cfg.mask_min_space + + self.mask_channel_prob = cfg.mask_channel_prob + self.mask_channel_before = cfg.mask_channel_before + self.mask_channel_selection = cfg.mask_channel_selection + self.mask_channel_other = cfg.mask_channel_other + self.mask_channel_length = cfg.mask_channel_length + self.no_mask_channel_overlap = cfg.no_mask_channel_overlap + self.mask_channel_min_space = cfg.mask_channel_min_space + + self.dropout_input = nn.Dropout(cfg.dropout_input) + self.dropout_features = nn.Dropout(cfg.dropout_features) + + self.feature_grad_mult = cfg.feature_grad_mult + + self.quantizer = None + self.input_quantizer = None + + self.n_negatives = cfg.num_negatives + self.cross_sample_negatives = cfg.cross_sample_negatives + self.codebook_negatives = cfg.codebook_negatives + self.negatives_from_everywhere = cfg.negatives_from_everywhere + + self.logit_temp = cfg.logit_temp + + final_dim = cfg.final_dim if cfg.final_dim > 0 else cfg.encoder_embed_dim + + if cfg.quantize_targets: + vq_dim = cfg.latent_dim if cfg.latent_dim > 0 else final_dim + self.quantizer = GumbelVectorQuantizer( + dim=self.embed, + num_vars=cfg.latent_vars, + temp=cfg.latent_temp, + groups=cfg.latent_groups, + combine_groups=False, + vq_dim=vq_dim, + time_first=True, + weight_proj_depth=cfg.quantizer_depth, + weight_proj_factor=cfg.quantizer_factor, + ) + self.project_q = nn.Linear(vq_dim, final_dim) + else: + self.project_q = nn.Linear(self.embed, final_dim) + + if cfg.quantize_input: + if cfg.same_quantizer and self.quantizer is not None: + vq_dim = final_dim + self.input_quantizer = self.quantizer + else: + vq_dim = cfg.latent_dim if cfg.latent_dim > 0 else cfg.encoder_embed_dim + self.input_quantizer = GumbelVectorQuantizer( + dim=self.embed, + num_vars=cfg.latent_vars, + temp=cfg.latent_temp, + groups=cfg.latent_groups, + combine_groups=False, + vq_dim=vq_dim, + time_first=True, + weight_proj_depth=cfg.quantizer_depth, + weight_proj_factor=cfg.quantizer_factor, + ) + self.project_inp = nn.Linear(vq_dim, cfg.encoder_embed_dim) + + self.mask_emb = nn.Parameter( + torch.FloatTensor(cfg.encoder_embed_dim).uniform_() + ) + encoder_cls = TransformerEncoder + if cfg.layer_type == "conformer" and cfg.pos_enc_type in ["rel_pos", "rope"]: + encoder_cls = ConformerEncoder + + self.encoder = encoder_cls(cfg) + self.layer_norm = LayerNorm(self.embed) + + self.target_glu = None + if cfg.target_glu: + self.target_glu = nn.Sequential( + nn.Linear(final_dim, final_dim * 2), nn.GLU() + ) + + self.final_proj = nn.Linear(cfg.encoder_embed_dim, final_dim) + + def upgrade_state_dict_named(self, state_dict, name): + super().upgrade_state_dict_named(state_dict, name) + """Upgrade a (possibly old) state dict for new versions of fairseq.""" + return state_dict + + @classmethod + def build_model(cls, cfg: Wav2Vec2Config, task=None): + """Build a new model instance.""" + + return cls(cfg) + + def apply_mask( + self, + x, + padding_mask, + mask_indices=None, + mask_channel_indices=None, + ): + B, T, C = x.shape + + if self.mask_channel_prob > 0 and self.mask_channel_before: + mask_channel_indices = compute_mask_indices( + (B, C), + None, + self.mask_channel_prob, + self.mask_channel_length, + self.mask_channel_selection, + self.mask_channel_other, + no_overlap=self.no_mask_channel_overlap, + min_space=self.mask_channel_min_space, + ) + mask_channel_indices = ( + torch.from_numpy(mask_channel_indices) + .to(x.device) + .unsqueeze(1) + .expand(-1, T, -1) + ) + x[mask_channel_indices] = 0 + + if self.mask_prob > 0: + if mask_indices is None: + mask_indices = compute_mask_indices( + (B, T), + padding_mask, + self.mask_prob, + self.mask_length, + self.mask_selection, + self.mask_other, + min_masks=2, + no_overlap=self.no_mask_overlap, + min_space=self.mask_min_space, + require_same_masks=self.cfg.require_same_masks, + mask_dropout=self.cfg.mask_dropout, + ) + mask_indices = torch.from_numpy(mask_indices).to(x.device) + x = index_put(x, mask_indices, self.mask_emb) + else: + mask_indices = None + + if self.mask_channel_prob > 0 and not self.mask_channel_before: + if mask_channel_indices is None: + mask_channel_indices = compute_mask_indices( + (B, C), + None, + self.mask_channel_prob, + self.mask_channel_length, + self.mask_channel_selection, + self.mask_channel_other, + no_overlap=self.no_mask_channel_overlap, + min_space=self.mask_channel_min_space, + ) + mask_channel_indices = ( + torch.from_numpy(mask_channel_indices) + .to(x.device) + .unsqueeze(1) + .expand(-1, T, -1) + ) + x = index_put(x, mask_channel_indices, 0) + + return x, mask_indices + + def sample_negatives(self, y, num, padding_count=None): + + if self.n_negatives == 0 and self.cross_sample_negatives == 0: + return y.new(0) + + bsz, tsz, fsz = y.shape + y = y.view(-1, fsz) # BTC => (BxT)C + + # FIXME: what happens if padding_count is specified? + cross_high = tsz * bsz + high = tsz - (padding_count or 0) + with torch.no_grad(): + assert high > 1, f"{bsz,tsz,fsz}" + + if self.n_negatives > 0: + tszs = ( + buffered_arange(num) + .unsqueeze(-1) + .expand(-1, self.n_negatives) + .flatten() + ) + + neg_idxs = torch.randint( + low=0, high=high - 1, size=(bsz, self.n_negatives * num) + ) + neg_idxs[neg_idxs >= tszs] += 1 + + if self.cross_sample_negatives > 0: + tszs = ( + buffered_arange(num) + .unsqueeze(-1) + .expand(-1, self.cross_sample_negatives) + .flatten() + ) + + cross_neg_idxs = torch.randint( + low=0, + high=cross_high - 1, + size=(bsz, self.cross_sample_negatives * num), + ) + cross_neg_idxs[cross_neg_idxs >= tszs] += 1 + + if self.n_negatives > 0: + neg_idxs = neg_idxs + (torch.arange(bsz).unsqueeze(1) * high) + else: + neg_idxs = cross_neg_idxs + + if self.cross_sample_negatives > 0 and self.n_negatives > 0: + neg_idxs = torch.cat([neg_idxs, cross_neg_idxs], dim=1) + + negs = y[neg_idxs.view(-1)] + negs = negs.view( + bsz, num, self.n_negatives + self.cross_sample_negatives, fsz + ).permute( + 2, 0, 1, 3 + ) # to NxBxTxC + return negs, neg_idxs + + def compute_preds(self, x, y, negatives): + + neg_is_pos = (y == negatives).all(-1) + y = y.unsqueeze(0) + targets = torch.cat([y, negatives], dim=0) + + logits = torch.cosine_similarity(x.float(), targets.float(), dim=-1) + logits = logits / self.logit_temp + logits = logits.type_as(x) + + if is_xla_tensor(logits) or neg_is_pos.any(): + if not hasattr(self, "_inftensor"): + fillval = -float(2**30) + self._inftensor = ( + torch.tensor(fillval).to(x.device) + if is_xla_tensor(logits) + else float("-inf") + ) + logits[1:] = index_put(logits[1:], neg_is_pos, self._inftensor) + + return logits + + def _get_feat_extract_output_lengths(self, input_lengths: torch.LongTensor): + """ + Computes the output length of the convolutional layers + """ + + def _conv_out_length(input_length, kernel_size, stride): + return torch.floor((input_length - kernel_size) / stride + 1) + + conv_cfg_list = eval(self.cfg.conv_feature_layers) + + for i in range(len(conv_cfg_list)): + input_lengths = _conv_out_length( + input_lengths, conv_cfg_list[i][1], conv_cfg_list[i][2] + ) + + return input_lengths.to(torch.long) + + def forward( + self, + source, + padding_mask=None, + mask=True, + features_only=False, + layer=None, + mask_indices=None, + mask_channel_indices=None, + padding_count=None, + corpus_key=None, + ): + + if self.feature_grad_mult > 0: + features = self.feature_extractor(source) + if self.feature_grad_mult != 1.0: + features = GradMultiply.apply(features, self.feature_grad_mult) + else: + with torch.no_grad(): + features = self.feature_extractor(source) + + features_pen = features.float().pow(2).mean() + + features = features.transpose(1, 2) + features = self.layer_norm(features) + unmasked_features = features.clone() + + if padding_mask is not None and padding_mask.any(): + input_lengths = (1 - padding_mask.long()).sum(-1) + # apply conv formula to get real output_lengths + output_lengths = self._get_feat_extract_output_lengths(input_lengths) + + padding_mask = torch.zeros( + features.shape[:2], dtype=features.dtype, device=features.device + ) + + # these two operations makes sure that all values + # before the output lengths indices are attended to + padding_mask[ + ( + torch.arange(padding_mask.shape[0], device=padding_mask.device), + output_lengths - 1, + ) + ] = 1 + padding_mask = (1 - padding_mask.flip([-1]).cumsum(-1).flip([-1])).bool() + else: + padding_mask = None + + time_steps_to_drop = features.size(1) % self.crop_seq_to_multiple + if time_steps_to_drop != 0: + features = features[:, :-time_steps_to_drop] + unmasked_features = unmasked_features[:, :-time_steps_to_drop] + if padding_mask is not None: + padding_mask = padding_mask[:, :-time_steps_to_drop] + + if self.post_extract_proj is not None: + features = self.post_extract_proj(features) + + features = self.dropout_input(features) + unmasked_features = self.dropout_features(unmasked_features) + + num_vars = None + code_ppl = None + prob_ppl = None + curr_temp = None + + if self.input_quantizer: + q = self.input_quantizer(features, produce_targets=False) + features = q["x"] + num_vars = q["num_vars"] + code_ppl = q["code_perplexity"] + prob_ppl = q["prob_perplexity"] + curr_temp = q["temp"] + features = self.project_inp(features) + + if mask: + x, mask_indices = self.apply_mask( + features, + padding_mask, + mask_indices=mask_indices, + mask_channel_indices=mask_channel_indices, + ) + if not is_xla_tensor(x) and mask_indices is not None: + # tpu-comment: reducing the size in a dynamic way causes + # too many recompilations on xla. + y = unmasked_features[mask_indices].view( + unmasked_features.size(0), -1, unmasked_features.size(-1) + ) + else: + y = unmasked_features + else: + x = features + y = unmasked_features + mask_indices = None + + x, layer_results = self.encoder( + x, padding_mask=padding_mask, layer=layer, corpus_key=corpus_key + ) + + if features_only: + return { + "x": x, + "padding_mask": padding_mask, + "features": unmasked_features, + "layer_results": layer_results, + } + + if self.quantizer: + if self.negatives_from_everywhere: + q = self.quantizer(unmasked_features, produce_targets=False) + y = q["x"] + num_vars = q["num_vars"] + code_ppl = q["code_perplexity"] + prob_ppl = q["prob_perplexity"] + curr_temp = q["temp"] + y = self.project_q(y) + + negs, _ = self.sample_negatives( + y, + mask_indices[0].sum(), + padding_count=padding_count, + ) + y = y[mask_indices].view(y.size(0), -1, y.size(-1)) + + else: + q = self.quantizer(y, produce_targets=False) + y = q["x"] + num_vars = q["num_vars"] + code_ppl = q["code_perplexity"] + prob_ppl = q["prob_perplexity"] + curr_temp = q["temp"] + + y = self.project_q(y) + + negs, _ = self.sample_negatives( + y, + y.size(1), + padding_count=padding_count, + ) + + if self.codebook_negatives > 0: + cb_negs = self.quantizer.sample_from_codebook( + y.size(0) * y.size(1), self.codebook_negatives + ) + cb_negs = cb_negs.view( + self.codebook_negatives, y.size(0), y.size(1), -1 + ) # order doesnt matter + cb_negs = self.project_q(cb_negs) + negs = torch.cat([negs, cb_negs], dim=0) + else: + y = self.project_q(y) + + if self.negatives_from_everywhere: + negs, _ = self.sample_negatives( + unmasked_features, + y.size(1), + padding_count=padding_count, + ) + negs = self.project_q(negs) + else: + negs, _ = self.sample_negatives( + y, + y.size(1), + padding_count=padding_count, + ) + + if not is_xla_tensor(x): + # tpu-comment: reducing the size in a dynamic way causes + # too many recompilations on xla. + x = x[mask_indices].view(x.size(0), -1, x.size(-1)) + + if self.target_glu: + y = self.target_glu(y) + negs = self.target_glu(negs) + + x = self.final_proj(x) + x = self.compute_preds(x, y, negs) + + result = { + "x": x, + "padding_mask": padding_mask, + "features_pen": features_pen, + } + + if prob_ppl is not None: + result["prob_perplexity"] = prob_ppl + result["code_perplexity"] = code_ppl + result["num_vars"] = num_vars + result["temp"] = curr_temp + + return result + + def quantize(self, x): + assert self.quantizer is not None + x = self.feature_extractor(x) + x = x.transpose(1, 2) + x = self.layer_norm(x) + return self.quantizer.forward_idx(x) + + def extract_features( + self, source, padding_mask, mask=False, layer=None, corpus_key=None + ): + res = self.forward( + source, + padding_mask, + mask=mask, + features_only=True, + layer=layer, + corpus_key=corpus_key, + ) + return res + + def get_logits(self, net_output): + logits = net_output["x"] + logits = logits.transpose(0, 2) + logits = logits.reshape(-1, logits.size(-1)) + return logits + + def get_targets(self, sample, net_output, expand_steps=True): + x = net_output["x"] + return x.new_zeros(x.size(1) * x.size(2), dtype=torch.long) + + def get_extra_losses(self, net_output): + pen = [] + + if "prob_perplexity" in net_output: + pen.append( + (net_output["num_vars"] - net_output["prob_perplexity"]) + / net_output["num_vars"] + ) + + if "features_pen" in net_output: + pen.append(net_output["features_pen"]) + + return pen + + def remove_pretraining_modules(self, last_layer=None): + self.quantizer = None + self.project_q = None + self.target_glu = None + self.final_proj = None + + if last_layer is not None: + self.encoder.layers = nn.ModuleList( + l for i, l in enumerate(self.encoder.layers) if i <= last_layer + ) + + +class ConvFeatureExtractionModel(nn.Module): + def __init__( + self, + conv_layers: List[Tuple[int, int, int]], + dropout: float = 0.0, + mode: str = "default", + conv_bias: bool = False, + ): + super().__init__() + + assert mode in {"default", "layer_norm"} + + def block( + n_in, + n_out, + k, + stride, + is_layer_norm=False, + is_group_norm=False, + conv_bias=False, + ): + def make_conv(): + conv = nn.Conv1d(n_in, n_out, k, stride=stride, bias=conv_bias) + nn.init.kaiming_normal_(conv.weight) + return conv + + assert ( + is_layer_norm and is_group_norm + ) == False, "layer norm and group norm are exclusive" + + if is_layer_norm: + return nn.Sequential( + make_conv(), + nn.Dropout(p=dropout), + nn.Sequential( + TransposeLast(), + Fp32LayerNorm(dim, elementwise_affine=True), + TransposeLast(), + ), + nn.GELU(), + ) + elif is_group_norm: + return nn.Sequential( + make_conv(), + nn.Dropout(p=dropout), + Fp32GroupNorm(dim, dim, affine=True), + nn.GELU(), + ) + else: + return nn.Sequential(make_conv(), nn.Dropout(p=dropout), nn.GELU()) + + in_d = 1 + self.conv_layers = nn.ModuleList() + for i, cl in enumerate(conv_layers): + assert len(cl) == 3, "invalid conv definition: " + str(cl) + (dim, k, stride) = cl + + self.conv_layers.append( + block( + in_d, + dim, + k, + stride, + is_layer_norm=mode == "layer_norm", + is_group_norm=mode == "default" and i == 0, + conv_bias=conv_bias, + ) + ) + in_d = dim + + def forward(self, x): + + # BxT -> BxCxT + x = x.unsqueeze(1) + + for conv in self.conv_layers: + x = conv(x) + + return x + + +def make_conv_pos(e, k, g, is_batch_norm=False): + pos_conv = nn.Conv1d( + e, + e, + kernel_size=k, + padding=k // 2, + groups=g, + ) + dropout = 0 + std = math.sqrt((4 * (1.0 - dropout)) / (k * e)) + nn.init.normal_(pos_conv.weight, mean=0, std=std) + nn.init.constant_(pos_conv.bias, 0) + + if not is_batch_norm: + pos_conv = nn.utils.weight_norm(pos_conv, name="weight", dim=2) + pos_conv = nn.Sequential(pos_conv, SamePad(k), nn.GELU()) + else: + batch_norm = nn.BatchNorm1d(e) + pos_conv = nn.Sequential(batch_norm, pos_conv, SamePad(k), nn.GELU()) + + return pos_conv + + +class TransformerEncoder(nn.Module): + def build_encoder_layer(self, args: Wav2Vec2Config, **kwargs): + if args.layer_type == "transformer": + layer = TransformerSentenceEncoderLayer( + embedding_dim=self.embedding_dim, + ffn_embedding_dim=args.encoder_ffn_embed_dim, + num_attention_heads=args.encoder_attention_heads, + dropout=self.dropout, + attention_dropout=args.attention_dropout, + activation_dropout=args.activation_dropout, + activation_fn=args.activation_fn, + layer_norm_first=args.layer_norm_first, + ) + elif args.layer_type == "conformer": + layer = ConformerWav2Vec2EncoderLayer( + embed_dim=self.embedding_dim, + ffn_embed_dim=args.encoder_ffn_embed_dim, + attention_heads=args.encoder_attention_heads, + dropout=args.dropout, + depthwise_conv_kernel_size=args.depthwise_conv_kernel_size, + activation_fn="swish", + attn_type=args.attn_type, + use_fp16=args.fp16, + pos_enc_type="abs", + ) + elif args.layer_type == "trf_adp": + use_adp = False + if args.adp_trf_idx == "all": + use_adp = True + else: + adp_trf_idx = list(range(*[int(g) for g in args.adp_trf_idx.split(":")])) + if kwargs.get("layer_idx", None) in adp_trf_idx: + use_adp = True + if use_adp: + layer = TransformerSentenceEncoderWithAdapterLayer( + embedding_dim=self.embedding_dim, + ffn_embedding_dim=args.encoder_ffn_embed_dim, + num_attention_heads=args.encoder_attention_heads, + dropout=self.dropout, + attention_dropout=args.attention_dropout, + activation_dropout=args.activation_dropout, + activation_fn=args.activation_fn, + layer_norm_first=args.layer_norm_first, + adapter_num=args.adp_num, + adapter_dim=args.adp_dim, + adapter_act_fn=args.adp_act_fn, + ) + else: + layer = TransformerSentenceEncoderLayer( + embedding_dim=self.embedding_dim, + ffn_embedding_dim=args.encoder_ffn_embed_dim, + num_attention_heads=args.encoder_attention_heads, + dropout=self.dropout, + attention_dropout=args.attention_dropout, + activation_dropout=args.activation_dropout, + activation_fn=args.activation_fn, + layer_norm_first=args.layer_norm_first, + ) + + layer = fsdp_wrap(layer) + if args.checkpoint_activations: + layer = checkpoint_wrapper(layer) + return layer + + def __init__(self, args: Wav2Vec2Config, skip_pos_conv: bool = False, override_encoder_layer: int = None): + super().__init__() + + self.dropout = args.dropout + self.embedding_dim = args.encoder_embed_dim + self.required_seq_len_multiple = args.required_seq_len_multiple + + pos_conv_depth = getattr(args, "pos_conv_depth", 1) + if pos_conv_depth > 1: + num_layers = args.pos_conv_depth + k = max(3, args.conv_pos // num_layers) + + def make_conv_block(e, k, g, l): + return nn.Sequential( + *[ + nn.Sequential( + nn.Conv1d( + e, + e, + kernel_size=k, + padding=k // 2, + groups=g, + ), + SamePad(k), + TransposeLast(), + LayerNorm(e, elementwise_affine=False), + TransposeLast(), + nn.GELU(), + ) + for _ in range(l) + ] + ) + + self.pos_conv = make_conv_block( + self.embedding_dim, k, args.conv_pos_groups, num_layers + ) + elif skip_pos_conv: + self.pos_conv = None + else: + self.pos_conv = make_conv_pos( + self.embedding_dim, + args.conv_pos, + args.conv_pos_groups, + is_batch_norm=args.conv_pos_batch_norm + if hasattr(args, "conv_pos_batch_norm") + else False, + ) + + if override_encoder_layer is None: + encoder_layers = args.encoder_layers + else: + encoder_layers = override_encoder_layer + + self.layers = nn.ModuleList( + [self.build_encoder_layer(args, layer_idx=ii) for ii in range(encoder_layers)] + ) + self.layer_norm_first = args.layer_norm_first + self.layer_norm = LayerNorm(self.embedding_dim) + self.layerdrop = args.encoder_layerdrop + + self.apply(init_bert_params) + + def forward(self, x, padding_mask=None, layer=None, corpus_key=None): + x, layer_results = self.extract_features( + x, padding_mask, layer, corpus_key=corpus_key + ) + + if self.layer_norm_first and layer is None: + x = self.layer_norm(x) + + return x, layer_results + + def extract_features( + self, + x, + padding_mask=None, + tgt_layer=None, + min_layer=0, + corpus_key=None, + ): + + if padding_mask is not None: + x = index_put(x, padding_mask, 0) + + if self.pos_conv is not None: + x_conv = self.pos_conv(x.transpose(1, 2)) + x_conv = x_conv.transpose(1, 2) + x = x + x_conv + + if not self.layer_norm_first: + x = self.layer_norm(x) + + # pad to the sequence length dimension + x, pad_length = pad_to_multiple( + x, self.required_seq_len_multiple, dim=-2, value=0 + ) + if pad_length > 0 and padding_mask is None: + padding_mask = x.new_zeros((x.size(0), x.size(1)), dtype=torch.bool) + padding_mask[:, -pad_length:] = True + else: + padding_mask, _ = pad_to_multiple( + padding_mask, self.required_seq_len_multiple, dim=-1, value=True + ) + x = F.dropout(x, p=self.dropout, training=self.training) + + # B x T x C -> T x B x C + x = x.transpose(0, 1) + + layer_results = [] + r = None + + for i, layer in enumerate(self.layers): + dropout_probability = np.random.random() if self.layerdrop > 0 else 1 + if not self.training or (dropout_probability > self.layerdrop): + layer_check = layer + if isinstance(layer, FullyShardedDataParallel): + layer_check = layer.unwrapped_module + if (corpus_key is None) or ( + not isinstance(layer_check, ( + TransformerSentenceEncoderWithAdapterLayer, + ) + ) + ): + x, (z, lr) = layer( + x, self_attn_padding_mask=padding_mask, need_weights=False + ) + else: + x, (z, lr) = layer( + x, + self_attn_padding_mask=padding_mask, + need_weights=False, + corpus_key=corpus_key, + ) + if i >= min_layer: + layer_results.append((x, z, lr)) + if i == tgt_layer: + r = x + break + + if r is not None: + x = r + + # T x B x C -> B x T x C + x = x.transpose(0, 1) + + # undo paddding + if pad_length > 0: + x = x[:, :-pad_length] + + def undo_pad(a, b, c): + return ( + a[:-pad_length], + b[:-pad_length] if b is not None else b, + c[:-pad_length], + ) + + layer_results = [undo_pad(*u) for u in layer_results] + + return x, layer_results + + def max_positions(self): + """Maximum output length supported by the encoder.""" + return self.args.max_positions + + def upgrade_state_dict_named(self, state_dict, name): + """Upgrade a (possibly old) state dict for new versions of fairseq.""" + return state_dict + + +class ConformerEncoder(TransformerEncoder): + def build_encoder_layer(self, args): + layer = ConformerWav2Vec2EncoderLayer( + embed_dim=self.embedding_dim, + ffn_embed_dim=args.encoder_ffn_embed_dim, + attention_heads=args.encoder_attention_heads, + dropout=args.dropout, + depthwise_conv_kernel_size=args.depthwise_conv_kernel_size, + activation_fn="swish", + attn_type=args.attn_type, + pos_enc_type=args.pos_enc_type, + use_fp16=args.fp16, # only used for rope + ) + layer = fsdp_wrap(layer) + if args.checkpoint_activations: + layer = checkpoint_wrapper(layer) + return layer + + def __init__(self, args): + super().__init__(args) + self.args = args + self.dropout = args.dropout + self.embedding_dim = args.encoder_embed_dim + self.pos_enc_type = args.pos_enc_type + max_source_positions = self.max_positions() + + if self.pos_enc_type == "rel_pos": + self.embed_positions = RelPositionalEncoding( + max_source_positions, self.embedding_dim + ) + elif self.pos_enc_type == "rope": + self.embed_positions = None + else: + raise Exception("Unsupported positional encoding type") + + self.layers = nn.ModuleList( + [self.build_encoder_layer(args) for _ in range(args.encoder_layers)] + ) + self.layer_norm_first = args.layer_norm_first + self.layer_norm = LayerNorm(self.embedding_dim) + self.layerdrop = args.encoder_layerdrop + + self.apply(init_bert_params) + + def extract_features(self, x, padding_mask=None, tgt_layer=None): + if padding_mask is not None: + x = index_put(x, padding_mask, 0) + + # B x T x C -> T x B x C + x = x.transpose(0, 1) + + # B X T X C here + position_emb = None + if self.pos_enc_type == "rel_pos": + position_emb = self.embed_positions(x) + + if not self.layer_norm_first: + x = self.layer_norm(x) + + x = F.dropout(x, p=self.dropout, training=self.training) + + layer_results = [] + r = None + for i, layer in enumerate(self.layers): + dropout_probability = np.random.random() + if not self.training or (dropout_probability > self.layerdrop): + x, z = layer( + x, + self_attn_padding_mask=padding_mask, + need_weights=False, + position_emb=position_emb, + ) + if tgt_layer is not None: + layer_results.append((x, z)) + if i == tgt_layer: + r = x + break + + if r is not None: + x = r + + # T x B x C -> B x T x C + x = x.transpose(0, 1) + + return x, layer_results + + +class TransformerSentenceEncoderLayer(nn.Module): + """ + Implements a Transformer Encoder Layer used in BERT/XLM style pre-trained + models. + """ + + def __init__( + self, + embedding_dim: float = 768, + ffn_embedding_dim: float = 3072, + num_attention_heads: int = 8, + dropout: float = 0.1, + attention_dropout: float = 0.1, + activation_dropout: float = 0.1, + activation_fn: str = "relu", + layer_norm_first: bool = False, + ) -> None: + + super().__init__() + # Initialize parameters + self.embedding_dim = embedding_dim + self.dropout = dropout + self.activation_dropout = activation_dropout + + # Initialize blocks + self.activation_fn = utils.get_activation_fn(activation_fn) + self.self_attn = MultiheadAttention( + self.embedding_dim, + num_attention_heads, + dropout=attention_dropout, + self_attention=True, + ) + + self.dropout1 = nn.Dropout(dropout) + self.dropout2 = nn.Dropout(self.activation_dropout) + self.dropout3 = nn.Dropout(dropout) + + self.layer_norm_first = layer_norm_first + + # layer norm associated with the self attention layer + self.self_attn_layer_norm = LayerNorm(self.embedding_dim) + self.fc1 = nn.Linear(self.embedding_dim, ffn_embedding_dim) + self.fc2 = nn.Linear(ffn_embedding_dim, self.embedding_dim) + + # layer norm associated with the position wise feed-forward NN + self.final_layer_norm = LayerNorm(self.embedding_dim) + + def forward( + self, + x: torch.Tensor, + self_attn_mask: torch.Tensor = None, + self_attn_padding_mask: torch.Tensor = None, + need_weights: bool = False, + att_args=None, + ): + """ + LayerNorm is applied either before or after the self-attention/ffn + modules similar to the original Transformer imlementation. + """ + residual = x + + if self.layer_norm_first: + x = self.self_attn_layer_norm(x) + x, attn = self.self_attn( + query=x, + key=x, + value=x, + key_padding_mask=self_attn_padding_mask, + attn_mask=self_attn_mask, + need_weights=False, + ) + x = self.dropout1(x) + x = residual + x + + residual = x + x = self.final_layer_norm(x) + x = self.activation_fn(self.fc1(x)) + x = self.dropout2(x) + x = self.fc2(x) + + layer_result = x + + x = self.dropout3(x) + x = residual + x + else: + x, attn = self.self_attn( + query=x, + key=x, + value=x, + key_padding_mask=self_attn_padding_mask, + need_weights=False, + ) + + x = self.dropout1(x) + x = residual + x + + x = self.self_attn_layer_norm(x) + + residual = x + x = self.activation_fn(self.fc1(x)) + x = self.dropout2(x) + x = self.fc2(x) + + layer_result = x + + x = self.dropout3(x) + x = residual + x + x = self.final_layer_norm(x) + + return x, (attn, layer_result) + + +class AdapterFast(nn.Module): + def __init__(self, adapter_num, input_dim, hidden_dim, act_fn): + """ + Implements adapter modules directly with 3D tensor weight as parameters + and without using ModuleList orto speed up training throughput. + """ + super().__init__() + + self.adapter_num = adapter_num + self.input_dim = input_dim + self.hidden_dim = hidden_dim + self.W_a = nn.Parameter(torch.empty(adapter_num, hidden_dim, input_dim)) + self.W_b = nn.Parameter(torch.empty(adapter_num, input_dim, hidden_dim)) + self.b_a = nn.Parameter(torch.empty(adapter_num, hidden_dim)) + self.b_b = nn.Parameter(torch.empty(adapter_num, input_dim)) + + self.ln_W = nn.Parameter(torch.empty(adapter_num, input_dim)) + self.ln_b = nn.Parameter(torch.empty(adapter_num, input_dim)) + self.act_fn = nn.Identity() + if act_fn == "relu": + self.act_fn = nn.ReLU() + elif act_fn == "gelu": + self.act_fn = nn.GELU() + elif act_fn == "selu": + self.act_fn = nn.SELU() + else: + raise ValueError(f"unsupported {act_fn}") + + + self.input_dim = input_dim + self.reset_parameters() + + def reset_parameters(self): + for ii in range(self.adapter_num): + nn.init.kaiming_uniform_(self.W_a[ii], a=math.sqrt(5)) + nn.init.kaiming_uniform_(self.W_b[ii], a=math.sqrt(5)) + fan_in, _ = nn.init._calculate_fan_in_and_fan_out(self.W_a[ii]) + bound = 1 / math.sqrt(fan_in) if fan_in > 0 else 0 + nn.init.uniform_(self.b_a[ii], -bound, bound) + fan_in, _ = nn.init._calculate_fan_in_and_fan_out(self.W_b[ii]) + bound = 1 / math.sqrt(fan_in) if fan_in > 0 else 0 + nn.init.uniform_(self.b_b[ii], -bound, bound) + + nn.init.ones_(self.ln_W) + nn.init.zeros_(self.ln_b) + + def forward(self, x, adapter_id): + ii = adapter_id + h = x + h = F.layer_norm(h, (self.input_dim, ), self.ln_W[ii], self.ln_b[ii]) + h = F.linear(h, self.W_a[ii], self.b_a[ii]) + h = self.act_fn(h) + h = F.linear(h, self.W_b[ii], self.b_b[ii]) + outputs = h + return outputs + + def extra_repr(self): + return ('adapter={}, input_dim={}, hidden_dim={}'.format(self.adapter_num, self.input_dim, self.hidden_dim)) + + + +class TransformerSentenceEncoderWithAdapterLayer(TransformerSentenceEncoderLayer): + """ + Implements a Transformer Encoder Layer with adapters used in BERT/XLM style pre-trained + models. An adapter module is added along with vanilla Transformer module. + """ + + def __init__( + self, + embedding_dim: float = 768, + ffn_embedding_dim: float = 3072, + num_attention_heads: int = 8, + dropout: float = 0.1, + attention_dropout: float = 0.1, + activation_dropout: float = 0.1, + activation_fn: str = "relu", + layer_norm_first: bool = False, + adapter_num=201, + adapter_dim=64, + adapter_act_fn="relu", + ) -> None: + + super().__init__( + embedding_dim=embedding_dim, + ffn_embedding_dim=ffn_embedding_dim, + num_attention_heads=num_attention_heads, + dropout=dropout, + attention_dropout=attention_dropout, + activation_dropout=activation_dropout, + activation_fn=activation_fn, + layer_norm_first=layer_norm_first, + + ) + + self.adapter_num = adapter_num + self.adapter_dim = adapter_dim + self.adapter_layer = AdapterFast(adapter_num, self.embedding_dim, self.adapter_dim, adapter_act_fn) + + def forward( + self, + x: torch.Tensor, + self_attn_mask: torch.Tensor = None, + self_attn_padding_mask: torch.Tensor = None, + need_weights: bool = False, + att_args=None, + corpus_key=None, + ): + + x, (attn, layer_result) = super().forward( + x=x, + self_attn_mask=self_attn_mask, + self_attn_padding_mask=self_attn_padding_mask, + need_weights=need_weights, + att_args=att_args, + ) + assert corpus_key is not None + assert len(set(corpus_key)) == 1, f"corpus_key items are not same {corpus_key}" + y = self.adapter_layer(x, corpus_key[0]) + x = x + y + return x, (attn, layer_result) diff --git a/data/fairseq/fairseq/models/wav2vec/wav2vec2_asr.py b/data/fairseq/fairseq/models/wav2vec/wav2vec2_asr.py new file mode 100644 index 0000000000000000000000000000000000000000..0403efebb9b4b48c4d8f518e127e7663395329e8 --- /dev/null +++ b/data/fairseq/fairseq/models/wav2vec/wav2vec2_asr.py @@ -0,0 +1,878 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import contextlib +import copy +import logging +import math +import re +from argparse import Namespace +from dataclasses import dataclass, field +from typing import Any, Optional + +import numpy as np +import torch +import torch.nn as nn +import torch.nn.functional as F +from omegaconf import II, MISSING, open_dict + +from fairseq import checkpoint_utils, tasks, utils +from fairseq.dataclass import FairseqDataclass +from fairseq.dataclass.utils import convert_namespace_to_omegaconf +from fairseq.models import ( + BaseFairseqModel, + FairseqEncoder, + FairseqEncoderDecoderModel, + FairseqIncrementalDecoder, + register_model, +) +from fairseq.models.wav2vec.wav2vec2 import MASKING_DISTRIBUTION_CHOICES, LAYER_TYPE_CHOICES, AdapterFast +from fairseq.modules import LayerNorm, PositionalEmbedding, TransformerDecoderLayer +from fairseq.tasks import FairseqTask + +logger = logging.getLogger(__name__) + + +@dataclass +class Wav2Vec2AsrConfig(FairseqDataclass): + w2v_path: str = field( + default=MISSING, metadata={"help": "path to wav2vec 2.0 model"} + ) + no_pretrained_weights: bool = field( + default=False, metadata={"help": "if true, does not load pretrained weights"} + ) + dropout_input: float = field( + default=0.0, + metadata={"help": "dropout to apply to the input (after feat extr)"}, + ) + + final_dropout: float = field( + default=0.0, + metadata={"help": "dropout after transformer and before final projection"}, + ) + dropout: float = field( + default=0.0, metadata={"help": "dropout probability inside wav2vec 2.0 model"} + ) + attention_dropout: float = field( + default=0.0, + metadata={ + "help": "dropout probability for attention weights inside wav2vec 2.0 model" + }, + ) + activation_dropout: float = field( + default=0.0, + metadata={ + "help": "dropout probability after activation in FFN inside wav2vec 2.0 model" + }, + ) + + # masking + apply_mask: bool = field( + default=False, metadata={"help": "apply masking during fine-tuning"} + ) + mask_length: int = field( + default=10, metadata={"help": "repeat the mask indices multiple times"} + ) + mask_prob: float = field( + default=0.5, + metadata={ + "help": "probability of replacing a token with mask (normalized by length)" + }, + ) + mask_selection: MASKING_DISTRIBUTION_CHOICES = field( + default="static", metadata={"help": "how to choose masks"} + ) + mask_other: float = field( + default=0, + metadata={ + "help": "secondary mask argument (used for more complex distributions), " + "see help in compute_mask_indices" + }, + ) + no_mask_overlap: bool = field( + default=False, metadata={"help": "whether to allow masks to overlap"} + ) + mask_min_space: Optional[int] = field( + default=1, + metadata={"help": "min space between spans (if no overlap is enabled)"}, + ) + require_same_masks: bool = field( + default=True, + metadata={ + "help": "whether to number of masked timesteps must be the same across all " + "examples in a batch" + }, + ) + mask_dropout: float = field( + default=0.0, + metadata={"help": "percent of masks to unmask for each sample"}, + ) + + # channel masking + mask_channel_length: int = field( + default=10, metadata={"help": "length of the mask for features (channels)"} + ) + mask_channel_prob: float = field( + default=0.0, metadata={"help": "probability of replacing a feature with 0"} + ) + mask_channel_selection: MASKING_DISTRIBUTION_CHOICES = field( + default="static", + metadata={"help": "how to choose mask length for channel masking"}, + ) + mask_channel_other: float = field( + default=0, + metadata={ + "help": "secondary mask argument (used for more complex distributions), " + "see help in compute_mask_indicesh" + }, + ) + no_mask_channel_overlap: bool = field( + default=False, metadata={"help": "whether to allow channel masks to overlap"} + ) + freeze_finetune_updates: int = field( + default=0, metadata={"help": "dont finetune wav2vec for this many updates"} + ) + feature_grad_mult: float = field( + default=0.0, metadata={"help": "reset feature grad mult in wav2vec 2.0 to this"} + ) + layerdrop: float = field( + default=0.0, metadata={"help": "probability of dropping a layer in wav2vec 2.0"} + ) + drop_path: float = 0 + mask_channel_min_space: Optional[int] = field( + default=1, + metadata={"help": "min space between spans (if no overlap is enabled)"}, + ) + mask_channel_before: bool = False + normalize: bool = II("task.normalize") + update_alibi: bool = True + data: str = II("task.data") + # this holds the loaded wav2vec args + w2v_args: Any = None + offload_activations: bool = field( + default=False, metadata={"help": "offload_activations"} + ) + min_params_to_wrap: int = field( + default=int(1e8), + metadata={ + "help": "minimum number of params for a layer to be wrapped with FSDP() when " + "training with --ddp-backend=fully_sharded. Smaller values will " + "improve memory efficiency, but may make torch.distributed " + "communication less efficient due to smaller input sizes. This option " + "is set to 0 (i.e., always wrap) when --checkpoint-activations or " + "--offload-activations are passed." + }, + ) + + checkpoint_activations: bool = field( + default=False, + metadata={"help": "recompute activations and save memory for extra compute"}, + ) + ddp_backend: str = II("distributed_training.ddp_backend") + + zero_mask: bool = False + load_ema: bool = False + + layer_decay: float = 1 + + + layer_type: LAYER_TYPE_CHOICES = field( + default="transformer", metadata={"help": "layer type in encoder"} + ) + # Adapter num + adp_num: int = field( + default=-1 + ) + adp_dim: int = field( + default=64 + ) + adp_act_fn: str = field( + default="relu" + ) + adp_trf_idx: str = field( + default="all", + ) + + freeze_regex: Optional[str] = field( + default=None, + ) + +@dataclass +class Wav2Vec2CtcConfig(Wav2Vec2AsrConfig): + blank_weight: float = 0 + blank_mode: str = "add" + + +@register_model("wav2vec_ctc", dataclass=Wav2Vec2CtcConfig) +class Wav2VecCtc(BaseFairseqModel): + def __init__(self, cfg: Wav2Vec2CtcConfig, w2v_encoder: BaseFairseqModel): + super().__init__() + self.cfg = cfg + self.w2v_encoder = w2v_encoder + self.blank_weight = cfg.blank_weight + self.blank_mode = cfg.blank_mode + + def upgrade_state_dict_named(self, state_dict, name): + super().upgrade_state_dict_named(state_dict, name) + return state_dict + + @classmethod + def build_model(cls, cfg: Wav2Vec2CtcConfig, task: FairseqTask): + """Build a new model instance.""" + w2v_encoder = Wav2VecEncoder(cfg, len(task.target_dictionary)) + return cls(cfg, w2v_encoder) + + def get_logits(self, net_output, normalize=False): + logits = net_output["encoder_out"] + if self.blank_weight != 0: + if self.blank_mode == "add": + logits[..., 0] += self.blank_weight + elif self.blank_mode == "set": + logits[..., 0] = self.blank_weight + else: + raise Exception(f"invalid blank mode {self.blank_mode}") + + if net_output["padding_mask"] is not None and net_output["padding_mask"].any(): + number_of_classes = logits.size(-1) + masking_tensor = torch.ones( + number_of_classes, device=logits.device + ) * float("-inf") + masking_tensor[0] = 0 + + if logits.size(0) > net_output["padding_mask"].size(1): + net_output["padding_mask"] = F.pad( + net_output["padding_mask"], (1, 0), value=False + ) + + logits[net_output["padding_mask"].T] = masking_tensor.type_as(logits) + + if normalize: + logits = utils.log_softmax(logits.float(), dim=-1) + + return logits + + def get_normalized_probs(self, net_output, log_probs): + """Get normalized probabilities (or log probs) from a net's output.""" + + logits = self.get_logits(net_output) + + if log_probs: + return utils.log_softmax(logits.float(), dim=-1) + else: + return utils.softmax(logits.float(), dim=-1) + + def forward(self, **kwargs): + x = self.w2v_encoder(**kwargs) + return x + + +@dataclass +class Wav2Vec2Seq2SeqConfig(Wav2Vec2AsrConfig): + decoder_embed_dim: int = field( + default=768, metadata={"help": "decoder embedding dimension"} + ) + decoder_ffn_embed_dim: int = field( + default=3072, metadata={"help": "decoder embedding dimension for FFN"} + ) + decoder_layers: int = field(default=6, metadata={"help": "num of decoder layers"}) + decoder_layerdrop: float = field( + default=0.0, metadata={"help": "decoder layerdrop chance"} + ) + decoder_attention_heads: int = field( + default=4, metadata={"help": "num decoder attention heads"} + ) + decoder_learned_pos: bool = field( + default=False, + metadata={"help": "use learned positional embeddings in the decoder"}, + ) + decoder_normalize_before: bool = field( + default=False, metadata={"help": "apply layernorm before each decoder block"} + ) + no_token_positional_embeddings: bool = field( + default=False, + metadata={ + "help": "if set, disables positional embeddings (outside self attention)" + }, + ) + decoder_dropout: float = field( + default=0.0, metadata={"help": "dropout probability in the decoder"} + ) + decoder_attention_dropout: float = field( + default=0.0, + metadata={ + "help": "dropout probability for attention weights inside the decoder" + }, + ) + decoder_activation_dropout: float = field( + default=0.0, + metadata={ + "help": "dropout probability after activation in FFN inside the decoder" + }, + ) + max_target_positions: int = field( + default=2048, metadata={"help": "max target positions"} + ) + share_decoder_input_output_embed: bool = field( + default=False, metadata={"help": "share decoder input and output embeddings"} + ) + autoregressive: bool = II("task.autoregressive") + + +@register_model("wav2vec_seq2seq", dataclass=Wav2Vec2Seq2SeqConfig) +class Wav2Vec2Seq2SeqModel(FairseqEncoderDecoderModel): + def __init__(self, encoder, decoder): + super().__init__(encoder, decoder) + + @classmethod + def build_model(cls, cfg: Wav2Vec2Seq2SeqConfig, task: FairseqTask): + """Build a new model instance.""" + + assert ( + cfg.autoregressive + ), "Please set task.autoregressive=true for seq2seq asr models" + + src_dict, tgt_dict = task.source_dictionary, task.target_dictionary + + def build_embedding(dictionary, embed_dim): + num_embeddings = len(dictionary) + padding_idx = dictionary.pad() + emb = Embedding(num_embeddings, embed_dim, padding_idx) + return emb + + decoder_embed_tokens = build_embedding(tgt_dict, cfg.decoder_embed_dim) + + encoder = cls.build_encoder(cfg) + decoder = cls.build_decoder(cfg, tgt_dict, decoder_embed_tokens) + + return Wav2Vec2Seq2SeqModel(encoder, decoder) + + @classmethod + def build_encoder(cls, cfg: Wav2Vec2AsrConfig): + return Wav2VecEncoder(cfg) + + @classmethod + def build_decoder(cls, cfg: Wav2Vec2Seq2SeqConfig, tgt_dict, embed_tokens): + return TransformerDecoder(cfg, tgt_dict, embed_tokens) + + def forward(self, **kwargs): + encoder_out = self.encoder(**kwargs) + decoder_out = self.decoder(encoder_out=encoder_out, **kwargs) + return decoder_out + + def upgrade_state_dict_named(self, state_dict, name): + super().upgrade_state_dict_named(state_dict, name) + return state_dict + + +class Wav2VecEncoder(FairseqEncoder): + def __init__(self, cfg: Wav2Vec2AsrConfig, output_size=None): + self.apply_mask = cfg.apply_mask + + arg_overrides = { + "dropout": cfg.dropout, + "activation_dropout": cfg.activation_dropout, + "dropout_input": cfg.dropout_input, + "attention_dropout": cfg.attention_dropout, + "mask_length": cfg.mask_length, + "mask_prob": cfg.mask_prob, + "require_same_masks": getattr(cfg, "require_same_masks", True), + "pct_holes": getattr(cfg, "mask_dropout", 0), + "mask_selection": cfg.mask_selection, + "mask_other": cfg.mask_other, + "no_mask_overlap": cfg.no_mask_overlap, + "mask_channel_length": cfg.mask_channel_length, + "mask_channel_prob": cfg.mask_channel_prob, + "mask_channel_before": cfg.mask_channel_before, + "mask_channel_selection": cfg.mask_channel_selection, + "mask_channel_other": cfg.mask_channel_other, + "no_mask_channel_overlap": cfg.no_mask_channel_overlap, + "encoder_layerdrop": cfg.layerdrop, + "feature_grad_mult": cfg.feature_grad_mult, + "checkpoint_activations": cfg.checkpoint_activations, + "offload_activations": cfg.offload_activations, + "min_params_to_wrap": cfg.min_params_to_wrap, + # d2v multi args + "encoder_dropout": cfg.dropout, + "drop_path": getattr(cfg, "drop_path", 0), + "mask_dropout": getattr(cfg, "mask_dropout", 0), + "zero_mask": getattr(cfg, "zero_mask", False), + "local_grad_mult": cfg.feature_grad_mult, + "layerdrop": cfg.layerdrop, + "prenet_layerdrop": cfg.layerdrop, + "prenet_dropout": cfg.dropout, + "post_mlp_drop": cfg.dropout, + "encoder_zero_mask": getattr(cfg, "zero_mask", False), + "inverse_mask": False, + "learned_alibi_scale": getattr(cfg, "update_alibi", True), + } + + if cfg.w2v_args is None: + state = checkpoint_utils.load_checkpoint_to_cpu(cfg.w2v_path, arg_overrides) + w2v_args = state.get("cfg", None) + if w2v_args is None: + w2v_args = convert_namespace_to_omegaconf(state["args"]) + w2v_args.criterion = None + w2v_args.lr_scheduler = None + + cfg.w2v_args = w2v_args + + logger.info(w2v_args) + + else: + state = None + w2v_args = cfg.w2v_args + if isinstance(w2v_args, Namespace): + cfg.w2v_args = w2v_args = convert_namespace_to_omegaconf(w2v_args) + + self.is_d2v_multi = "data2vec_multi" in w2v_args.model.get("_name", None) + + if not self.is_d2v_multi: + model_normalized = w2v_args.task.get( + "normalize", w2v_args.model.get("normalize", False) + ) + assert cfg.normalize == model_normalized, ( + "Fine-tuning works best when data normalization is the same. " + "Please check that --normalize is set or unset for both pre-training and here" + ) + + with open_dict(w2v_args): + args_replacement = ["checkpoint_activations", "layer_type", + "adp_num", "adp_dim", + "adp_act_fn", "adp_trf_idx"] + for _args in args_replacement: + if hasattr(cfg, _args) and getattr(cfg, _args, None) is not None: + w2v_args.model[_args] = getattr(cfg, _args, None) + + if hasattr(cfg, "checkpoint_activations") and cfg.checkpoint_activations: + with open_dict(w2v_args): + w2v_args.model.checkpoint_activations = cfg.checkpoint_activations + + w2v_args.task.data = cfg.data + task = tasks.setup_task(w2v_args.task, from_checkpoint=True) + model = task.build_model(w2v_args.model, from_checkpoint=True) + model.remove_pretraining_modules() + d = w2v_args.model.encoder_embed_dim + else: + assert cfg.normalize + + if hasattr(w2v_args.task, "audio"): + w2v_args.task.audio.data = cfg.data + else: + w2v_args.task.data = cfg.data + task = tasks.setup_task(w2v_args.task, from_checkpoint=True) + + model = task.build_model(w2v_args.model, from_checkpoint=True) + + model.remove_pretraining_modules(modality="audio") + d = w2v_args.model.embed_dim + + if state is not None and not cfg.no_pretrained_weights: + if cfg.load_ema: + assert "_ema" in state["model"] + for k in state["model"]["_ema"]: + mk = "encoder." + k + assert mk in state["model"], mk + state["model"][mk] = state["model"]["_ema"][k] + self.load_model_weights(state, model, cfg) + + super().__init__(task.source_dictionary) + + self.w2v_model = model + + self.final_dropout = nn.Dropout(cfg.final_dropout) + self.freeze_finetune_updates = cfg.freeze_finetune_updates + self.num_updates = 0 + + targ_d = None + self.proj = None + + if output_size is not None: + targ_d = output_size + elif getattr(cfg, "decoder_embed_dim", d) != d: + targ_d = cfg.decoder_embed_dim + + if targ_d is not None: + self.proj = Linear(d, targ_d) + + if cfg.freeze_regex is not None: + self.freeze_regex(cfg.freeze_regex) + + layer_decay = getattr(cfg, "layer_decay", 1) + if layer_decay < 1: + mod_encs = list(model.modality_encoders.values()) + assert len(mod_encs) == 1, len(mod_encs) + blocks = list(mod_encs[0].context_encoder.blocks) + list(model.blocks) + num_layers = len(blocks) + 1 + layer_scales = list( + layer_decay ** (num_layers - i) for i in range(num_layers + 1) + ) + + for i, b in enumerate(blocks): + lid = i + 1 + if layer_scales[lid] == 1.0: + continue + + for n, p in b.named_parameters(): + optim_override = getattr(p, "optim_overrides", {}) + if "optimizer" not in optim_override: + optim_override["optimizer"] = {} + + optim_override["optimizer"]["lr_scale"] = layer_scales[lid] + p.optim_overrides = optim_override + + def freeze_regex(self, pattern): + unfrozen_names = [] + for name, param in self.named_parameters(): + if re.fullmatch(pattern, name) is not None: + param.requires_grad_(False) + else: + unfrozen_names.append(name) + + def load_model_weights(self, state, model, cfg): + if cfg.ddp_backend == "fully_sharded": + from fairseq.distributed import FullyShardedDataParallel + + for name, module in model.named_modules(): + if "encoder.layers" in name and len(name.split(".")) == 3: + # Only for layers, we do a special handling and load the weights one by one + # We dont load all weights together as that wont be memory efficient and may + # cause oom + new_dict = { + k.replace(name + ".", ""): v + for (k, v) in state["model"].items() + if name + "." in k + } + assert isinstance(module, FullyShardedDataParallel) + with module.summon_full_params(): + module.load_state_dict(new_dict, strict=True) + module._reset_lazy_init() + + # Once layers are loaded, filter them out and load everything else. + r = re.compile("encoder.layers.\d.") + filtered_list = list(filter(r.match, state["model"].keys())) + + new_big_dict = { + k: v for (k, v) in state["model"].items() if k not in filtered_list + } + + model.load_state_dict(new_big_dict, strict=False) + else: + to_delete = {"_ema", "target_proj", "decoder"} + for k in to_delete: + if k in state["model"]: + del state["model"][k] + + if hasattr(model, "modality_encoders"): + if "modality_encoders.AUDIO.encoder_mask" not in state["model"]: + model.modality_encoders["AUDIO"].encoder_mask = None + elif not cfg.zero_mask: + model.modality_encoders["AUDIO"].encoder_mask = None + del state["model"]["modality_encoders.AUDIO.encoder_mask"] + + for k in list(state["model"].keys()): + if k.startswith("modality_encoders.") and not k.startswith( + "modality_encoders.AUDIO" + ): + del state["model"][k] + + print(model) + model.load_state_dict(state["model"], strict=True) + + def set_num_updates(self, num_updates): + """Set the number of parameters updates.""" + super().set_num_updates(num_updates) + self.num_updates = num_updates + + def forward(self, source, padding_mask, **kwargs): + + w2v_args = { + "source": source, + "padding_mask": padding_mask, + "mask": self.apply_mask and self.training, + } + if "corpus_key" in kwargs: + w2v_args["corpus_key"] = kwargs["corpus_key"] + + if self.is_d2v_multi: + w2v_args["mode"] = "AUDIO" + + ft = self.freeze_finetune_updates <= self.num_updates + + with torch.no_grad() if not ft else contextlib.ExitStack(): + res = self.w2v_model.extract_features(**w2v_args) + + x = res["x"] + padding_mask = res["padding_mask"] + + # B x T x C -> T x B x C + x = x.transpose(0, 1) + + x = self.final_dropout(x) + + if self.proj: + x = self.proj(x) + + return { + "encoder_out": x, # T x B x C + "padding_mask": padding_mask, # B x T, + "layer_results": res["layer_results"], + } + + def forward_torchscript(self, net_input): + if torch.jit.is_scripting(): + return self.forward(net_input["source"], net_input["padding_mask"]) + else: + return self.forward_non_torchscript(net_input) + + def reorder_encoder_out(self, encoder_out, new_order): + if encoder_out["encoder_out"] is not None: + encoder_out["encoder_out"] = encoder_out["encoder_out"].index_select( + 1, new_order + ) + if encoder_out["padding_mask"] is not None: + encoder_out["padding_mask"] = encoder_out["padding_mask"].index_select( + 0, new_order + ) + return encoder_out + + def max_positions(self): + """Maximum input length supported by the encoder.""" + return None + + def upgrade_state_dict_named(self, state_dict, name): + return state_dict + + +class TransformerDecoder(FairseqIncrementalDecoder): + """ + Transformer decoder consisting of *args.decoder_layers* layers. Each layer + is a :class:`TransformerDecoderLayer`. + + Args: + args (argparse.Namespace): parsed command-line arguments + dictionary (~fairseq.data.Dictionary): decoding dictionary + embed_tokens (torch.nn.Embedding): output embedding + no_encoder_attn (bool, optional): whether to attend to encoder outputs + (default: False). + """ + + def __init__( + self, + cfg: Wav2Vec2Seq2SeqConfig, + dictionary, + embed_tokens, + no_encoder_attn=False, + ): + super().__init__(dictionary) + + self.dropout = cfg.decoder_dropout + self.share_input_output_embed = cfg.share_decoder_input_output_embed + + input_embed_dim = embed_tokens.embedding_dim + embed_dim = cfg.decoder_embed_dim + self.output_embed_dim = cfg.decoder_embed_dim + + self.layerdrop = cfg.decoder_layerdrop + + self.padding_idx = embed_tokens.padding_idx + self.max_target_positions = cfg.max_target_positions + + self.embed_tokens = embed_tokens + self.embed_scale = math.sqrt(embed_dim) # todo: try with input_embed_dim + + self.project_in_dim = ( + Linear(input_embed_dim, embed_dim, bias=False) + if embed_dim != input_embed_dim + else None + ) + + self.embed_positions = ( + PositionalEmbedding( + cfg.max_target_positions, + embed_dim, + self.padding_idx, + learned=cfg.decoder_learned_pos, + ) + if not cfg.no_token_positional_embeddings + else None + ) + + # TODO: update this when transformer gets converted to dataclass configs + transformer_cfg = copy.deepcopy(cfg) + with open_dict(transformer_cfg): + transformer_cfg.dropout = transformer_cfg.decoder_dropout + transformer_cfg.attention_dropout = ( + transformer_cfg.decoder_attention_dropout + ) + transformer_cfg.activation_dropout = ( + transformer_cfg.decoder_activation_dropout + ) + + self.layers = nn.ModuleList([]) + self.layers.extend( + [ + TransformerDecoderLayer(transformer_cfg, no_encoder_attn) + for _ in range(transformer_cfg.decoder_layers) + ] + ) + + if not self.share_input_output_embed: + self.embed_out = nn.Parameter( + torch.Tensor(len(dictionary), self.output_embed_dim) + ) + nn.init.normal_(self.embed_out, mean=0, std=self.output_embed_dim**-0.5) + + if transformer_cfg.decoder_normalize_before: + self.layer_norm = LayerNorm(embed_dim) + else: + self.layer_norm = None + + def forward( + self, prev_output_tokens, encoder_out=None, incremental_state=None, **unused + ): + """ + Args: + prev_output_tokens (LongTensor): previous decoder outputs of shape + `(batch, tgt_len)`, for teacher forcing + encoder_out (Tensor, optional): output from the encoder, used for + encoder-side attention + incremental_state (dict): dictionary used for storing state during + :ref:`Incremental decoding` + + Returns: + tuple: + - the decoder's output of shape `(batch, tgt_len, vocab)` + - a dictionary with any model-specific outputs + """ + + if type(prev_output_tokens) == list: + max_len = max((len(x) for x in prev_output_tokens)) + tmp = torch.zeros( + [len(prev_output_tokens), max_len], device=prev_output_tokens[0].device + ) + for (i, p) in enumerate(prev_output_tokens): + tmp[i, : len(p)] = p + prev_output_tokens = tmp + + prev_output_tokens = prev_output_tokens.long() + x, extra = self.extract_features( + prev_output_tokens, encoder_out, incremental_state + ) + x = self.output_layer(x) + return x, extra + + def extract_features( + self, prev_output_tokens, encoder_out=None, incremental_state=None, **unused + ): + """ + Similar to *forward* but only return features. + + Returns: + tuple: + - the decoder's features of shape `(batch, tgt_len, embed_dim)` + - a dictionary with any model-specific outputs + """ + + # embed positions + positions = ( + self.embed_positions( + prev_output_tokens, incremental_state=incremental_state + ) + if self.embed_positions is not None + else None + ) + + if incremental_state is not None: + prev_output_tokens = prev_output_tokens[:, -1:] + if positions is not None: + positions = positions[:, -1:] + + # embed tokens and positions + x = self.embed_scale * self.embed_tokens(prev_output_tokens) + + if self.project_in_dim is not None: + x = self.project_in_dim(x) + + if positions is not None: + x += positions + x = F.dropout(x, p=self.dropout, training=self.training) + + # B x T x C -> T x B x C + x = x.transpose(0, 1) + attn = None + + inner_states = [x] + + # decoder layers + self_attn_padding_mask = None + if prev_output_tokens.eq(self.padding_idx).any(): + self_attn_padding_mask = prev_output_tokens.eq(self.padding_idx) + for layer in self.layers: + dropout_probability = np.random.random() + if not self.training or (dropout_probability > self.layerdrop): + x, attn, _ = layer( + x, + encoder_out["encoder_out"] if encoder_out is not None else None, + encoder_out["padding_mask"] if encoder_out is not None else None, + incremental_state, + self_attn_mask=self.buffered_future_mask(x) + if incremental_state is None + else None, + self_attn_padding_mask=self_attn_padding_mask, + ) + inner_states.append(x) + + if self.layer_norm: + x = self.layer_norm(x) + + # T x B x C -> B x T x C + x = x.transpose(0, 1) + + return x, {"attn": attn, "inner_states": inner_states} + + def output_layer(self, features, **kwargs): + """Project features to the vocabulary size.""" + # project back to size of vocabulary + if self.share_input_output_embed: + return F.linear(features, self.embed_tokens.weight) + else: + return F.linear(features, self.embed_out) + + def max_positions(self): + """Maximum output length supported by the decoder.""" + if self.embed_positions is None: + return self.max_target_positions + return min(self.max_target_positions, self.embed_positions.max_positions) + + def buffered_future_mask(self, tensor): + dim = tensor.size(0) + if ( + not hasattr(self, "_future_mask") + or self._future_mask is None + or self._future_mask.device != tensor.device + or self._future_mask.size(0) < dim + ): + self._future_mask = torch.triu( + utils.fill_with_neg_inf(tensor.new(dim, dim)), 1 + ) + return self._future_mask[:dim, :dim] + + def upgrade_state_dict_named(self, state_dict, name): + return state_dict + + +def Embedding(num_embeddings, embedding_dim, padding_idx): + m = nn.Embedding(num_embeddings, embedding_dim, padding_idx=padding_idx) + nn.init.normal_(m.weight, mean=0, std=embedding_dim**-0.5) + nn.init.constant_(m.weight[padding_idx], 0) + return m + + +def Linear(in_features, out_features, bias=True): + m = nn.Linear(in_features, out_features, bias) + nn.init.xavier_uniform_(m.weight) + if bias: + nn.init.constant_(m.bias, 0.0) + return m diff --git a/data/fairseq/fairseq/models/wav2vec/wav2vec2_classification.py b/data/fairseq/fairseq/models/wav2vec/wav2vec2_classification.py new file mode 100644 index 0000000000000000000000000000000000000000..c9bbaab28e16e2a8eab153b16d7f003979934d03 --- /dev/null +++ b/data/fairseq/fairseq/models/wav2vec/wav2vec2_classification.py @@ -0,0 +1,348 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import contextlib +import logging +from argparse import Namespace +from dataclasses import dataclass, field +from typing import Any, Optional + +import torch +import torch.nn as nn +import torch.nn.functional as F +from omegaconf import II, MISSING, open_dict + +from fairseq import checkpoint_utils, tasks, utils +from fairseq.dataclass import ChoiceEnum, FairseqDataclass +from fairseq.dataclass.utils import convert_namespace_to_omegaconf +from fairseq.models import BaseFairseqModel, FairseqEncoder, register_model +from fairseq.models.wav2vec.wav2vec2 import MASKING_DISTRIBUTION_CHOICES, Wav2Vec2Config +from fairseq.models.wav2vec.wav2vec2_asr import Embedding, Linear, Wav2VecEncoder, Wav2Vec2AsrConfig +from fairseq.tasks import FairseqTask + +logging.basicConfig(level=logging.DEBUG) + + +@dataclass +class Wav2Vec2ClassificationConfig(Wav2Vec2AsrConfig): + latent_embed_dim: Optional[int] = field( + default=None, metadata={"help": "latent dim (encoder w2v -> latent -> class"} + ) + pooling: str = field( + default="first_token", + metadata={"help": "pooling layer choices"}, + ) + activation_fn: ChoiceEnum(utils.get_available_activation_fns()) = field( + default="gelu", metadata={"help": "activation function to use"} + ) + + +@register_model("wav2vec_classification", dataclass=Wav2Vec2ClassificationConfig) +class Wav2VecClassification(BaseFairseqModel): + # TODO: Can be shared/merged with ASR model class as w2v_encoder params are common. + def __init__( + self, + cfg: Wav2Vec2ClassificationConfig, + w2v_encoder: BaseFairseqModel, + pooling_layer, + ): + super().__init__() + self.cfg = cfg + self.w2v_encoder = w2v_encoder + self.pooling_layer = pooling_layer + + def upgrade_state_dict_named(self, state_dict, name): + super().upgrade_state_dict_named(state_dict, name) + return state_dict + + @classmethod + def build_model(cls, cfg: Wav2Vec2ClassificationConfig, task: FairseqTask): + """Build a new model instance.""" + w2v_encoder = Wav2VecEncoder(cfg, None) + pooling_layer = get_pooling_layer( + cfg, + w2v_encoder.w2v_model.encoder.layers[-1].embedding_dim, + len(task.target_dictionary), + len(w2v_encoder.w2v_model.encoder.layers), + ) + return cls(cfg, w2v_encoder, pooling_layer) + + def get_normalized_probs(self, net_output, log_probs): + """Get normalized probabilities (or log probs) from a net's output.""" + logits = net_output + + if log_probs: + return utils.log_softmax(logits.float(), dim=-1) + else: + return utils.softmax(logits.float(), dim=-1) + + def get_logits(self, net_output): + return net_output + + def forward(self, **kwargs): + encoder_out_dict = self.w2v_encoder(**kwargs) + w2v_encoder_out = encoder_out_dict["encoder_out"] # TxBxC + w2v_encoder_padding_mask = encoder_out_dict["padding_mask"] # BxT + # w2v_encoder_layer_results = encoder_out_dict["layer_results"] + return self.pooling_layer( + last_layer_feats=w2v_encoder_out, + padding_mask=w2v_encoder_padding_mask, + # all_layer_feats=w2v_encoder_layer_results, + ) + + # def forward_latent(self, **kwargs): + # encoder_out_dict = self.w2v_encoder(**kwargs) + # w2v_encoder_out = encoder_out_dict["encoder_out"] + # w2v_encoder_padding_mask = encoder_out_dict["encoder_padding_mask"] + # w2v_encoder_layer_results = encoder_out_dict["layer_results"] + # return self.pooling_layer.forward_latent( + # last_layer_feats=w2v_encoder_out, + # padding_mask=w2v_encoder_padding_mask, + # all_layer_feats=w2v_encoder_layer_results, + # ) + + +def get_pooling_layer( + cfg: Wav2Vec2ClassificationConfig, + encoder_embed_dim: int, + num_targets: int, + encoder_layers: int, +): + assert cfg.pooling == 'mean' + if cfg.pooling == "first_token": + return FirstToken(cfg, encoder_embed_dim, num_targets) + # elif cfg.pooling == "mean": + # return MeanPooling(cfg, encoder_embed_dim, num_targets) + elif cfg.pooling == "mean": + return MeanPoolingFast(cfg, encoder_embed_dim, num_targets) + elif cfg.pooling == "mean_amsoftmax": + return MeanPoolingFastAMSoftmax(cfg, encoder_embed_dim, num_targets) + elif cfg.pooling == "max": + return MaxPoolingFast(cfg, encoder_embed_dim, num_targets) + elif cfg.pooling == "elmo": + return LayerWeightedMeanPooling( + cfg, encoder_embed_dim, num_targets, encoder_layers + ) + else: + raise NotImplementedError(f"{cfg.pooling} has not been implemented yet.") + + +class Pooling(nn.Module): + def __init__( + self, + cfg: Wav2Vec2ClassificationConfig, + encoder_embed_dim: int, + num_targets: int, + ): + super().__init__() + self.projection = Linear(encoder_embed_dim, num_targets) + + def forward(self, last_layer_feats, **kwargs): + raise NotImplementedError() + + +class FirstToken(Pooling): + def __init__(self, *args, **kwargs): + super().__init__(*args, **kwargs) + + def forward(self, last_layer_feats, **kwargs): + return self.projection(last_layer_feats[:, 0]) + + +# class MeanPooling(Pooling): +# def __init__( +# self, +# cfg: Wav2VecClassificationConfig, +# encoder_embed_dim: int, +# num_targets: int, +# **kwargs, +# ): +# super().__init__(cfg, encoder_embed_dim, num_targets) +# self.activation_fn = utils.get_activation_fn(cfg.activation_fn) +# self.linear = Linear(encoder_embed_dim, encoder_embed_dim) + +# def forward(self, last_layer_feats, padding_mask, **kwargs): +# # last_layer_feats: [BxTxD] +# # padding_mask: [BxT] +# last_layer_feats = self.linear(self.activation_fn(last_layer_feats)) +# input_lengths = (1 - padding_mask.long()).sum(-1) +# pooled_feature_list = [] +# for i in range(len(last_layer_feats)): +# length = input_lengths[i] +# pooled_feature = torch.mean(last_layer_feats[i][:length], dim=0) +# pooled_feature_list.append(pooled_feature) +# return self.projection(torch.stack(pooled_feature_list)) + + +def fn_mean(x, mask): + """ + Args: + x: TxBxD + mask: BxT + Return: + y: BxD + """ + if mask is not None: + mask = mask.t()[:, :, None] + return (x * mask).sum(0) / mask.sum(0) + else: + return x.sum(0) / x.shape[0] + + +class MeanPoolingFast(nn.Module): + def __init__( + self, + cfg: Wav2Vec2ClassificationConfig, + encoder_embed_dim: int, + num_targets: int, + **kwargs, + ): + super().__init__() + self.activation_fn = utils.get_activation_fn(cfg.activation_fn) + self.latent_embed_dim = ( + cfg.latent_embed_dim + if cfg.latent_embed_dim is not None + else encoder_embed_dim + ) + logging.debug(f"| {self.latent_embed_dim=}") + self.linear = Linear(encoder_embed_dim, self.latent_embed_dim) + self.projection = Linear(self.latent_embed_dim, num_targets) + + def forward(self, last_layer_feats, padding_mask, **kwargs): + """ + Arguments + features - [TxBxD] Acoustic feature with shape + padding_mask - [BxT] Padding Mask + """ + if padding_mask is not None: + feat_mask = (~padding_mask).to(last_layer_feats.dtype) + else: + feat_mask = None + feat = self.linear(last_layer_feats) + feat = fn_mean(feat, feat_mask) + feat = self.activation_fn(feat) + return self.projection(feat) + + def forward_latent(self, last_layer_feats, padding_mask, **kwargs): + """ + Arguments + features - [TxBxD] Acoustic feature with shape + padding_mask - [BxT] Padding Mask + """ + if padding_mask is not None: + feat_mask = (~padding_mask).to(last_layer_feats.dtype) + else: + feat_mask = None + feat = self.linear(last_layer_feats) + feat = fn_mean(feat, feat_mask) + return feat + + +class MeanPoolingFastAMSoftmax(MeanPoolingFast): + def __init__( + self, + cfg: Wav2Vec2ClassificationConfig, + encoder_embed_dim: int, + num_targets: int, + **kwargs, + ): + super().__init__(cfg, encoder_embed_dim, num_targets, **kwargs) + self.projection = Linear(self.latent_embed_dim, num_targets, bias=False) + nn.init.xavier_normal_(self.projection.weight, gain=1) + + def forward(self, last_layer_feats, padding_mask, **kwargs): + + """ + Arguments + features - [BxTxD] Acoustic feature with shape + padding_mask - [BxT] Padding Mask + """ + feat_mask = (~padding_mask).to(last_layer_feats.dtype) # T,B -> B,T + feat = self.linear(last_layer_feats) # B,T,D + feat = fn_mean(feat, feat_mask) # B,D + feat = self.activation_fn(feat) + # normalize feat + feat_norm = F.normalize(feat, p=2, dim=-1) # B,D + weight_norm = F.normalize(self.projection.weight.t(), p=2, dim=-1) # D,K + cos_fw = feat_norm @ weight_norm + return cos_fw + + +def fn_max(x, mask): + """ + Args: + x: TxBxD + mask: BxT + Return: + y: BxD + """ + mask = mask.t()[:, :, None].to(torch.bool) + return x.masked_fill(~mask, -1e-8).max(0)[0] + + +class MaxPoolingFast(Pooling): + def __init__( + self, + cfg: Wav2Vec2ClassificationConfig, + encoder_embed_dim: int, + num_targets: int, + **kwargs, + ): + super().__init__(cfg, encoder_embed_dim, num_targets) + self.activation_fn = utils.get_activation_fn(cfg.activation_fn) + self.linear = Linear(encoder_embed_dim, encoder_embed_dim) + + def forward(self, last_layer_feats, padding_mask, **kwargs): + + """ + Arguments + features - [TxBxD] Acoustic feature with shape + padding_mask - [BxT] Padding Mask + """ + feat_mask = (~padding_mask).to(last_layer_feats.dtype) + feat = self.linear(last_layer_feats) + feat = fn_max(feat, feat_mask) + feat = self.activation_fn(feat) + return self.projection(feat) + + +class LayerWeightedMeanPooling(MeanPoolingFast): + """Elmo-style weighted average representation.""" + + def __init__( + self, + cfg: Wav2Vec2ClassificationConfig, + encoder_embed_dim: int, + num_targets: int, + encoder_layers: int, + ): + super().__init__(cfg, encoder_embed_dim, num_targets) + self.num_layers = encoder_layers + self.weights = nn.Parameter(torch.ones(encoder_layers)) + + def forward(self, last_layer_feats, padding_mask, all_layer_feats): + # last_layer_feats: [BxTxD] + # padding_mask: [BxT] + if not self.training: + msg = ( + f"Number of layers in input features = {len(all_layer_feats)}." + f" Expected {self.num_layers} layers." + ) + assert len(all_layer_feats) == self.num_layers, msg + + # Stack up all layers and reshape to (num_layers, features) + all_layer_feats_stacked = torch.stack(all_layer_feats, dim=0) + num_layers, *original_feat_shape = all_layer_feats_stacked.shape + all_layer_feats_stacked_flat = all_layer_feats_stacked.view(num_layers, -1) + + # Weighted average + normalized_weights = F.softmax(self.weights, dim=-1) + weighted_avg_features = ( + normalized_weights.unsqueeze(-1) * all_layer_feats_stacked_flat + ).sum(dim=0) + weighted_avg_features = weighted_avg_features.view(*original_feat_shape) + + # Mean Pooling on weighted average features. + return super().forward(weighted_avg_features, padding_mask) \ No newline at end of file diff --git a/data/fairseq/fairseq/models/wav2vec/wav2vec2_laser.py b/data/fairseq/fairseq/models/wav2vec/wav2vec2_laser.py new file mode 100644 index 0000000000000000000000000000000000000000..ff89759d38cc1837270a8571d279f11cf1edfa75 --- /dev/null +++ b/data/fairseq/fairseq/models/wav2vec/wav2vec2_laser.py @@ -0,0 +1,39 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from fairseq.models import BaseFairseqModel, register_model +from fairseq.models.wav2vec.wav2vec2_asr import ( + Wav2Vec2CtcConfig, + Wav2VecCtc, + Wav2VecEncoder, +) +from fairseq.tasks import FairseqTask + + +@register_model("wav2vec2_laser", dataclass=Wav2Vec2CtcConfig) +class Wav2VecLaser(Wav2VecCtc): + def __init__(self, cfg: Wav2Vec2CtcConfig, w2v_encoder: BaseFairseqModel): + super().__init__(cfg, w2v_encoder) + self.num_updates = 0 + self.freeze_finetune_updates = cfg.freeze_finetune_updates + + @classmethod + def build_model(cls, cfg: Wav2Vec2CtcConfig, task: FairseqTask): + """Build a new model instance.""" + w2v_encoder = Wav2VecEncoder(cfg, 1024) + return cls(cfg, w2v_encoder) + + def forward(self, **kwargs): + output = super().forward(**kwargs) + x_out = output["encoder_out"] * 0.01 + out_pad_mask = output["padding_mask"] + # Set padded outputs to -inf so they are not selected by max-pooling + if out_pad_mask is not None and out_pad_mask.any(): + x_out = ( + x_out.float() + .masked_fill_(out_pad_mask.T.unsqueeze(-1), float("-inf")) + .type_as(x_out) + ) + return x_out.max(dim=0)[0] diff --git a/data/fairseq/fairseq/models/xmod/__init__.py b/data/fairseq/fairseq/models/xmod/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..bbf7694920eb00bed27e17dac272611be1ab44f9 --- /dev/null +++ b/data/fairseq/fairseq/models/xmod/__init__.py @@ -0,0 +1,7 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from .model import * # noqa +from .transformer_layer_xmod import * # noqa diff --git a/data/fairseq/fairseq/models/xmod/__pycache__/__init__.cpython-38.pyc b/data/fairseq/fairseq/models/xmod/__pycache__/__init__.cpython-38.pyc new file mode 100644 index 0000000000000000000000000000000000000000..106038831f4224138c2f226f23ddb54dbac7f351 Binary files /dev/null and b/data/fairseq/fairseq/models/xmod/__pycache__/__init__.cpython-38.pyc differ diff --git a/data/fairseq/fairseq/models/xmod/__pycache__/hub_interface.cpython-38.pyc b/data/fairseq/fairseq/models/xmod/__pycache__/hub_interface.cpython-38.pyc new file mode 100644 index 0000000000000000000000000000000000000000..ac6e93f00ba4a9f33db173ad010f51cb416919ab Binary files /dev/null and b/data/fairseq/fairseq/models/xmod/__pycache__/hub_interface.cpython-38.pyc differ diff --git a/data/fairseq/fairseq/models/xmod/__pycache__/model.cpython-38.pyc b/data/fairseq/fairseq/models/xmod/__pycache__/model.cpython-38.pyc new file mode 100644 index 0000000000000000000000000000000000000000..c1eea329edc8116769e56b6302f1b7cc9068b22f Binary files /dev/null and b/data/fairseq/fairseq/models/xmod/__pycache__/model.cpython-38.pyc differ diff --git a/data/fairseq/fairseq/models/xmod/__pycache__/transformer_layer_xmod.cpython-38.pyc b/data/fairseq/fairseq/models/xmod/__pycache__/transformer_layer_xmod.cpython-38.pyc new file mode 100644 index 0000000000000000000000000000000000000000..3954119f966280cc4b3606679203e01e2b78cabf Binary files /dev/null and b/data/fairseq/fairseq/models/xmod/__pycache__/transformer_layer_xmod.cpython-38.pyc differ diff --git a/data/fairseq/fairseq/models/xmod/hub_interface.py b/data/fairseq/fairseq/models/xmod/hub_interface.py new file mode 100644 index 0000000000000000000000000000000000000000..909bb423cac449b7f48f11e47126730aeadee919 --- /dev/null +++ b/data/fairseq/fairseq/models/xmod/hub_interface.py @@ -0,0 +1,51 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + + +from fairseq.models.roberta.hub_interface import RobertaHubInterface +import torch +import torch.nn.functional as F + + +class XMODHubInterface(RobertaHubInterface): + def extract_features( + self, + tokens: torch.LongTensor, + return_all_hiddens: bool = False, + lang_id=None, + ) -> torch.Tensor: + if tokens.dim() == 1: + tokens = tokens.unsqueeze(0) + if tokens.size(-1) > self.model.max_positions(): + raise ValueError( + "tokens exceeds maximum length: {} > {}".format( + tokens.size(-1), self.model.max_positions() + ) + ) + features, extra = self.model( + tokens.to(device=self.device), + features_only=True, + return_all_hiddens=return_all_hiddens, + lang_id=lang_id, + ) + if return_all_hiddens: + # convert from T x B x C -> B x T x C + inner_states = extra["inner_states"] + return [inner_state.transpose(0, 1) for inner_state in inner_states] + else: + return features # just the last layer's features + + def predict( + self, + head: str, + tokens: torch.LongTensor, + return_logits: bool = False, + lang_id=None, + ): + features = self.extract_features(tokens.to(device=self.device), lang_id=lang_id) + logits = self.model.classification_heads[head](features) + if return_logits: + return logits + return F.log_softmax(logits, dim=-1) diff --git a/data/fairseq/fairseq/models/xmod/model.py b/data/fairseq/fairseq/models/xmod/model.py new file mode 100644 index 0000000000000000000000000000000000000000..fb6c7a8deac1656855e6cdca9d4cf156942e10b2 --- /dev/null +++ b/data/fairseq/fairseq/models/xmod/model.py @@ -0,0 +1,742 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from ..roberta.model_xlmr import XLMRModel +from fairseq.models.xmod.transformer_layer_xmod import XMODTransformerEncoderLayerBase +from ..roberta.model import base_architecture, RobertaEncoder +from fairseq.models.transformer import TransformerEncoder +from fairseq.modules.transformer_sentence_encoder import init_bert_params +from typing import Optional +from fairseq.models.xmod.hub_interface import XMODHubInterface +import torch +from fairseq.distributed import fsdp_wrap +from fairseq.models import ( + register_model, + register_model_architecture, +) + +from fairseq.modules.checkpoint_activations import checkpoint_wrapper + +DEFAULT_MIN_PARAMS_TO_WRAP = int(1e8) + + +@register_model("xmod") +class XMODModel(XLMRModel): + @classmethod + def hub_models(cls): + return { + "xmod.base": "https://dl.fbaipublicfiles.com/fairseq/models/xmod/xmod.base.81.1M.tar.gz", + "xmod.large.prenorm": "https://dl.fbaipublicfiles.com/fairseq/models/xmod/xmod.large.prenorm.81.500k.tar.gz", + "xmod.base.13.125k": "https://dl.fbaipublicfiles.com/fairseq/models/xmod/xmod.base.13.125k.tar.gz", + "xmod.base.30.125k": "https://dl.fbaipublicfiles.com/fairseq/models/xmod/xmod.base.30.125k.tar.gz", + "xmod.base.30.195k": "https://dl.fbaipublicfiles.com/fairseq/models/xmod/xmod.base.30.195k.tar.gz", + "xmod.base.60.125k": "https://dl.fbaipublicfiles.com/fairseq/models/xmod/xmod.base.60.125k.tar.gz", + "xmod.base.60.265k": "https://dl.fbaipublicfiles.com/fairseq/models/xmod/xmod.base.60.265k.tar.gz", + "xmod.base.75.125k": "https://dl.fbaipublicfiles.com/fairseq/models/xmod/xmod.base.75.125k.tar.gz", + "xmod.base.75.269k": "https://dl.fbaipublicfiles.com/fairseq/models/xmod/xmod.base.75.269k.tar.gz", + } + + @classmethod + def from_pretrained( + cls, + model_name_or_path, + checkpoint_file="model.pt", + data_name_or_path=".", + bpe="sentencepiece", + **kwargs, + ): + from fairseq import hub_utils + + x = hub_utils.from_pretrained( + model_name_or_path, + checkpoint_file, + data_name_or_path, + archive_map=cls.hub_models(), + bpe=bpe, + load_checkpoint_heads=True, + **kwargs, + ) + return XMODHubInterface(x["args"], x["task"], x["models"][0]) + + @classmethod + def build_model(cls, args, task): + """Build a new model instance.""" + + from omegaconf import OmegaConf + + if OmegaConf.is_config(args): + OmegaConf.set_struct(args, False) + + # make sure all arguments are present + base_architecture(args) + + if not hasattr(args, "max_positions"): + if not hasattr(args, "tokens_per_sample"): + args.tokens_per_sample = task.max_positions() + args.max_positions = args.tokens_per_sample + + encoder = XMODEncoder(args, task.source_dictionary) + + if OmegaConf.is_config(args): + OmegaConf.set_struct(args, True) + + return cls(args, encoder) + + def forward( + self, + src_tokens, + features_only=False, + return_all_hiddens=False, + classification_head_name=None, + lang_id=None, + **kwargs, + ): + if classification_head_name is not None: + features_only = True + x, extra = self.encoder( + src_tokens, features_only, return_all_hiddens, lang_id=lang_id, **kwargs + ) + + if classification_head_name is not None: + x = self.classification_heads[classification_head_name](x) + return x, extra + + +class XMODEncoder(RobertaEncoder): + """XMOD encoder.""" + + def build_encoder(self, args, dictionary, embed_tokens): + encoder = XMODTransformerEncoder(args, dictionary, embed_tokens) + encoder.apply(init_bert_params) + return encoder + + def forward( + self, + src_tokens, + features_only=False, + return_all_hiddens=False, + masked_tokens=None, + lang_id=None, + **unused, + ): + """ + Args: + src_tokens (LongTensor): input tokens of shape `(batch, src_len)` + features_only (bool, optional): skip LM head and just return + features. If True, the output will be of shape + `(batch, src_len, embed_dim)`. + return_all_hiddens (bool, optional): also return all of the + intermediate hidden states (default: False). + + Returns: + tuple: + - the LM output of shape `(batch, src_len, vocab)` + - a dictionary of additional data, where 'inner_states' + is a list of hidden states. Note that the hidden + states have shape `(src_len, batch, vocab)`. + """ + x, extra = self.extract_features( + src_tokens, return_all_hiddens=return_all_hiddens, lang_id=lang_id + ) + if not features_only: + x = self.output_layer(x, masked_tokens=masked_tokens) + return x, extra + + def extract_features( + self, src_tokens, return_all_hiddens=False, lang_id=None, **kwargs + ): + encoder_out = self.sentence_encoder( + src_tokens, + return_all_hiddens=return_all_hiddens, + lang_id=lang_id, + token_embeddings=kwargs.get("token_embeddings", None), + ) + # T x B x C -> B x T x C + features = encoder_out["encoder_out"][0].transpose(0, 1) + inner_states = encoder_out["encoder_states"] if return_all_hiddens else None + return features, {"inner_states": inner_states} + + +class XMODTransformerEncoder(TransformerEncoder): + def build_encoder_layer(self, cfg): + layer = XMODTransformerEncoderLayerBase(cfg) + checkpoint = cfg.checkpoint_activations + if checkpoint: + offload_to_cpu = cfg.offload_activations + layer = checkpoint_wrapper(layer, offload_to_cpu=offload_to_cpu) + # if we are checkpointing, enforce that FSDP always wraps the + # checkpointed layer, regardless of layer size + min_params_to_wrap = cfg.min_params_to_wrap if not checkpoint else 0 + layer = fsdp_wrap(layer, min_num_params=min_params_to_wrap) + return layer + + def forward( + self, + src_tokens, + src_lengths: Optional[torch.Tensor] = None, + return_all_hiddens: bool = False, + token_embeddings: Optional[torch.Tensor] = None, + lang_id=None, + ): + """ + Args: + src_tokens (LongTensor): tokens in the source language of shape + `(batch, src_len)` + src_lengths (torch.LongTensor): lengths of each source sentence of + shape `(batch)` + return_all_hiddens (bool, optional): also return all of the + intermediate hidden states (default: False). + token_embeddings (torch.Tensor, optional): precomputed embeddings + default `None` will recompute embeddings + + Returns: + dict: + - **encoder_out** (Tensor): the last encoder layer's output of + shape `(src_len, batch, embed_dim)` + - **encoder_padding_mask** (ByteTensor): the positions of + padding elements of shape `(batch, src_len)` + - **encoder_embedding** (Tensor): the (scaled) embedding lookup + of shape `(batch, src_len, embed_dim)` + - **encoder_states** (List[Tensor]): all intermediate + hidden states of shape `(src_len, batch, embed_dim)`. + Only populated if *return_all_hiddens* is True. + """ + return self.forward_scriptable( + src_tokens, + src_lengths, + return_all_hiddens, + token_embeddings, + lang_id=lang_id, + ) + # TorchScript doesn't support super() method so that the scriptable Subclass + # can't access the base class model in Torchscript. + # Current workaround is to add a helper function with different name and + # call the helper function from scriptable Subclass. + + def forward_scriptable( + self, + src_tokens, + src_lengths: Optional[torch.Tensor] = None, + return_all_hiddens: bool = False, + token_embeddings: Optional[torch.Tensor] = None, + lang_id=None, + ): + """ + Args: + src_tokens (LongTensor): tokens in the source language of shape + `(batch, src_len)` + src_lengths (torch.LongTensor): lengths of each source sentence of + shape `(batch)` + return_all_hiddens (bool, optional): also return all of the + intermediate hidden states (default: False). + token_embeddings (torch.Tensor, optional): precomputed embeddings + default `None` will recompute embeddings + + Returns: + dict: + - **encoder_out** (Tensor): the last encoder layer's output of + shape `(src_len, batch, embed_dim)` + - **encoder_padding_mask** (ByteTensor): the positions of + padding elements of shape `(batch, src_len)` + - **encoder_embedding** (Tensor): the (scaled) embedding lookup + of shape `(batch, src_len, embed_dim)` + - **encoder_states** (List[Tensor]): all intermediate + hidden states of shape `(src_len, batch, embed_dim)`. + Only populated if *return_all_hiddens* is True. + """ + # compute padding mask + encoder_padding_mask = src_tokens.eq(self.padding_idx) + has_pads = src_tokens.device.type == "xla" or encoder_padding_mask.any() + + x, encoder_embedding = self.forward_embedding(src_tokens, token_embeddings) + + # account for padding while computing the representation + if has_pads: + x = x * (1 - encoder_padding_mask.unsqueeze(-1).type_as(x)) + + # B x T x C -> T x B x C + x = x.transpose(0, 1) + + encoder_states = [] + + if return_all_hiddens: + encoder_states.append(x) + + # encoder layers + for layer in self.layers: + x = layer( + x, + encoder_padding_mask=encoder_padding_mask if has_pads else None, + lang_id=lang_id, + ) + if return_all_hiddens: + assert encoder_states is not None + encoder_states.append(x) + + if self.layer_norm is not None: + x = self.layer_norm(x) + + # The Pytorch Mobile lite interpreter does not supports returning NamedTuple in + # `forward` so we use a dictionary instead. + # TorchScript does not support mixed values so the values are all lists. + # The empty list is equivalent to None. + src_lengths = ( + src_tokens.ne(self.padding_idx) + .sum(dim=1, dtype=torch.int32) + .reshape(-1, 1) + .contiguous() + ) + return { + "encoder_out": [x], # T x B x C + "encoder_padding_mask": [encoder_padding_mask], # B x T + "encoder_embedding": [encoder_embedding], # B x T x C + "encoder_states": encoder_states, # List[T x B x C] + "src_tokens": [], + "src_lengths": [src_lengths], + } + + +@register_model_architecture("xmod", "xmod_base_13") +def roberta_base_architecture(args): + args.ffn_modules = getattr(args, "ffn_modules", False) + args.adapter_modules = getattr(args, "adapter_modules", True) + args.adapter_layer_norm = getattr(args, "adapter_layer_norm", False) + args.adapter_reuse_layer_norm = getattr(args, "adapter_reuse_layer_norm", True) + args.ln_before_adapter = getattr(args, "ln_before_adapter", True) + args.languages = getattr( + args, + "languages", + [ + "ar_AR", + "en_XX", + "fi_FI", + "fr_XX", + "hi_IN", + "id_ID", + "ka_GE", + "ko_KR", + "ru_RU", + "sw_KE", + "ta_IN", + "th_TH", + "vi_VN", + ], + ) + base_architecture(args) + + +@register_model_architecture("xmod", "xmod_base_30") +def roberta_base_architecture(args): + args.ffn_modules = getattr(args, "ffn_modules", False) + args.adapter_modules = getattr(args, "adapter_modules", True) + args.adapter_layer_norm = getattr(args, "adapter_layer_norm", False) + args.adapter_reuse_layer_norm = getattr(args, "adapter_reuse_layer_norm", True) + args.ln_before_adapter = getattr(args, "ln_before_adapter", True) + args.languages = getattr( + args, + "languages", + [ + "ar_AR", + "cs_CZ", + "en_XX", + "eu_ES", + "fi_FI", + "fr_XX", + "hi_IN", + "hr_HR", + "hu_HU", + "hy_AM", + "id_ID", + "it_IT", + "ka_GE", + "ko_KR", + "lt_LT", + "ml_IN", + "mn_MN", + "ms_MY", + "pl_PL", + "ro_RO", + "ru_RU", + "si_LK", + "sk_SK", + "sq_AL", + "sv_SE", + "sw_KE", + "ta_IN", + "th_TH", + "tl_XX", + "vi_VN", + ], + ) + base_architecture(args) + + +@register_model_architecture("xmod", "xmod_base_60") +def roberta_base_architecture(args): + args.ffn_modules = getattr(args, "ffn_modules", False) + args.adapter_modules = getattr(args, "adapter_modules", True) + args.adapter_layer_norm = getattr(args, "adapter_layer_norm", False) + args.adapter_reuse_layer_norm = getattr(args, "adapter_reuse_layer_norm", True) + args.ln_before_adapter = getattr(args, "ln_before_adapter", True) + args.languages = getattr( + args, + "languages", + [ + "af_ZA", + "am_ET", + "ar_AR", + "be_BY", + "bn_IN", + "ca_ES", + "cs_CZ", + "cy_GB", + "da_DK", + "en_XX", + "eo_EO", + "et_EE", + "eu_ES", + "fa_IR", + "fi_FI", + "fr_XX", + "ga_IE", + "gl_ES", + "gu_IN", + "ha_NG", + "hi_IN", + "hr_HR", + "hu_HU", + "hy_AM", + "id_ID", + "is_IS", + "it_IT", + "ka_GE", + "ko_KR", + "ku_TR", + "la_VA", + "lt_LT", + "lv_LV", + "mk_MK", + "ml_IN", + "mn_MN", + "ms_MY", + "ne_NP", + "nl_XX", + "no_XX", + "pl_PL", + "ps_AF", + "pt_XX", + "ro_RO", + "ru_RU", + "sa_IN", + "sd_PK", + "si_LK", + "sk_SK", + "sl_SI", + "so_SO", + "sq_AL", + "sr_RS", + "sv_SE", + "sw_KE", + "ta_IN", + "te_IN", + "th_TH", + "tl_XX", + "vi_VN", + ], + ) + base_architecture(args) + + +@register_model_architecture("xmod", "xmod_base_75") +def roberta_base_architecture(args): + args.ffn_modules = getattr(args, "ffn_modules", False) + args.adapter_modules = getattr(args, "adapter_modules", True) + args.adapter_layer_norm = getattr(args, "adapter_layer_norm", False) + args.adapter_reuse_layer_norm = getattr(args, "adapter_reuse_layer_norm", True) + args.ln_before_adapter = getattr(args, "ln_before_adapter", True) + args.languages = getattr( + args, + "languages", + [ + "af_ZA", + "am_ET", + "ar_AR", + "as_IN", + "be_BY", + "bn_IN", + "br_FR", + "bs_BA", + "ca_ES", + "cs_CZ", + "cy_GB", + "da_DK", + "en_XX", + "eo_EO", + "et_EE", + "eu_ES", + "fa_IR", + "fi_FI", + "fr_XX", + "fy_NL", + "ga_IE", + "gd_GB", + "gl_ES", + "gu_IN", + "ha_NG", + "hi_IN", + "hr_HR", + "hu_HU", + "hy_AM", + "id_ID", + "is_IS", + "it_IT", + "jv_ID", + "ka_GE", + "kn_IN", + "ko_KR", + "ku_TR", + "la_VA", + "lt_LT", + "lv_LV", + "mg_MG", + "mk_MK", + "ml_IN", + "mn_MN", + "mr_IN", + "ms_MY", + "ne_NP", + "nl_XX", + "no_XX", + "om_KE", + "or_IN", + "pa_IN", + "pl_PL", + "ps_AF", + "pt_XX", + "ro_RO", + "ru_RU", + "sa_IN", + "sd_PK", + "si_LK", + "sk_SK", + "sl_SI", + "so_SO", + "sq_AL", + "sr_RS", + "su_ID", + "sv_SE", + "sw_KE", + "ta_IN", + "te_IN", + "th_TH", + "tl_XX", + "vi_VN", + "xh_ZA", + "yi_DE", + ], + ) + base_architecture(args) + + +@register_model_architecture("xmod", "xmod_base") +def roberta_base_architecture(args): + args.ffn_modules = getattr(args, "ffn_modules", False) + args.adapter_modules = getattr(args, "adapter_modules", True) + args.adapter_layer_norm = getattr(args, "adapter_layer_norm", False) + args.adapter_reuse_layer_norm = getattr(args, "adapter_reuse_layer_norm", True) + args.ln_before_adapter = getattr(args, "ln_before_adapter", True) + args.languages = getattr( + args, + "languages", + [ + "en_XX", + "id_ID", + "vi_VN", + "ru_RU", + "fa_IR", + "sv_SE", + "ja_XX", + "fr_XX", + "de_DE", + "ro_RO", + "ko_KR", + "hu_HU", + "es_XX", + "fi_FI", + "uk_UA", + "da_DK", + "pt_XX", + "no_XX", + "th_TH", + "pl_PL", + "bg_BG", + "nl_XX", + "zh_CN", + "he_IL", + "el_GR", + "it_IT", + "sk_SK", + "hr_HR", + "tr_TR", + "ar_AR", + "cs_CZ", + "lt_LT", + "hi_IN", + "zh_TW", + "ca_ES", + "ms_MY", + "sl_SI", + "lv_LV", + "ta_IN", + "bn_IN", + "et_EE", + "az_AZ", + "sq_AL", + "sr_RS", + "kk_KZ", + "ka_GE", + "tl_XX", + "ur_PK", + "is_IS", + "hy_AM", + "ml_IN", + "mk_MK", + "be_BY", + "la_VA", + "te_IN", + "eu_ES", + "gl_ES", + "mn_MN", + "kn_IN", + "ne_NP", + "sw_KE", + "si_LK", + "mr_IN", + "af_ZA", + "gu_IN", + "cy_GB", + "eo_EO", + "km_KH", + "ky_KG", + "uz_UZ", + "ps_AF", + "pa_IN", + "ga_IE", + "ha_NG", + "am_ET", + "lo_LA", + "ku_TR", + "so_SO", + "my_MM", + "or_IN", + "sa_IN", + ], + ) + base_architecture(args) + + +@register_model_architecture("xmod", "xmod_large_prenorm") +def roberta_base_architecture(args): + args.ffn_modules = getattr(args, "ffn_modules", False) + args.adapter_modules = getattr(args, "adapter_modules", True) + args.adapter_layer_norm = getattr(args, "adapter_layer_norm", True) + args.adapter_reuse_layer_norm = getattr(args, "adapter_reuse_layer_norm", False) + args.ln_before_adapter = getattr(args, "ln_before_adapter", False) + # args.bottleneck = getattr(args, "bottleneck", 8) + args.bottleneck = getattr(args, "bottleneck", 4) + args.languages = getattr( + args, + "languages", + [ + "en_XX", + "id_ID", + "vi_VN", + "ru_RU", + "fa_IR", + "sv_SE", + "ja_XX", + "fr_XX", + "de_DE", + "ro_RO", + "ko_KR", + "hu_HU", + "es_XX", + "fi_FI", + "uk_UA", + "da_DK", + "pt_XX", + "no_XX", + "th_TH", + "pl_PL", + "bg_BG", + "nl_XX", + "zh_CN", + "he_IL", + "el_GR", + "it_IT", + "sk_SK", + "hr_HR", + "tr_TR", + "ar_AR", + "cs_CZ", + "lt_LT", + "hi_IN", + "zh_TW", + "ca_ES", + "ms_MY", + "sl_SI", + "lv_LV", + "ta_IN", + "bn_IN", + "et_EE", + "az_AZ", + "sq_AL", + "sr_RS", + "kk_KZ", + "ka_GE", + "tl_XX", + "ur_PK", + "is_IS", + "hy_AM", + "ml_IN", + "mk_MK", + "be_BY", + "la_VA", + "te_IN", + "eu_ES", + "gl_ES", + "mn_MN", + "kn_IN", + "ne_NP", + "sw_KE", + "si_LK", + "mr_IN", + "af_ZA", + "gu_IN", + "cy_GB", + "eo_EO", + "km_KH", + "ky_KG", + "uz_UZ", + "ps_AF", + "pa_IN", + "ga_IE", + "ha_NG", + "am_ET", + "lo_LA", + "ku_TR", + "so_SO", + "my_MM", + "or_IN", + "sa_IN", + ], + ) + + args.encoder_normalize_before = getattr(args, "encoder_normalize_before", True) + args.encoder_layers = getattr(args, "encoder_layers", 24) + args.encoder_embed_dim = getattr(args, "encoder_embed_dim", 1024) + args.encoder_ffn_embed_dim = getattr(args, "encoder_ffn_embed_dim", 4096) + args.encoder_attention_heads = getattr(args, "encoder_attention_heads", 16) + base_architecture(args) diff --git a/data/fairseq/fairseq/models/xmod/transformer_layer_xmod.py b/data/fairseq/fairseq/models/xmod/transformer_layer_xmod.py new file mode 100644 index 0000000000000000000000000000000000000000..47a91cdc2339f235efe0cecc85d6a7a260d19c6a --- /dev/null +++ b/data/fairseq/fairseq/models/xmod/transformer_layer_xmod.py @@ -0,0 +1,179 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from fairseq.modules.transformer_layer import TransformerEncoderLayer +from typing import Optional +import torch +import torch.nn as nn +from fairseq import utils +from fairseq.modules import LayerNorm +from fairseq.modules.fairseq_dropout import FairseqDropout +from fairseq.modules.quant_noise import quant_noise +from torch import Tensor + + +class Adapter(nn.Module): + def __init__(self, cfg, red_fac=2): + super(Adapter, self).__init__() + self.cfg = cfg + self.embed_dim = cfg.encoder_embed_dim + self.quant_noise = getattr(cfg, "quant_noise_pq", 0) + self.quant_noise_block_size = getattr(cfg, "quant_noise_pq_block_size", 8) or 8 + self.activation_fn = utils.get_activation_fn( + activation=getattr(cfg, "activation_fn", "relu") or "relu" + ) + self.fc1 = quant_noise( + nn.Linear(self.embed_dim, self.embed_dim // red_fac), + p=self.quant_noise, + block_size=self.quant_noise_block_size, + ) + self.fc2 = quant_noise( + nn.Linear(self.embed_dim // red_fac, self.embed_dim), + p=self.quant_noise, + block_size=self.quant_noise_block_size, + ) + activation_dropout_p = getattr(cfg, "activation_dropout", 0) or 0 + if activation_dropout_p == 0: + # for backwards compatibility with models that use cfg.relu_dropout + activation_dropout_p = getattr(cfg, "relu_dropout", 0) or 0 + self.activation_dropout_module = FairseqDropout( + float(activation_dropout_p), module_name=self.__class__.__name__ + ) + + def forward(self, x): + x = self.activation_fn(self.fc1(x)) + if not hasattr(self.cfg, "adapter_dropout") or self.cfg.adapter_dropout: + x = self.activation_dropout_module(x) + x = self.fc2(x) + return x + + +class XMODTransformerEncoderLayerBase(TransformerEncoderLayer): + """Encoder layer block. + + In the original paper each operation (multi-head attention or FFN) is + postprocessed with: `dropout -> add residual -> layernorm`. In the + tensor2tensor code they suggest that learning is more robust when + preprocessing each layer with layernorm and postprocessing with: + `dropout -> add residual`. We default to the approach in the paper, but the + tensor2tensor approach can be enabled by setting + *cfg.encoder.normalize_before* to ``True``. + + Args: + args (argparse.Namespace): parsed command-line arguments + """ + + def __init__(self, cfg): + super().__init__(cfg) + if hasattr(cfg, "adapter_modules") and cfg.adapter_modules: + export = getattr(cfg, "export", False) + if cfg.adapter_layer_norm: + self.adapter_layer_norm = LayerNorm(self.embed_dim, export=export) + self.adapter_modules = nn.ModuleDict(dict()) + if hasattr(self.cfg, "bottleneck"): + bottleneck = self.cfg.bottleneck + else: + bottleneck = 2 + for language in cfg.languages: + self.adapter_modules[str(language)] = Adapter(cfg, red_fac=bottleneck) + + def lang_adapter(self, lang_id, x): + # If language adapters exist pass throught them + if hasattr(self.cfg, "adapter_modules") and self.cfg.adapter_modules: + if lang_id is None: + lang_id = ["en_XX"] * x.shape[1] + d_langs = [lang_id[0]] + lang_lengths = [1] + for lang in lang_id[1:]: + if lang == d_langs[-1]: + lang_lengths[-1] += 1 + else: + d_langs.append(lang) + lang_lengths.append(1) + + if ( + not hasattr(self.cfg, "ln_before_adapter") + or not self.cfg.ln_before_adapter + ): + residual = x + if self.cfg.adapter_layer_norm: + x = self.adapter_layer_norm(x) + elif self.cfg.adapter_reuse_layer_norm: + x = self.final_layer_norm(x) + if hasattr(self.cfg, "ln_before_adapter") and self.cfg.ln_before_adapter: + residual = x + + split_x = torch.split(x, lang_lengths, 1) + x_ = [] + for i, (lang, s_x) in enumerate(zip(d_langs, split_x)): + lang = lang.replace("_rom", "").replace("_zaw", "") + x_.append(self.adapter_modules[str(lang)](s_x)) + x = torch.cat(x_, 1) + + x = self.dropout_module(x) + x = self.residual_connection(x, residual) + + return x + + def forward( + self, + x, + encoder_padding_mask: Optional[Tensor], + attn_mask: Optional[Tensor] = None, + lang_id: Optional[list] = None, + ): + """ + Args: + x (Tensor): input to the layer of shape `(seq_len, batch, embed_dim)` + encoder_padding_mask (ByteTensor): binary ByteTensor of shape + `(batch, seq_len)` where padding elements are indicated by ``1``. + attn_mask (ByteTensor): binary tensor of shape `(tgt_len, src_len)`, + where `tgt_len` is the length of output and `src_len` is the + length of input, though here both are equal to `seq_len`. + `attn_mask[tgt_i, src_j] = 1` means that when calculating the + embedding for `tgt_i`, we exclude (mask out) `src_j`. This is + useful for strided self-attention. + + Returns: + encoded output of shape `(seq_len, batch, embed_dim)` + """ + # anything in original attn_mask = 1, becomes -1e8 + # anything in original attn_mask = 0, becomes 0 + # Note that we cannot use -inf here, because at some edge cases, + # the attention weight (before softmax) for some padded element in query + # will become -inf, which results in NaN in model parameters + if attn_mask is not None: + attn_mask = attn_mask.masked_fill(attn_mask.to(torch.bool), -1e8) + + residual = x + if self.normalize_before: + x = self.self_attn_layer_norm(x) + x, _ = self.self_attn( + query=x, + key=x, + value=x, + key_padding_mask=encoder_padding_mask, + need_weights=False, + attn_mask=attn_mask, + ) + x = self.dropout_module(x) + x = self.residual_connection(x, residual) + if not self.normalize_before: + x = self.self_attn_layer_norm(x) + + residual = x + if self.normalize_before: + x = self.final_layer_norm(x) + x = self.activation_fn(self.fc1(x)) + x = self.activation_dropout_module(x) + x = self.fc2(x) + x = self.dropout_module(x) + x = self.residual_connection(x, residual) + + x = self.lang_adapter(lang_id, x) + + if not self.normalize_before: + x = self.final_layer_norm(x) + return x diff --git a/data/fairseq/fairseq/modules/__init__.py b/data/fairseq/fairseq/modules/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..dcfda9b82aa74819d99dc9305fcad437fd90456f --- /dev/null +++ b/data/fairseq/fairseq/modules/__init__.py @@ -0,0 +1,106 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. +"""isort:skip_file""" + +from .adaptive_input import AdaptiveInput +from .adaptive_softmax import AdaptiveSoftmax +from .base_layer import BaseLayer +from .beamable_mm import BeamableMM +from .character_token_embedder import CharacterTokenEmbedder +from .conv_tbc import ConvTBC +from .cross_entropy import cross_entropy +from .downsampled_multihead_attention import DownsampledMultiHeadAttention +from .dynamic_convolution import DynamicConv, DynamicConv1dTBC, DynamicConv_scripatable +from .dynamic_crf_layer import DynamicCRF +from .ema_module import EMAModuleConfig, EMAModule +from .fairseq_dropout import FairseqDropout +from .fp32_batch_norm import Fp32BatchNorm +from .fp32_group_norm import Fp32GroupNorm +from .fp32_instance_norm import Fp32InstanceNorm +from .gelu import gelu, gelu_accurate +from .grad_multiply import GradMultiply +from .gumbel_vector_quantizer import GumbelVectorQuantizer +from .kmeans_vector_quantizer import KmeansVectorQuantizer +from .layer_drop import LayerDropModuleList +from .layer_norm import Fp32LayerNorm, LayerNorm +from .learned_positional_embedding import LearnedPositionalEmbedding +from .lightweight_convolution import LightweightConv, LightweightConv1dTBC +from .linearized_convolution import LinearizedConvolution +from .location_attention import LocationAttention +from .lstm_cell_with_zoneout import LSTMCellWithZoneOut +from .multihead_attention import MultiheadAttention +from .positional_embedding import PositionalEmbedding +from .same_pad import SamePad, SamePad2d +from .scalar_bias import ScalarBias +from .sinusoidal_positional_embedding import SinusoidalPositionalEmbedding +from .transformer_sentence_encoder_layer import TransformerSentenceEncoderLayer +from .transformer_sentence_encoder import TransformerSentenceEncoder +from .transpose_last import TransposeLast +from .unfold import unfold1d +from .transformer_layer import TransformerDecoderLayer, TransformerEncoderLayer +from .vggblock import VGGBlock +from .espnet_multihead_attention import ( + ESPNETMultiHeadedAttention, + RelPositionMultiHeadedAttention, + RotaryPositionMultiHeadedAttention, +) +from .rotary_positional_embedding import RotaryPositionalEmbedding +from .positional_encoding import ( + RelPositionalEncoding, +) + +__all__ = [ + "AdaptiveInput", + "AdaptiveSoftmax", + "BaseLayer", + "BeamableMM", + "CharacterTokenEmbedder", + "ConvTBC", + "cross_entropy", + "DownsampledMultiHeadAttention", + "DynamicConv1dTBC", + "DynamicConv", + "DynamicConv_scripatable", + "DynamicCRF", + "EMAModule", + "EMAModuleConfig", + "FairseqDropout", + "Fp32BatchNorm", + "Fp32GroupNorm", + "Fp32LayerNorm", + "Fp32InstanceNorm", + "gelu", + "gelu_accurate", + "GradMultiply", + "GumbelVectorQuantizer", + "KmeansVectorQuantizer", + "LayerDropModuleList", + "LayerNorm", + "LearnedPositionalEmbedding", + "LightweightConv1dTBC", + "LightweightConv", + "LinearizedConvolution", + "LocationAttention", + "LSTMCellWithZoneOut", + "MultiheadAttention", + "PositionalEmbedding", + "SamePad", + "SamePad2d", + "ScalarBias", + "SinusoidalPositionalEmbedding", + "TransformerSentenceEncoderLayer", + "TransformerSentenceEncoder", + "TransformerDecoderLayer", + "TransformerEncoderLayer", + "TransposeLast", + "VGGBlock", + "unfold1d", + "ESPNETMultiheadedAttention", + "PositionalEmbedding", + "RelPositionMultiHeadedAttention", + "RelPositionalEncoding", + "RotaryPositionalEmbedding", + "RotaryPositionMultiHeadedAttention", +] diff --git a/data/fairseq/fairseq/modules/__pycache__/__init__.cpython-38.pyc b/data/fairseq/fairseq/modules/__pycache__/__init__.cpython-38.pyc new file mode 100644 index 0000000000000000000000000000000000000000..ca77b3629ee1177516b7d66f464a6ab507ae008c Binary files /dev/null and b/data/fairseq/fairseq/modules/__pycache__/__init__.cpython-38.pyc differ diff --git a/data/fairseq/fairseq/modules/__pycache__/adaptive_input.cpython-38.pyc b/data/fairseq/fairseq/modules/__pycache__/adaptive_input.cpython-38.pyc new file mode 100644 index 0000000000000000000000000000000000000000..a7f0cf31d6742c8ffa42266efdaaf8e056e256da Binary files /dev/null and b/data/fairseq/fairseq/modules/__pycache__/adaptive_input.cpython-38.pyc differ diff --git a/data/fairseq/fairseq/modules/__pycache__/adaptive_softmax.cpython-38.pyc b/data/fairseq/fairseq/modules/__pycache__/adaptive_softmax.cpython-38.pyc new file mode 100644 index 0000000000000000000000000000000000000000..22d9f21ad1ea2e2d8cf36f84720d878df3a0e2ed Binary files /dev/null and b/data/fairseq/fairseq/modules/__pycache__/adaptive_softmax.cpython-38.pyc differ diff --git a/data/fairseq/fairseq/modules/__pycache__/base_layer.cpython-38.pyc b/data/fairseq/fairseq/modules/__pycache__/base_layer.cpython-38.pyc new file mode 100644 index 0000000000000000000000000000000000000000..2f640892d6439debc8f8e9df6a210bbf849d654b Binary files /dev/null and b/data/fairseq/fairseq/modules/__pycache__/base_layer.cpython-38.pyc differ diff --git a/data/fairseq/fairseq/modules/__pycache__/beamable_mm.cpython-38.pyc b/data/fairseq/fairseq/modules/__pycache__/beamable_mm.cpython-38.pyc new file mode 100644 index 0000000000000000000000000000000000000000..5ea700128b144b7de4d3f002483196c8fd54b8fa Binary files /dev/null and b/data/fairseq/fairseq/modules/__pycache__/beamable_mm.cpython-38.pyc differ diff --git a/data/fairseq/fairseq/modules/__pycache__/character_token_embedder.cpython-38.pyc b/data/fairseq/fairseq/modules/__pycache__/character_token_embedder.cpython-38.pyc new file mode 100644 index 0000000000000000000000000000000000000000..6f9a4425f0049a789e8be03f3d6a49cf34102bef Binary files /dev/null and b/data/fairseq/fairseq/modules/__pycache__/character_token_embedder.cpython-38.pyc differ diff --git a/data/fairseq/fairseq/modules/__pycache__/checkpoint_activations.cpython-38.pyc b/data/fairseq/fairseq/modules/__pycache__/checkpoint_activations.cpython-38.pyc new file mode 100644 index 0000000000000000000000000000000000000000..651d900bccdca867f637bbc626a931f98a262343 Binary files /dev/null and b/data/fairseq/fairseq/modules/__pycache__/checkpoint_activations.cpython-38.pyc differ diff --git a/data/fairseq/fairseq/modules/__pycache__/conformer_layer.cpython-38.pyc b/data/fairseq/fairseq/modules/__pycache__/conformer_layer.cpython-38.pyc new file mode 100644 index 0000000000000000000000000000000000000000..aa9b61519768d2217e2e810fe2cdfb6474c97d49 Binary files /dev/null and b/data/fairseq/fairseq/modules/__pycache__/conformer_layer.cpython-38.pyc differ diff --git a/data/fairseq/fairseq/modules/__pycache__/conv_tbc.cpython-38.pyc b/data/fairseq/fairseq/modules/__pycache__/conv_tbc.cpython-38.pyc new file mode 100644 index 0000000000000000000000000000000000000000..183c4da2b37f134dc058e09e54fff3cfb222fcb6 Binary files /dev/null and b/data/fairseq/fairseq/modules/__pycache__/conv_tbc.cpython-38.pyc differ diff --git a/data/fairseq/fairseq/modules/__pycache__/cross_entropy.cpython-38.pyc b/data/fairseq/fairseq/modules/__pycache__/cross_entropy.cpython-38.pyc new file mode 100644 index 0000000000000000000000000000000000000000..ba685afda54c74f9e43f409fd973d94261566dda Binary files /dev/null and b/data/fairseq/fairseq/modules/__pycache__/cross_entropy.cpython-38.pyc differ diff --git a/data/fairseq/fairseq/modules/__pycache__/downsampled_multihead_attention.cpython-38.pyc b/data/fairseq/fairseq/modules/__pycache__/downsampled_multihead_attention.cpython-38.pyc new file mode 100644 index 0000000000000000000000000000000000000000..85931c7b5647272a1ca5732a21b954baf6002967 Binary files /dev/null and b/data/fairseq/fairseq/modules/__pycache__/downsampled_multihead_attention.cpython-38.pyc differ diff --git a/data/fairseq/fairseq/modules/__pycache__/dynamic_convolution.cpython-38.pyc b/data/fairseq/fairseq/modules/__pycache__/dynamic_convolution.cpython-38.pyc new file mode 100644 index 0000000000000000000000000000000000000000..33d22921a453e84a2c2bb38e913ec828e26cd807 Binary files /dev/null and b/data/fairseq/fairseq/modules/__pycache__/dynamic_convolution.cpython-38.pyc differ diff --git a/data/fairseq/fairseq/modules/__pycache__/dynamic_crf_layer.cpython-38.pyc b/data/fairseq/fairseq/modules/__pycache__/dynamic_crf_layer.cpython-38.pyc new file mode 100644 index 0000000000000000000000000000000000000000..ab821018272012c83a4b3b01d54fa78ce46571f0 Binary files /dev/null and b/data/fairseq/fairseq/modules/__pycache__/dynamic_crf_layer.cpython-38.pyc differ diff --git a/data/fairseq/fairseq/modules/__pycache__/ema_module.cpython-38.pyc b/data/fairseq/fairseq/modules/__pycache__/ema_module.cpython-38.pyc new file mode 100644 index 0000000000000000000000000000000000000000..1e610c97e0f6d585a162929a4848638ebfe93918 Binary files /dev/null and b/data/fairseq/fairseq/modules/__pycache__/ema_module.cpython-38.pyc differ diff --git a/data/fairseq/fairseq/modules/__pycache__/espnet_multihead_attention.cpython-38.pyc b/data/fairseq/fairseq/modules/__pycache__/espnet_multihead_attention.cpython-38.pyc new file mode 100644 index 0000000000000000000000000000000000000000..a04a2d048391156b32171a9030acaa895deb9c53 Binary files /dev/null and b/data/fairseq/fairseq/modules/__pycache__/espnet_multihead_attention.cpython-38.pyc differ diff --git a/data/fairseq/fairseq/modules/__pycache__/fairseq_dropout.cpython-38.pyc b/data/fairseq/fairseq/modules/__pycache__/fairseq_dropout.cpython-38.pyc new file mode 100644 index 0000000000000000000000000000000000000000..5c433e7149397f20612026572cd46ab719c91e40 Binary files /dev/null and b/data/fairseq/fairseq/modules/__pycache__/fairseq_dropout.cpython-38.pyc differ diff --git a/data/fairseq/fairseq/modules/__pycache__/fp32_batch_norm.cpython-38.pyc b/data/fairseq/fairseq/modules/__pycache__/fp32_batch_norm.cpython-38.pyc new file mode 100644 index 0000000000000000000000000000000000000000..8b0d14228ec1bf7357f0d7701a38757ed76a96bb Binary files /dev/null and b/data/fairseq/fairseq/modules/__pycache__/fp32_batch_norm.cpython-38.pyc differ diff --git a/data/fairseq/fairseq/modules/__pycache__/fp32_group_norm.cpython-38.pyc b/data/fairseq/fairseq/modules/__pycache__/fp32_group_norm.cpython-38.pyc new file mode 100644 index 0000000000000000000000000000000000000000..cb0b868853de2797ae06e054561722aefef8341f Binary files /dev/null and b/data/fairseq/fairseq/modules/__pycache__/fp32_group_norm.cpython-38.pyc differ diff --git a/data/fairseq/fairseq/modules/__pycache__/fp32_instance_norm.cpython-38.pyc b/data/fairseq/fairseq/modules/__pycache__/fp32_instance_norm.cpython-38.pyc new file mode 100644 index 0000000000000000000000000000000000000000..92f78bf336f36a3720fb7efd68931ff1b7e3880e Binary files /dev/null and b/data/fairseq/fairseq/modules/__pycache__/fp32_instance_norm.cpython-38.pyc differ diff --git a/data/fairseq/fairseq/modules/__pycache__/gelu.cpython-38.pyc b/data/fairseq/fairseq/modules/__pycache__/gelu.cpython-38.pyc new file mode 100644 index 0000000000000000000000000000000000000000..db6a7430ad9164e8d650867a95d533668fe508db Binary files /dev/null and b/data/fairseq/fairseq/modules/__pycache__/gelu.cpython-38.pyc differ diff --git a/data/fairseq/fairseq/modules/__pycache__/grad_multiply.cpython-38.pyc b/data/fairseq/fairseq/modules/__pycache__/grad_multiply.cpython-38.pyc new file mode 100644 index 0000000000000000000000000000000000000000..22ec36adad6e91a3aa28884415e809893cdb941b Binary files /dev/null and b/data/fairseq/fairseq/modules/__pycache__/grad_multiply.cpython-38.pyc differ diff --git a/data/fairseq/fairseq/modules/__pycache__/gumbel_vector_quantizer.cpython-38.pyc b/data/fairseq/fairseq/modules/__pycache__/gumbel_vector_quantizer.cpython-38.pyc new file mode 100644 index 0000000000000000000000000000000000000000..1f2e5d7b0a90fd4c91a2d3565e48a84f08c6f6ee Binary files /dev/null and b/data/fairseq/fairseq/modules/__pycache__/gumbel_vector_quantizer.cpython-38.pyc differ diff --git a/data/fairseq/fairseq/modules/__pycache__/kmeans_vector_quantizer.cpython-38.pyc b/data/fairseq/fairseq/modules/__pycache__/kmeans_vector_quantizer.cpython-38.pyc new file mode 100644 index 0000000000000000000000000000000000000000..16ed3425a73b1a3ac553094e903f03a40d189d4a Binary files /dev/null and b/data/fairseq/fairseq/modules/__pycache__/kmeans_vector_quantizer.cpython-38.pyc differ diff --git a/data/fairseq/fairseq/modules/__pycache__/layer_drop.cpython-38.pyc b/data/fairseq/fairseq/modules/__pycache__/layer_drop.cpython-38.pyc new file mode 100644 index 0000000000000000000000000000000000000000..950efc2e5eedf9b233f50e9d9dea15950db606b0 Binary files /dev/null and b/data/fairseq/fairseq/modules/__pycache__/layer_drop.cpython-38.pyc differ diff --git a/data/fairseq/fairseq/modules/__pycache__/layer_norm.cpython-38.pyc b/data/fairseq/fairseq/modules/__pycache__/layer_norm.cpython-38.pyc new file mode 100644 index 0000000000000000000000000000000000000000..ca9330de6119ca11a12bc1e1695a329f11a18bae Binary files /dev/null and b/data/fairseq/fairseq/modules/__pycache__/layer_norm.cpython-38.pyc differ diff --git a/data/fairseq/fairseq/modules/__pycache__/learned_positional_embedding.cpython-38.pyc b/data/fairseq/fairseq/modules/__pycache__/learned_positional_embedding.cpython-38.pyc new file mode 100644 index 0000000000000000000000000000000000000000..905af7cd6d0fe34d04b9185556f726262bfe103f Binary files /dev/null and b/data/fairseq/fairseq/modules/__pycache__/learned_positional_embedding.cpython-38.pyc differ diff --git a/data/fairseq/fairseq/modules/__pycache__/lightweight_convolution.cpython-38.pyc b/data/fairseq/fairseq/modules/__pycache__/lightweight_convolution.cpython-38.pyc new file mode 100644 index 0000000000000000000000000000000000000000..2f74dc31d34f94042d42d8dcf1a292326bfff173 Binary files /dev/null and b/data/fairseq/fairseq/modules/__pycache__/lightweight_convolution.cpython-38.pyc differ diff --git a/data/fairseq/fairseq/modules/__pycache__/linearized_convolution.cpython-38.pyc b/data/fairseq/fairseq/modules/__pycache__/linearized_convolution.cpython-38.pyc new file mode 100644 index 0000000000000000000000000000000000000000..4e998a7df29f0b3d4cb034a490ac94ba4cccd988 Binary files /dev/null and b/data/fairseq/fairseq/modules/__pycache__/linearized_convolution.cpython-38.pyc differ diff --git a/data/fairseq/fairseq/modules/__pycache__/location_attention.cpython-38.pyc b/data/fairseq/fairseq/modules/__pycache__/location_attention.cpython-38.pyc new file mode 100644 index 0000000000000000000000000000000000000000..3100064aa7c2abf04b693ee2d94e0b9c9569e05d Binary files /dev/null and b/data/fairseq/fairseq/modules/__pycache__/location_attention.cpython-38.pyc differ diff --git a/data/fairseq/fairseq/modules/__pycache__/lstm_cell_with_zoneout.cpython-38.pyc b/data/fairseq/fairseq/modules/__pycache__/lstm_cell_with_zoneout.cpython-38.pyc new file mode 100644 index 0000000000000000000000000000000000000000..3c1f1eb8ca66f903f03a7d71f78c7a20080d35b0 Binary files /dev/null and b/data/fairseq/fairseq/modules/__pycache__/lstm_cell_with_zoneout.cpython-38.pyc differ diff --git a/data/fairseq/fairseq/modules/__pycache__/multihead_attention.cpython-38.pyc b/data/fairseq/fairseq/modules/__pycache__/multihead_attention.cpython-38.pyc new file mode 100644 index 0000000000000000000000000000000000000000..4b55b0fab1c7563ce7de3c226f0b60064797d096 Binary files /dev/null and b/data/fairseq/fairseq/modules/__pycache__/multihead_attention.cpython-38.pyc differ diff --git a/data/fairseq/fairseq/modules/__pycache__/positional_embedding.cpython-38.pyc b/data/fairseq/fairseq/modules/__pycache__/positional_embedding.cpython-38.pyc new file mode 100644 index 0000000000000000000000000000000000000000..3878624b0b9dc32704e5159918b964eea7f1ab4a Binary files /dev/null and b/data/fairseq/fairseq/modules/__pycache__/positional_embedding.cpython-38.pyc differ diff --git a/data/fairseq/fairseq/modules/__pycache__/positional_encoding.cpython-38.pyc b/data/fairseq/fairseq/modules/__pycache__/positional_encoding.cpython-38.pyc new file mode 100644 index 0000000000000000000000000000000000000000..fddf507700b203a1b699901d2b32b14ebab3d3fc Binary files /dev/null and b/data/fairseq/fairseq/modules/__pycache__/positional_encoding.cpython-38.pyc differ diff --git a/data/fairseq/fairseq/modules/__pycache__/quant_noise.cpython-38.pyc b/data/fairseq/fairseq/modules/__pycache__/quant_noise.cpython-38.pyc new file mode 100644 index 0000000000000000000000000000000000000000..90188c307ae8eab31e7807014e1aa26484ddeb54 Binary files /dev/null and b/data/fairseq/fairseq/modules/__pycache__/quant_noise.cpython-38.pyc differ diff --git a/data/fairseq/fairseq/modules/__pycache__/rotary_positional_embedding.cpython-38.pyc b/data/fairseq/fairseq/modules/__pycache__/rotary_positional_embedding.cpython-38.pyc new file mode 100644 index 0000000000000000000000000000000000000000..d2ed2ccf6602d771a8349139266b49e400cbc884 Binary files /dev/null and b/data/fairseq/fairseq/modules/__pycache__/rotary_positional_embedding.cpython-38.pyc differ diff --git a/data/fairseq/fairseq/modules/__pycache__/same_pad.cpython-38.pyc b/data/fairseq/fairseq/modules/__pycache__/same_pad.cpython-38.pyc new file mode 100644 index 0000000000000000000000000000000000000000..de32fd1fc4f687379e64a23b8513f1daede80f98 Binary files /dev/null and b/data/fairseq/fairseq/modules/__pycache__/same_pad.cpython-38.pyc differ diff --git a/data/fairseq/fairseq/modules/__pycache__/scalar_bias.cpython-38.pyc b/data/fairseq/fairseq/modules/__pycache__/scalar_bias.cpython-38.pyc new file mode 100644 index 0000000000000000000000000000000000000000..17b9b4af89da5049ffc132c12a049537853127fc Binary files /dev/null and b/data/fairseq/fairseq/modules/__pycache__/scalar_bias.cpython-38.pyc differ diff --git a/data/fairseq/fairseq/modules/__pycache__/sinusoidal_positional_embedding.cpython-38.pyc b/data/fairseq/fairseq/modules/__pycache__/sinusoidal_positional_embedding.cpython-38.pyc new file mode 100644 index 0000000000000000000000000000000000000000..3333a90594a43960a1cb3366b760389bfe2d6ef1 Binary files /dev/null and b/data/fairseq/fairseq/modules/__pycache__/sinusoidal_positional_embedding.cpython-38.pyc differ diff --git a/data/fairseq/fairseq/modules/__pycache__/transformer_layer.cpython-38.pyc b/data/fairseq/fairseq/modules/__pycache__/transformer_layer.cpython-38.pyc new file mode 100644 index 0000000000000000000000000000000000000000..9a886bc70d80dd344a041f1bedd86ec160a0c3bf Binary files /dev/null and b/data/fairseq/fairseq/modules/__pycache__/transformer_layer.cpython-38.pyc differ diff --git a/data/fairseq/fairseq/modules/__pycache__/transformer_layer_aug.cpython-38.pyc b/data/fairseq/fairseq/modules/__pycache__/transformer_layer_aug.cpython-38.pyc new file mode 100644 index 0000000000000000000000000000000000000000..056f0eb2a7befb571117d14e5e971d9f13bbf2f7 Binary files /dev/null and b/data/fairseq/fairseq/modules/__pycache__/transformer_layer_aug.cpython-38.pyc differ diff --git a/data/fairseq/fairseq/modules/__pycache__/transformer_sentence_encoder.cpython-38.pyc b/data/fairseq/fairseq/modules/__pycache__/transformer_sentence_encoder.cpython-38.pyc new file mode 100644 index 0000000000000000000000000000000000000000..c20e9ab8da43d6e48d258e8db5151cffb01675ae Binary files /dev/null and b/data/fairseq/fairseq/modules/__pycache__/transformer_sentence_encoder.cpython-38.pyc differ diff --git a/data/fairseq/fairseq/modules/__pycache__/transformer_sentence_encoder_layer.cpython-38.pyc b/data/fairseq/fairseq/modules/__pycache__/transformer_sentence_encoder_layer.cpython-38.pyc new file mode 100644 index 0000000000000000000000000000000000000000..7d530062feda9e58b7b8090431595ce5aad75c18 Binary files /dev/null and b/data/fairseq/fairseq/modules/__pycache__/transformer_sentence_encoder_layer.cpython-38.pyc differ diff --git a/data/fairseq/fairseq/modules/__pycache__/transpose_last.cpython-38.pyc b/data/fairseq/fairseq/modules/__pycache__/transpose_last.cpython-38.pyc new file mode 100644 index 0000000000000000000000000000000000000000..cbf48fb3b197f5b63b14c44a6b686c8466898e8d Binary files /dev/null and b/data/fairseq/fairseq/modules/__pycache__/transpose_last.cpython-38.pyc differ diff --git a/data/fairseq/fairseq/modules/__pycache__/unfold.cpython-38.pyc b/data/fairseq/fairseq/modules/__pycache__/unfold.cpython-38.pyc new file mode 100644 index 0000000000000000000000000000000000000000..01bb4c8ef864cff824a9f3764b9a55349fda8715 Binary files /dev/null and b/data/fairseq/fairseq/modules/__pycache__/unfold.cpython-38.pyc differ diff --git a/data/fairseq/fairseq/modules/__pycache__/vggblock.cpython-38.pyc b/data/fairseq/fairseq/modules/__pycache__/vggblock.cpython-38.pyc new file mode 100644 index 0000000000000000000000000000000000000000..a7e4f1c574df2078e2e10c2e402c6d145178181e Binary files /dev/null and b/data/fairseq/fairseq/modules/__pycache__/vggblock.cpython-38.pyc differ diff --git a/data/fairseq/fairseq/modules/adaptive_input.py b/data/fairseq/fairseq/modules/adaptive_input.py new file mode 100644 index 0000000000000000000000000000000000000000..01ac4accaccc3e4625ee2ca6a17cf88df1db0ca1 --- /dev/null +++ b/data/fairseq/fairseq/modules/adaptive_input.py @@ -0,0 +1,81 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + + +from typing import List + +import torch +from torch import nn + +from fairseq.modules.quant_noise import quant_noise + + +class AdaptiveInput(nn.Module): + def __init__( + self, + vocab_size: int, + padding_idx: int, + initial_dim: int, + factor: float, + output_dim: int, + cutoff: List[int], + q_noise: float = 0, + qn_block_size: int = 8, + ): + super().__init__() + + if vocab_size > cutoff[-1]: + cutoff = cutoff + [vocab_size] + else: + assert ( + vocab_size == cutoff[-1] + ), "cannot specify cutoff larger than vocab size" + + self.cutoff = cutoff + self.embedding_dim = output_dim + self.padding_idx = padding_idx + + self.embeddings = nn.ModuleList() + for i in range(len(self.cutoff)): + prev = self.cutoff[i - 1] if i > 0 else 0 + size = self.cutoff[i] - prev + dim = int(initial_dim // (factor**i)) + seq = nn.Sequential( + nn.Embedding(size, dim, self.padding_idx), + quant_noise( + nn.Linear(dim, output_dim, bias=False), q_noise, qn_block_size + ), + ) + + self.embeddings.append(seq) + self.padding_idx = None + self.padding_idx = padding_idx + + def init_weights(m): + if isinstance(m, nn.Embedding): + nn.init.normal_(m.weight, mean=0, std=m.weight.shape[1] ** -0.5) + nn.init.constant_(m.weight[padding_idx], 0) + elif hasattr(m, "weight"): + nn.init.xavier_uniform_(m.weight) + + self.apply(init_weights) + + self.register_buffer("_float_tensor", torch.FloatTensor(1)) + + def weights_for_band(self, band: int): + return self.embeddings[band][0].weight, self.embeddings[band][1].weight + + def forward(self, input: torch.Tensor): + result = self._float_tensor.new(input.shape + (self.embedding_dim,)) + for i in range(len(self.cutoff)): + mask = input.lt(self.cutoff[i]) + if i > 0: + mask.mul_(input.ge(self.cutoff[i - 1])) + chunk_input = input[mask] - self.cutoff[i - 1] + else: + chunk_input = input[mask] + if mask.any(): + result[mask] = self.embeddings[i](chunk_input) + return result diff --git a/data/fairseq/fairseq/modules/adaptive_softmax.py b/data/fairseq/fairseq/modules/adaptive_softmax.py new file mode 100644 index 0000000000000000000000000000000000000000..ae0c77ba0f6ee98501306d66cbc4a948b4ade0f7 --- /dev/null +++ b/data/fairseq/fairseq/modules/adaptive_softmax.py @@ -0,0 +1,268 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import functools +import operator + +import torch +import torch.nn.functional as F +from fairseq.modules.fairseq_dropout import FairseqDropout +from fairseq.modules.quant_noise import quant_noise +from torch import nn + + +class TiedLinear(nn.Module): + def __init__(self, weight, transpose): + super().__init__() + self.weight = weight + self.transpose = transpose + + def forward(self, input): + return F.linear(input, self.weight.t() if self.transpose else self.weight) + + +class TiedHeadModule(nn.Module): + def __init__(self, weights, input_dim, num_classes, q_noise, qn_block_size): + super().__init__() + tied_emb, _ = weights + self.num_words, emb_dim = tied_emb.size() + + self.word_proj = quant_noise( + TiedLinear(tied_emb, transpose=False), q_noise, qn_block_size + ) + if input_dim != emb_dim: + self.word_proj = nn.Sequential( + quant_noise( + nn.Linear(input_dim, emb_dim, bias=False), q_noise, qn_block_size + ), + self.word_proj, + ) + + self.class_proj = quant_noise( + nn.Linear(input_dim, num_classes, bias=False), q_noise, qn_block_size + ) + self.out_dim = self.num_words + num_classes + + self.register_buffer("_float_tensor", torch.FloatTensor(1)) + + def forward(self, input): + inp_sz = functools.reduce(operator.mul, input.shape[:-1], 1) + out = self._float_tensor.new(inp_sz, self.out_dim) + out[:, : self.num_words] = self.word_proj(input.view(inp_sz, -1)) + out[:, self.num_words :] = self.class_proj(input.view(inp_sz, -1)) + return out + + +class AdaptiveSoftmax(nn.Module): + """ + This is an implementation of the efficient softmax approximation for + graphical processing units (GPU), described in the paper "Efficient softmax + approximation for GPUs" (http://arxiv.org/abs/1609.04309). + """ + + def __init__( + self, + vocab_size, + input_dim, + cutoff, + dropout, + factor=4.0, + adaptive_inputs=None, + tie_proj=False, + q_noise=0, + qn_block_size=8, + ): + super().__init__() + + if vocab_size > cutoff[-1]: + cutoff = cutoff + [vocab_size] + else: + assert ( + vocab_size == cutoff[-1] + ), "cannot specify cutoff larger than vocab size" + + output_dim = cutoff[0] + len(cutoff) - 1 + + self.vocab_size = vocab_size + self.cutoff = cutoff + self.dropout_module = FairseqDropout( + dropout, module_name=self.__class__.__name__ + ) + self.input_dim = input_dim + self.factor = factor + self.q_noise = q_noise + self.qn_block_size = qn_block_size + + self.lsm = nn.LogSoftmax(dim=1) + + if adaptive_inputs is not None: + self.head = TiedHeadModule( + adaptive_inputs.weights_for_band(0), + input_dim, + len(cutoff) - 1, + self.q_noise, + self.qn_block_size, + ) + else: + self.head = quant_noise( + nn.Linear(input_dim, output_dim, bias=False), + self.q_noise, + self.qn_block_size, + ) + + self._make_tail(adaptive_inputs, tie_proj) + + def init_weights(m): + if ( + hasattr(m, "weight") + and not isinstance(m, TiedLinear) + and not isinstance(m, TiedHeadModule) + ): + nn.init.xavier_uniform_(m.weight) + + self.apply(init_weights) + + self.register_buffer("version", torch.LongTensor([1])) + + def _make_tail(self, adaptive_inputs=None, tie_proj=False): + self.tail = nn.ModuleList() + for i in range(len(self.cutoff) - 1): + dim = int(self.input_dim // self.factor ** (i + 1)) + + tied_emb, tied_proj = ( + adaptive_inputs.weights_for_band(i + 1) + if adaptive_inputs is not None + else (None, None) + ) + + if tied_proj is not None: + if tie_proj: + proj = quant_noise( + TiedLinear(tied_proj, transpose=True), + self.q_noise, + self.qn_block_size, + ) + else: + proj = quant_noise( + nn.Linear(tied_proj.size(0), tied_proj.size(1), bias=False), + self.q_noise, + self.qn_block_size, + ) + else: + proj = quant_noise( + nn.Linear(self.input_dim, dim, bias=False), + self.q_noise, + self.qn_block_size, + ) + + if tied_emb is None: + out_proj = nn.Linear( + dim, self.cutoff[i + 1] - self.cutoff[i], bias=False + ) + else: + out_proj = TiedLinear(tied_emb, transpose=False) + + m = nn.Sequential( + proj, + nn.Dropout(self.dropout_module.p), + quant_noise(out_proj, self.q_noise, self.qn_block_size), + ) + + self.tail.append(m) + + def upgrade_state_dict_named(self, state_dict, name): + version_name = name + ".version" + if version_name not in state_dict: + raise Exception("This version of the model is no longer supported") + + def adapt_target(self, target): + """ + In order to be efficient, the AdaptiveSoftMax does not compute the + scores for all the word of the vocabulary for all the examples. It is + thus necessary to call the method adapt_target of the AdaptiveSoftMax + layer inside each forward pass. + """ + + target = target.view(-1) + new_target = [target.clone()] + target_idxs = [] + + for i in range(len(self.cutoff) - 1): + mask = target.ge(self.cutoff[i]).mul(target.lt(self.cutoff[i + 1])) + new_target[0][mask] = self.cutoff[0] + i + + if mask.any(): + target_idxs.append(mask.nonzero(as_tuple=False).squeeze(1)) + new_target.append(target[mask].add(-self.cutoff[i])) + else: + target_idxs.append(None) + new_target.append(None) + + return new_target, target_idxs + + def forward(self, input, target): + """ + Args: + input: (b x t x d) + target: (b x t) + Returns: + 2 lists: output for each cutoff section and new targets by cut off + """ + + input = input.contiguous().view(-1, input.size(-1)) + input = self.dropout_module(input) + + new_target, target_idxs = self.adapt_target(target) + output = [self.head(input)] + + for i in range(len(target_idxs)): + if target_idxs[i] is not None: + output.append(self.tail[i](input.index_select(0, target_idxs[i]))) + else: + output.append(None) + + return output, new_target + + def get_log_prob(self, input, target): + """ + Computes the log probabilities for all the words of the vocabulary, + given a 2D tensor of hidden vectors. + """ + + bsz, length, dim = input.size() + input = input.contiguous().view(-1, dim) + + if target is not None: + _, target_idxs = self.adapt_target(target) + else: + target_idxs = None + + head_y = self.head(input) + log_probs = head_y.new_zeros(input.size(0), self.vocab_size) + + head_sz = self.cutoff[0] + len(self.tail) + log_probs[:, :head_sz] = self.lsm(head_y) + tail_priors = log_probs[:, self.cutoff[0] : head_sz].clone() + + for i in range(len(self.tail)): + start = self.cutoff[i] + end = self.cutoff[i + 1] + + if target_idxs is None: + tail_out = log_probs[:, start:end] + tail_out.copy_(self.tail[i](input)) + log_probs[:, start:end] = self.lsm(tail_out).add_( + tail_priors[:, i, None] + ) + elif target_idxs[i] is not None: + idxs = target_idxs[i] + tail_out = log_probs[idxs, start:end] + tail_out.copy_(self.tail[i](input[idxs])) + log_probs[idxs, start:end] = self.lsm(tail_out).add_( + tail_priors[idxs, i, None] + ) + + log_probs = log_probs.view(bsz, length, -1) + return log_probs diff --git a/data/fairseq/fairseq/modules/base_layer.py b/data/fairseq/fairseq/modules/base_layer.py new file mode 100644 index 0000000000000000000000000000000000000000..e823f7bae2c2e9b7e2277b0161cf26b453e7216c --- /dev/null +++ b/data/fairseq/fairseq/modules/base_layer.py @@ -0,0 +1,170 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import torch.nn as nn +import torch +import sys +from fairseq import utils +from fairseq.distributed import utils as distributed_utils +from fairseq.modules.layer_norm import LayerNorm + + +class BaseLayer(nn.Module): + def __init__(self, args): + super().__init__() + self.num_workers = distributed_utils.get_data_parallel_world_size() + expert_centroids = torch.empty(self.num_workers, args.decoder_embed_dim) + torch.nn.init.orthogonal_(expert_centroids, gain=0.1) + self.register_parameter( + "expert_centroids", torch.nn.Parameter(expert_centroids) + ) + self.expert_network = nn.Sequential( + *([BaseSublayer(args) for _ in range(args.base_sublayers)]) + ) + self.expert_id = distributed_utils.get_data_parallel_rank() + self.shuffle = args.base_shuffle + self.cpp = self.load_assignment() + + # Add a special attribute to the expert parameters, so we know not to sync their gradients + for param in self.expert_network.parameters(): + param.expert = True + + def forward(self, input_features, *args, **kwargs): + features = input_features.reshape(-1, input_features.size(-1)) + is_training = input_features.requires_grad + + if self.shuffle and is_training: + # Send each token to a random worker, to break correlations within the batch + shuffle_sort = torch.randperm(features.size(0), device=features.device) + features = All2All.apply(features[shuffle_sort]) + + with torch.no_grad(): + # Compute similarity of each token to each expert, for routing + token_expert_affinities = features.matmul( + self.expert_centroids.transpose(0, 1) + ) + + # Compute which token goes to which expert + sort_by_expert, input_splits, output_splits = ( + self.balanced_assignment(token_expert_affinities) + if is_training + else self.greedy_assignment(token_expert_affinities) + ) + # Swap these tokens for the right ones for our expert + routed_features = All2All.apply( + features[sort_by_expert], output_splits, input_splits + ) + + if routed_features.size(0) > 0: + # Mix in the expert network based on how appropriate it is for these tokens + alpha = torch.sigmoid( + routed_features.mv(self.expert_centroids[self.expert_id]) + ).unsqueeze(1) + routed_features = ( + alpha * self.expert_network(routed_features) + + (1 - alpha) * routed_features + ) + # Return to original worker and ordering + result = All2All.apply(routed_features, input_splits, output_splits)[ + self.inverse_sort(sort_by_expert) + ] + + if self.shuffle and is_training: + # Undo shuffling + result = All2All.apply(result)[self.inverse_sort(shuffle_sort)] + + # Return additional Nones for compatibility with TransformerDecoderLayer + return result.view(input_features.size()), None, None + + def inverse_sort(self, order): + # Creates an index that undoes a sort: xs==xs[order][inverse_sort(order)] + return torch.empty_like(order).scatter_( + 0, order, torch.arange(0, order.size(0), device=order.device) + ) + + def balanced_assignment(self, scores): + ok = scores.isfinite() + if not ok.all(): + # NaNs here can break the assignment algorithm + scores[~ok] = scores[ok].min() + return self.cpp.balanced_assignment(scores), None, None + + # Assigns each token to the top k experts + def greedy_assignment(self, scores, k=1): + token_to_workers = torch.topk(scores, dim=1, k=k, largest=True).indices.view(-1) + token_to_workers, sort_ordering = torch.sort(token_to_workers) + worker2token = sort_ordering // k + + # Find how many tokens we're sending to each other worker (being careful for sending 0 tokens to some workers) + output_splits = torch.zeros( + (self.num_workers,), dtype=torch.long, device=scores.device + ) + workers, counts = torch.unique_consecutive(token_to_workers, return_counts=True) + output_splits[workers] = counts + # Tell other workers how many tokens to expect from us + input_splits = All2All.apply(output_splits) + return worker2token, input_splits.tolist(), output_splits.tolist() + + def load_assignment(self): + try: + from fairseq import libbase + + return libbase + + except ImportError as e: + sys.stderr.write( + "ERROR: missing libbase. run `python setup.py build_ext --inplace`\n" + ) + raise e + + +class BaseSublayer(nn.Module): + def __init__(self, args): + super().__init__() + self.activation_fn = utils.get_activation_fn( + activation=getattr(args, "activation_fn", "relu") or "relu" + ) + self.norm = LayerNorm(args.decoder_embed_dim, export=False) + self.ff1 = torch.nn.Linear(args.decoder_embed_dim, args.decoder_ffn_embed_dim) + self.ff2 = torch.nn.Linear(args.decoder_ffn_embed_dim, args.decoder_embed_dim) + self.ff2.weight.data.zero_() + + def forward(self, xs): + return xs + self.ff2(self.activation_fn(self.ff1(self.norm(xs)))) + + +# Wraps torch.distributed.all_to_all_single as a function that supports autograd +class All2All(torch.autograd.Function): + @staticmethod + def forward(ctx, xs, input_splits=None, output_splits=None): + ctx.input_splits = input_splits + ctx.output_splits = output_splits + + ys = ( + torch.empty_like(xs) + if output_splits is None + else xs.new_empty(size=[sum(output_splits)] + list(xs.size()[1:])) + ) + torch.distributed.all_to_all_single( + ys, xs, output_split_sizes=output_splits, input_split_sizes=input_splits + ) + return ys + + @staticmethod + def backward(ctx, grad_output): + result = ( + torch.empty_like(grad_output) + if ctx.input_splits is None + else grad_output.new_empty( + size=[sum(ctx.input_splits)] + list(grad_output.size()[1:]) + ) + ) + torch.distributed.all_to_all_single( + result, + grad_output, + output_split_sizes=ctx.input_splits, + input_split_sizes=ctx.output_splits, + ) + return result, None, None diff --git a/data/fairseq/fairseq/modules/beamable_mm.py b/data/fairseq/fairseq/modules/beamable_mm.py new file mode 100644 index 0000000000000000000000000000000000000000..eff1a4607f600c71210e6b914985dc48731aae86 --- /dev/null +++ b/data/fairseq/fairseq/modules/beamable_mm.py @@ -0,0 +1,49 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import torch +import torch.nn as nn + + +class BeamableMM(nn.Module): + """This module provides an optimized MM for beam decoding with attention. + + It leverage the fact that the source-side of the input is replicated beam + times and the target-side of the input is of width one. This layer speeds up + inference by replacing the inputs {(bsz x 1 x nhu), (bsz x sz2 x nhu)} + with smaller inputs {(bsz/beam x beam x nhu), (bsz/beam x sz2 x nhu)}. + """ + + def __init__(self, beam_size=None): + super(BeamableMM, self).__init__() + self.beam_size = beam_size + + def forward(self, input1, input2): + if ( + not self.training + and self.beam_size is not None # test mode + and input1.dim() == 3 # beam size is set + and input1.size(1) # only support batched input + == 1 # single time step update + ): + bsz, beam = input1.size(0), self.beam_size + + # bsz x 1 x nhu --> bsz/beam x beam x nhu + input1 = input1[:, 0, :].unfold(0, beam, beam).transpose(2, 1) + + # bsz x sz2 x nhu --> bsz/beam x sz2 x nhu + input2 = input2.unfold(0, beam, beam)[:, :, :, 0] + + # use non batched operation if bsz = beam + if input1.size(0) == 1: + output = torch.mm(input1[0, :, :], input2[0, :, :]) + else: + output = input1.bmm(input2) + return output.view(bsz, 1, -1) + else: + return input1.bmm(input2) + + def set_beam_size(self, beam_size): + self.beam_size = beam_size diff --git a/data/fairseq/fairseq/modules/character_token_embedder.py b/data/fairseq/fairseq/modules/character_token_embedder.py new file mode 100644 index 0000000000000000000000000000000000000000..181221b61b9f76453b67e3b848b198620dce912c --- /dev/null +++ b/data/fairseq/fairseq/modules/character_token_embedder.py @@ -0,0 +1,214 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import logging +from typing import List, Tuple + +import torch +import torch.nn.functional as F +from fairseq.data import Dictionary +from torch import nn + + +CHAR_PAD_IDX = 0 +CHAR_EOS_IDX = 257 + + +logger = logging.getLogger(__name__) + + +class CharacterTokenEmbedder(torch.nn.Module): + def __init__( + self, + vocab: Dictionary, + filters: List[Tuple[int, int]], + char_embed_dim: int, + word_embed_dim: int, + highway_layers: int, + max_char_len: int = 50, + char_inputs: bool = False, + ): + super(CharacterTokenEmbedder, self).__init__() + + self.onnx_trace = False + self.embedding_dim = word_embed_dim + self.max_char_len = max_char_len + self.char_embeddings = nn.Embedding(257, char_embed_dim, padding_idx=0) + self.symbol_embeddings = nn.Parameter(torch.FloatTensor(2, word_embed_dim)) + self.eos_idx, self.unk_idx = 0, 1 + self.char_inputs = char_inputs + + self.convolutions = nn.ModuleList() + for width, out_c in filters: + self.convolutions.append( + nn.Conv1d(char_embed_dim, out_c, kernel_size=width) + ) + + last_dim = sum(f[1] for f in filters) + + self.highway = Highway(last_dim, highway_layers) if highway_layers > 0 else None + + self.projection = nn.Linear(last_dim, word_embed_dim) + + assert ( + vocab is not None or char_inputs + ), "vocab must be set if not using char inputs" + self.vocab = None + if vocab is not None: + self.set_vocab(vocab, max_char_len) + + self.reset_parameters() + + def prepare_for_onnx_export_(self): + self.onnx_trace = True + + def set_vocab(self, vocab, max_char_len): + word_to_char = torch.LongTensor(len(vocab), max_char_len) + + truncated = 0 + for i in range(len(vocab)): + if i < vocab.nspecial: + char_idxs = [0] * max_char_len + else: + chars = vocab[i].encode() + # +1 for padding + char_idxs = [c + 1 for c in chars] + [0] * (max_char_len - len(chars)) + if len(char_idxs) > max_char_len: + truncated += 1 + char_idxs = char_idxs[:max_char_len] + word_to_char[i] = torch.LongTensor(char_idxs) + + if truncated > 0: + logger.info( + "truncated {} words longer than {} characters".format( + truncated, max_char_len + ) + ) + + self.vocab = vocab + self.word_to_char = word_to_char + + @property + def padding_idx(self): + return Dictionary().pad() if self.vocab is None else self.vocab.pad() + + def reset_parameters(self): + nn.init.xavier_normal_(self.char_embeddings.weight) + nn.init.xavier_normal_(self.symbol_embeddings) + nn.init.xavier_uniform_(self.projection.weight) + + nn.init.constant_( + self.char_embeddings.weight[self.char_embeddings.padding_idx], 0.0 + ) + nn.init.constant_(self.projection.bias, 0.0) + + def forward( + self, + input: torch.Tensor, + ): + if self.char_inputs: + chars = input.view(-1, self.max_char_len) + pads = chars[:, 0].eq(CHAR_PAD_IDX) + eos = chars[:, 0].eq(CHAR_EOS_IDX) + if eos.any(): + if self.onnx_trace: + chars = torch.where(eos.unsqueeze(1), chars.new_zeros(1), chars) + else: + chars[eos] = 0 + + unk = None + else: + flat_words = input.view(-1) + chars = self.word_to_char[flat_words.type_as(self.word_to_char)].type_as( + input + ) + pads = flat_words.eq(self.vocab.pad()) + eos = flat_words.eq(self.vocab.eos()) + unk = flat_words.eq(self.vocab.unk()) + + word_embs = self._convolve(chars) + if self.onnx_trace: + if pads.any(): + word_embs = torch.where( + pads.unsqueeze(1), word_embs.new_zeros(1), word_embs + ) + if eos.any(): + word_embs = torch.where( + eos.unsqueeze(1), self.symbol_embeddings[self.eos_idx], word_embs + ) + if unk is not None and unk.any(): + word_embs = torch.where( + unk.unsqueeze(1), self.symbol_embeddings[self.unk_idx], word_embs + ) + else: + if pads.any(): + word_embs[pads] = 0 + if eos.any(): + word_embs[eos] = self.symbol_embeddings[self.eos_idx] + if unk is not None and unk.any(): + word_embs[unk] = self.symbol_embeddings[self.unk_idx] + + return word_embs.view(input.size()[:2] + (-1,)) + + def _convolve( + self, + char_idxs: torch.Tensor, + ): + char_embs = self.char_embeddings(char_idxs) + char_embs = char_embs.transpose(1, 2) # BTC -> BCT + + conv_result = [] + + for conv in self.convolutions: + x = conv(char_embs) + x, _ = torch.max(x, -1) + x = F.relu(x) + conv_result.append(x) + + x = torch.cat(conv_result, dim=-1) + + if self.highway is not None: + x = self.highway(x) + x = self.projection(x) + + return x + + +class Highway(torch.nn.Module): + """ + A `Highway layer `_. + Adopted from the AllenNLP implementation. + """ + + def __init__(self, input_dim: int, num_layers: int = 1): + super(Highway, self).__init__() + self.input_dim = input_dim + self.layers = nn.ModuleList( + [nn.Linear(input_dim, input_dim * 2) for _ in range(num_layers)] + ) + self.activation = nn.ReLU() + + self.reset_parameters() + + def reset_parameters(self): + for layer in self.layers: + # As per comment in AllenNLP: + # We should bias the highway layer to just carry its input forward. We do that by + # setting the bias on `B(x)` to be positive, because that means `g` will be biased to + # be high, so we will carry the input forward. The bias on `B(x)` is the second half + # of the bias vector in each Linear layer. + nn.init.constant_(layer.bias[self.input_dim :], 1) + + nn.init.constant_(layer.bias[: self.input_dim], 0) + nn.init.xavier_normal_(layer.weight) + + def forward(self, x: torch.Tensor): + for layer in self.layers: + projection = layer(x) + proj_x, gate = projection.chunk(2, dim=-1) + proj_x = self.activation(proj_x) + gate = torch.sigmoid(gate) + x = gate * x + (gate.new_tensor([1]) - gate) * proj_x + return x diff --git a/data/fairseq/fairseq/modules/checkpoint_activations.py b/data/fairseq/fairseq/modules/checkpoint_activations.py new file mode 100644 index 0000000000000000000000000000000000000000..aa0b5929a39d5acec91038d7335276b4e5098aed --- /dev/null +++ b/data/fairseq/fairseq/modules/checkpoint_activations.py @@ -0,0 +1,242 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import functools +from typing import Any, Dict, List, Tuple, Union + +import torch +import torch.utils.checkpoint as checkpoint +from fairseq import utils + + +def checkpoint_wrapper(m, offload_to_cpu=False): + """ + A friendlier wrapper for performing activation checkpointing. + + Compared to the PyTorch version, this version: + - wraps an nn.Module, so that all subsequent calls will use checkpointing + - handles keyword arguments in the forward + - handles non-Tensor outputs from the forward + + Usage:: + + checkpointed_module = checkpoint_wrapper(my_module, offload_to_cpu=True) + a, b = checkpointed_module(x, y=3, z=torch.Tensor([1])) + """ + # should I check whether original_forward has already been set? + assert not hasattr( + m, "precheckpoint_forward" + ), "checkpoint function has already been applied?" + m.precheckpoint_forward = m.forward + m.forward = functools.partial( + _checkpointed_forward, + m.precheckpoint_forward, # original_forward + offload_to_cpu, + ) + return m + + +def unwrap_checkpoint(m: torch.nn.Module): + """ + unwrap a module and its children from checkpoint_wrapper + """ + for module in m.modules(): + if hasattr(module, "precheckpoint_forward"): + module.forward = module.precheckpoint_forward + del module.precheckpoint_forward + if hasattr(module, "old_deepcopy_method"): + module.__deepcopy__ = module.old_deepcopy_method + del module.old_deepcopy_method + return m + + +def _checkpointed_forward(original_forward, offload_to_cpu, *args, **kwargs): + # Autograd Functions in PyTorch work best with positional args, since + # the backward must return gradients (or None) for every input argument. + # We can flatten keyword arguments to make this easier. + kwarg_keys, flat_args = pack_kwargs(*args, **kwargs) + parent_ctx_dict = {"offload": offload_to_cpu} + output = CheckpointFunction.apply( + original_forward, parent_ctx_dict, kwarg_keys, *flat_args + ) + if isinstance(output, torch.Tensor): + return output + else: + packed_non_tensor_outputs = parent_ctx_dict["packed_non_tensor_outputs"] + if packed_non_tensor_outputs: + output = unpack_non_tensors(output, packed_non_tensor_outputs) + return output + + +def pack_kwargs(*args, **kwargs) -> Tuple[List[str], List[Any]]: + """ + Usage:: + + kwarg_keys, flat_args = pack_kwargs(1, 2, a=3, b=4) + args, kwargs = unpack_kwargs(kwarg_keys, flat_args) + assert args == [1, 2] + assert kwargs == {"a": 3, "b": 4} + """ + kwarg_keys = [] + flat_args = list(args) + for k, v in kwargs.items(): + kwarg_keys.append(k) + flat_args.append(v) + return kwarg_keys, flat_args + + +def unpack_kwargs( + kwarg_keys: List[str], flat_args: List[Any] +) -> Tuple[List[Any], Dict[str, Any]]: + if len(kwarg_keys) == 0: + return flat_args, {} + args = flat_args[: -len(kwarg_keys)] + kwargs = {k: v for k, v in zip(kwarg_keys, flat_args[-len(kwarg_keys) :])} + return args, kwargs + + +def split_non_tensors( + mixed: Union[torch.Tensor, Tuple[Any]] +) -> Tuple[Tuple[torch.Tensor], Dict[str, List[Any]]]: + """ + Usage:: + + x = torch.Tensor([1]) + y = torch.Tensor([2]) + tensors, packed_non_tensors = split_non_tensors((x, y, None, 3)) + recon = unpack_non_tensors(tensors, packed_non_tensors) + assert recon == (x, y, None, 3) + """ + if isinstance(mixed, torch.Tensor): + return (mixed,), None + tensors = [] + packed_non_tensors = {"is_tensor": [], "objects": []} + for o in mixed: + if isinstance(o, torch.Tensor): + packed_non_tensors["is_tensor"].append(True) + tensors.append(o) + else: + packed_non_tensors["is_tensor"].append(False) + packed_non_tensors["objects"].append(o) + return tuple(tensors), packed_non_tensors + + +def unpack_non_tensors( + tensors: Tuple[torch.Tensor], + packed_non_tensors: Dict[str, List[Any]], +) -> Tuple[Any]: + if packed_non_tensors is None: + return tensors + assert isinstance(packed_non_tensors, dict) + mixed = [] + is_tensor_list = packed_non_tensors["is_tensor"] + objects = packed_non_tensors["objects"] + assert len(tensors) + len(objects) == len(is_tensor_list) + obj_i = tnsr_i = 0 + for is_tensor in is_tensor_list: + if is_tensor: + mixed.append(tensors[tnsr_i]) + tnsr_i += 1 + else: + mixed.append(objects[obj_i]) + obj_i += 1 + return tuple(mixed) + + +class CheckpointFunction(torch.autograd.Function): + """Similar to the torch version, but support non-Tensor outputs. + + The caller is expected to provide a dict (*parent_ctx_dict*) that will hold + the non-Tensor outputs. These should be combined with the Tensor *outputs* + by calling ``unpack_non_tensors``. + """ + + @staticmethod + def forward(ctx, run_function, parent_ctx_dict, kwarg_keys, *args): + if torch.is_grad_enabled(): # grad may be disabled, e.g., during validation + checkpoint.check_backward_validity(args) + + ctx.run_function = run_function + ctx.kwarg_keys = kwarg_keys + ctx.fwd_rng_state = utils.get_rng_state() + + tensor_inputs, packed_non_tensor_inputs = split_non_tensors(args) + if parent_ctx_dict["offload"]: + ctx.fwd_device = tuple(x.device for x in tensor_inputs) + ctx.grad_requirements = tuple(x.requires_grad for x in tensor_inputs) + tensor_inputs = tuple( + x.to(torch.device("cpu"), non_blocking=True) for x in tensor_inputs + ) + + else: + ctx.fwd_device, ctx.grad_requirements = None, None + + ctx.save_for_backward(*tensor_inputs) + ctx.packed_non_tensor_inputs = packed_non_tensor_inputs + + with torch.no_grad(): + unpacked_args, unpacked_kwargs = unpack_kwargs(kwarg_keys, args) + outputs = run_function(*unpacked_args, **unpacked_kwargs) + + if isinstance(outputs, torch.Tensor): + return outputs + else: + # Autograd Functions don't like non-Tensor outputs. We can split the + # non-Tensor and Tensor outputs, returning the former by reference + # through *parent_ctx_dict* and returning the latter directly. + outputs, packed_non_tensor_outputs = split_non_tensors(outputs) + parent_ctx_dict["packed_non_tensor_outputs"] = packed_non_tensor_outputs + return outputs + + @staticmethod + def backward(ctx, *args): + if not torch.autograd._is_checkpoint_valid(): + raise RuntimeError( + "Checkpointing is not compatible with .grad(), please use .backward() if possible" + ) + + tensor_inputs: Tuple = ctx.saved_tensors + tensor_inputs = checkpoint.detach_variable(tensor_inputs) + if ctx.fwd_device is not None: + tensor_inputs = [ + t.to(ctx.fwd_device[i], non_blocking=True) + for i, t in enumerate(tensor_inputs) + ] + for i, need_grad in enumerate(ctx.grad_requirements): + tensor_inputs[i].requires_grad = need_grad + inputs = unpack_non_tensors(tensor_inputs, ctx.packed_non_tensor_inputs) + + # Store the current states. + bwd_rng_state = utils.get_rng_state() + + # Set the states to what it used to be before the forward pass. + utils.set_rng_state(ctx.fwd_rng_state) + + with torch.enable_grad(): + unpacked_args, unpacked_kwargs = unpack_kwargs(ctx.kwarg_keys, inputs) + outputs = ctx.run_function(*unpacked_args, **unpacked_kwargs) + tensor_outputs, _ = split_non_tensors(outputs) + # Set the states back to what it was at the start of this function. + utils.set_rng_state(bwd_rng_state) + + # Run backward() with only Tensors that require grad + outputs_with_grad = [] + args_with_grad = [] + for i in range(len(tensor_outputs)): + if tensor_outputs[i].requires_grad: + outputs_with_grad.append(tensor_outputs[i]) + args_with_grad.append(args[i]) + if len(outputs_with_grad) == 0: + raise RuntimeError( + "None of the outputs have requires_grad=True, " + "this checkpoint() is not necessary" + ) + + torch.autograd.backward(outputs_with_grad, args_with_grad) + + grads = tuple( + inp.grad if isinstance(inp, torch.Tensor) else None for inp in inputs + ) + return (None, None, None) + grads diff --git a/data/fairseq/fairseq/modules/conformer_layer.py b/data/fairseq/fairseq/modules/conformer_layer.py new file mode 100644 index 0000000000000000000000000000000000000000..964af243ec8fb5e09e0b61da688a9657277d39c1 --- /dev/null +++ b/data/fairseq/fairseq/modules/conformer_layer.py @@ -0,0 +1,301 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + + +from typing import Optional + +import torch + +from fairseq.modules import ( + ESPNETMultiHeadedAttention, + LayerNorm, + MultiheadAttention, + RelPositionMultiHeadedAttention, + RotaryPositionMultiHeadedAttention, +) +from fairseq.utils import get_activation_fn + + +class ConvolutionModule(torch.nn.Module): + """Convolution block used in the conformer block""" + + def __init__( + self, + embed_dim, + channels, + depthwise_kernel_size, + dropout, + activation_fn="swish", + bias=False, + export=False, + ): + """ + Args: + embed_dim: Embedding dimension + channels: Number of channels in depthwise conv layers + depthwise_kernel_size: Depthwise conv layer kernel size + dropout: dropout value + activation_fn: Activation function to use after depthwise convolution kernel + bias: If bias should be added to conv layers + export: If layernorm should be exported to jit + """ + super(ConvolutionModule, self).__init__() + assert ( + depthwise_kernel_size - 1 + ) % 2 == 0, "kernel_size should be a odd number for 'SAME' padding" + self.layer_norm = LayerNorm(embed_dim, export=export) + self.pointwise_conv1 = torch.nn.Conv1d( + embed_dim, + 2 * channels, + kernel_size=1, + stride=1, + padding=0, + bias=bias, + ) + self.glu = torch.nn.GLU(dim=1) + self.depthwise_conv = torch.nn.Conv1d( + channels, + channels, + depthwise_kernel_size, + stride=1, + padding=(depthwise_kernel_size - 1) // 2, + groups=channels, + bias=bias, + ) + self.batch_norm = torch.nn.BatchNorm1d(channels) + self.activation = get_activation_fn(activation_fn)(channels) + self.pointwise_conv2 = torch.nn.Conv1d( + channels, + embed_dim, + kernel_size=1, + stride=1, + padding=0, + bias=bias, + ) + self.dropout = torch.nn.Dropout(dropout) + + def forward(self, x): + """ + Args: + x: Input of shape B X T X C + Returns: + Tensor of shape B X T X C + """ + x = self.layer_norm(x) + # exchange the temporal dimension and the feature dimension + x = x.transpose(1, 2) + + # GLU mechanism + x = self.pointwise_conv1(x) # (batch, 2*channel, dim) + x = self.glu(x) # (batch, channel, dim) + + # 1D Depthwise Conv + x = self.depthwise_conv(x) + x = self.batch_norm(x) + x = self.activation(x) + + x = self.pointwise_conv2(x) + x = self.dropout(x) + return x.transpose(1, 2) + + +class FeedForwardModule(torch.nn.Module): + """Positionwise feed forward layer used in conformer""" + + def __init__( + self, + input_feat, + hidden_units, + dropout1, + dropout2, + activation_fn="swish", + bias=True, + ): + """ + Args: + input_feat: Input feature dimension + hidden_units: Hidden unit dimension + dropout1: dropout value for layer1 + dropout2: dropout value for layer2 + activation_fn: Name of activation function + bias: If linear layers should have bias + """ + + super(FeedForwardModule, self).__init__() + self.layer_norm = LayerNorm(input_feat) + self.w_1 = torch.nn.Linear(input_feat, hidden_units, bias=bias) + self.w_2 = torch.nn.Linear(hidden_units, input_feat, bias=bias) + self.dropout1 = torch.nn.Dropout(dropout1) + self.dropout2 = torch.nn.Dropout(dropout2) + self.activation = get_activation_fn(activation_fn)(hidden_units) + + def forward(self, x): + """ + Args: + x: Input Tensor of shape T X B X C + Returns: + Tensor of shape T X B X C + """ + x = self.layer_norm(x) + x = self.w_1(x) + x = self.activation(x) + x = self.dropout1(x) + x = self.w_2(x) + return self.dropout2(x) + + +class ConformerEncoderLayer(torch.nn.Module): + """Conformer block based on https://arxiv.org/abs/2005.08100. We currently don't support relative positional encoding in MHA""" + + def __init__( + self, + embed_dim, + ffn_embed_dim, + attention_heads, + dropout, + use_fp16, + depthwise_conv_kernel_size=31, + activation_fn="swish", + attn_type=None, + pos_enc_type="abs", + ): + """ + Args: + embed_dim: Input embedding dimension + ffn_embed_dim: FFN layer dimension + attention_heads: Number of attention heads in MHA + dropout: dropout value + depthwise_conv_kernel_size: Size of kernel in depthwise conv layer in convolution module + activation_fn: Activation function name to use in convulation block and feed forward block + attn_type: MHA implementation from ESPNET vs fairseq + pos_enc_type: Positional encoding type - abs, rope, rel_pos + """ + self.pos_enc_type = pos_enc_type + super(ConformerEncoderLayer, self).__init__() + + self.ffn1 = FeedForwardModule( + embed_dim, + ffn_embed_dim, + dropout, + dropout, + ) + + self.self_attn_layer_norm = LayerNorm(embed_dim, export=False) + self.self_attn_dropout = torch.nn.Dropout(dropout) + if attn_type == "espnet": + if self.pos_enc_type == "rel_pos": + self.self_attn = RelPositionMultiHeadedAttention( + embed_dim, + attention_heads, + dropout=dropout, + ) + elif self.pos_enc_type == "rope": + self.self_attn = RotaryPositionMultiHeadedAttention( + embed_dim, attention_heads, dropout=dropout, precision=use_fp16 + ) + elif self.pos_enc_type == "abs": + self.self_attn = ESPNETMultiHeadedAttention( + embed_dim, + attention_heads, + dropout=dropout, + ) + else: + raise Exception(f"Unsupported attention type {self.pos_enc_type}") + else: + # Default to fairseq MHA + self.self_attn = MultiheadAttention( + embed_dim, + attention_heads, + dropout=dropout, + ) + + self.conv_module = ConvolutionModule( + embed_dim=embed_dim, + channels=embed_dim, + depthwise_kernel_size=depthwise_conv_kernel_size, + dropout=dropout, + activation_fn=activation_fn, + ) + + self.ffn2 = FeedForwardModule( + embed_dim, + ffn_embed_dim, + dropout, + dropout, + activation_fn=activation_fn, + ) + self.final_layer_norm = LayerNorm(embed_dim, export=False) + + def forward( + self, + x, + encoder_padding_mask: Optional[torch.Tensor], + position_emb: Optional[torch.Tensor] = None, + ): + """ + Args: + x: Tensor of shape T X B X C + encoder_padding_mask: Optional mask tensor + positions: + Returns: + Tensor of shape T X B X C + """ + residual = x + x = self.ffn1(x) + x = x * 0.5 + residual + residual = x + x = self.self_attn_layer_norm(x) + if self.pos_enc_type == "rel_pos": + x, attn = self.self_attn( + query=x, + key=x, + value=x, + key_padding_mask=encoder_padding_mask, + pos_emb=position_emb, + need_weights=False, + ) + else: + x, attn = self.self_attn( + query=x, + key=x, + value=x, + key_padding_mask=encoder_padding_mask, + need_weights=False, + ) + x = self.self_attn_dropout(x) + x = x + residual + + residual = x + # TBC to BTC + x = x.transpose(0, 1) + x = self.conv_module(x) + # BTC to TBC + x = x.transpose(0, 1) + x = residual + x + + residual = x + x = self.ffn2(x) + + layer_result = x + + x = x * 0.5 + residual + + x = self.final_layer_norm(x) + return x, (attn, layer_result) + + +class ConformerWav2Vec2EncoderLayer(ConformerEncoderLayer): + """Encoder layer for Wav2vec2 encoder""" + + def forward( + self, + x: torch.Tensor, + self_attn_mask: torch.Tensor = None, + self_attn_padding_mask: torch.Tensor = None, + need_weights: bool = False, + att_args=None, + position_emb=None, + ): + return super().forward(x, self_attn_padding_mask, position_emb) diff --git a/data/fairseq/fairseq/modules/conv_tbc.py b/data/fairseq/fairseq/modules/conv_tbc.py new file mode 100644 index 0000000000000000000000000000000000000000..65e17ec94f7e595cb657b3d2daaa1052a95d0677 --- /dev/null +++ b/data/fairseq/fairseq/modules/conv_tbc.py @@ -0,0 +1,53 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import torch +from torch import nn +from torch.nn.modules.utils import _single +from torch import Tensor + + +class ConvTBC(torch.nn.Module): + """1D convolution over an input of shape (time x batch x channel) + + The implementation uses gemm to perform the convolution. This implementation + is faster than cuDNN for small kernel sizes. + """ + + def __init__(self, in_channels, out_channels, kernel_size, padding=0): + super(ConvTBC, self).__init__() + self.in_channels = in_channels + self.out_channels = out_channels + self.kernel_size = _single(kernel_size) + self.padding = _single(padding) + + self.weight = torch.nn.Parameter( + torch.Tensor(self.kernel_size[0], in_channels, out_channels) + ) + self.bias = torch.nn.Parameter(torch.Tensor(out_channels)) + + self.reset_parameters() + + def reset_parameters(self): + nn.init.xavier_normal_(self.weight) + nn.init.zeros_(self.bias) + + def conv_tbc(self, input: Tensor): + return torch.conv_tbc( + input.contiguous(), self.weight, self.bias, self.padding[0] + ) + + def forward(self, input: Tensor): + return self.conv_tbc(input) + + def __repr__(self): + s = ( + "{name}({in_channels}, {out_channels}, kernel_size={kernel_size}" + ", padding={padding}" + ) + if self.bias is None: + s += ", bias=False" + s += ")" + return s.format(name=self.__class__.__name__, **self.__dict__) diff --git a/data/fairseq/fairseq/modules/cross_entropy.py b/data/fairseq/fairseq/modules/cross_entropy.py new file mode 100644 index 0000000000000000000000000000000000000000..286c00eeccbc9d5596a793a20d9c7fdf3dd092f1 --- /dev/null +++ b/data/fairseq/fairseq/modules/cross_entropy.py @@ -0,0 +1,59 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import logging + +import torch +import torch.nn.functional as F + +logger = logging.getLogger(__name__) + + +def _cross_entropy_pytorch(logits, target, ignore_index=None, reduction="mean"): + lprobs = F.log_softmax(logits, dim=-1, dtype=torch.float32) + return F.nll_loss( + lprobs, + target, + ignore_index=ignore_index, + reduction=reduction, + ) + + +try: + import xentropy_cuda + from apex.contrib import xentropy + + def cross_entropy(logits, target, ignore_index=-100, reduction="mean"): + if logits.device == torch.device("cpu"): + return _cross_entropy_pytorch(logits, target, ignore_index, reduction) + else: + if not getattr(cross_entropy, "_has_logged_once", False): + logger.info("using fused cross entropy") + cross_entropy._has_logged_once = True + + half_to_float = logits.dtype == torch.half + losses = xentropy.SoftmaxCrossEntropyLoss.apply( + logits, + target, + 0.0, + ignore_index, + half_to_float, + ) + if reduction == "sum": + return losses.sum() + elif reduction == "mean": + if ignore_index >= 0: + return losses.sum() / target.ne(ignore_index).sum() + else: + return losses.mean() + elif reduction == "none": + return losses + else: + raise NotImplementedError + +except ImportError: + + def cross_entropy(logits, target, ignore_index=-100, reduction="mean"): + return _cross_entropy_pytorch(logits, target, ignore_index, reduction) diff --git a/data/fairseq/fairseq/modules/cuda_utils.cu b/data/fairseq/fairseq/modules/cuda_utils.cu new file mode 100644 index 0000000000000000000000000000000000000000..924f852758ee654e462546274db4b5e7199a9c90 --- /dev/null +++ b/data/fairseq/fairseq/modules/cuda_utils.cu @@ -0,0 +1,202 @@ +/** + * Copyright (c) Facebook, Inc. and its affiliates. + * + * This source code is licensed under the MIT license found in the + * LICENSE file in the root directory of this source tree. + */ + +template +constexpr __host__ __device__ auto divUp(U a, V b) -> decltype(a + b) { + return (a + b - 1) / b; +} + +template +__inline__ __device__ void zeroSharedMem(scalar_t* data) { + /* + Given an array of length FS + SB, zero out the first padding_l and last + (FS - padding_l) values in the array + */ + + int tid = threadIdx.x; + + if (FS < SB) { + // zero all if we have enough threads in a block to do all of them + if (tid < padding_l || tid > SB - FS + padding_l - 1) { + data[tid] = scalar_t(0.0); + } + } else { + // otherwise zero out one block at a time + const int numIterations = divUp(FS, SB); + for (int i = 0; i < numIterations; i++) { + int offset = i * SB; + if (tid + offset < padding_l) { + data[tid + offset] = scalar_t(0.0); + } else if (tid + offset < FS) { + data[SB + tid + offset] = scalar_t(0.0); + } + } + } +} + +template +__inline__ __device__ scalar_t warpReduce(scalar_t data) { + /* + Reduce an array within each warp. After processing all values in warp will + caontain the sum of all original values in that warp. + + data - pointer to data to reduce + */ + data += __shfl_xor_sync(SHFL_MASK, data, 16); + data += __shfl_xor_sync(SHFL_MASK, data, 8); + data += __shfl_xor_sync(SHFL_MASK, data, 4); + data += __shfl_xor_sync(SHFL_MASK, data, 2); + data += __shfl_xor_sync(SHFL_MASK, data, 1); + return data; +} + +template +__inline__ __device__ scalar_t blockReduce(scalar_t data) { + /* + Reduce an entire array on the block level. After processing, the + first value in the array will contain the reduced sum. + + data - pointer to data to reduce + */ + + static __shared__ scalar_t warpSum[32]; + const int tid = threadIdx.x; + int wid = tid / 32; + int lane = tid % 32; + + __syncthreads(); + + // reduce each warp then write to shared memory + scalar_t sum = warpReduce(data); + if (lane == 0) { + warpSum[wid] = sum; + } + + __syncthreads(); + + scalar_t v; + // perform final sum of partial warp sums + if (tid < blockDim.x / 32) { + v = warpSum[lane]; + } else { + v = scalar_t(0.0); + } + + if (wid == 0) { + v = warpReduce(v); + } + __syncthreads(); + + return v; +} + +void checkCudaStatus(cudaError_t status, int lineNumber = -1) { + if (status != cudaSuccess) { + std::cout << cudaGetErrorString(status) << " at line " << lineNumber + << std::endl; + std::cout << "Exiting" << std::endl; + exit(1); + } +} + +template +__device__ void load_input_to_shared( + const scalar_t* input, // global memory + int inputOffset, + int sequenceLength, + int iteration, + int numIterations, + bool no_prev, + scalar_t* output /* shared memory */) { + /* + Load a block size of input into shared memory with + right and left overhang of total size FS. If previously + loaded memory, overlap will be shifted over to reduce + global memory access + + input - pointer to start of channel sequence + inputOffset - how far in the sequence to start loading + sequenceLength - total length of sequence + iteration - which block of sequence we are loading + numIterations - total number of blocks to load + no_prev - whether to load the whole block if the previous block + wasn't loaded + output - shared memory to write input to + */ + + const int tid = threadIdx.x; + + // Load the left "overhang" of input + if (iteration > 0) { + if (padding_l < SB) { + // load all at once + if (tid < padding_l) { + output[tid] = + (no_prev) ? input[inputOffset - padding_l + tid] : output[tid + SB]; + } + } else { + // load in chunks of size SB + int numIterations = divUp(padding_l, SB); + for (int i = 0; i < numIterations; i++) { + int offset = i * SB; + if ((tid + offset) < padding_l) { + output[tid + offset] = (no_prev) + ? input[inputOffset - padding_l + tid + offset] + : output[tid + offset + SB]; + } + } + } + } + + // Load the right "overhang" of input + if (iteration < (numIterations - 1)) { + const int elementsLeft = sequenceLength - (iteration + 1) * SB; + + if ((FS - padding_l) < SB) { + // load all at once + if (tid < (FS - padding_l)) { + output[padding_l + SB + tid] = (tid < elementsLeft) + ? input[inputOffset + SB + tid] + : scalar_t(0.0); + } + } else { + // load in chunks of size SB + int numIterations = divUp(FS - padding_l, SB); + for (int i = 0; i < numIterations; i++) { + int offset = i * SB; + if ((tid + offset) < (FS - padding_l)) { + output[padding_l + SB + tid + offset] = + ((tid + offset) < elementsLeft) + ? input[inputOffset + SB + tid + offset] + : scalar_t(0.0); + } + } + } + } + + // We should also clear out the right "overhang" + if (iteration == (numIterations - 1)) { + if ((FS - padding_l) < SB) { + // clear out all at once + if (tid < (FS - padding_l)) { + output[padding_l + SB + tid] = scalar_t(0.0); + } + } else { + // clear in chunks of size SB + int numIterations = divUp(FS - padding_l, SB); + for (int i = 0; i < numIterations; i++) { + int offset = i * SB; + if ((tid + offset) < (FS - padding_l)) { + output[padding_l + SB + tid + offset] = scalar_t(0.0); + } + } + } + } + output[tid + padding_l] = ((inputOffset + tid) < sequenceLength) + ? input[inputOffset + tid] + : scalar_t(0.0); +} diff --git a/data/fairseq/fairseq/modules/downsampled_multihead_attention.py b/data/fairseq/fairseq/modules/downsampled_multihead_attention.py new file mode 100644 index 0000000000000000000000000000000000000000..5e42942a9f6c8a396ff848d181dd8c7885ee97bc --- /dev/null +++ b/data/fairseq/fairseq/modules/downsampled_multihead_attention.py @@ -0,0 +1,317 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. +# + +import math + +import torch +import torch.nn as nn +import torch.nn.functional as F + +from fairseq.modules.fairseq_dropout import FairseqDropout +from fairseq.modules.scalar_bias import scalar_bias + + +class SingleHeadAttention(nn.Module): + """ + Single-head attention that supports Gating and Downsampling + """ + + def __init__( + self, + out_channels, + embed_dim, + head_dim, + head_index, + dropout=0.0, + bias=True, + project_input=True, + gated=False, + downsample=False, + num_heads=1, + ): + super().__init__() + self.embed_dim = embed_dim + self.dropout_module = FairseqDropout( + dropout, module_name=self.__class__.__name__ + ) + self.head_index = head_index + self.head_dim = head_dim + self.project_input = project_input + self.gated = gated + self.downsample = downsample + self.num_heads = num_heads + self.projection = None + + k_layers = [] + v_layers = [] + if self.downsample: + k_layers.append(Downsample(self.head_index)) + v_layers.append(Downsample(self.head_index)) + out_proj_size = self.head_dim + else: + out_proj_size = self.head_dim * self.num_heads + if self.gated: + k_layers.append(GatedLinear(self.embed_dim, out_proj_size, bias=bias)) + self.in_proj_q = GatedLinear(self.embed_dim, out_proj_size, bias=bias) + v_layers.append(GatedLinear(self.embed_dim, out_proj_size, bias=bias)) + else: + k_layers.append(Linear(self.embed_dim, out_proj_size, bias=bias)) + self.in_proj_q = Linear(self.embed_dim, out_proj_size, bias=bias) + v_layers.append(Linear(self.embed_dim, out_proj_size, bias=bias)) + + self.in_proj_k = nn.Sequential(*k_layers) + self.in_proj_v = nn.Sequential(*v_layers) + + if self.downsample: + self.out_proj = Linear(out_proj_size, self.head_dim, bias=bias) + else: + self.out_proj = Linear(out_proj_size, out_channels, bias=bias) + + self.scaling = self.head_dim**-0.5 + + def forward( + self, + query, + key, + value, + mask_future_timesteps=False, + key_padding_mask=None, + use_scalar_bias=False, + ): + """Input shape: Time x Batch x Channel + Self-attention can be implemented by passing in the same arguments for + query, key and value. Future timesteps can be masked with the + `mask_future_timesteps` argument. Padding elements can be excluded from + the key by passing a binary ByteTensor (`key_padding_mask`) with shape: + batch x src_len, where padding elements are indicated by 1s. + """ + src_len, bsz, out_channels = key.size() + tgt_len = query.size(0) + assert list(query.size()) == [tgt_len, bsz, out_channels] + assert key.size() == value.size() + + if key_padding_mask is not None: + assert key_padding_mask.size(0) == bsz + assert key_padding_mask.size(1) == src_len + + if self.downsample: + size = bsz + else: + size = bsz * self.num_heads + + k = key + v = value + q = query + if self.project_input: + q = self.in_proj_q(q) + k = self.in_proj_k(k) + v = self.in_proj_v(v) + src_len = k.size()[0] + q *= self.scaling + + if not self.downsample: + q = q.view(tgt_len, size, self.head_dim) + k = k.view(src_len, size, self.head_dim) + v = v.view(src_len, size, self.head_dim) + + q = q.transpose(0, 1) + k = k.transpose(0, 1) + v = v.transpose(0, 1) + + attn_weights = torch.bmm(q, k.transpose(1, 2)) + if mask_future_timesteps: + assert ( + query.size() == key.size() + ), "mask_future_timesteps only applies to self-attention" + attn_weights *= torch.tril( + attn_weights.data.new([1]).expand(tgt_len, tgt_len).clone(), + diagonal=-1, + )[:, :: self.head_index + 1 if self.downsample else 1].unsqueeze(0) + attn_weights += torch.triu( + attn_weights.data.new([-math.inf]).expand(tgt_len, tgt_len).clone(), + diagonal=0, + )[:, :: self.head_index + 1 if self.downsample else 1].unsqueeze(0) + tgt_size = tgt_len + if use_scalar_bias: + attn_weights = scalar_bias(attn_weights, 2) + v = scalar_bias(v, 1) + tgt_size += 1 + + if key_padding_mask is not None: + # don't attend to padding symbols + if key_padding_mask.max() > 0: + if self.downsample: + attn_weights = attn_weights.view(bsz, 1, tgt_len, src_len) + else: + attn_weights = attn_weights.view( + size, self.num_heads, tgt_len, src_len + ) + attn_weights = attn_weights.masked_fill( + key_padding_mask.unsqueeze(1).unsqueeze(2), + -math.inf, + ) + attn_weights = attn_weights.view(size, tgt_len, src_len) + attn_weights = F.softmax(attn_weights, dim=-1) + attn_weights = self.dropout_module(attn_weights) + + attn = torch.bmm(attn_weights, v) + if self.downsample: + attn = attn.transpose(0, 1).contiguous().view(tgt_len, bsz, self.head_dim) + else: + attn = attn.transpose(0, 1).contiguous().view(tgt_len, bsz, self.embed_dim) + + attn = self.out_proj(attn) + + return attn, attn_weights + + +class DownsampledMultiHeadAttention(nn.ModuleList): + """ + Multi-headed attention with Gating and Downsampling + """ + + def __init__( + self, + out_channels, + embed_dim, + num_heads, + dropout=0.0, + bias=True, + project_input=True, + gated=False, + downsample=False, + ): + self.embed_dim = embed_dim + self.num_heads = num_heads + self.head_dim = embed_dim // num_heads + self.downsample = downsample + self.gated = gated + self.project_input = project_input + assert self.head_dim * num_heads == embed_dim + + if self.downsample: + attention_heads = [] + for index in range(self.num_heads): + attention_heads.append( + SingleHeadAttention( + out_channels, + self.embed_dim, + self.head_dim, + index, + dropout, + bias, + self.project_input, + self.gated, + self.downsample, + self.num_heads, + ) + ) + super().__init__(modules=attention_heads) + self.out_proj = Linear(embed_dim, out_channels, bias=bias) + else: + # either we have a list of attention heads, or just one attention head + # if not being downsampled, we can do the heads with one linear layer instead of separate ones + super().__init__() + self.attention_module = SingleHeadAttention( + out_channels, + self.embed_dim, + self.head_dim, + 1, + dropout, + bias, + self.project_input, + self.gated, + self.downsample, + self.num_heads, + ) + + def forward( + self, + query, + key, + value, + mask_future_timesteps=False, + key_padding_mask=None, + use_scalar_bias=False, + ): + src_len, bsz, embed_dim = key.size() + tgt_len = query.size(0) + assert embed_dim == self.embed_dim + assert list(query.size()) == [tgt_len, bsz, embed_dim] + assert key.size() == value.size() + + tgt_size = tgt_len + if use_scalar_bias: + tgt_size += 1 + + attn = [] + attn_weights = [] + if self.downsample: + for attention_head_number in range(self.num_heads): + # call the forward of each attention head + _attn, _attn_weight = self[attention_head_number]( + query, + key, + value, + mask_future_timesteps, + key_padding_mask, + use_scalar_bias, + ) + attn.append(_attn) + attn_weights.append(_attn_weight) + full_attn = torch.cat(attn, dim=2) + full_attn = self.out_proj(full_attn) + return full_attn, attn_weights[0].clone() + else: + _attn, _attn_weight = self.attention_module( + query, + key, + value, + mask_future_timesteps, + key_padding_mask, + use_scalar_bias, + ) + attn.append(_attn) + attn_weights.append(_attn_weight) + full_attn = torch.cat(attn, dim=2) + full_attn_weights = torch.cat(attn_weights) + full_attn_weights = full_attn_weights.view( + bsz, self.num_heads, tgt_size, src_len + ) + full_attn_weights = full_attn_weights.sum(dim=1) / self.num_heads + return full_attn, full_attn_weights + + +class Downsample(nn.Module): + """ + Selects every nth element, where n is the index + """ + + def __init__(self, index): + super().__init__() + self.index = index + + def forward(self, x): + return x[:: self.index + 1] + + +def Linear(in_features, out_features, dropout=0.0, bias=True): + """Weight-normalized Linear layer (input: B x T x C)""" + m = nn.Linear(in_features, out_features, bias=bias) + m.weight.data.normal_(mean=0, std=math.sqrt((1 - dropout) / in_features)) + m.bias.data.zero_() + return nn.utils.weight_norm(m) + + +def GatedLinear(in_features, out_features, dropout=0.0, bias=True): + """Weight-normalized Linear layer (input: B x T x C) with interspersed GLU units""" + return nn.Sequential( + Linear(in_features, out_features * 4, dropout, bias), + nn.GLU(), + Linear(out_features * 2, out_features * 2, dropout, bias), + nn.GLU(), + Linear(out_features, out_features, dropout, bias), + ) diff --git a/data/fairseq/fairseq/modules/dynamic_convolution.py b/data/fairseq/fairseq/modules/dynamic_convolution.py new file mode 100644 index 0000000000000000000000000000000000000000..0ff02cd62a27dd34232cc7b5f1fccfe57fa13041 --- /dev/null +++ b/data/fairseq/fairseq/modules/dynamic_convolution.py @@ -0,0 +1,526 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from typing import Dict, Optional + +import torch +import torch.nn as nn +import torch.nn.functional as F +from fairseq import utils +from fairseq.incremental_decoding_utils import ( + FairseqIncrementalState, + with_incremental_state, +) +from fairseq.modules.fairseq_dropout import FairseqDropout +from torch import Tensor + +from .unfold import unfold1d + + +def DynamicConv( + input_size, + kernel_size=1, + padding_l=None, + num_heads=1, + weight_dropout=0.0, + weight_softmax=False, + renorm_padding=False, + bias=False, + conv_bias=False, + query_size=None, + in_proj=False, +): + if torch.cuda.is_available(): + try: + from fairseq.modules.dynamicconv_layer import DynamicconvLayer + + return DynamicconvLayer( + input_size, + kernel_size=kernel_size, + padding_l=padding_l, + num_heads=num_heads, + weight_dropout=weight_dropout, + weight_softmax=weight_softmax, + renorm_padding=renorm_padding, + bias=bias, + conv_bias=conv_bias, + query_size=query_size, + ) + except ImportError as e: + print(e) + return DynamicConv1dTBC( + input_size, + kernel_size=kernel_size, + padding_l=padding_l, + num_heads=num_heads, + weight_dropout=weight_dropout, + weight_softmax=weight_softmax, + renorm_padding=renorm_padding, + bias=bias, + conv_bias=conv_bias, + query_size=query_size, + ) + + +def Linear(in_features, out_features, bias=True): + m = nn.Linear(in_features, out_features, bias) + nn.init.xavier_uniform_(m.weight) + if bias: + nn.init.constant_(m.bias, 0.0) + return m + + +@with_incremental_state +class DynamicConv1dTBC(nn.Module): + """Dynamic lightweight convolution taking T x B x C inputs + Args: + input_size: # of channels of the input + kernel_size: convolution channels + padding_l: padding to the left when using "same" padding + num_heads: number of heads used. The weight is of shape (num_heads, 1, kernel_size) + weight_dropout: the drop rate of the DropConnect to drop the weight + weight_softmax: normalize the weight with softmax before the convolution + renorm_padding: re-normalize the filters to ignore the padded part (only the non-padding parts sum up to 1) + bias: use bias + conv_bias: bias of the convolution + query_size: specified when feeding a different input as the query + in_proj: project the input and generate the filter together + + Shape: + Input: TxBxC, i.e. (timesteps, batch_size, input_size) + Output: TxBxC, i.e. (timesteps, batch_size, input_size) + + Attributes: + weight: the learnable weights of the module of shape + `(num_heads, 1, kernel_size)` + bias: the learnable bias of the module of shape `(input_size)` + """ + + def __init__( + self, + input_size, + kernel_size=1, + padding_l=None, + num_heads=1, + weight_dropout=0.0, + weight_softmax=False, + renorm_padding=False, + bias=False, + conv_bias=False, + query_size=None, + in_proj=False, + ): + super().__init__() + self.input_size = input_size + self.query_size = input_size if query_size is None else query_size + self.kernel_size = kernel_size + self.padding_l = padding_l + self.num_heads = num_heads + self.weight_dropout_module = FairseqDropout( + weight_dropout, module_name=self.__class__.__name__ + ) + self.weight_softmax = weight_softmax + self.renorm_padding = renorm_padding + + if in_proj: + self.weight_linear = Linear( + self.input_size, self.input_size + num_heads * kernel_size * 1 + ) + else: + self.weight_linear = Linear( + self.query_size, num_heads * kernel_size * 1, bias=bias + ) + if conv_bias: + self.conv_bias = nn.Parameter(torch.Tensor(input_size)) + else: + self.conv_bias = None + self.reset_parameters() + + @property + def in_proj(self): + return ( + self.weight_linear.out_features + == self.input_size + self.num_heads * self.kernel_size + ) + + def reset_parameters(self): + self.weight_linear.reset_parameters() + if self.conv_bias is not None: + nn.init.constant_(self.conv_bias, 0.0) + + def forward(self, x, incremental_state=None, query=None, unfold=None): + """Assuming the input, x, of the shape T x B x C and producing an output in the shape T x B x C + args: + x: Input of shape T x B x C, i.e. (timesteps, batch_size, input_size) + incremental_state: A dict to keep the state + unfold: unfold the input or not. If not, we use the matrix trick instead + query: use the specified query to predict the conv filters + """ + unfold = ( + x.size(0) > 512 if unfold is None else unfold + ) # use unfold mode as default for long sequence to save memory + unfold = unfold or (incremental_state is not None) + assert query is None or not self.in_proj + + if query is None: + query = x + if unfold: + output = self._forward_unfolded(x, incremental_state, query) + else: + output = self._forward_expanded(x, incremental_state, query) + + if self.conv_bias is not None: + output = output + self.conv_bias.view(1, 1, -1) + return output + + def _forward_unfolded(self, x, incremental_state, query): + """The conventional implementation of convolutions. + Unfolding the input by having a window shifting to the right.""" + T, B, C = x.size() + K, H = self.kernel_size, self.num_heads + R = C // H + assert R * H == C == self.input_size + + if self.in_proj: + proj = self.weight_linear(x) + x = proj.narrow(2, 0, self.input_size).contiguous() + weight = ( + proj.narrow(2, self.input_size, H * K).contiguous().view(T * B * H, -1) + ) + else: + weight = self.weight_linear(query).view(T * B * H, -1) + + # renorm_padding is only implemented in _forward_expanded + assert not self.renorm_padding or incremental_state is not None + + if incremental_state is not None: + input_buffer = self._get_input_buffer(incremental_state) + if input_buffer is None: + input_buffer = x.new() + x_unfold = torch.cat([input_buffer, x.unsqueeze(3)], dim=3) + if self.kernel_size > 1: + self._set_input_buffer( + incremental_state, x_unfold[:, :, :, -self.kernel_size + 1 :] + ) + x_unfold = x_unfold.view(T * B * H, R, -1) + else: + padding_l = self.padding_l + if K > T and padding_l == K - 1: + weight = weight.narrow(1, K - T, T) + K, padding_l = T, T - 1 + # unfold the input: T x B x C --> T' x B x C x K + x_unfold = unfold1d(x, K, padding_l, 0) + x_unfold = x_unfold.view(T * B * H, R, K) + + if self.weight_softmax and not self.renorm_padding: + weight = F.softmax(weight, dim=1) + weight = weight.narrow(1, 0, K) + + if incremental_state is not None: + weight = weight[:, -x_unfold.size(2) :] + K = weight.size(1) + + if self.weight_softmax and self.renorm_padding: + weight = F.softmax(weight, dim=1) + + weight = self.weight_dropout_module(weight, inplace=False) + + output = torch.bmm(x_unfold, weight.unsqueeze(2)) # T*B*H x R x 1 + output = output.view(T, B, C) + return output + + def _forward_expanded(self, x, incremental_stat, query): + """Turn the convolution filters into band matrices and do matrix multiplication. + This is faster when the sequence is short, but less memory efficient. + This is not used in the decoder during inference. + """ + T, B, C = x.size() + K, H = self.kernel_size, self.num_heads + R = C // H + assert R * H == C == self.input_size + if self.in_proj: + proj = self.weight_linear(x) + x = proj.narrow(2, 0, self.input_size).contiguous() + weight = ( + proj.narrow(2, self.input_size, H * K).contiguous().view(T * B * H, -1) + ) + else: + weight = self.weight_linear(query).view(T * B * H, -1) + + if not self.renorm_padding: + if self.weight_softmax: + weight = F.softmax(weight, dim=1) + weight = self.weight_dropout_module(weight, inplace=False) + weight = weight.narrow(1, 0, K).contiguous() + weight = weight.view(T, B * H, K).transpose(0, 1) + + x = x.view(T, B * H, R).transpose(0, 1) + if self.weight_softmax and self.renorm_padding: + # turn the convolution filters into band matrices + weight_expanded = weight.new(B * H, T, T + K - 1).fill_(float("-inf")) + weight_expanded.as_strided( + (B * H, T, K), (T * (T + K - 1), T + K, 1) + ).copy_(weight) + weight_expanded = weight_expanded.narrow(2, self.padding_l, T) + # normalize the weight over valid positions like self-attention + weight_expanded = F.softmax(weight_expanded, dim=2) + weight_expanded = self.weight_dropout_module(weight_expanded, inplace=False) + else: + P = self.padding_l + # For efficiency, we cut the kernel size and reduce the padding when the kernel is larger than the length + if K > T and P == K - 1: + weight = weight.narrow(2, K - T, T) + K, P = T, T - 1 + # turn the convolution filters into band matrices + weight_expanded = weight.new_zeros(B * H, T, T + K - 1, requires_grad=False) + weight_expanded.as_strided( + (B * H, T, K), (T * (T + K - 1), T + K, 1) + ).copy_(weight) + weight_expanded = weight_expanded.narrow(2, P, T) # B*H x T x T + output = torch.bmm(weight_expanded, x) + output = output.transpose(0, 1).contiguous().view(T, B, C) + return output + + def reorder_incremental_state(self, incremental_state, new_order): + input_buffer = self._get_input_buffer(incremental_state) + if input_buffer is not None: + input_buffer = input_buffer.index_select(1, new_order) + self._set_input_buffer(incremental_state, input_buffer) + + def _get_input_buffer(self, incremental_state): + return utils.get_incremental_state(self, incremental_state, "input_buffer") + + def _set_input_buffer(self, incremental_state, new_buffer): + return utils.set_incremental_state( + self, incremental_state, "input_buffer", new_buffer + ) + + def extra_repr(self): + s = "{}, kernel_size={}, padding_l={}, num_heads={}, weight_softmax={}, conv_bias={}, renorm_padding={}, in_proj={}".format( + self.input_size, + self.kernel_size, + self.padding_l, + self.num_heads, + self.weight_softmax, + self.conv_bias is not None, + self.renorm_padding, + self.in_proj, + ) + + if self.query_size != self.input_size: + s += ", query_size={}".format(self.query_size) + if self.weight_dropout_module.p > 0.0: + s += ", weight_dropout={}".format(self.weight_dropout_module.p) + return s + + +class DynamicConv_scripatable(nn.Module, FairseqIncrementalState): + """Dynamic lightweight convolution taking T x B x C inputs + Args: + input_size: # of channels of the input + kernel_size: convolution channels + padding_l: padding to the left when using "same" padding + num_heads: number of heads used. The weight is of shape (num_heads, 1, kernel_size) + weight_dropout: the drop rate of the DropConnect to drop the weight + weight_softmax: normalize the weight with softmax before the convolution + renorm_padding: re-normalize the filters to ignore the padded part (only the non-padding parts sum up to 1) + bias: use bias + conv_bias: bias of the convolution + query_size: specified when feeding a different input as the query + in_proj: project the input and generate the filter together + + Shape: + Input: TxBxC, i.e. (timesteps, batch_size, input_size) + Output: TxBxC, i.e. (timesteps, batch_size, input_size) + + Attributes: + weight: the learnable weights of the module of shape + `(num_heads, 1, kernel_size)` + bias: the learnable bias of the module of shape `(input_size)` + """ + + def __init__( + self, + input_size, + kernel_size=1, + padding_l=None, + num_heads=1, + weight_dropout=0.0, + weight_softmax=False, + renorm_padding=False, + bias=False, + conv_bias=False, + query_size=None, + in_proj=False, + ): + super().__init__() + self.input_size = input_size + self.query_size = input_size if query_size is None else query_size + self.kernel_size = kernel_size + self.padding_l = padding_l + self.num_heads = num_heads + self.weight_dropout_module = FairseqDropout( + weight_dropout, module_name=self.__class__.__name__ + ) + self.weight_softmax = weight_softmax + self.renorm_padding = renorm_padding + + if in_proj: + self.weight_linear = Linear( + self.input_size, self.input_size + num_heads * kernel_size * 1 + ) + else: + self.weight_linear = Linear( + self.query_size, num_heads * kernel_size * 1, bias=bias + ) + self.in_proj = ( + self.weight_linear.out_features + == self.input_size + self.num_heads * self.kernel_size + ) + self.has_conv_bias = conv_bias + self.conv_bias = nn.Parameter(torch.Tensor(input_size).view(1, 1, -1)) + self.init_incremental_state() + + self.reset_parameters() + + def reset_parameters(self): + self.weight_linear.reset_parameters() + if self.has_conv_bias: + nn.init.constant_(self.conv_bias, 0.0) + + def forward( + self, + x, + incremental_state: Optional[Dict[str, Dict[str, Optional[Tensor]]]] = None, + query: Optional[Tensor] = None, + ): + """Assuming the input, x, of the shape T x B x C and producing an output in the shape T x B x C + args: + x: Input of shape T x B x C, i.e. (timesteps, batch_size, input_size) + incremental_state: A dict to keep the state + unfold: unfold the input or not. If not, we use the matrix trick instead + query: use the specified query to predict the conv filters + """ + assert query is None or not self.in_proj + + if query is None: + query = x + + output = self._forward_unfolded(x, incremental_state, query) + + if self.has_conv_bias: + output = output + self.conv_bias + return output + + def _forward_unfolded( + self, + x, + incremental_state: Optional[Dict[str, Dict[str, Optional[Tensor]]]], + query, + ): + """The conventional implementation of convolutions. + Unfolding the input by having a window shifting to the right.""" + T, B, C = x.size() + K, H = self.kernel_size, self.num_heads + R = C // H + assert R * H == C == self.input_size + + TxBxH = T * B * H + + if self.in_proj: + proj = self.weight_linear(x) + x = proj.narrow(2, 0, self.input_size).contiguous() + weight = proj.narrow(2, self.input_size, H * K).contiguous().view(TxBxH, -1) + else: + weight = self.weight_linear(query).view(TxBxH, -1) + + # renorm_padding is only implemented in _forward_expanded + assert not self.renorm_padding or incremental_state is not None + + if incremental_state is not None: + input_buffer = self._get_input_buffer(incremental_state) + if input_buffer is not None: + x_unfold = torch.cat([input_buffer, x.unsqueeze(3)], dim=3) + else: + x_unfold = x.unsqueeze(3).clone() + if self.kernel_size > 1: + self._set_input_buffer( + incremental_state, x_unfold[:, :, :, -self.kernel_size + 1 :] + ) + x_unfold = x_unfold.view(TxBxH, R, -1) + else: + padding_l = self.padding_l + if K > T and padding_l == K - 1: + weight = weight.narrow(1, K - T, T) + K, padding_l = T, T - 1 + # unfold the input: T x B x C --> T' x B x C x K + x_unfold = unfold1d(x, K, padding_l, 0.0) + x_unfold = x_unfold.view(TxBxH, R, K) + + if self.weight_softmax and not self.renorm_padding: + weight = F.softmax(weight, dim=1) + weight = weight.narrow(1, 0, K) + + if incremental_state is not None: + weight = weight[:, -(x_unfold.size(2)) :] + K = weight.size(1) + + if self.weight_softmax and self.renorm_padding: + weight = F.softmax(weight, dim=1) + + weight = self.weight_dropout_module(weight, inplace=False) + + output = torch.bmm(x_unfold, weight.unsqueeze(2)) # T x B x H x R x 1 + output = output.view(T, B, C) + return output + + def reorder_incremental_state( + self, + incremental_state: Optional[Dict[str, Dict[str, Optional[Tensor]]]], + new_order: Tensor, + ): + input_buffer = self._get_input_buffer(incremental_state) + if input_buffer is not None: + input_buffer = input_buffer.index_select(1, new_order) + self._set_input_buffer(incremental_state, input_buffer) + + def _get_input_buffer( + self, incremental_state: Optional[Dict[str, Dict[str, Optional[Tensor]]]] + ): + result = self.get_incremental_state(incremental_state, "input_buffer") + if result is not None and "input_buffer" in result: + return result["input_buffer"] + else: + return None + + def _set_input_buffer( + self, + incremental_state: Optional[Dict[str, Dict[str, Optional[Tensor]]]], + new_buffer: Optional[Tensor], + ): + result = self.set_incremental_state( + incremental_state, "input_buffer", {"input_buffer": new_buffer} + ) + if result is not None: + incremental_state = result + return incremental_state + + def extra_repr(self): + s = "{}, kernel_size={}, padding_l={}, num_heads={}, weight_softmax={}, conv_bias={}, renorm_padding={}, in_proj={}".format( # noqa + self.input_size, + self.kernel_size, + self.padding_l, + self.num_heads, + self.weight_softmax, + self.conv_bias is not None, + self.renorm_padding, + self.in_proj, + ) + + if self.query_size != self.input_size: + s += ", query_size={}".format(self.query_size) + if self.weight_dropout_module.p > 0.0: + s += ", weight_dropout={}".format(self.weight_dropout_module.p) + return s diff --git a/data/fairseq/fairseq/modules/dynamic_crf_layer.py b/data/fairseq/fairseq/modules/dynamic_crf_layer.py new file mode 100644 index 0000000000000000000000000000000000000000..8fcc6b8d2672d2eacc6d01b9688bac44d5e1ce26 --- /dev/null +++ b/data/fairseq/fairseq/modules/dynamic_crf_layer.py @@ -0,0 +1,189 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +""" +This file is to re-implemented the low-rank and beam approximation of CRF layer +Proposed by: + +Sun, Zhiqing, et al. +Fast Structured Decoding for Sequence Models +https://arxiv.org/abs/1910.11555 + +The CRF implementation is mainly borrowed from +https://github.com/kmkurn/pytorch-crf/blob/master/torchcrf/__init__.py + +""" + +import numpy as np +import torch +import torch.nn as nn + + +def logsumexp(x, dim=1): + return torch.logsumexp(x.float(), dim=dim).type_as(x) + + +class DynamicCRF(nn.Module): + """Dynamic CRF layer is used to approximate the traditional + Conditional Random Fields (CRF) + $P(y | x) = 1/Z(x) exp(sum_i s(y_i, x) + sum_i t(y_{i-1}, y_i, x))$ + + where in this function, we assume the emition scores (s) are given, + and the transition score is a |V| x |V| matrix $M$ + + in the following two aspects: + (1) it used a low-rank approximation for the transition matrix: + $M = E_1 E_2^T$ + (2) it used a beam to estimate the normalizing factor Z(x) + """ + + def __init__(self, num_embedding, low_rank=32, beam_size=64): + super().__init__() + + self.E1 = nn.Embedding(num_embedding, low_rank) + self.E2 = nn.Embedding(num_embedding, low_rank) + + self.vocb = num_embedding + self.rank = low_rank + self.beam = beam_size + + def extra_repr(self): + return "vocab_size={}, low_rank={}, beam_size={}".format( + self.vocb, self.rank, self.beam + ) + + def forward(self, emissions, targets, masks, beam=None): + """ + Compute the conditional log-likelihood of a sequence of target tokens given emission scores + + Args: + emissions (`~torch.Tensor`): Emission score are usually the unnormalized decoder output + ``(batch_size, seq_len, vocab_size)``. We assume batch-first + targets (`~torch.LongTensor`): Sequence of target token indices + ``(batch_size, seq_len) + masks (`~torch.ByteTensor`): Mask tensor with the same size as targets + + Returns: + `~torch.Tensor`: approximated log-likelihood + """ + numerator = self._compute_score(emissions, targets, masks) + denominator = self._compute_normalizer(emissions, targets, masks, beam) + return numerator - denominator + + def forward_decoder(self, emissions, masks=None, beam=None): + """ + Find the most likely output sequence using Viterbi algorithm. + + Args: + emissions (`~torch.Tensor`): Emission score are usually the unnormalized decoder output + ``(batch_size, seq_len, vocab_size)``. We assume batch-first + masks (`~torch.ByteTensor`): Mask tensor with the same size as targets + + Returns: + `~torch.LongTensor`: decoded sequence from the CRF model + """ + return self._viterbi_decode(emissions, masks, beam) + + def _compute_score(self, emissions, targets, masks=None): + batch_size, seq_len = targets.size() + emission_scores = emissions.gather(2, targets[:, :, None])[:, :, 0] # B x T + transition_scores = (self.E1(targets[:, :-1]) * self.E2(targets[:, 1:])).sum(2) + + scores = emission_scores + scores[:, 1:] += transition_scores + + if masks is not None: + scores = scores * masks.type_as(scores) + return scores.sum(-1) + + def _compute_normalizer(self, emissions, targets=None, masks=None, beam=None): + # HACK: we include "target" which is a hueristic for training + # HACK: we use a beam of tokens to approximate the normalizing factor (which is bad?) + + beam = beam if beam is not None else self.beam + batch_size, seq_len = emissions.size()[:2] + if targets is not None: + _emissions = emissions.scatter(2, targets[:, :, None], np.float("inf")) + beam_targets = _emissions.topk(beam, 2)[1] + beam_emission_scores = emissions.gather(2, beam_targets) + else: + beam_emission_scores, beam_targets = emissions.topk(beam, 2) + beam_transition_score1 = self.E1(beam_targets[:, :-1]) # B x (T-1) x K x D + beam_transition_score2 = self.E2(beam_targets[:, 1:]) # B x (T-1) x K x D + beam_transition_matrix = torch.bmm( + beam_transition_score1.view(-1, beam, self.rank), + beam_transition_score2.view(-1, beam, self.rank).transpose(1, 2), + ) + beam_transition_matrix = beam_transition_matrix.view(batch_size, -1, beam, beam) + + # compute the normalizer in the log-space + score = beam_emission_scores[:, 0] # B x K + for i in range(1, seq_len): + next_score = score[:, :, None] + beam_transition_matrix[:, i - 1] + next_score = logsumexp(next_score, dim=1) + beam_emission_scores[:, i] + + if masks is not None: + score = torch.where(masks[:, i : i + 1], next_score, score) + else: + score = next_score + + # Sum (log-sum-exp) over all possible tags + return logsumexp(score, dim=1) + + def _viterbi_decode(self, emissions, masks=None, beam=None): + # HACK: we use a beam of tokens to approximate the normalizing factor (which is bad?) + + beam = beam if beam is not None else self.beam + batch_size, seq_len = emissions.size()[:2] + beam_emission_scores, beam_targets = emissions.topk(beam, 2) + beam_transition_score1 = self.E1(beam_targets[:, :-1]) # B x (T-1) x K x D + beam_transition_score2 = self.E2(beam_targets[:, 1:]) # B x (T-1) x K x D + beam_transition_matrix = torch.bmm( + beam_transition_score1.view(-1, beam, self.rank), + beam_transition_score2.view(-1, beam, self.rank).transpose(1, 2), + ) + beam_transition_matrix = beam_transition_matrix.view(batch_size, -1, beam, beam) + + traj_tokens, traj_scores = [], [] + finalized_tokens, finalized_scores = [], [] + + # compute the normalizer in the log-space + score = beam_emission_scores[:, 0] # B x K + dummy = ( + torch.arange(beam, device=score.device).expand(*score.size()).contiguous() + ) + + for i in range(1, seq_len): + traj_scores.append(score) + _score = score[:, :, None] + beam_transition_matrix[:, i - 1] + _score, _index = _score.max(dim=1) + _score = _score + beam_emission_scores[:, i] + + if masks is not None: + score = torch.where(masks[:, i : i + 1], _score, score) + index = torch.where(masks[:, i : i + 1], _index, dummy) + else: + score, index = _score, _index + traj_tokens.append(index) + + # now running the back-tracing and find the best + best_score, best_index = score.max(dim=1) + finalized_tokens.append(best_index[:, None]) + finalized_scores.append(best_score[:, None]) + + for idx, scs in zip(reversed(traj_tokens), reversed(traj_scores)): + previous_index = finalized_tokens[-1] + finalized_tokens.append(idx.gather(1, previous_index)) + finalized_scores.append(scs.gather(1, previous_index)) + + finalized_tokens.reverse() + finalized_tokens = torch.cat(finalized_tokens, 1) + finalized_tokens = beam_targets.gather(2, finalized_tokens[:, :, None])[:, :, 0] + + finalized_scores.reverse() + finalized_scores = torch.cat(finalized_scores, 1) + finalized_scores[:, 1:] = finalized_scores[:, 1:] - finalized_scores[:, :-1] + + return finalized_scores, finalized_tokens diff --git a/data/fairseq/fairseq/modules/dynamicconv_layer/__init__.py b/data/fairseq/fairseq/modules/dynamicconv_layer/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..22dc6f403d2a0ecdb1b9e7e69ed96bd560e93b2c --- /dev/null +++ b/data/fairseq/fairseq/modules/dynamicconv_layer/__init__.py @@ -0,0 +1,6 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from .dynamicconv_layer import DynamicconvLayer # noqa diff --git a/data/fairseq/fairseq/modules/dynamicconv_layer/cuda_function_gen.py b/data/fairseq/fairseq/modules/dynamicconv_layer/cuda_function_gen.py new file mode 100644 index 0000000000000000000000000000000000000000..9304f99eb8169a614f39babc830c84cac80e080b --- /dev/null +++ b/data/fairseq/fairseq/modules/dynamicconv_layer/cuda_function_gen.py @@ -0,0 +1,223 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + + +def gen_forward(): + + kernels = [3, 5, 7, 15, 31, 63, 127, 255] + blocks = [32, 64, 128, 256] + + head = """ +/** + * Copyright (c) Facebook, Inc. and its affiliates. + * + * This source code is licensed under the MIT license found in the + * LICENSE file in the root directory of this source tree. + */ + +#include "dynamicconv_cuda.cuh" + +std::vector dynamicconv_cuda_forward(at::Tensor input, at::Tensor weight, int padding_l) { + + at::DeviceGuard g(input.device()); + const auto minibatch = input.size(0); + const auto numFeatures = input.size(1); + const auto sequenceLength = input.size(2); + + const auto numHeads = weight.size(1); + const auto filterSize = weight.size(2); + + const auto numFiltersInBlock = numFeatures / numHeads; + const dim3 blocks(minibatch, numFeatures); + + auto output = at::zeros_like(input); + auto stream = at::cuda::getCurrentCUDAStream(); +""" + + switch = """ + switch(filterSize) { +""" + + case_k = """ + case {k}: +""" + + main_block = """ + if (padding_l == {pad}) {{ + AT_DISPATCH_FLOATING_TYPES_AND_HALF(input.scalar_type(), "dynamicconv_forward", ([&] {{ + dynamicconv_forward_kernel<{k}, {b_size}, {pad}, scalar_t> + <<>>( + input.data(), + weight.data(), + minibatch, + sequenceLength, + numFeatures, + numFiltersInBlock, + numHeads, + output.data()); + }})); + }} else +""" + + bad_padding = """ + { + std::cout << "WARNING: Unsupported padding size - skipping forward pass" << std::endl; + } + break;\n +""" + + end = """ + default: + std::cout << "WARNING: Unsupported filter length passed - skipping forward pass" << std::endl; + } + + return {output}; +} +""" + + with open("dynamicconv_cuda_forward.cu", "w") as forward: + forward.write(head) + forward.write(switch) + for k in kernels: + b_size = 32 + for b in blocks: + if b > k: + b_size = b + break + forward.write(case_k.format(k=k)) + for pad in [k // 2, k - 1]: + forward.write(main_block.format(k=k, b_size=b_size, pad=pad)) + forward.write(bad_padding) + forward.write(end) + + +def gen_backward(): + + kernels = [3, 5, 7, 15, 31, 63, 127, 255] + thresh = [512, 512, 512, 512, 512, 380, 256, 256] + min_block = [64, 64, 64, 64, 64, 64, 128, 256] + seqs = [32 * x for x in [1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16]] + + head = """ +/** + * Copyright (c) Facebook, Inc. and its affiliates. + * + * This source code is licensed under the MIT license found in the + * LICENSE file in the root directory of this source tree. + */ + +#include "dynamicconv_cuda.cuh" + +std::vector dynamicconv_cuda_backward(at::Tensor gradOutput, int padding_l, at::Tensor input, at::Tensor weight) { + + at::DeviceGuard g(input.device()); + const auto minibatch = input.size(0); + const auto numFeatures = input.size(1); + const auto sequenceLength = input.size(2); + + const auto numHeads = weight.size(1); + const auto filterSize = weight.size(2); + + const auto numFiltersInBlock = numFeatures / numHeads; + auto numChunks = 1; + + auto gradInput = at::zeros_like(input); + auto gradWeight = at::zeros_like(weight); + auto stream = at::cuda::getCurrentCUDAStream(); + + dim3 blocks(minibatch, numHeads, numChunks); +""" + + sequence_if = """ + if (sequenceLength < {seq}) {{ + switch(filterSize) {{ +""" + + case_k = """ + case {k}: +""" + + chunks_reset = """ + numChunks = int(ceilf(sequenceLength/float({b_size}))); + blocks = dim3(minibatch, numHeads, numChunks); +""" + + main_block = """ + if (padding_l == {p}) {{ + AT_DISPATCH_FLOATING_TYPES_AND_HALF(gradOutput.scalar_type(), "dynamicconv_backward", ([&] {{ + dynamicconv_backward_kernel<{k}, {b_size}, {p}, scalar_t> + <<>>( + gradOutput.data(), + input.data(), + weight.data(), + minibatch, + sequenceLength, + numFeatures, + numFiltersInBlock, + numHeads, + gradWeight.data(), + gradInput.data()); + }})); + }} else +""" + + bad_padding = """ + { + std::cout << "WARNING: Unsupported padding size - skipping backward pass" << std::endl; + } + break;\n +""" + + bad_filter = """ + default: + std::cout << "WARNING: Unsupported filter length passed - skipping backward pass" << std::endl; + } +""" + + con_else = """ + } else +""" + + final_else = """ + { + switch(filterSize) { +""" + + last_return = """ + } + return {gradInput, gradWeight}; +} +""" + + with open("dynamicconv_cuda_backward.cu", "w") as backward: + backward.write(head) + for seq in seqs: + backward.write(sequence_if.format(seq=seq)) + for k, t, m in zip(kernels, thresh, min_block): + backward.write(case_k.format(k=k)) + if seq <= t: + b_size = seq + else: + b_size = m + backward.write(chunks_reset.format(b_size=b_size)) + for p in [k // 2, k - 1]: + backward.write(main_block.format(k=k, b_size=b_size, p=p)) + backward.write(bad_padding) + backward.write(bad_filter) + backward.write(con_else) + backward.write(final_else) + for k, m in zip(kernels, min_block): + backward.write(case_k.format(k=k)) + backward.write(chunks_reset.format(b_size=m)) + for p in [k // 2, k - 1]: + backward.write(main_block.format(k=k, b_size=m, p=p)) + backward.write(bad_padding) + backward.write(bad_filter) + backward.write(last_return) + + +if __name__ == "__main__": + gen_forward() + gen_backward() diff --git a/data/fairseq/fairseq/modules/dynamicconv_layer/dynamicconv_cuda.cpp b/data/fairseq/fairseq/modules/dynamicconv_layer/dynamicconv_cuda.cpp new file mode 100644 index 0000000000000000000000000000000000000000..744c363e550231b8e0fbb94f998d46039daf5c00 --- /dev/null +++ b/data/fairseq/fairseq/modules/dynamicconv_layer/dynamicconv_cuda.cpp @@ -0,0 +1,51 @@ +/** + * Copyright (c) Facebook, Inc. and its affiliates. + * + * This source code is licensed under the MIT license found in the + * LICENSE file in the root directory of this source tree. + */ + +#include +#include + +std::vector +dynamicconv_cuda_forward(at::Tensor input, at::Tensor filters, int padding_l); + +std::vector dynamicconv_cuda_backward( + at::Tensor gradOutput, + int padding_l, + at::Tensor input, + at::Tensor filters); + +#define CHECK_CUDA(x) \ + AT_ASSERTM(x.type().is_cuda(), #x " must be a CUDA tensor") +#define CHECK_CONTIGUOUS(x) \ + AT_ASSERTM(x.is_contiguous(), #x " must be contiguous") +#define CHECK_INPUT(x) \ + CHECK_CUDA(x); \ + CHECK_CONTIGUOUS(x) + +std::vector +dynamicconv_forward(at::Tensor input, at::Tensor filters, int padding_l) { + CHECK_INPUT(input); + CHECK_INPUT(filters); + + return dynamicconv_cuda_forward(input, filters, padding_l); +} + +std::vector dynamicconv_backward( + at::Tensor gradOutput, + int padding_l, + at::Tensor input, + at::Tensor filters) { + CHECK_INPUT(gradOutput); + CHECK_INPUT(input); + CHECK_INPUT(filters); + + return dynamicconv_cuda_backward(gradOutput, padding_l, input, filters); +} + +PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) { + m.def("forward", &dynamicconv_forward, "dynamicconv forward (CUDA)"); + m.def("backward", &dynamicconv_backward, "dynamicconv backward (CUDA)"); +} diff --git a/data/fairseq/fairseq/modules/dynamicconv_layer/dynamicconv_cuda.cuh b/data/fairseq/fairseq/modules/dynamicconv_layer/dynamicconv_cuda.cuh new file mode 100644 index 0000000000000000000000000000000000000000..44baf21bdd2d4a7a692ae6f7953a413ea6513268 --- /dev/null +++ b/data/fairseq/fairseq/modules/dynamicconv_layer/dynamicconv_cuda.cuh @@ -0,0 +1,50 @@ +/** + * Copyright (c) Facebook, Inc. and its affiliates. + * + * This source code is licensed under the MIT license found in the + * LICENSE file in the root directory of this source tree. + */ + +#include +#include + +#include +#include +#include + +#include +#include +#include +#include +#include +#include + +#include +#include +#include + +#define SHFL_MASK 0xffffffff + +template +__global__ void dynamicconv_forward_kernel( + const scalar_t* input, + const scalar_t* weight, + int minibatch, + int sequenceLength, + int numFeatures, + int numFiltersInBlock, + int numHeads, + scalar_t* output); + +template +__global__ void dynamicconv_backward_kernel( + const scalar_t* gradOutput, // B * C * T + const scalar_t* input, // B * C * T + const scalar_t* weight, + int minibatch, + int sequenceLength, + int numFeatures, + int numFiltersInBlock, + int numHeads, + scalar_t* gradWeight, + scalar_t* gradInput); // B * H * k * T diff --git a/data/fairseq/fairseq/modules/dynamicconv_layer/dynamicconv_cuda_kernel.cu b/data/fairseq/fairseq/modules/dynamicconv_layer/dynamicconv_cuda_kernel.cu new file mode 100644 index 0000000000000000000000000000000000000000..4630f1e9826aea61973f2e82feb57ac0a4390735 --- /dev/null +++ b/data/fairseq/fairseq/modules/dynamicconv_layer/dynamicconv_cuda_kernel.cu @@ -0,0 +1,176 @@ +/** + * Copyright (c) Facebook, Inc. and its affiliates. + * + * This source code is licensed under the MIT license found in the + * LICENSE file in the root directory of this source tree. + */ + +#include "../cuda_utils.cu" +#include "dynamicconv_cuda.cuh" +#include "dynamicconv_cuda_backward.cu" +#include "dynamicconv_cuda_forward.cu" + +// FS is filter size and kernels are specialized for filter sizes +template +__global__ void dynamicconv_forward_kernel( + const scalar_t* input, + const scalar_t* weight, + int minibatch, + int sequenceLength, + int numFeatures, + int numFiltersInBlock, + int numHeads, + scalar_t* output) { + assert(blockDim.x == SB); + + const int tid = threadIdx.x; + const int batchIdx = blockIdx.x; + const int featureIdx = blockIdx.y; + const int head = featureIdx / numFiltersInBlock; + + const int IOOffset = + batchIdx * numFeatures * sequenceLength + featureIdx * sequenceLength; + const scalar_t* inputFeature = &input[IOOffset]; + scalar_t* outputFeature = &output[IOOffset]; + + scalar_t filter[FS]; + + __shared__ scalar_t tempInput[SB + FS]; + zeroSharedMem(tempInput); + + const int numIterations = divUp(sequenceLength, SB); + + for (int i = 0; i < numIterations; ++i) { + __syncthreads(); + const int inputOffset = i * SB; + load_input_to_shared( + inputFeature, + inputOffset, + sequenceLength, + i, + numIterations, + false, + tempInput); + __syncthreads(); + if (inputOffset + tid < sequenceLength) { +#pragma unroll + for (int k = 0; k < FS; ++k) { + const int filterOffset = batchIdx * numHeads * FS * sequenceLength + + head * FS * sequenceLength + k * sequenceLength + i * SB + tid; + filter[k] = weight[filterOffset]; + } + + scalar_t out = scalar_t(0.0); +#pragma unroll + for (int k = 0; k < FS; ++k) { + out += filter[k] * tempInput[tid + k]; + } + + outputFeature[inputOffset + tid] = out; + } + } +} + +template +__global__ void dynamicconv_backward_kernel( + const scalar_t* gradOutput, // B * C * T + const scalar_t* input, // B * C * T + const scalar_t* weight, + int minibatch, + int sequenceLength, + int numFeatures, + int numFiltersInBlock, + int numHeads, + scalar_t* gradWeight, + scalar_t* gradInput) { // B * H * k * T + + assert(blockDim.x == SB); + + // each block operates on a single batch and filter head + const int tid = threadIdx.x; + const int batchIdx = blockIdx.x; + const int headIdx = blockIdx.y; + const int chunkIdx = blockIdx.z; + + const int numChunks = divUp(sequenceLength, SB); + const int inputOffset = chunkIdx * SB; + + // initialize shared memory for output gradient and input + __shared__ scalar_t tempGradOutput[SB + FS]; + __shared__ scalar_t tempInput[SB + FS]; + const int padding = FS - padding_l - 1; + + zeroSharedMem(tempGradOutput); + zeroSharedMem(tempInput); + + // initialize local filter and weight gradient sum arrays + scalar_t tempGradSum[FS]; + scalar_t bfilter[FS]; + for (int k = 0; k < FS; ++k) { + tempGradSum[k] = scalar_t(0.0); + + int idxOffset = inputOffset + tid + k - padding; + if (idxOffset >= 0 && idxOffset < sequenceLength) { + int bfilterOffset = batchIdx * numHeads * FS * sequenceLength + + headIdx * FS * sequenceLength + (FS - k - 1) * sequenceLength + + idxOffset; + bfilter[k] = weight[bfilterOffset]; + } else { + bfilter[k] = scalar_t(0.0); + } + } + + // iterate over filter block + for (int featureIdx = 0; featureIdx < numFiltersInBlock; ++featureIdx) { + __syncthreads(); + + // load input and output gradient for this channel and chunk + const int IOOffset = batchIdx * numFeatures * sequenceLength + + (headIdx * numFiltersInBlock + featureIdx) * sequenceLength; + const scalar_t* inputFeature = &input[IOOffset]; + const scalar_t* gradOutputFeature = &gradOutput[IOOffset]; + scalar_t* gradInputFeature = &gradInput[IOOffset]; + + load_input_to_shared( + gradOutputFeature, + inputOffset, + sequenceLength, + chunkIdx, + numChunks, + true, + tempGradOutput); + load_input_to_shared( + inputFeature, + inputOffset, + sequenceLength, + chunkIdx, + numChunks, + true, + tempInput); + __syncthreads(); + + // sum input and weight gradients + scalar_t out = scalar_t(0.0); +#pragma unroll + for (int k = 0; k < FS; ++k) { + tempGradSum[k] += tempInput[tid + k] * tempGradOutput[tid + padding]; + out += bfilter[k] * tempGradOutput[tid + k]; + } + + if (inputOffset + tid < sequenceLength) { + gradInputFeature[inputOffset + tid] = out; + } + } + + const int gradOffset = + batchIdx * numHeads * FS * sequenceLength + headIdx * FS * sequenceLength; + scalar_t* gradWeightFeature = &gradWeight[gradOffset]; + + // write weight gradient + if (inputOffset + tid < sequenceLength) { + for (int k = 0; k < FS; ++k) { + const int outputOffset = k * sequenceLength + inputOffset + tid; + gradWeightFeature[outputOffset] = tempGradSum[k]; + } + } +} diff --git a/data/fairseq/fairseq/modules/dynamicconv_layer/dynamicconv_layer.py b/data/fairseq/fairseq/modules/dynamicconv_layer/dynamicconv_layer.py new file mode 100644 index 0000000000000000000000000000000000000000..711ed03483f4089dbe91964a89021b49eeffbedc --- /dev/null +++ b/data/fairseq/fairseq/modules/dynamicconv_layer/dynamicconv_layer.py @@ -0,0 +1,227 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import dynamicconv_cuda +import torch +import torch.nn.functional as F +from fairseq import utils +from fairseq.incremental_decoding_utils import with_incremental_state +from fairseq.modules.fairseq_dropout import FairseqDropout +from fairseq.modules.unfold import unfold1d +from torch import nn +from torch.autograd import Function + + +class dynamicconvFunction(Function): + @staticmethod + def forward(ctx, x, weights, padding_l): + ctx.padding_l = padding_l + outputs = dynamicconv_cuda.forward(x, weights, padding_l) + variables = [x, weights] + ctx.save_for_backward(*variables) + return outputs[0] + + @staticmethod + def backward(ctx, grad_output): + outputs = dynamicconv_cuda.backward( + grad_output.contiguous(), ctx.padding_l, *ctx.saved_tensors + ) + grad_input, grad_weights = outputs + return grad_input, grad_weights, None + + +@with_incremental_state +class DynamicconvLayer(nn.Module): + def __init__( + self, + input_size, + kernel_size=1, + padding_l=None, + weight_softmax=False, + num_heads=1, + weight_dropout=0.0, + bias=False, + renorm_padding=False, + conv_bias=False, + query_size=None, + ): + + super(DynamicconvLayer, self).__init__() + self.input_size = input_size + self.query_size = input_size if query_size is None else query_size + self.kernel_size = kernel_size + self.padding_l = padding_l + self.num_heads = num_heads + self.weight_softmax = weight_softmax + self.weight_dropout_module = FairseqDropout( + weight_dropout, module_name=self.__class__.__name__ + ) + self.renorm_padding = renorm_padding + self.bias = bias + + self.weight_linear = nn.Linear(input_size, num_heads * kernel_size, bias) + if conv_bias: + self.conv_bias = nn.Parameter(torch.Tensor(input_size)) + else: + self.conv_bias = None + self.reset_parameters() + + def reset_parameters(self): + nn.init.xavier_uniform_(self.weight_linear.weight) + if self.conv_bias is not None: + nn.init.constant_(self.conv_bias, 0.0) + nn.init.constant_(self.weight_linaer.bias, 0.0) + + def forward(self, x, incremental_state=None, query=None, unfold=None): + + T, B, C = x.size() + K, H = self.kernel_size, self.num_heads + # R = C // H + + # during inference time, incremental BMM is faster + if incremental_state is not None: + unfold = ( + x.size(0) > 512 if unfold is None else unfold + ) # use unfold mode as default for long sequence to save memory + unfold = unfold or (incremental_state is not None) + assert query is None + + if query is None: + query = x + if unfold: + output = self._forward_unfolded(x, incremental_state, query) + else: + output = self._forward_expanded(x, incremental_state, query) + + if self.conv_bias is not None: + output = output + self.conv_bias.view(1, 1, -1) + + return output + + # during training time, use CUDA kernel + else: + weight = self.weight_linear(x).view(T, B, H, K) + if self.weight_softmax: + weight = F.softmax(weight, dim=-1) + if self.weight_dropout_module.p: + weight = self.weight_dropout_module(weight) + + weight = weight.permute(1, 2, 3, 0).contiguous() + self.filters = weight + x = x.permute(1, 2, 0).contiguous() + output = dynamicconvFunction.apply(x, weight, self.padding_l).permute( + 2, 0, 1 + ) + if self.conv_bias is not None: + output = output + self.conv_bias.view(1, 1, -1) + return output + + def reorder_incremental_state(self, incremental_state, new_order): + input_buffer = self._get_input_buffer(incremental_state) + if input_buffer is not None: + input_buffer = input_buffer.index_select(1, new_order) + self._set_input_buffer(incremental_state, input_buffer) + + def _get_input_buffer(self, incremental_state): + return utils.get_incremental_state(self, incremental_state, "input_buffer") + + def _set_input_buffer(self, incremental_state, new_buffer): + return utils.set_incremental_state( + self, incremental_state, "input_buffer", new_buffer + ) + + def _forward_unfolded(self, x, incremental_state, query): + """The conventional implementation of convolutions. + Unfolding the input by having a window shifting to the right.""" + T, B, C = x.size() + K, H = self.kernel_size, self.num_heads + R = C // H + assert R * H == C == self.input_size + + weight = self.weight_linear(query).view(T * B * H, -1) + + # renorm_padding is only implemented in _forward_expanded + assert not self.renorm_padding or incremental_state is not None + + if incremental_state is not None: + input_buffer = self._get_input_buffer(incremental_state) + if input_buffer is None: + input_buffer = x.new() + x_unfold = torch.cat([input_buffer, x.unsqueeze(3)], dim=3) + if self.kernel_size > 1: + self._set_input_buffer( + incremental_state, x_unfold[:, :, :, -self.kernel_size + 1 :] + ) + x_unfold = x_unfold.view(T * B * H, R, -1) + else: + padding_l = self.padding_l + if K > T and padding_l == K - 1: + weight = weight.narrow(1, K - T, T) + K, padding_l = T, T - 1 + # unfold the input: T x B x C --> T' x B x C x K + x_unfold = unfold1d(x, K, padding_l, 0) + x_unfold = x_unfold.view(T * B * H, R, K) + + if self.weight_softmax and not self.renorm_padding: + weight = F.softmax(weight, dim=1) + weight = weight.narrow(1, 0, K) + + if incremental_state is not None: + weight = weight[:, -x_unfold.size(2) :] + K = weight.size(1) + + if self.weight_softmax and self.renorm_padding: + weight = F.softmax(weight, dim=1) + + weight = self.weight_dropout_module(weight, inplace=False) + + output = torch.bmm(x_unfold, weight.unsqueeze(2)) # T*B*H x R x 1 + output = output.view(T, B, C) + return output + + def _forward_expanded(self, x, incremental_stat, query): + """Turn the convolution filters into band matrices and do matrix multiplication. + This is faster when the sequence is short, but less memory efficient. + This is not used in the decoder during inference. + """ + T, B, C = x.size() + K, H = self.kernel_size, self.num_heads + R = C // H + assert R * H == C == self.input_size + weight = self.weight_linear(query).view(T * B * H, -1) + + if not self.renorm_padding: + if self.weight_softmax: + weight = F.softmax(weight, dim=1) + weight = self.weight_dropout_module(weight, inplace=False) + weight = weight.narrow(1, 0, K).contiguous() + weight = weight.view(T, B * H, K).transpose(0, 1) + + x = x.view(T, B * H, R).transpose(0, 1) + if self.weight_softmax and self.renorm_padding: + # turn the convolution filters into band matrices + weight_expanded = weight.new(B * H, T, T + K - 1).fill_(float("-inf")) + weight_expanded.as_strided( + (B * H, T, K), (T * (T + K - 1), T + K, 1) + ).copy_(weight) + weight_expanded = weight_expanded.narrow(2, self.padding_l, T) + # normalize the weight over valid positions like self-attention + weight_expanded = F.softmax(weight_expanded, dim=2) + weight_expanded = self.weight_dropout_module(weight_expanded, inplace=False) + else: + P = self.padding_l + # For efficiency, we cut the kernel size and reduce the padding when the kernel is larger than the length + if K > T and P == K - 1: + weight = weight.narrow(2, K - T, T) + K, P = T, T - 1 + # turn the convolution filters into band matrices + weight_expanded = weight.new_zeros(B * H, T, T + K - 1, requires_grad=False) + weight_expanded.as_strided( + (B * H, T, K), (T * (T + K - 1), T + K, 1) + ).copy_(weight) + weight_expanded = weight_expanded.narrow(2, P, T) # B*H x T x T + output = torch.bmm(weight_expanded, x) + output = output.transpose(0, 1).contiguous().view(T, B, C) + return output diff --git a/data/fairseq/fairseq/modules/dynamicconv_layer/dynamiconv_cpu.cpp b/data/fairseq/fairseq/modules/dynamicconv_layer/dynamiconv_cpu.cpp new file mode 100644 index 0000000000000000000000000000000000000000..d7e57c859085f98ec10960330ca763ae2764585a --- /dev/null +++ b/data/fairseq/fairseq/modules/dynamicconv_layer/dynamiconv_cpu.cpp @@ -0,0 +1,29 @@ +#include +#include + +std::vector +dynamicconv_cpu_forward(float* input, float* filters, int padding_l); + +std::vector dynamicconv_cpu_backward( + float* gradOutput, + int padding_l, + float* input, + float* filters); + +std::vector +dynamicconv_forward(float* input, float* filters, int padding_l) { + return dynamicconv_cpu_forward(input, filters, padding_l); +} + +std::vector dynamicconv_backward( + float* gradOutput, + int padding_l, + float* input, + float* filters) { + return dynamicconv_cpu_backward(gradOutput, padding_l, input, filters); +} + +PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) { + m.def("forward", &dynamicconv_forward, "dynamicconv forward (CPU)"); + m.def("backward", &dynamicconv_backward, "dynamicconv backward (CPU)"); +} diff --git a/data/fairseq/fairseq/modules/dynamicconv_layer/setup.py b/data/fairseq/fairseq/modules/dynamicconv_layer/setup.py new file mode 100644 index 0000000000000000000000000000000000000000..6a21f7e2ee0840a3b251522275a0b32a856951d7 --- /dev/null +++ b/data/fairseq/fairseq/modules/dynamicconv_layer/setup.py @@ -0,0 +1,23 @@ +#!/usr/bin/env python3 +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from setuptools import setup +from torch.utils.cpp_extension import BuildExtension, CUDAExtension + + +setup( + name="dynamicconv_layer", + ext_modules=[ + CUDAExtension( + name="dynamicconv_cuda", + sources=[ + "dynamicconv_cuda.cpp", + "dynamicconv_cuda_kernel.cu", + ], + ), + ], + cmdclass={"build_ext": BuildExtension}, +) diff --git a/data/fairseq/fairseq/modules/ema_module.py b/data/fairseq/fairseq/modules/ema_module.py new file mode 100644 index 0000000000000000000000000000000000000000..f0ece842d435597cba5ed287de1961e0929639e0 --- /dev/null +++ b/data/fairseq/fairseq/modules/ema_module.py @@ -0,0 +1,215 @@ +#!/usr/bin/env python3 + +""" +Used for EMA tracking a given pytorch module. The user is responsible for calling step() +and setting the appropriate decay +""" + +import copy +from dataclasses import dataclass, field +import logging + +import torch + +from omegaconf import II +from fairseq.dataclass import FairseqDataclass + +try: + from amp_C import multi_tensor_l2norm + + multi_tensor_l2norm_available = True +except ImportError: + multi_tensor_l2norm_available = False + +logger = logging.getLogger(__name__) + + +@dataclass +class EMAModuleConfig(FairseqDataclass): + ema_decay: float = field( + default=0.9999, metadata={"help": "decay for exponential moving average model"} + ) + ema_fp32: bool = field( + default=False, + metadata={"help": "If true, store EMA model in fp32 even if model is in fp16"}, + ) + add_missing_params: bool = True + log_norms: bool = False + + +class EMAModule: + """Exponential Moving Average of Fairseq Models""" + + def __init__( + self, + model, + config: EMAModuleConfig, + copy_model=True, + device=None, + skip_keys=None, + ): + """ + @param model model to initialize the EMA with + @param config EMAConfig object with configuration like + ema_decay, ema_update_freq, ema_fp32 + @param device If provided, copy EMA to this device (e.g. gpu). + Otherwise EMA is in the same device as the model. + """ + + self.config = config + + if copy_model: + self.model = copy.deepcopy(model) + self.model.requires_grad_(False) + else: + self.model = model + + self.config = config + self.decay = config.ema_decay + self.skip_keys = skip_keys or set() + self.add_missing_params = config.add_missing_params + self.fp32_params = {} + + if device is not None: + logging.info(f"Copying EMA model to device {device}") + self.model = self.model.to(device=device) + + if self.config.ema_fp32: + self.build_fp32_params() + + self.log_norms = config.log_norms and multi_tensor_l2norm_available + self.logs = {} + + def build_fp32_params(self, state_dict=None): + """ + Store a copy of the EMA params in fp32. + If state dict is passed, the EMA params is copied from + the provided state dict. Otherwise, it is copied from the + current EMA model parameters. + """ + if not self.config.ema_fp32: + raise RuntimeError( + "build_fp32_params should not be called if ema_fp32=False. " + "Use ema_fp32=True if this is really intended." + ) + + if state_dict is None: + state_dict = self.model.state_dict() + + def _to_float(t): + return t.float() if torch.is_floating_point(t) else t + + for param_key in state_dict: + if param_key in self.fp32_params: + if param_key == "__sq_mom": + self.fp32_params[param_key] = state_dict[param_key] + else: + self.fp32_params[param_key].copy_(state_dict[param_key]) + else: + self.fp32_params[param_key] = _to_float(state_dict[param_key]) + if "__sq_mom" in self.fp32_params: + self.fp32_params["__sq_mom"][param_key] = torch.zeros_like( + self.fp32_params[param_key] + ) + + def restore(self, state_dict, build_fp32_params=False): + """Load data from a model spec into EMA model""" + self.model.load_state_dict(state_dict, strict=False) + if build_fp32_params: + self.build_fp32_params(state_dict) + + def set_decay(self, decay, weight_decay=None): + self.decay = decay + if weight_decay is not None: + self.weight_decay = weight_decay + + def get_decay(self): + return self.decay + + def _step_internal(self, new_model): + """One update of the EMA model based on new model weights""" + decay = self.decay + + ema_state_dict = {} + ema_params = ( + self.fp32_params if self.config.ema_fp32 else self.model.state_dict() + ) + + new_p = [] + ema_p = [] + + for key, param in new_model.named_parameters(): + if isinstance(param, dict): + continue + + if not self.add_missing_params and key not in ema_params: + continue + + try: + ema_param = ema_params[key] + except KeyError: + ema_param = ( + param.float().clone() if param.ndim == 1 else copy.deepcopy(param) + ) + ema_params[key] = ema_param + + if param.shape != ema_param.shape: + raise ValueError( + "incompatible tensor shapes between model param and ema param" + + "{} vs. {}".format(param.shape, ema_param.shape) + ) + + if "version" in key: + # Do not decay a model.version pytorch param + continue + + lr = 1 - decay + + if key in self.skip_keys or not param.requires_grad: + ema_params[key].copy_(param.to(dtype=ema_param.dtype).data) + ema_param = ema_params[key] + else: + if self.log_norms: + new_p.append(param) + ema_p.append(ema_param) + + ema_param.mul_(1 - lr) + ema_param.add_(param.data.to(dtype=ema_param.dtype), alpha=lr) + + ema_state_dict[key] = ema_param + + for key, param in new_model.named_buffers(): + ema_state_dict[key] = param + + if self.log_norms: + if "model_norm" in self.logs: + self.prev_model_norm = self.logs["model_norm"] + + chunk_size = 2048 * 32 + has_inf = torch.zeros( + (1, 1), dtype=torch.int, device=next(new_model.parameters()).device + ) + + new_norm = multi_tensor_l2norm(chunk_size, has_inf, [new_p], False) + old_norm = multi_tensor_l2norm(chunk_size, has_inf, [ema_p], False) + + self.logs["model_norm"] = new_norm[0] + self.logs["ema_norm"] = old_norm[0] + + self.restore(ema_state_dict, build_fp32_params=False) + + @torch.no_grad() + def step(self, new_model): + self._step_internal(new_model) + + def reverse(self, model): + """ + Load the model parameters from EMA model. + Useful for inference or fine-tuning from the EMA model. + """ + d = self.model.state_dict() + if "_ema" in d: + del d["_ema"] + + model.load_state_dict(d, strict=False) + return model diff --git a/data/fairseq/fairseq/modules/espnet_multihead_attention.py b/data/fairseq/fairseq/modules/espnet_multihead_attention.py new file mode 100644 index 0000000000000000000000000000000000000000..82bc0d7b452b47b3135548f89c098272c8699c0a --- /dev/null +++ b/data/fairseq/fairseq/modules/espnet_multihead_attention.py @@ -0,0 +1,256 @@ +#!/usr/bin/env python3 +# -*- coding: utf-8 -*- + +# Copyright 2019 Shigeki Karita +# Apache 2.0 (http://www.apache.org/licenses/LICENSE-2.0) + +"""Multi-Head Attention layer definition.""" + +import math + +import torch +from torch import nn + +from fairseq.modules.rotary_positional_embedding import ( + RotaryPositionalEmbedding, + apply_rotary_pos_emb, +) + + +class ESPNETMultiHeadedAttention(nn.Module): + """Multi-Head Attention layer. + Args: + n_head: The number of heads. + n_feat: The number of features. + dropout: Dropout rate. + """ + + def __init__(self, n_feat, n_head, dropout): + """Construct an MultiHeadedAttention object.""" + super(ESPNETMultiHeadedAttention, self).__init__() + assert n_feat % n_head == 0 + # We assume d_v always equals d_k + self.d_k = n_feat // n_head + self.h = n_head + self.linear_q = nn.Linear(n_feat, n_feat) + self.linear_k = nn.Linear(n_feat, n_feat) + self.linear_v = nn.Linear(n_feat, n_feat) + self.linear_out = nn.Linear(n_feat, n_feat) + self.attn = None + self.dropout = nn.Dropout(p=dropout) + + def forward_qkv(self, query, key, value, **kwargs): + """Transform query, key and value. + Args: + query: Query tensor B X T1 X C + key: Key tensor B X T2 X C + value: Value tensor B X T2 X C + Returns: + torch.Tensor: Transformed query tensor B X n_head X T1 X d_k + torch.Tensor: Transformed key tensor B X n_head X T2 X d_k + torch.Tensor: Transformed value tensor B X n_head X T2 X d_k + """ + n_batch = query.size(0) + q = self.linear_q(query).view(n_batch, -1, self.h, self.d_k) + k = self.linear_k(key).view(n_batch, -1, self.h, self.d_k) + v = self.linear_v(value).view(n_batch, -1, self.h, self.d_k) + q = q.transpose(1, 2) # (batch, head, time1, d_k) + k = k.transpose(1, 2) # (batch, head, time2, d_k) + v = v.transpose(1, 2) # (batch, head, time2, d_k) + return q, k, v + + def forward_attention(self, value, scores, mask): + """Compute attention context vector. + Args: + value: Transformed value B X n_head X T2 X d_k. + scores: Attention score B X n_head X T1 X T2 + mask: Mask T2 X B + Returns: + torch.Tensor: Transformed value B X T1 X d_model + weighted by the attention score B X T1 X T2 + """ + n_batch = value.size(0) + if mask is not None: + scores = scores.masked_fill( + mask.unsqueeze(1).unsqueeze(2).to(bool), + float("-inf"), # (batch, head, time1, time2) + ) + self.attn = torch.softmax(scores, dim=-1) # (batch, head, time1, time2) + + else: + self.attn = torch.softmax(scores, dim=-1) # (batch, head, time1, time2) + p_attn = self.dropout(self.attn) + x = torch.matmul(p_attn, value) # (batch, head, time1, d_k) + x = ( + x.transpose(1, 2).contiguous().view(n_batch, -1, self.h * self.d_k) + ) # (batch, time1, d_model) + + return self.linear_out(x) # (batch, time1, d_model) + + def forward(self, query, key, value, key_padding_mask=None, **kwargs): + """Compute scaled dot product attention. + Args: + query (torch.Tensor): Query tensor T X B X C + key (torch.Tensor): Key tensor T X B X C + value (torch.Tensor): Value tensor T X B X C + mask (torch.Tensor): Mask tensor T X B + Returns: + torch.Tensor: Output tensor T X B X D. + """ + query = query.transpose(0, 1) + key = key.transpose(0, 1) + value = value.transpose(0, 1) + + q, k, v = self.forward_qkv(query, key, value) + scores = torch.matmul(q, k.transpose(-2, -1)) / math.sqrt(self.d_k) + scores = self.forward_attention(v, scores, key_padding_mask) + scores = scores.transpose(0, 1) + return scores, None + + +class RelPositionMultiHeadedAttention(ESPNETMultiHeadedAttention): + """Multi-Head Attention layer with relative position encoding. + Paper: https://arxiv.org/abs/1901.02860 + Args: + n_head: The number of heads. + n_feat: The number of features. + dropout: Dropout rate. + zero_triu: Whether to zero the upper triangular part of attention matrix. + """ + + def __init__(self, n_feat, n_head, dropout, zero_triu=False): + """Construct an RelPositionMultiHeadedAttention object.""" + super().__init__(n_feat, n_head, dropout) + self.zero_triu = zero_triu + # linear transformation for positional encoding + self.linear_pos = nn.Linear(n_feat, n_feat, bias=False) + # these two learnable bias are used in matrix c and matrix d + # as described in https://arxiv.org/abs/1901.02860 Section 3.3 + self.pos_bias_u = nn.Parameter(torch.zeros(self.h, self.d_k)) + self.pos_bias_v = nn.Parameter(torch.zeros(self.h, self.d_k)) + torch.nn.init.xavier_uniform_(self.pos_bias_u) + torch.nn.init.xavier_uniform_(self.pos_bias_v) + + def rel_shift(self, x): + """Compute relative positional encoding. + Args: + x: Input tensor B X n_head X T X 2T-1 + Returns: + torch.Tensor: Output tensor. + """ + zero_pad = torch.zeros((*x.size()[:3], 1), device=x.device, dtype=x.dtype) + x_padded = torch.cat([zero_pad, x], dim=-1) + + x_padded = x_padded.view(*x.size()[:2], x.size(3) + 1, x.size(2)) + x = x_padded[:, :, 1:].view_as(x)[ + :, :, :, : x.size(-1) // 2 + 1 + ] # only keep the positions from 0 to time2 + + if self.zero_triu: + ones = torch.ones((x.size(2), x.size(3)), device=x.device) + x = x * torch.tril(ones, x.size(3) - x.size(2))[None, None, :, :] + + return x + + def forward(self, query, key, value, pos_emb, key_padding_mask=None, **kwargs): + """Compute scaled dot product attention. + Args: + query: Query tensor T X B X C + key: Key tensor T X B X C + value: Value tensor T X B X C + pos_emb: Positional embedding tensor B X 2T-1 X C + key_padding_mask: Mask tensor T X B + Returns: + torch.Tensor: Output tensor T X B X C. + """ + query = query.transpose(0, 1) + key = key.transpose(0, 1) + value = value.transpose(0, 1) + pos_emb = pos_emb.transpose(0, 1) + q, k, v = self.forward_qkv(query, key, value) + q = q.transpose(1, 2) # (batch, time1, head, d_k) + n_batch_pos = pos_emb.size(0) + p = self.linear_pos(pos_emb).view(n_batch_pos, -1, self.h, self.d_k) + p = p.transpose(1, 2) # (batch, head, 2*time1-1, d_k) + + # (batch, head, time1, d_k) + q_with_bias_u = (q + self.pos_bias_u).transpose(1, 2) + # (batch, head, time1, d_k) + q_with_bias_v = (q + self.pos_bias_v).transpose(1, 2) + + # compute attention score + # first compute matrix a and matrix c + # as described in https://arxiv.org/abs/1901.02860 Section 3.3 + # (batch, head, time1, time2) + matrix_ac = torch.matmul(q_with_bias_u, k.transpose(-2, -1)) + + # compute matrix b and matrix d + # (batch, head, time1, 2*time1-1) + matrix_bd = torch.matmul(q_with_bias_v, p.transpose(-2, -1)) + matrix_bd = self.rel_shift(matrix_bd) + + scores = (matrix_ac + matrix_bd) / math.sqrt( + self.d_k + ) # (batch, head, time1, time2) + + scores = self.forward_attention(v, scores, key_padding_mask) + scores = scores.transpose(0, 1) + return scores, None + + +class RotaryPositionMultiHeadedAttention(ESPNETMultiHeadedAttention): + def __init__( + self, + n_feat, + n_head, + dropout, + precision, + rotary_emd_base=10000, + ): + """Construct an RotaryPositionMultiHeadedAttention object.""" + super().__init__(n_feat, n_head, dropout) + precision = torch.float + self.rotary_ndims = self.d_k # also try self.d_k//2 + if precision == "fp16": + precision = torch.half + + self.rotary_emb = RotaryPositionalEmbedding( + self.rotary_ndims, base=rotary_emd_base, precision=precision + ) + + def forward(self, query, key, value, key_padding_mask=None, **kwargs): + """Compute rotary position attention. + Args: + query: Query tensor T X B X C + key: Key tensor T X B X C + value: Value tensor T X B X C + key_padding_mask: Mask tensor T X B + Returns: + torch.Tensor: Output tensor T X B X D. + Notes: + Assumes self attn + """ + + T, B, C = value.size() + query = query.view(T, B, self.h, self.d_k) + key = key.view(T, B, self.h, self.d_k) + value = value.view(T, B, self.h, self.d_k) + cos, sin = self.rotary_emb(value, seq_len=T) + query, key = apply_rotary_pos_emb( + query, key, cos, sin, offset=0 + ) # offset is based on layer_past + + query = query.view(T, B, self.h * self.d_k) + key = key.view(T, B, self.h * self.d_k) + value = value.view(T, B, self.h * self.d_k) + + # TBD to BTD + query = query.transpose(0, 1) + key = key.transpose(0, 1) + value = value.transpose(0, 1) + + q, k, v = self.forward_qkv(query, key, value) + scores = torch.matmul(q, k.transpose(-2, -1)) / math.sqrt(self.d_k) + scores = self.forward_attention(v, scores, key_padding_mask) + scores = scores.transpose(0, 1) + return scores, None diff --git a/data/fairseq/fairseq/modules/fairseq_dropout.py b/data/fairseq/fairseq/modules/fairseq_dropout.py new file mode 100644 index 0000000000000000000000000000000000000000..3cddca77186f5ddd5cfb9c0ed6def9bafdf3bf1e --- /dev/null +++ b/data/fairseq/fairseq/modules/fairseq_dropout.py @@ -0,0 +1,51 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import logging +from typing import List, Optional + +import torch.nn as nn +import torch.nn.functional as F + + +logger = logging.getLogger(__name__) + + +class FairseqDropout(nn.Module): + def __init__(self, p, module_name=None): + super().__init__() + self.p = p + self.module_name = module_name + self.apply_during_inference = False + + def forward(self, x, inplace: bool = False): + if self.p > 0 and (self.training or self.apply_during_inference): + return F.dropout(x, p=self.p, training=True, inplace=inplace) + else: + return x + + def make_generation_fast_( + self, + name: str, + retain_dropout: bool = False, + retain_dropout_modules: Optional[List[str]] = None, + **kwargs + ): + if retain_dropout: + if retain_dropout_modules is not None and self.module_name is None: + logger.warning( + "Cannot enable dropout during inference for module {} " + "because module_name was not set".format(name) + ) + elif ( + retain_dropout_modules is None # if None, apply to all modules + or self.module_name in retain_dropout_modules + ): + logger.info( + "Enabling dropout during inference for module: {}".format(name) + ) + self.apply_during_inference = True + else: + logger.info("Disabling dropout for module: {}".format(name)) diff --git a/data/fairseq/fairseq/modules/fp32_batch_norm.py b/data/fairseq/fairseq/modules/fp32_batch_norm.py new file mode 100644 index 0000000000000000000000000000000000000000..c560f338fdb79870c49aa70ec3f65fc9d6184989 --- /dev/null +++ b/data/fairseq/fairseq/modules/fp32_batch_norm.py @@ -0,0 +1,44 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. +""" +batch norm done in fp32 (for fp16 training) +""" +import torch +import torch.nn as nn + + +class Fp32BatchNorm(nn.Module): + def __init__(self, sync=False, *args, **kwargs): + super().__init__() + + if sync: + from fairseq.distributed import utils + + if utils.get_global_world_size() == 1: + sync = False + + if sync: + self.bn = nn.SyncBatchNorm(*args, **kwargs) + else: + self.bn = nn.BatchNorm1d(*args, **kwargs) + + self.sync = sync + + def forward(self, input): + if self.bn.running_mean.dtype != torch.float: + if self.sync: + self.bn.running_mean = self.bn.running_mean.float() + self.bn.running_var = self.bn.running_var.float() + if self.bn.affine: + try: + self.bn.weight = self.bn.weight.float() + self.bn.bias = self.bn.bias.float() + except: + self.bn.float() + else: + self.bn.float() + + output = self.bn(input.float()) + return output.type_as(input) diff --git a/data/fairseq/fairseq/modules/fp32_group_norm.py b/data/fairseq/fairseq/modules/fp32_group_norm.py new file mode 100644 index 0000000000000000000000000000000000000000..d03aac022e30c8c14a600062d1d86429504ba003 --- /dev/null +++ b/data/fairseq/fairseq/modules/fp32_group_norm.py @@ -0,0 +1,25 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. +""" +Layer norm done in fp32 (for fp16 training) +""" + +import torch.nn as nn +import torch.nn.functional as F + + +class Fp32GroupNorm(nn.GroupNorm): + def __init__(self, *args, **kwargs): + super().__init__(*args, **kwargs) + + def forward(self, input): + output = F.group_norm( + input.float(), + self.num_groups, + self.weight.float() if self.weight is not None else None, + self.bias.float() if self.bias is not None else None, + self.eps, + ) + return output.type_as(input) diff --git a/data/fairseq/fairseq/modules/fp32_instance_norm.py b/data/fairseq/fairseq/modules/fp32_instance_norm.py new file mode 100644 index 0000000000000000000000000000000000000000..30a54496ded633ee005cf11cbdbd2cc2ebb29ca0 --- /dev/null +++ b/data/fairseq/fairseq/modules/fp32_instance_norm.py @@ -0,0 +1,35 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. +""" +Layer norm done in fp32 (for fp16 training) +""" + +import torch.nn as nn +import torch.nn.functional as F + + +class Fp32InstanceNorm(nn.InstanceNorm1d): + def __init__(self, *args, **kwargs): + self.transpose_last = "transpose_last" in kwargs and kwargs["transpose_last"] + if "transpose_last" in kwargs: + del kwargs["transpose_last"] + super().__init__(*args, **kwargs) + + def forward(self, input): + if self.transpose_last: + input = input.transpose(1, 2) + output = F.instance_norm( + input.float(), + running_mean=self.running_mean, + running_var=self.running_var, + weight=self.weight.float() if self.weight is not None else None, + bias=self.bias.float() if self.bias is not None else None, + use_input_stats=self.training or not self.track_running_stats, + momentum=self.momentum, + eps=self.eps, + ) + if self.transpose_last: + output = output.transpose(1, 2) + return output.type_as(input) diff --git a/data/fairseq/fairseq/modules/gelu.py b/data/fairseq/fairseq/modules/gelu.py new file mode 100644 index 0000000000000000000000000000000000000000..a2f1ecff4a3ae3de3eb7d327b9163c46b18a15ed --- /dev/null +++ b/data/fairseq/fairseq/modules/gelu.py @@ -0,0 +1,25 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. +""" +See "Gaussian Error Linear Units (GELUs)" by Dan Hendrycks and Kevin Gimpel with +the corresponding GitHub repo: https://github.com/hendrycks/GELUs +""" + +import math + +import torch +import torch.nn as nn + + +def gelu_accurate(x): + if not hasattr(gelu_accurate, "_a"): + gelu_accurate._a = math.sqrt(2 / math.pi) + return ( + 0.5 * x * (1 + torch.tanh(gelu_accurate._a * (x + 0.044715 * torch.pow(x, 3)))) + ) + + +def gelu(x: torch.Tensor) -> torch.Tensor: + return torch.nn.functional.gelu(x.float()).type_as(x) diff --git a/data/fairseq/fairseq/modules/grad_multiply.py b/data/fairseq/fairseq/modules/grad_multiply.py new file mode 100644 index 0000000000000000000000000000000000000000..08d15f55dfda9c61a1cf8641ea31424fe1d97f57 --- /dev/null +++ b/data/fairseq/fairseq/modules/grad_multiply.py @@ -0,0 +1,18 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import torch + + +class GradMultiply(torch.autograd.Function): + @staticmethod + def forward(ctx, x, scale): + ctx.scale = scale + res = x.new(x) + return res + + @staticmethod + def backward(ctx, grad): + return grad * ctx.scale, None diff --git a/data/fairseq/fairseq/modules/gumbel_vector_quantizer.py b/data/fairseq/fairseq/modules/gumbel_vector_quantizer.py new file mode 100644 index 0000000000000000000000000000000000000000..867b019f676d72a51db8f8ea54e08fab2b535bfc --- /dev/null +++ b/data/fairseq/fairseq/modules/gumbel_vector_quantizer.py @@ -0,0 +1,212 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import torch +import torch.nn as nn +import torch.nn.functional as F + + +class GumbelVectorQuantizer(nn.Module): + def __init__( + self, + dim, + num_vars, + temp, + groups, + combine_groups, + vq_dim, + time_first, + activation=nn.GELU(), + weight_proj_depth=1, + weight_proj_factor=1, + hard=True, + std=0, + ): + """Vector quantization using gumbel softmax + + Args: + dim: input dimension (channels) + num_vars: number of quantized vectors per group + temp: temperature for training. this should be a tuple of 3 elements: (start, stop, decay factor) + groups: number of groups for vector quantization + combine_groups: whether to use the vectors for all groups + vq_dim: dimensionality of the resulting quantized vector + time_first: if true, expect input in BxTxC format, otherwise in BxCxT + activation: what activation to use (should be a module). this is only used if weight_proj_depth is > 1 + weight_proj_depth: number of layers (with activation in between) to project input before computing logits + weight_proj_factor: this is used only if weight_proj_depth is > 1. scales the inner dimensionality of + projections by this factor + """ + super().__init__() + + self.groups = groups + self.combine_groups = combine_groups + self.input_dim = dim + self.num_vars = num_vars + self.time_first = time_first + self.hard = hard + + assert ( + vq_dim % groups == 0 + ), f"dim {vq_dim} must be divisible by groups {groups} for concatenation" + + var_dim = vq_dim // groups + num_groups = groups if not combine_groups else 1 + + self.vars = nn.Parameter(torch.FloatTensor(1, num_groups * num_vars, var_dim)) + if std == 0: + nn.init.uniform_(self.vars) + else: + nn.init.normal_(self.vars, mean=0, std=std) + + if weight_proj_depth > 1: + + def block(input_dim, output_dim): + return nn.Sequential(nn.Linear(input_dim, output_dim), activation) + + inner_dim = self.input_dim * weight_proj_factor + self.weight_proj = nn.Sequential( + *[ + block(self.input_dim if i == 0 else inner_dim, inner_dim) + for i in range(weight_proj_depth - 1) + ], + nn.Linear(inner_dim, groups * num_vars), + ) + else: + self.weight_proj = nn.Linear(self.input_dim, groups * num_vars) + nn.init.normal_(self.weight_proj.weight, mean=0, std=1) + nn.init.zeros_(self.weight_proj.bias) + + if isinstance(temp, str): + import ast + + temp = ast.literal_eval(temp) + assert len(temp) == 3, f"{temp}, {len(temp)}" + + self.max_temp, self.min_temp, self.temp_decay = temp + self.curr_temp = self.max_temp + self.codebook_indices = None + + def set_num_updates(self, num_updates): + self.curr_temp = max( + self.max_temp * self.temp_decay**num_updates, self.min_temp + ) + + def get_codebook_indices(self): + if self.codebook_indices is None: + from itertools import product + + p = [range(self.num_vars)] * self.groups + inds = list(product(*p)) + self.codebook_indices = torch.tensor( + inds, dtype=torch.long, device=self.vars.device + ).flatten() + + if not self.combine_groups: + self.codebook_indices = self.codebook_indices.view( + self.num_vars**self.groups, -1 + ) + for b in range(1, self.groups): + self.codebook_indices[:, b] += self.num_vars * b + self.codebook_indices = self.codebook_indices.flatten() + return self.codebook_indices + + def codebook(self): + indices = self.get_codebook_indices() + return ( + self.vars.squeeze(0) + .index_select(0, indices) + .view(self.num_vars**self.groups, -1) + ) + + def sample_from_codebook(self, b, n): + indices = self.get_codebook_indices() + indices = indices.view(-1, self.groups) + cb_size = indices.size(0) + assert ( + n < cb_size + ), f"sample size {n} is greater than size of codebook {cb_size}" + sample_idx = torch.randint(low=0, high=cb_size, size=(b * n,)) + indices = indices[sample_idx] + + z = self.vars.squeeze(0).index_select(0, indices.flatten()).view(b, n, -1) + return z + + def to_codebook_index(self, indices): + res = indices.new_full(indices.shape[:-1], 0) + for i in range(self.groups): + exponent = self.groups - i - 1 + res += indices[..., i] * (self.num_vars**exponent) + return res + + def forward_idx(self, x): + res = self.forward(x, produce_targets=True) + return res["x"], res["targets"] + + def forward(self, x, produce_targets=False): + + result = {"num_vars": self.num_vars * self.groups} + + if not self.time_first: + x = x.transpose(1, 2) + + bsz, tsz, fsz = x.shape + x = x.reshape(-1, fsz) + x = self.weight_proj(x) + x = x.view(bsz * tsz * self.groups, -1) + + with torch.no_grad(): + _, k = x.max(-1) + hard_x = ( + x.new_zeros(*x.shape) + .scatter_(-1, k.view(-1, 1), 1.0) + .view(bsz * tsz, self.groups, -1) + ) + hard_probs = torch.mean(hard_x.float(), dim=0) + result["code_perplexity"] = torch.exp( + -torch.sum(hard_probs * torch.log(hard_probs + 1e-7), dim=-1) + ).sum() + + avg_probs = torch.softmax( + x.view(bsz * tsz, self.groups, -1).float(), dim=-1 + ).mean(dim=0) + result["prob_perplexity"] = torch.exp( + -torch.sum(avg_probs * torch.log(avg_probs + 1e-7), dim=-1) + ).sum() + + result["temp"] = self.curr_temp + + if self.training: + x = F.gumbel_softmax(x.float(), tau=self.curr_temp, hard=self.hard).type_as( + x + ) + else: + x = hard_x + + x = x.view(bsz * tsz, -1) + + vars = self.vars + if self.combine_groups: + vars = vars.repeat(1, self.groups, 1) + + if produce_targets: + result["targets"] = ( + x.view(bsz * tsz * self.groups, -1) + .argmax(dim=-1) + .view(bsz, tsz, self.groups) + .detach() + ) + + x = x.unsqueeze(-1) * vars + x = x.view(bsz * tsz, self.groups, self.num_vars, -1) + x = x.sum(-2) + x = x.view(bsz, tsz, -1) + + if not self.time_first: + x = x.transpose(1, 2) # BTC -> BCT + + result["x"] = x + + return result diff --git a/data/fairseq/fairseq/modules/kmeans_attention.py b/data/fairseq/fairseq/modules/kmeans_attention.py new file mode 100644 index 0000000000000000000000000000000000000000..0088d1ebdca0331cc51fa785bab9eefce939b782 --- /dev/null +++ b/data/fairseq/fairseq/modules/kmeans_attention.py @@ -0,0 +1,744 @@ +import math +from functools import reduce, wraps +from inspect import isfunction +from operator import mul + +import torch +import torch.nn as nn +import torch.nn.functional as F +from aml.multimodal_video.utils.einops.lib import rearrange, repeat +from aml.multimodal_video.utils.einops.lib.layers.torch import Rearrange + +from fairseq.modules.local_attention import LocalAttention + +# constants + +TOKEN_SELF_ATTN_VALUE = -5e4 +KMEAN_INIT_ITERS = 10 + +# helper functions + + +def exists(val): + return val is not None + + +def identity(x, *args, **kwargs): + return x + + +def default(x, d): + if not exists(x): + return d if not isfunction(d) else d() + return x + + +def cast_tuple(x): + return x if isinstance(x, tuple) else (x,) + + +def cache_fn(f): + cache = None + + @wraps(f) + def cached_fn(*args, **kwargs): + nonlocal cache + if exists(cache): + return cache + cache = f(*args, **kwargs) + return cache + + return cached_fn + + +def to(t): + return {"device": t.device, "dtype": t.dtype} + + +def find_modules(nn_module, type): + return [module for module in nn_module.modules() if isinstance(module, type)] + + +def is_empty(t): + return t.nelement() == 0 + + +def max_neg_value(tensor): + return -torch.finfo(tensor.dtype).max + + +def batched_index_select(values, indices): + last_dim = values.shape[-1] + return values.gather(2, expand_dim(indices, -1, last_dim)) + + +def merge_dims(ind_from, ind_to, tensor): + shape = list(tensor.shape) + arr_slice = slice(ind_from, ind_to + 1) + shape[arr_slice] = [reduce(mul, shape[arr_slice])] + return tensor.reshape(*shape) + + +def expand_dim(t, dim, k): + t = t.unsqueeze(dim) + expand_shape = [-1] * len(t.shape) + expand_shape[dim] = k + return t.expand(*expand_shape) + + +def scatter_mean(src, t, index, dim, eps=1e-5): + numer = src.scatter_add(dim, index, t) + denom = src.scatter_add(dim, index, torch.ones_like(t)) + return numer / (denom + eps) + + +def split_at_index(dim, index, t): + pre_slices = (slice(None),) * dim + l = (*pre_slices, slice(None, index)) + r = (*pre_slices, slice(index, None)) + return t[l], t[r] + + +def reshape_dim(t, dim, split_dims): + shape = list(t.shape) + num_dims = len(shape) + dim = (dim + num_dims) % num_dims + shape[dim : dim + 1] = split_dims + return t.reshape(shape) + + +def ema(old, new, decay): + if not exists(old): + return new + return old * decay + new * (1 - decay) + + +def ema_inplace(moving_avg, new, decay): + if is_empty(moving_avg): + moving_avg.data.copy_(new) + return + moving_avg.data.mul_(decay).add_(new, alpha=(1 - decay)) + + +# helper classes + + +def map_first_tuple_or_el(x, fn): + if isinstance(x, tuple): + return (fn(x[0]),) + x[1:] + return fn(x) + + +class Chunk(nn.Module): + def __init__(self, chunks, fn, along_dim=-1): + super().__init__() + self.dim = along_dim + self.chunks = chunks + self.fn = fn + + def forward(self, x, **kwargs): + if self.chunks <= 1: + return self.fn(x, **kwargs) + chunks = x.chunk(self.chunks, dim=self.dim) + return torch.cat([self.fn(c, **kwargs) for c in chunks], dim=self.dim) + + +class PreNorm(nn.ModuleList): + def __init__(self, norm_class, dim, fn): + super().__init__() + self.norm = norm_class(dim) + self.fn = fn + + def forward(self, x, **kwargs): + x = self.norm(x) + return self.fn(x, **kwargs) + + +class ReZero(nn.Module): + def __init__(self, fn): + super().__init__() + self.residual_weight = nn.Parameter(torch.zeros(1)) + self.fn = fn + + def forward(self, x, **kwargs): + x = self.fn(x, **kwargs) + return map_first_tuple_or_el(x, lambda t: t * self.residual_weight) + + +class ScaleNorm(nn.Module): + def __init__(self, dim, eps=1e-5): + super().__init__() + self.g = nn.Parameter(torch.ones(1)) + self.eps = eps + + def forward(self, x): + def norm(t): + n = torch.norm(t, dim=-1, keepdim=True).clamp(min=self.eps) + return t / n * self.g + + return map_first_tuple_or_el(x, norm) + + +class ProjectInOut(nn.Module): + def __init__(self, fn, dim_in, dim_out, project_out=True): + super().__init__() + self.fn = fn + self.project_in = nn.Linear(dim_in, dim_out) + self.project_out = nn.Linear(dim_out, dim_in) if project_out else identity + + def forward(self, x, **kwargs): + x = self.project_in(x) + x, loss = self.fn(x, **kwargs) + x = self.project_out(x) + return x, loss + + +class MatrixMultiply(nn.Module): + def __init__(self, tensor, transpose=False): + super().__init__() + self.tensor = tensor + self.transpose = transpose + + def forward(self, x): + tensor = self.tensor + if self.transpose: + tensor = tensor.t() + return x @ tensor + + +# positional embeddings + + +class DepthWiseConv1d(nn.Module): + def __init__(self, dim_in, dim_out, kernel_size, stride=1, bias=True, causal=False): + super().__init__() + self.padding = ( + ((kernel_size - 1), 0) if causal else (kernel_size // 2, kernel_size // 2) + ) + + self.net = nn.Sequential( + nn.Conv1d( + dim_in, + dim_in, + kernel_size=kernel_size, + groups=dim_in, + stride=stride, + bias=bias, + ), + nn.Conv1d(dim_in, dim_out, 1, bias=bias), + ) + + def forward(self, x): + x = F.pad(x, self.padding, value=0.0) + return self.net(x) + + +class FixedPositionalEmbedding(nn.Module): + def __init__(self, dim, max_seq_len): + super().__init__() + inv_freq = 1.0 / (10000 ** (torch.arange(0, dim, 2).float() / dim)) + position = torch.arange(0, max_seq_len, dtype=torch.float) + sinusoid_inp = torch.einsum("i,j->ij", position, inv_freq) + emb = torch.cat((sinusoid_inp.sin(), sinusoid_inp.cos()), dim=-1) + self.register_buffer("emb", emb) + + def forward(self, x): + return self.emb[None, : x.shape[1], :].to(x) + + +def rotate_every_two(x): + x = rearrange(x, "... (d j) -> ... d j", j=2) + x1, x2 = x.unbind(dim=-1) + x = torch.stack((-x2, x1), dim=-1) + return rearrange(x, "... d j -> ... (d j)") + + +def apply_rotary_pos_emb(q, k, sinu_pos): + sinu_pos = rearrange(sinu_pos, "() n (j d) -> n j d", j=2) + sin, cos = sinu_pos.unbind(dim=-2) + sin, cos = map(lambda t: repeat(t, "b n -> b (n j)", j=2), (sin, cos)) + q, k = map(lambda t: (t * cos) + (rotate_every_two(t) * sin), (q, k)) + return q, k + + +# kmeans related function and class + + +def update_kmeans_on_backwards(module): + module.kmean_modules = find_modules(module, Kmeans) + + def hook(_, grad_in, grad_out): + for m in module.kmean_modules: + m.update() + + return module.register_backward_hook(hook) + + +def similarity(x, means): + return torch.einsum("bhld,hcd->bhlc", x, means) + + +def dists_and_buckets(x, means): + dists = similarity(x, means) + _, buckets = torch.max(dists, dim=-1) + return dists, buckets + + +def batched_bincount(index, num_classes, dim=-1): + shape = list(index.shape) + shape[dim] = num_classes + out = index.new_zeros(shape) + out.scatter_add_(dim, index, torch.ones_like(index, dtype=index.dtype)) + return out + + +def kmeans_iter(x, means, buckets=None): + b, h, _, d, dtype, num_clusters = *x.shape, x.dtype, means.shape[1] + + if not exists(buckets): + _, buckets = dists_and_buckets(x, means) + + bins = batched_bincount(buckets, num_clusters).sum(0, keepdim=True) + zero_mask = bins.long() == 0 + + means_ = buckets.new_zeros(b, h, num_clusters, d, dtype=dtype) + means_.scatter_add_(-2, expand_dim(buckets, -1, d), x) + means_ = F.normalize(means_.sum(0, keepdim=True), dim=-1).type(dtype) + + means = torch.where(zero_mask.unsqueeze(-1), means, means_) + means = means.squeeze(0) + return means + + +def distribution(dists, window_size): + _, topk_indices = dists.topk(k=window_size, dim=-2) + indices = topk_indices.transpose(-2, -1) + return indices.reshape(*indices.size()[:2], -1) + + +class Kmeans(nn.Module): + def __init__( + self, num_heads, head_dim, num_clusters, ema_decay=0.999, commitment=1e-4 + ): + super().__init__() + self.commitment = commitment + self.ema_decay = ema_decay + + self.register_buffer("means", torch.randn(num_heads, num_clusters, head_dim)) + self.register_buffer("initted", torch.tensor(False)) + self.num_new_means = 0 + self.new_means = None + + @torch.no_grad() + def init(self, x): + if self.initted: + return + _, h, _, d, device, _ = *x.shape, x.device, x.dtype + + num_clusters = self.means.shape[1] + + means = x.transpose(0, 1).contiguous().view(h, -1, d) + num_samples = means.shape[1] + + if num_samples >= num_clusters: + indices = torch.randperm(num_samples, device=device)[:num_clusters] + else: + indices = torch.randint(0, num_samples, (num_clusters,), device=device) + + means = means[:, indices] + + for _ in range(KMEAN_INIT_ITERS): + means = kmeans_iter(x, means) + + self.num_new_means = 0 + self.means.data.copy_(means) + self.initted.data.copy_(torch.tensor(True)) + + @torch.no_grad() + def update(self, new_means=None): + new_means = default(new_means, self.new_means) + assert exists(new_means), "new kmeans has not been supplied" + ema_inplace(self.means, new_means, self.ema_decay) + + del self.new_means + self.new_means = None + self.num_new_means = 0 + + def forward(self, x, update_means=False): + self.init(x) + + b, dtype = x.shape[0], x.dtype + means = self.means.type(dtype) + x = F.normalize(x, 2, dim=-1).type(dtype) + + with torch.no_grad(): + dists, buckets = dists_and_buckets(x, means) + + routed_means = batched_index_select(expand_dim(means, 0, b), buckets) + loss = F.mse_loss(x, routed_means) * self.commitment + + if update_means: + with torch.no_grad(): + means = kmeans_iter(x, means, buckets) + self.new_means = ema( + self.new_means, means, self.num_new_means / (self.num_new_means + 1) + ) + self.num_new_means += 1 + + return dists, loss + + +# kmeans attention class + + +class KmeansAttention(nn.Module): + def __init__( + self, + num_clusters, + window_size, + num_heads, + head_dim, + causal=False, + dropout=0.0, + ema_decay=0.999, + commitment=1e-4, + context_window_size=None, + receives_context=False, + num_mem_kv=0, + shared_qk=False, + ): + super().__init__() + self.num_heads = num_heads + self.num_clusters = num_clusters + self.head_dim = head_dim + + self.window_size = window_size + self.context_window_size = default(context_window_size, window_size) + self.causal = causal + + self.shared_qk = shared_qk + self.receives_context = receives_context + self.kmeans = Kmeans(num_heads, head_dim, num_clusters, ema_decay, commitment) + self.dropout = nn.Dropout(dropout) + + self.num_mem_kv = max(num_mem_kv, 1 if causal and not shared_qk else 0) + self.mem_key = nn.Parameter( + torch.randn(num_heads, num_clusters, self.num_mem_kv, head_dim) + ) + self.mem_value = nn.Parameter( + torch.randn(num_heads, num_clusters, self.num_mem_kv, head_dim) + ) + + def forward(self, q, k, v, query_mask=None, key_mask=None, **kwargs): + b, h, t, d, kv_t, wsz, c_wsz, nc, device, dtype = ( + *q.shape, + k.shape[2], + self.window_size, + self.context_window_size, + self.num_clusters, + q.device, + q.dtype, + ) + is_reverse = kwargs.pop("_reverse", False) + + out = torch.zeros_like(q, dtype=dtype) + + update_kmeans = self.training and not is_reverse + + key_mask = ( + default(key_mask, query_mask) if not self.receives_context else key_mask + ) + kv_wsz = wsz if not self.receives_context else c_wsz + + wsz = min(wsz, t) + kv_wsz = min(kv_wsz, kv_t) + + if not self.shared_qk or self.receives_context: + dists, aux_loss = self.kmeans(torch.cat((q, k), dim=2), update_kmeans) + q_dists, k_dists = split_at_index(2, t, dists) + indices = distribution(q_dists, wsz) + kv_indices = distribution(k_dists, kv_wsz) + else: + dists, aux_loss = self.kmeans(q, update_kmeans) + k = F.normalize(k, dim=-1).to(q) + indices = distribution(dists, wsz) + kv_indices = indices + + q = batched_index_select(q, indices) + k = batched_index_select(k, kv_indices) + v = batched_index_select(v, kv_indices) + + reshape_with_window = lambda x: x.reshape(b, h, nc, -1, d) + q, k, v = map(reshape_with_window, (q, k, v)) + + m_k, m_v = map( + lambda x: expand_dim(x, 0, b).to(q), (self.mem_key, self.mem_value) + ) + k, v = map(lambda x: torch.cat(x, dim=3), ((m_k, k), (m_v, v))) + + dots = torch.einsum("bhnid,bhnjd->bhnij", q, k) * (d**-0.5) + + mask_value = max_neg_value(dots) + + if exists(query_mask) or exists(key_mask): + query_mask = default( + query_mask, lambda: torch.ones((b, t), device=device).bool() + ) + key_mask = default( + key_mask, lambda: torch.ones((b, kv_t), device=device).bool() + ) + + q_mask = expand_dim(query_mask, 1, h).gather(2, indices) + kv_mask = expand_dim(key_mask, 1, h).gather(2, kv_indices) + q_mask, kv_mask = map(lambda t: t.reshape(b, h, nc, -1), (q_mask, kv_mask)) + mask = q_mask[:, :, :, :, None] * kv_mask[:, :, :, None, :] + mask = F.pad(mask, (self.num_mem_kv, 0), value=1) + dots.masked_fill_(~mask, mask_value) + del mask + + if self.causal: + q_mask, kv_mask = map( + lambda t: t.reshape(b, h, nc, -1), (indices, kv_indices) + ) + mask = q_mask[:, :, :, :, None] >= kv_mask[:, :, :, None, :] + mask = F.pad(mask, (self.num_mem_kv, 0), value=1) + dots.masked_fill_(~mask, mask_value) + del mask + + if self.shared_qk: + q_mask, kv_mask = map( + lambda t: t.reshape(b, h, nc, -1), (indices, kv_indices) + ) + mask = q_mask[:, :, :, :, None] == kv_mask[:, :, :, None, :] + mask = F.pad(mask, (self.num_mem_kv, 0), value=0) + dots.masked_fill_(mask, TOKEN_SELF_ATTN_VALUE) + del mask + + dots = dots.softmax(dim=-1) + dots = self.dropout(dots) + + bo = torch.einsum("bhcij,bhcjd->bhcid", dots, v) + so = torch.reshape(bo, (b, h, -1, bo.shape[-1])).type(dtype) + out = scatter_mean(out, so, indices.unsqueeze(-1).expand_as(so), -2) + return out, aux_loss + + +# feedforward + + +class GELU_(nn.Module): + def forward(self, x): + return ( + 0.5 + * x + * ( + 1 + + torch.tanh(math.sqrt(2 / math.pi) * (x + 0.044715 * torch.pow(x, 3))) + ) + ) + + +GELU = nn.GELU if hasattr(nn, "GELU") else GELU_ + + +class FeedForward(nn.Module): + def __init__(self, dim, mult=4, dropout=0.0, activation=None, glu=False): + super().__init__() + activation = default(activation, GELU) + + self.glu = glu + self.w1 = nn.Linear(dim, dim * mult * (2 if glu else 1)) + self.act = activation() + self.dropout = nn.Dropout(dropout) + self.w2 = nn.Linear(dim * mult, dim) + + def forward(self, x, **kwargs): + if not self.glu: + x = self.w1(x) + x = self.act(x) + else: + x, v = self.w1(x).chunk(2, dim=-1) + x = self.act(x) * v + + x = self.dropout(x) + x = self.w2(x) + return x + + +# self attention + + +class SelfAttention(nn.Module): + def __init__( + self, + dim, + max_seq_len, + heads, + local_attn_heads, + window_size, + dim_head=None, + local_attn_window_size=None, + local_attn_radius_blocks=1, + causal=False, + attn_dropout=0.0, + dropout=0.0, + kmeans_ema_decay=0.999, + commitment_factor=1e-4, + receives_context=False, + context_window_size=None, + rel_pos_emb=True, + num_mem_kv=0, + shared_qk=False, + conv_query_kernel=9, + ): + super().__init__() + assert ( + dim_head or (dim % heads) == 0 + ), "hidden dimension must be divisible by number of heads" + assert ( + max_seq_len % window_size + ) == 0, "maximum sequence length must be divisible by the target window size" + assert ( + local_attn_heads <= heads + ), "number of local attention heads must be less than total heads" + assert not ( + receives_context and local_attn_heads > 0 + ), "local attention cannot be used for self attention with context" + assert not ( + receives_context and causal + ), "contextual attention layer cannot be causal" + + local_attn_window_size = default(local_attn_window_size, window_size) + context_window_size = default(context_window_size, window_size) + + self.shared_qk = shared_qk + self.receives_context = receives_context + self.heads = heads + self.local_attn_heads = local_attn_heads + self.global_attn_heads = heads - local_attn_heads + + self.causal = causal + self.window_size = window_size + + dim_head = default(dim_head, dim // heads) + dim_heads = dim_head * heads + self.dim_head = dim_head + + num_clusters = max_seq_len // window_size + + # local + + local_dim_heads = dim_head * self.local_attn_heads + + if self.local_attn_heads > 0: + rel_pos_emb_config = (dim_head, local_attn_heads) if rel_pos_emb else None + self.local_attn = LocalAttention( + dim=dim_head, + window_size=local_attn_window_size, + causal=causal, + dropout=attn_dropout, + rel_pos_emb_config=rel_pos_emb_config, + look_backward=local_attn_radius_blocks, + look_forward=0 if causal else local_attn_radius_blocks, + ) + self.local_to_qkv = nn.Linear(dim, 3 * local_dim_heads) + + # global + + global_dim_heads = dim_head * self.global_attn_heads + + if self.global_attn_heads > 0: + self.global_attn = KmeansAttention( + num_clusters, + window_size, + self.global_attn_heads, + dim_head, + causal=causal, + dropout=attn_dropout, + ema_decay=kmeans_ema_decay, + commitment=commitment_factor, + receives_context=receives_context, + num_mem_kv=num_mem_kv, + shared_qk=shared_qk, + ) + + self.to_q = nn.Sequential( + Rearrange("b n c -> b c n"), + DepthWiseConv1d(dim, global_dim_heads, conv_query_kernel, causal=causal), + Rearrange("b c n -> b n c"), + ) + + self.to_v = nn.Linear(dim, global_dim_heads, bias=False) + + if not self.shared_qk: + self.to_k = nn.Linear(dim, global_dim_heads, bias=False) + + # out + + self.to_out = nn.Linear(dim_heads, dim, bias=False) + self.dropout = nn.Dropout(dropout) + + def forward( + self, + query, + key, + value, + context=None, + key_padding_mask=None, + context_mask=None, + pos_emb=None, + **kwargs + ): + assert not ( + self.receives_context and not exists(context) + ), "context must be passed if self attention is set to receive context" + input_mask = key_padding_mask + x = query.transpose(0, 1) + b, t, _, h, dh = *x.shape, self.heads, self.dim_head + has_local, has_global = map( + lambda x: x > 0, (self.local_attn_heads, self.global_attn_heads) + ) + + split_heads = ( + lambda v: reshape_dim(v, -1, (-1, dh)).transpose(1, 2).contiguous() + ) + + if has_local: + local_qkv = self.local_to_qkv(x).chunk(3, dim=-1) + lq, lk, lv = map(split_heads, local_qkv) + + if has_global: + kv_input = x if not self.receives_context else context + + q, v = self.to_q(x), self.to_v(kv_input) + + if not self.shared_qk: + k = self.to_k(kv_input) + else: + k = self.to_q(kv_input) if self.receives_context else q + + q, k, v = map(split_heads, (q, k, v)) + + out = [] + total_loss = torch.tensor(0.0, requires_grad=True, **to(x)) + + if has_local: + local_out = self.local_attn(lq, lk, lv, input_mask=input_mask) + out.append(local_out) + + if has_global: + if not self.receives_context and exists(pos_emb): + q, k = apply_rotary_pos_emb(q, k, pos_emb) + + global_out, loss = self.global_attn( + q, k, v, query_mask=input_mask, key_mask=context_mask + ) + total_loss = total_loss + loss + + out.append(global_out) + + out = torch.cat(out, dim=1) + out = out.reshape(b, h, t, -1).transpose(1, 2).reshape(b, t, -1) + out = self.dropout(out.transpose(0, 1)) + # out = self.to_out(out) + return out, total_loss diff --git a/data/fairseq/fairseq/modules/kmeans_vector_quantizer.py b/data/fairseq/fairseq/modules/kmeans_vector_quantizer.py new file mode 100644 index 0000000000000000000000000000000000000000..1015c389995d16e433d08303e3f37ce6f29653da --- /dev/null +++ b/data/fairseq/fairseq/modules/kmeans_vector_quantizer.py @@ -0,0 +1,128 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import torch +import torch.nn as nn +from fairseq.modules import Fp32GroupNorm + + +class KmeansVectorQuantizer(nn.Module): + def __init__( + self, dim, num_vars, groups, combine_groups, vq_dim, time_first, gamma=0.25 + ): + """Vector quantization using straight pass-through estimator (i.e. kmeans) + + Args: + dim: input dimension (channels) + num_vars: number of quantized vectors per group + groups: number of groups for vector quantization + combine_groups: whether to use the vectors for all groups + vq_dim: dimensionality of the resulting quantized vector + time_first: if true, expect input in BxTxC format, otherwise in BxCxT + gamma: commitment loss coefficient + """ + super().__init__() + + self.groups = groups + self.combine_groups = combine_groups + self.input_dim = dim + self.num_vars = num_vars + self.vq_dim = vq_dim + self.time_first = time_first + + assert ( + vq_dim % groups == 0 + ), f"dim {vq_dim} must be divisible by groups {groups} for concatenation" + + self.var_dim = vq_dim // groups + num_groups = groups if not combine_groups else 1 + + self.embedding = nn.Parameter( + 0.01 * torch.randn(num_vars, num_groups, self.var_dim) + ) + self.projection = nn.Sequential( + nn.Conv1d(dim, dim, kernel_size=1, groups=groups, bias=False), + Fp32GroupNorm(groups, dim), + ) + self.gamma = gamma + self.mse_mean = nn.MSELoss(reduction="mean") + + def _pass_grad(self, x, y): + """Manually set gradient for backward pass. + for y = f(x), ensure that during the backward pass, + dL/dy = dL/dx regardless of f(x). + Returns: + y, with the gradient forced to be dL/dy = dL/dx. + """ + + return y.detach() + (x - x.detach()) + + @property + def expand_embedding(self): + if self.combine_groups: + return self.embedding.expand(self.num_vars, self.groups, self.var_dim) + return self.embedding + + def forward_idx(self, x): + res = self.forward(x, produce_targets=True) + return res["x"], res["targets"] + + def forward(self, x, produce_targets=False): + + result = {"num_vars": self.num_vars} + + if self.time_first: + x = x.transpose(1, 2) + + bsz, fsz, tsz = x.shape + + ze = self.projection(x) + ze_ = ze.view(bsz, self.groups, self.var_dim, tsz).permute(0, 3, 1, 2) + d = ( + (ze_.unsqueeze(0) - self.expand_embedding.unsqueeze(1).unsqueeze(1)) + .view(self.num_vars, bsz, tsz, self.groups, -1) + .norm(dim=-1, p=2) + ) + idx = d.argmin(dim=0) + zq = ( + torch.stack( + [ + self.expand_embedding[idx[..., group], group] + for group in range(self.groups) + ], + dim=-2, + ) + .view(bsz, tsz, self.groups * self.var_dim) + .permute(0, 2, 1) + ) + assert ze.shape == zq.shape, (ze.shape, zq.shape) + x = self._pass_grad(ze, zq) + + with torch.no_grad(): + hard_x = ( + idx.new_zeros(bsz * tsz * self.groups, self.num_vars) + .scatter_(-1, idx.view(-1, 1), 1.0) + .view(bsz * tsz, self.groups, -1) + ) + hard_probs = torch.mean(hard_x.float(), dim=0) + result["code_perplexity"] = torch.exp( + -torch.sum(hard_probs * torch.log(hard_probs + 1e-7), dim=-1) + ).sum() + + if produce_targets: + result["targets"] = idx + + if self.time_first: + x = x.transpose(1, 2) # BCT -> BTC + result["x"] = x + + ze = ze.float() + zq = zq.float() + latent_loss = self.mse_mean(zq, ze.detach()) + commitment_loss = self.mse_mean(ze, zq.detach()) + + result["kmeans_loss"] = latent_loss + self.gamma * commitment_loss + + return result diff --git a/data/fairseq/fairseq/modules/layer_drop.py b/data/fairseq/fairseq/modules/layer_drop.py new file mode 100644 index 0000000000000000000000000000000000000000..8961d8bcbc492c40c6b30973234416ce5a414f5a --- /dev/null +++ b/data/fairseq/fairseq/modules/layer_drop.py @@ -0,0 +1,44 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. +""" +LayerDrop as described in https://arxiv.org/abs/1909.11556. +""" + +import torch +import torch.nn as nn + + +class LayerDropModuleList(nn.ModuleList): + """ + A LayerDrop implementation based on :class:`torch.nn.ModuleList`. + + We refresh the choice of which layers to drop every time we iterate + over the LayerDropModuleList instance. During evaluation we always + iterate over all layers. + + Usage:: + + layers = LayerDropList(p=0.5, modules=[layer1, layer2, layer3]) + for layer in layers: # this might iterate over layers 1 and 3 + x = layer(x) + for layer in layers: # this might iterate over all layers + x = layer(x) + for layer in layers: # this might not iterate over any layers + x = layer(x) + + Args: + p (float): probability of dropping out each layer + modules (iterable, optional): an iterable of modules to add + """ + + def __init__(self, p, modules=None): + super().__init__(modules) + self.p = p + + def __iter__(self): + dropout_probs = torch.empty(len(self)).uniform_() + for i, m in enumerate(super().__iter__()): + if not self.training or (dropout_probs[i] > self.p): + yield m diff --git a/data/fairseq/fairseq/modules/layer_norm.py b/data/fairseq/fairseq/modules/layer_norm.py new file mode 100644 index 0000000000000000000000000000000000000000..0b276ce02fc6bcb9619c9e8a0f7ec10cd28bc420 --- /dev/null +++ b/data/fairseq/fairseq/modules/layer_norm.py @@ -0,0 +1,48 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import torch +import torch.nn as nn +import torch.nn.functional as F + +try: + from apex.normalization import FusedLayerNorm as _FusedLayerNorm + + has_fused_layernorm = True + + class FusedLayerNorm(_FusedLayerNorm): + @torch.jit.unused + def forward(self, x): + if not x.is_cuda: + return super().forward(x) + else: + with torch.cuda.device(x.device): + return super().forward(x) + +except ImportError: + has_fused_layernorm = False + + +def LayerNorm(normalized_shape, eps=1e-5, elementwise_affine=True, export=False): + if torch.jit.is_scripting() or torch.jit.is_tracing(): + export = True + if not export and torch.cuda.is_available() and has_fused_layernorm: + return FusedLayerNorm(normalized_shape, eps, elementwise_affine) + return torch.nn.LayerNorm(normalized_shape, eps, elementwise_affine) + + +class Fp32LayerNorm(nn.LayerNorm): + def __init__(self, *args, **kwargs): + super().__init__(*args, **kwargs) + + def forward(self, input): + output = F.layer_norm( + input.float(), + self.normalized_shape, + self.weight.float() if self.weight is not None else None, + self.bias.float() if self.bias is not None else None, + self.eps, + ) + return output.type_as(input) diff --git a/data/fairseq/fairseq/modules/learned_positional_embedding.py b/data/fairseq/fairseq/modules/learned_positional_embedding.py new file mode 100644 index 0000000000000000000000000000000000000000..378d0f707183dd344dbb9288dda394b11053acf0 --- /dev/null +++ b/data/fairseq/fairseq/modules/learned_positional_embedding.py @@ -0,0 +1,61 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from typing import Dict, Optional + +import torch +import torch.nn as nn +import torch.nn.functional as F +from fairseq import utils +from torch import Tensor + + +class LearnedPositionalEmbedding(nn.Embedding): + """ + This module learns positional embeddings up to a fixed maximum size. + Padding ids are ignored by either offsetting based on padding_idx + or by setting padding_idx to None and ensuring that the appropriate + position ids are passed to the forward function. + """ + + def __init__(self, num_embeddings: int, embedding_dim: int, padding_idx: int): + super().__init__(num_embeddings, embedding_dim, padding_idx) + self.onnx_trace = False + if self.padding_idx is not None: + self.max_positions = self.num_embeddings - self.padding_idx - 1 + else: + self.max_positions = self.num_embeddings + + def forward( + self, + input: Tensor, + incremental_state: Optional[Dict[str, Dict[str, Optional[Tensor]]]] = None, + positions: Optional[Tensor] = None, + ): + """Input is expected to be of size [bsz x seqlen].""" + assert (positions is None) or ( + self.padding_idx is None + ), "If positions is pre-computed then padding_idx should not be set." + + if positions is None: + if incremental_state is not None: + # positions is the same for every token when decoding a single step + # Without the int() cast, it doesn't work in some cases when exporting to ONNX + positions = torch.zeros( + (1, 1), device=input.device, dtype=input.dtype + ).fill_(int(self.padding_idx + input.size(1))) + else: + positions = utils.make_positions( + input, self.padding_idx, onnx_trace=self.onnx_trace + ) + return F.embedding( + positions, + self.weight, + self.padding_idx, + self.max_norm, + self.norm_type, + self.scale_grad_by_freq, + self.sparse, + ) diff --git a/data/fairseq/fairseq/modules/lightconv_layer/__init__.py b/data/fairseq/fairseq/modules/lightconv_layer/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..3b2a99c1227f827768911e5e22e79f6865ffbfd3 --- /dev/null +++ b/data/fairseq/fairseq/modules/lightconv_layer/__init__.py @@ -0,0 +1,6 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from .lightconv_layer import LightconvLayer # noqa diff --git a/data/fairseq/fairseq/modules/lightconv_layer/cuda_function_gen.py b/data/fairseq/fairseq/modules/lightconv_layer/cuda_function_gen.py new file mode 100644 index 0000000000000000000000000000000000000000..a25433dd8edae2f0b52d7d0eeeb829cabc6b4b89 --- /dev/null +++ b/data/fairseq/fairseq/modules/lightconv_layer/cuda_function_gen.py @@ -0,0 +1,289 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + + +def gen_forward(): + + kernels = [3, 5, 7, 15, 31, 63, 127, 255] + seqs = [32 * x for x in [1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16]] + + head = """ +/** + * Copyright (c) Facebook, Inc. and its affiliates. + * + * This source code is licensed under the MIT license found in the + * LICENSE file in the root directory of this source tree. + */ + +#include "lightconv_cuda.cuh" + +std::vector lightconv_cuda_forward(at::Tensor input, at::Tensor filters, int padding_l) { + + at::DeviceGuard g(input.device()); + const auto minibatch = input.size(0); + const auto numFeatures = input.size(1); + const auto sequenceLength = input.size(2); + + const auto numHeads = filters.size(0); + const auto filterSize = filters.size(1); + + const auto numFiltersInBlock = numFeatures / numHeads; + + const dim3 blocks(minibatch, numFeatures); + + auto output = at::zeros_like(input); + auto stream = at::cuda::getCurrentCUDAStream(); +""" + + sequence_if = """ + if (sequenceLength <= {seq}) {{ + switch(filterSize) {{ +""" + + case_k = """ + case {k}: +""" + + main_block = """ + if (padding_l == {pad}) {{ + AT_DISPATCH_FLOATING_TYPES_AND_HALF(input.scalar_type(), "lightconv_forward", ([&] {{ + lightconv_forward_kernel<{k}, {b_size}, {pad}, scalar_t> + <<>>( + input.data(), + filters.data(), + minibatch, + sequenceLength, + numFeatures, + numFiltersInBlock, + output.data()); + }})); + }} else +""" + + bad_padding = """ + { + std::cout << "WARNING: Unsupported padding size - skipping forward pass" << std::endl; + } + break; +""" + + bad_filter = """ + default: + std::cout << "WARNING: Unsupported filter length passed - skipping forward pass" << std::endl; + } +""" + + con_else = """ + } else +""" + + final_else = """ + { + switch(filterSize) { +""" + + final_return = """ + } + + return {output}; +} +""" + + with open("lightconv_cuda_forward.cu", "w") as forward: + forward.write(head) + for seq in seqs: + forward.write(sequence_if.format(seq=seq)) + for k in kernels: + forward.write(case_k.format(k=k)) + for pad in [k // 2, k - 1]: + forward.write(main_block.format(k=k, b_size=seq, pad=pad)) + forward.write(bad_padding) + forward.write(bad_filter) + forward.write(con_else) + + forward.write(final_else) + for k in kernels: + forward.write(case_k.format(k=k)) + for pad in [k // 2, k - 1]: + forward.write(main_block.format(k=k, b_size=seq, pad=pad)) + forward.write(bad_padding) + forward.write(bad_filter) + forward.write(final_return) + + +def gen_backward(): + + head = """ +/** + * Copyright (c) Facebook, Inc. and its affiliates. + * + * This source code is licensed under the MIT license found in the + * LICENSE file in the root directory of this source tree. + */ + +#include "lightconv_cuda.cuh" + +std::vector lightconv_cuda_backward( + at::Tensor gradOutput, + int padding_l, + at::Tensor input, + at::Tensor filters) { + + // gradWrtInput + const int minibatch = input.size(0); + const int numFeatures = input.size(1); + const int sequenceLength = input.size(2); + + const int numHeads = filters.size(0); + const int filterSize = filters.size(1); + + const dim3 gradBlocks(minibatch, numFeatures); + const dim3 weightGradFirstpassShortBlocks(minibatch, numHeads); + const dim3 weightGradSecondpassBlocks(numHeads, filterSize); + + const int numFiltersInBlock = numFeatures / numHeads; + + auto gradInput = at::zeros_like(input); + auto gradFilters = at::zeros_like(filters); + + at::DeviceGuard g(input.device()); + auto stream = at::cuda::getCurrentCUDAStream(); + + switch(filterSize) { +""" + + sequence_if = """ + if (sequenceLength <= {seq}) {{ +""" + + case_k = """ + case {k}: +""" + + main_block = """ + if (padding_l == {p}) {{ + AT_DISPATCH_FLOATING_TYPES_AND_HALF(input.scalar_type(), "lightconv_backward", ([&] {{ + lightconv_grad_wrt_input_kernel<{k}, {b_size}, {p}, scalar_t> + <<>>( + gradOutput.data(), + filters.data(), + minibatch, + sequenceLength, + numFeatures, + numFiltersInBlock, + gradInput.data()); + +""" + + weight_grad_short = """ + at::Tensor tempSumGradFilters = at::zeros({{minibatch, numHeads, filterSize}}, input.options().dtype(at::kFloat)); + lightconv_grad_wrt_weights_firstpass_short_kernel<{k}, {b_size}, {p}, scalar_t> + <<>>( + input.data(), + gradOutput.data(), + minibatch, + sequenceLength, + numFeatures, + numFiltersInBlock, + numHeads, + tempSumGradFilters.data() + ); + + lightconv_grad_wrt_weights_secondpass_short_kernel<{k}, {b_size}, scalar_t> + <<>>( + tempSumGradFilters.data(), + minibatch, + numFiltersInBlock, + gradFilters.data() + ); + }})); + }} else +""" + + weight_grad = """ + at::Tensor tempSumGradFilters = at::zeros({{minibatch, numFeatures, filterSize}}, input.options().dtype(at::kFloat)); + lightconv_grad_wrt_weights_firstpass_kernel<{k}, {b_size}, {p}, scalar_t> + <<>>( + input.data(), + gradOutput.data(), + minibatch, + sequenceLength, + numFeatures, + numFiltersInBlock, + tempSumGradFilters.data() + ); + + lightconv_grad_wrt_weights_secondpass_kernel<{k}, {b_size}, scalar_t> + <<>>( + tempSumGradFilters.data(), + minibatch, + numFiltersInBlock, + gradFilters.data() + ); + }})); + }} else +""" + + bad_padding = """ + { + std::cout << "WARNING: Unsupported padding size - skipping backward pass" << std::endl; + } +""" + + breakout = """ + break; +""" + + bad_filter = """ + default: + std::cout << "WARNING: Unsupported filter length passed - skipping backward pass" << std::endl; +""" + + con_else = """ + } else +""" + + final_else = """ + { + switch(filterSize) { +""" + + last_return = """ + } + return {gradInput, gradFilters}; +} +""" + + kernels = [3, 5, 7, 15, 31, 63, 127, 255] + seqs = [32 * x for x in [1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16]] + thresh = [32, 32, 64, 128, 256, -1, -1, -1] + max_mem = [-1, -1, -1, -1, -1, 192, 96, 64] + + with open("lightconv_cuda_backward.cu", "w") as backward: + backward.write(head) + for (k, t, mem) in zip(kernels, thresh, max_mem): + backward.write(case_k.format(k=k)) + for seq in seqs: + if (t == -1 or seq <= t) and (mem == -1 or seq < mem): + backward.write(sequence_if.format(seq=seq)) + for p in [k // 2, k - 1]: + backward.write(main_block.format(k=k, b_size=seq, p=p)) + backward.write(weight_grad_short.format(k=k, b_size=seq, p=p)) + backward.write(bad_padding) + else: + for p in [k // 2, k - 1]: + backward.write(main_block.format(k=k, b_size=32, p=p)) + backward.write(weight_grad.format(k=k, b_size=32, p=p)) + backward.write(bad_padding) + backward.write(breakout) + break + backward.write(con_else) + backward.write(bad_filter) + backward.write(last_return) + + +if __name__ == "__main__": + gen_forward() + gen_backward() diff --git a/data/fairseq/fairseq/modules/lightconv_layer/lightconv_cuda.cpp b/data/fairseq/fairseq/modules/lightconv_layer/lightconv_cuda.cpp new file mode 100644 index 0000000000000000000000000000000000000000..ece47a8d908b93cec102743070c9057986d39d3f --- /dev/null +++ b/data/fairseq/fairseq/modules/lightconv_layer/lightconv_cuda.cpp @@ -0,0 +1,51 @@ +/** + * Copyright (c) Facebook, Inc. and its affiliates. + * + * This source code is licensed under the MIT license found in the + * LICENSE file in the root directory of this source tree. + */ + +#include +#include + +std::vector +lightconv_cuda_forward(at::Tensor input, at::Tensor filters, int padding_l); + +std::vector lightconv_cuda_backward( + at::Tensor gradOutput, + int padding_l, + at::Tensor input, + at::Tensor filters); + +#define CHECK_CUDA(x) \ + AT_ASSERTM(x.type().is_cuda(), #x " must be a CUDA tensor") +#define CHECK_CONTIGUOUS(x) \ + AT_ASSERTM(x.is_contiguous(), #x " must be contiguous") +#define CHECK_INPUT(x) \ + CHECK_CUDA(x); \ + CHECK_CONTIGUOUS(x) + +std::vector +lightconv_forward(at::Tensor input, at::Tensor filters, int padding_l) { + CHECK_INPUT(input); + CHECK_INPUT(filters); + + return lightconv_cuda_forward(input, filters, padding_l); +} + +std::vector lightconv_backward( + at::Tensor gradOutput, + int padding_l, + at::Tensor input, + at::Tensor filters) { + CHECK_INPUT(gradOutput); + CHECK_INPUT(input); + CHECK_INPUT(filters); + + return lightconv_cuda_backward(gradOutput, padding_l, input, filters); +} + +PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) { + m.def("forward", &lightconv_forward, "lighconv forward (CUDA)"); + m.def("backward", &lightconv_backward, "lighconv backward (CUDA)"); +} diff --git a/data/fairseq/fairseq/modules/lightconv_layer/lightconv_cuda.cuh b/data/fairseq/fairseq/modules/lightconv_layer/lightconv_cuda.cuh new file mode 100644 index 0000000000000000000000000000000000000000..610ab399e9b201cd8b0fb87a91e09b6f7aab9803 --- /dev/null +++ b/data/fairseq/fairseq/modules/lightconv_layer/lightconv_cuda.cuh @@ -0,0 +1,79 @@ +/** + * Copyright (c) Facebook, Inc. and its affiliates. + * + * This source code is licensed under the MIT license found in the + * LICENSE file in the root directory of this source tree. + */ + +#include +#include + +#include +#include + +#include +#include +#include +#include +#include +#include + +#include +#include + +#define SHFL_MASK 0xffffffff + +template +__global__ void lightconv_forward_kernel( + const scalar_t* input, + const scalar_t* filters, + int minibatch, + int sequenceLength, + int numFeatures, + int numFiltersInBlock, + scalar_t* output); + +template +__global__ void lightconv_grad_wrt_input_kernel( + const scalar_t* input, + const scalar_t* filters, + int minibatch, + int sequenceLength, + int numFeatures, + int numFiltersInBlock, + scalar_t* output); + +template +__global__ void lightconv_grad_wrt_weights_firstpass_short_kernel( + const scalar_t* input, + const scalar_t* gradInput, + int minibatch, + int sequenceLength, + int numFeatures, + int numFiltersInBlock, + int numHeads, + float* output); + +template +__global__ void lightconv_grad_wrt_weights_secondpass_short_kernel( + const float* input, + const int minibatch, + const int numFiltersInBlock, + scalar_t* output); + +template +__global__ void lightconv_grad_wrt_weights_firstpass_kernel( + const scalar_t* input, + const scalar_t* gradInput, + int minibatch, + int sequenceLength, + int numFeatures, + int numFiltersInBlock, + float* output); + +template +__global__ void lightconv_grad_wrt_weights_secondpass_kernel( + const float* input, + const int minibatch, + const int numFiltersInBlock, + scalar_t* output); diff --git a/data/fairseq/fairseq/modules/lightconv_layer/lightconv_cuda_kernel.cu b/data/fairseq/fairseq/modules/lightconv_layer/lightconv_cuda_kernel.cu new file mode 100644 index 0000000000000000000000000000000000000000..cdf31d5d2df2d3433c66e167f098e48a99f96db2 --- /dev/null +++ b/data/fairseq/fairseq/modules/lightconv_layer/lightconv_cuda_kernel.cu @@ -0,0 +1,400 @@ +/** + * Copyright (c) Facebook, Inc. and its affiliates. + * + * This source code is licensed under the MIT license found in the + * LICENSE file in the root directory of this source tree. + */ + +#include "../cuda_utils.cu" +#include "lightconv_cuda.cuh" +#include "lightconv_cuda_backward.cu" +#include "lightconv_cuda_forward.cu" + +template +__global__ void lightconv_forward_kernel( + const scalar_t* input, + const scalar_t* filters, + int minibatch, + int sequenceLength, + int numFeatures, + int numFiltersInBlock, + scalar_t* output) { + const int tid = threadIdx.x; + const int batchIdx = blockIdx.x; + const int featureIdx = blockIdx.y; + const int filterIdx = featureIdx / numFiltersInBlock; + + const int IOOffset = + numFeatures * sequenceLength * batchIdx + featureIdx * sequenceLength; + const scalar_t* inputFeature = &input[IOOffset]; + scalar_t* outputFeature = &output[IOOffset]; + const scalar_t* inputFilter = &filters[filterIdx * FS]; + + assert(blockDim.x == SB); + + scalar_t filter[FS]; +#pragma unroll + for (int i = 0; i < FS; ++i) { + filter[i] = inputFilter[i]; + } + + __shared__ scalar_t temp[SB + FS]; + zeroSharedMem(temp); + + const int numIterations = divUp(sequenceLength, SB); + + for (int i = 0; i < numIterations; ++i) { + // Read input into shared memory + const int inputOffset = i * SB; + + load_input_to_shared( + inputFeature, + inputOffset, + sequenceLength, + i, + numIterations, + (numIterations == 1), + temp); + + __syncthreads(); + + scalar_t out = 0; +#pragma unroll + for (int j = 0; j < FS; ++j) { + out += filter[j] * temp[tid + j]; + } + + // Write output + const int outputOffset = inputOffset; + if ((outputOffset + tid) < sequenceLength) { + outputFeature[outputOffset + tid] = out; + } + + __syncthreads(); + } +} + +template +__global__ void lightconv_grad_wrt_input_kernel( + const scalar_t* input, + const scalar_t* filters, + int minibatch, + int sequenceLength, + int numFeatures, + int numFiltersInBlock, + scalar_t* output) { + // input grad kernel is similar to forward kernel + const int tid = threadIdx.x; + const int batchIdx = blockIdx.x; + const int featureIdx = blockIdx.y; + const int filterIdx = featureIdx / numFiltersInBlock; + + const int IOOffset = + numFeatures * sequenceLength * batchIdx + featureIdx * sequenceLength; + const scalar_t* inputFeature = &input[IOOffset]; + scalar_t* outputFeature = &output[IOOffset]; + const scalar_t* inputFilter = &filters[filterIdx * FS]; + + assert(blockDim.x == SB); + + scalar_t filter[FS]; + +// The only change is loading the filter in reverse +#pragma unroll + for (int i = 0; i < FS; ++i) { + filter[i] = inputFilter[FS - i - 1]; + } + + __shared__ scalar_t temp[SB + FS]; + const int padding = FS - padding_l - 1; + zeroSharedMem(temp); + + __syncthreads(); + + const int numIterations = divUp(sequenceLength, SB); + + for (int i = 0; i < numIterations; ++i) { + // Read input into shared memory + const int inputOffset = i * SB; + + load_input_to_shared( + inputFeature, + inputOffset, + sequenceLength, + i, + numIterations, + false, + temp); + + __syncthreads(); + + scalar_t out = 0; +#pragma unroll + for (int j = 0; j < FS; ++j) { + out += filter[j] * temp[tid + j]; + } + + // Write output + const int outputOffset = inputOffset; + if ((outputOffset + tid) < sequenceLength) { + outputFeature[outputOffset + tid] = out; + } + + __syncthreads(); + } +} + +// This is by far the most expensive kernel in terms of time taken. +// Can be 16x slower than the forward or grad_wrt_input when filter size is 31 +template +__global__ void lightconv_grad_wrt_weights_firstpass_short_kernel( + const scalar_t* input, + const scalar_t* gradInput, + int minibatch, + int sequenceLength, + int numFeatures, + int numFiltersInBlock, + int numHeads, + float* output) { + const int tid = threadIdx.x; + const int batchIdx = blockIdx.x; + const int filterIdx = blockIdx.y; + + const int numIterations = divUp(sequenceLength, SB); + + float* tempOutputGradWeight = &output[filterIdx * FS * minibatch]; + + assert(blockDim.x == SB); + + __shared__ scalar_t tempInput[SB + FS]; + __shared__ scalar_t tempGradInput[SB + FS]; + + // local weight accumulation + float accumWeights[FS]; + + // Initialize memory + for (int i = 0; i < FS; ++i) { + accumWeights[i] = float(0.0); + } + + // loop over each sequence within filterblock + for (int idxInFilterBlock = 0; idxInFilterBlock < numFiltersInBlock; + ++idxInFilterBlock) { + const int featureOffset = batchIdx * numFeatures * sequenceLength + + (filterIdx * numFiltersInBlock + idxInFilterBlock) * sequenceLength; + const scalar_t* inputFeature = &input[featureOffset]; + const scalar_t* gradInputFeature = &gradInput[featureOffset]; + + zeroSharedMem(tempInput); + zeroSharedMem(tempGradInput); + __syncthreads(); + + for (int i = 0; i < numIterations; ++i) { + const int inputOffset = i * SB; + + load_input_to_shared( + inputFeature, + inputOffset, + sequenceLength, + i, + numIterations, + false, + tempInput); + load_input_to_shared( + gradInputFeature, + inputOffset, + sequenceLength, + i, + numIterations, + false, + tempGradInput); + + __syncthreads(); + + const int gradIndex = (FS / 2) + tid; + scalar_t tempGrad = tempGradInput[gradIndex]; + +#pragma unroll + for (int j = 0; j < FS; j++) { + const int inputIndex = tid + j; + accumWeights[j] += tempInput[inputIndex] * tempGrad; + } + + __syncthreads(); + } + } + + // Row-major sum + for (int filterWeightIdx = 0; filterWeightIdx < FS; ++filterWeightIdx) { + float temp; + if (tid < sequenceLength) { + temp = accumWeights[filterWeightIdx]; + } else { + temp = float(0.0); + } + + const int outputOffset = filterWeightIdx * minibatch + batchIdx; + + temp = blockReduce(temp); + + if (tid == 0) { + tempOutputGradWeight[outputOffset] = temp; + } + } +} + +template +__global__ void lightconv_grad_wrt_weights_secondpass_short_kernel( + const float* input, + const int minibatch, + const int numFiltersInBlock, + scalar_t* output) { + assert(blockDim.x == SB); + + const int tid = threadIdx.x; + + const int filterIdx = blockIdx.x; + const int filterWeightIdx = blockIdx.y; + + const int inputOffset = + filterIdx * FS * minibatch + filterWeightIdx * minibatch; + const float* tempInput = &input[inputOffset]; + + // read into shared memory for reduction + int readIndex = tid; + + float sum = 0.0; + while (readIndex < minibatch) { + sum += tempInput[readIndex]; + readIndex += SB; + } + + float temp = blockReduce(sum); + + if (tid == 0) { + output[blockIdx.x * FS + blockIdx.y] = temp; + } +} + +// This is by far the most expensive kernel in terms of time taken. +// Can be 16x slower than the forward or grad_wrt_input when filter size is 31 +template +__global__ void lightconv_grad_wrt_weights_firstpass_kernel( + const scalar_t* input, + const scalar_t* gradInput, + int minibatch, + int sequenceLength, + int numFeatures, + int numFiltersInBlock, + float* output) { + assert(blockDim.x == SB); + + const int tid = threadIdx.x; + const int batchIdx = blockIdx.x; + const int featureIdx = blockIdx.y; + const int filterIdx = featureIdx / numFiltersInBlock; + const int idxInFilterBlock = featureIdx % numFiltersInBlock; + + const int numIterations = divUp(sequenceLength, SB); + + float temp; + + __shared__ scalar_t tempInput[SB + FS]; + __shared__ scalar_t tempGradInput[SB + FS]; + zeroSharedMem(tempInput); + zeroSharedMem(tempGradInput); + __syncthreads(); + + float accumWeights[FS]; + + for (int i = 0; i < FS; ++i) { + accumWeights[i] = float(0.0); + } + + const int IOOffset = + batchIdx * numFeatures * sequenceLength + featureIdx * sequenceLength; + const scalar_t* inputFeature = &input[IOOffset]; + const scalar_t* gradInputFeature = &gradInput[IOOffset]; + float* tempOutputGradWeight = + &output[filterIdx * FS * minibatch * numFiltersInBlock]; + + for (int i = 0; i < numIterations; ++i) { + const int inputOffset = i * SB; + + load_input_to_shared( + inputFeature, + inputOffset, + sequenceLength, + i, + numIterations, + false, + tempInput); + load_input_to_shared( + gradInputFeature, + inputOffset, + sequenceLength, + i, + numIterations, + false, + tempGradInput); + __syncthreads(); + +#pragma unroll + for (int j = 0; j < FS; ++j) { + accumWeights[j] += tempInput[tid + j] * tempGradInput[tid + (FS / 2)]; + } + + __syncthreads(); + } + + // Row-major sum + for (int filterWeightIdx = 0; filterWeightIdx < FS; ++filterWeightIdx) { + // Write to shared memory before reduction + if (tid < sequenceLength) { + temp = accumWeights[filterWeightIdx]; + } else { + temp = float(0.0); + } + + temp = blockReduce(temp); + + const int outputOffset = filterWeightIdx * minibatch * numFiltersInBlock + + batchIdx * numFiltersInBlock + idxInFilterBlock; + + if (tid == 0) { + tempOutputGradWeight[outputOffset] = temp; + } + } +} + +template +__global__ void lightconv_grad_wrt_weights_secondpass_kernel( + const float* input, + const int minibatch, + const int numFiltersInBlock, + scalar_t* output) { + assert(blockDim.x == SB); + const int tid = threadIdx.x; + + // What is the id within a minibatch + const int filterIdx = blockIdx.x; + const int filterWeightIdx = blockIdx.y; + + const int inputOffset = filterIdx * FS * minibatch * numFiltersInBlock + + filterWeightIdx * minibatch * numFiltersInBlock; + const float* tempInput = &input[inputOffset]; + + int readIndex = tid; + + float sum = float(0.0); + while (readIndex < (minibatch * numFiltersInBlock)) { + sum += tempInput[readIndex]; + readIndex += SB; + } + + float temp = blockReduce(sum); + + if (tid == 0) { + output[blockIdx.x * FS + blockIdx.y] = temp; + } +} diff --git a/data/fairseq/fairseq/modules/lightconv_layer/lightconv_layer.py b/data/fairseq/fairseq/modules/lightconv_layer/lightconv_layer.py new file mode 100644 index 0000000000000000000000000000000000000000..e7e597f4749c591b057d776aacec39b44d99c037 --- /dev/null +++ b/data/fairseq/fairseq/modules/lightconv_layer/lightconv_layer.py @@ -0,0 +1,137 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import lightconv_cuda +import torch +import torch.nn.functional as F +from fairseq import utils +from fairseq.incremental_decoding_utils import with_incremental_state +from fairseq.modules.fairseq_dropout import FairseqDropout +from torch import nn +from torch.autograd import Function + + +class lightconvFunction(Function): + @staticmethod + def forward(ctx, x, weights, padding_l): + ctx.padding_l = padding_l + outputs = lightconv_cuda.forward(x, weights, padding_l) + variables = [x, weights] + ctx.save_for_backward(*variables) + return outputs[0] + + @staticmethod + def backward(ctx, grad_output): + outputs = lightconv_cuda.backward( + grad_output.contiguous(), ctx.padding_l, *ctx.saved_tensors + ) + grad_input, grad_weights = outputs + return grad_input, grad_weights, None + + +@with_incremental_state +class LightconvLayer(nn.Module): + def __init__( + self, + input_size, + kernel_size=1, + padding_l=None, + weight_softmax=False, + num_heads=1, + weight_dropout=0.0, + bias=False, + ): + super(LightconvLayer, self).__init__() + self.input_size = input_size + self.kernel_size = kernel_size + self.padding_l = padding_l + self.num_heads = num_heads + self.weight_softmax = weight_softmax + self.weight_dropout_module = FairseqDropout( + weight_dropout, module_name=self.__class__.__name__ + ) + + self.weight = nn.Parameter(torch.Tensor(num_heads, kernel_size)) + if bias: + self.bias = nn.Parameter(torch.Tensor(input_size)) + else: + self.bias = None + self.reset_parameters() + + def upgrade_state_dict_named(self, state_dict, name): + prefix = name + "." if name != "" else "" + for k, v in state_dict.items(): + if k.endswith(prefix + "weight"): + if v.dim() == 3 and v.size(1) == 1: + state_dict[k] = v.squeeze(1) + + def reset_parameters(self): + nn.init.xavier_uniform_(self.weight) + if self.bias is not None: + nn.init.constant_(self.bias, 0.0) + + def forward(self, x, incremental_state=None): + + # during inference time, incremental BMM is faster + if incremental_state is not None: + T, B, C = x.size() + K, H = self.kernel_size, self.num_heads + R = C // H + input_buffer = self._get_input_buffer(incremental_state) + if input_buffer is None: + input_buffer = x.new() + x_unfold = torch.cat([input_buffer, x.unsqueeze(3)], dim=3) + if self.kernel_size > 1: + self._set_input_buffer( + incremental_state, x_unfold[:, :, :, -self.kernel_size + 1 :] + ) + x_unfold = x_unfold.view(T * B * H, R, -1) + + weight = self.weight + if self.weight_softmax: + weight = F.softmax(weight.float(), dim=1).type_as(weight) + + weight = weight[:, -x_unfold.size(2) :] + + K = weight.size(1) + + weight = ( + weight.view(1, H, K) + .expand(T * B, H, K) + .contiguous() + .view(T * B * H, K, 1) + ) + + weight = self.weight_dropout_module(weight) + output = torch.bmm(x_unfold, weight) # T*B*H x R x 1 + output = output.view(T, B, C) + return output + + # during training time, use CUDA kernel + else: + x = x.permute(1, 2, 0).contiguous() + weight = self.weight + if self.weight_softmax: + weight = F.softmax(self.weight, -1) + if self.weight_dropout_module.p: + weight = self.weight_dropout_module(weight) + return lightconvFunction.apply(x, weight, self.padding_l).permute(2, 0, 1) + + def reorder_incremental_state(self, incremental_state, new_order): + input_buffer = self._get_input_buffer(incremental_state) + if input_buffer is not None: + input_buffer = input_buffer.index_select(1, new_order) + self._set_input_buffer(incremental_state, input_buffer) + + def _get_input_buffer(self, incremental_state): + return utils.get_incremental_state(self, incremental_state, "input_buffer") + + def _set_input_buffer(self, incremental_state, new_buffer): + return utils.set_incremental_state( + self, incremental_state, "input_buffer", new_buffer + ) + + def half(self): + return self._apply(lambda t: t.half() if t.is_floating_point() else t) diff --git a/data/fairseq/fairseq/modules/lightconv_layer/setup.py b/data/fairseq/fairseq/modules/lightconv_layer/setup.py new file mode 100644 index 0000000000000000000000000000000000000000..052635be79b466d0ad56cf5cf607bd10c2297ecf --- /dev/null +++ b/data/fairseq/fairseq/modules/lightconv_layer/setup.py @@ -0,0 +1,23 @@ +#!/usr/bin/env python3 +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from setuptools import setup +from torch.utils.cpp_extension import BuildExtension, CUDAExtension + + +setup( + name="lightconv_layer", + ext_modules=[ + CUDAExtension( + "lightconv_cuda", + [ + "lightconv_cuda.cpp", + "lightconv_cuda_kernel.cu", + ], + ), + ], + cmdclass={"build_ext": BuildExtension}, +) diff --git a/data/fairseq/fairseq/modules/lightweight_convolution.py b/data/fairseq/fairseq/modules/lightweight_convolution.py new file mode 100644 index 0000000000000000000000000000000000000000..ec11a9507951c9e8f3564753841dd9c74a4900e0 --- /dev/null +++ b/data/fairseq/fairseq/modules/lightweight_convolution.py @@ -0,0 +1,310 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import torch +import torch.nn as nn +import torch.nn.functional as F +from fairseq import utils +from fairseq.incremental_decoding_utils import with_incremental_state +from fairseq.modules.fairseq_dropout import FairseqDropout +from fairseq.modules.unfold import unfold1d + + +def LightweightConv( + input_size, + kernel_size=1, + padding_l=None, + num_heads=1, + weight_dropout=0.0, + weight_softmax=False, + bias=False, +): + if torch.cuda.is_available(): + try: + from fairseq.modules.lightconv_layer import LightconvLayer + + return LightconvLayer( + input_size, + kernel_size=kernel_size, + padding_l=padding_l, + num_heads=num_heads, + weight_dropout=weight_dropout, + weight_softmax=weight_softmax, + bias=bias, + ) + except ImportError as e: + print(e) + return LightweightConv1dTBC( + input_size, + kernel_size=kernel_size, + padding_l=padding_l, + num_heads=num_heads, + weight_dropout=weight_dropout, + weight_softmax=weight_softmax, + bias=bias, + ) + + +class LightweightConv1d(nn.Module): + """Lightweight Convolution assuming the input is BxCxT + This is just an example that explains LightConv clearer than the TBC version. + We don't use this module in the model. + + Args: + input_size: # of channels of the input and output + kernel_size: convolution channels + padding: padding + num_heads: number of heads used. The weight is of shape + `(num_heads, 1, kernel_size)` + weight_softmax: normalize the weight with softmax before the convolution + + Shape: + Input: BxCxT, i.e. (batch_size, input_size, timesteps) + Output: BxCxT, i.e. (batch_size, input_size, timesteps) + + Attributes: + weight: the learnable weights of the module of shape + `(num_heads, 1, kernel_size)` + bias: the learnable bias of the module of shape `(input_size)` + """ + + def __init__( + self, + input_size, + kernel_size=1, + padding=0, + num_heads=1, + weight_softmax=False, + bias=False, + weight_dropout=0.0, + ): + super().__init__() + self.input_size = input_size + self.kernel_size = kernel_size + self.num_heads = num_heads + self.padding = padding + self.weight_softmax = weight_softmax + self.weight = nn.Parameter(torch.Tensor(num_heads, 1, kernel_size)) + + if bias: + self.bias = nn.Parameter(torch.Tensor(input_size)) + else: + self.bias = None + self.weight_dropout_module = FairseqDropout( + weight_dropout, module_name=self.__class__.__name__ + ) + self.reset_parameters() + + def reset_parameters(self): + nn.init.xavier_uniform_(self.weight) + if self.bias is not None: + nn.init.constant_(self.bias, 0.0) + + def forward(self, input): + """ + input size: B x C x T + output size: B x C x T + """ + B, C, T = input.size() + H = self.num_heads + + weight = self.weight + if self.weight_softmax: + weight = F.softmax(weight, dim=-1) + + weight = self.weight_dropout_module(weight) + # Merge every C/H entries into the batch dimension (C = self.input_size) + # B x C x T -> (B * C/H) x H x T + # One can also expand the weight to C x 1 x K by a factor of C/H + # and do not reshape the input instead, which is slow though + input = input.view(-1, H, T) + output = F.conv1d(input, weight, padding=self.padding, groups=self.num_heads) + output = output.view(B, C, T) + if self.bias is not None: + output = output + self.bias.view(1, -1, 1) + + return output + + +@with_incremental_state +class LightweightConv1dTBC(nn.Module): + """Lightweight Convolution assuming the input is TxBxC + Args: + input_size: # of channels of the input + kernel_size: convolution channels + padding_l: padding to the left when using "same" padding + num_heads: number of heads used. The weight is of shape (num_heads, 1, kernel_size) + weight_dropout: the drop rate of the DropConnect to drop the weight + weight_softmax: normalize the weight with softmax before the convolution + bias: use bias + + Shape: + Input: TxBxC, i.e. (timesteps, batch_size, input_size) + Output: TxBxC, i.e. (timesteps, batch_size, input_size) + + Attributes: + weight: the learnable weights of the module of shape + `(num_heads, 1, kernel_size)` + bias: the learnable bias of the module of shape `(input_size)` + """ + + def __init__( + self, + input_size, + kernel_size=1, + padding_l=None, + num_heads=1, + weight_dropout=0.0, + weight_softmax=False, + bias=False, + ): + super().__init__() + self.input_size = input_size + self.kernel_size = kernel_size + self.padding_l = padding_l + self.num_heads = num_heads + self.weight_dropout_module = FairseqDropout( + weight_dropout, module_name=self.__class__.__name__ + ) + self.weight_softmax = weight_softmax + + self.weight = nn.Parameter(torch.Tensor(num_heads, 1, kernel_size)) + if bias: + self.bias = nn.Parameter(torch.Tensor(input_size)) + else: + self.bias = None + + self.reset_parameters() + self.onnx_trace = False + + def reset_parameters(self): + nn.init.xavier_uniform_(self.weight) + if self.bias is not None: + nn.init.constant_(self.bias, 0.0) + + def forward(self, x, incremental_state=None, unfold=False): + """Assuming the input, x, of the shape T x B x C and producing an output in the shape T x B x C + args: + x: Input of shape T x B x C, i.e. (timesteps, batch_size, input_size) + incremental_state: A dict to keep the state + unfold: unfold the input or not. If not, we use the matrix trick instead + """ + unfold = unfold or (incremental_state is not None) + + if unfold: + output = self._forward_unfolded(x, incremental_state) + else: + output = self._forward_expanded(x, incremental_state) + + if self.bias is not None: + output = output + self.bias.view(1, 1, -1) + return output + + def prepare_for_onnx_export_(self): + self.onnx_trace = True + + def _forward_unfolded(self, x, incremental_state): + """The conventional implementation of convolutions. + Unfolding the input by having a window shifting to the right.""" + T, B, C = x.size() + K, H = self.kernel_size, self.num_heads + R = C // H + assert R * H == C == self.input_size + + weight = self.weight.view(H, K) + if incremental_state is not None: + input_buffer = self._get_input_buffer(incremental_state) + if input_buffer is None: + input_buffer = x.new() + x_unfold = torch.cat([input_buffer, x.unsqueeze(3)], dim=3) + if self.kernel_size > 1: + self._set_input_buffer( + incremental_state, x_unfold[:, :, :, -self.kernel_size + 1 :] + ) + x_unfold = x_unfold.view(T * B * H, R, -1) + else: + # unfold the input: T x B x C --> T' x B x C x K + x_unfold = unfold1d(x, self.kernel_size, self.padding_l, 0) + x_unfold = x_unfold.view(T * B * H, R, K) + + if self.weight_softmax: + weight = utils.softmax(weight, dim=1, onnx_trace=self.onnx_trace).type_as( + weight + ) + + if incremental_state is not None: + weight = weight[:, -x_unfold.size(2) :] + K = weight.size(1) + + weight = ( + weight.view(1, H, K).expand(T * B, H, K).contiguous().view(T * B * H, K, 1) + ) + + weight = self.weight_dropout_module(weight) + output = torch.bmm(x_unfold, weight) # T*B*H x R x 1 + output = output.view(T, B, C) + return output + + def _forward_expanded(self, x, incremental_state): + """Turn the convolution filters into band matrices and do matrix multiplication. + This is faster when the sequence is short, but less memory efficient. + This is not used in the decoder during inference. + """ + T, B, C = x.size() + K, H = self.kernel_size, self.num_heads + R = C // H + assert R * H == C == self.input_size + + weight = self.weight.view(H, K) + if self.weight_softmax: + weight = utils.softmax(weight, dim=1, onnx_trace=self.onnx_trace).type_as( + weight + ) + weight = weight.view(1, H, K).expand(T * B, H, K).contiguous() + weight = weight.view(T, B * H, K).transpose(0, 1) + + x = x.view(T, B * H, R).transpose(0, 1) + P = self.padding_l + if K > T and P == K - 1: + weight = weight.narrow(2, K - T, T) + K, P = T, T - 1 + # turn the convolution filters into band matrices + weight_expanded = weight.new_zeros(B * H, T, T + K - 1, requires_grad=False) + weight_expanded.as_strided((B * H, T, K), (T * (T + K - 1), T + K, 1)).copy_( + weight + ) + weight_expanded = weight_expanded.narrow(2, P, T) + weight_expanded = self.weight_dropout_module(weight_expanded) + + output = torch.bmm(weight_expanded, x) + output = output.transpose(0, 1).contiguous().view(T, B, C) + return output + + def reorder_incremental_state(self, incremental_state, new_order): + input_buffer = self._get_input_buffer(incremental_state) + if input_buffer is not None: + input_buffer = input_buffer.index_select(1, new_order) + self._set_input_buffer(incremental_state, input_buffer) + + def _get_input_buffer(self, incremental_state): + return utils.get_incremental_state(self, incremental_state, "input_buffer") + + def _set_input_buffer(self, incremental_state, new_buffer): + return utils.set_incremental_state( + self, incremental_state, "input_buffer", new_buffer + ) + + def extra_repr(self): + s = "{}, kernel_size={}, padding_l={}, num_heads={}, weight_softmax={}, bias={}".format( + self.input_size, + self.kernel_size, + self.padding_l, + self.num_heads, + self.weight_softmax, + self.bias is not None, + ) + if self.weight_dropout_module.p > 0.0: + s += ", weight_dropout={}".format(self.weight_dropout_module.p) + return s diff --git a/data/fairseq/fairseq/modules/linearized_convolution.py b/data/fairseq/fairseq/modules/linearized_convolution.py new file mode 100644 index 0000000000000000000000000000000000000000..1c7a9f09acd734eaee7a4c825eb99c4357f9cd05 --- /dev/null +++ b/data/fairseq/fairseq/modules/linearized_convolution.py @@ -0,0 +1,125 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import torch +import torch.nn.functional as F +from fairseq import utils +from fairseq.incremental_decoding_utils import with_incremental_state + +from .conv_tbc import ConvTBC + +from typing import Dict, Optional +from torch import Tensor + + +@with_incremental_state +class LinearizedConvolution(ConvTBC): + """An optimized version of nn.Conv1d. + + At training time, this module uses ConvTBC, which is an optimized version + of Conv1d. At inference time, it optimizes incremental generation (i.e., + one time step at a time) by replacing the convolutions with linear layers. + Note that the input order changes from training to inference. + """ + + def __init__(self, in_channels, out_channels, kernel_size, **kwargs): + super().__init__(in_channels, out_channels, kernel_size, **kwargs) + self._linearized_weight = None + self.register_backward_hook(self._clear_linearized_weight) + + def state_dict(self, destination=None, prefix="", keep_vars=False): + state = ConvTBC.state_dict(self, destination, prefix, keep_vars=keep_vars) + # don't store redundant _linearized_weight in checkpoints + if prefix + "_linearized_weight" in state: + del state[prefix + "_linearized_weight"] + return state + + def upgrade_state_dict_named(self, state_dict, name): + prefix = name + "." if name != "" else "" + if prefix + "_linearized_weight" in state_dict: + del state_dict[prefix + "_linearized_weight"] + + @torch.jit.export + def forward( + self, + input, + incremental_state: Optional[Dict[str, Dict[str, Optional[Tensor]]]] = None, + ): + """ + Args: + incremental_state: Used to buffer signal; if not None, then input is + expected to contain a single frame. If the input order changes + between time steps, call reorder_incremental_state. + Input: + Time x Batch x Channel during training + Batch x Time x Channel during inference + """ + if incremental_state is None: + output = self.conv_tbc(input) + if self.kernel_size[0] > 1 and self.padding[0] > 0: + # remove future timesteps added by padding + output = output[: -self.padding[0], :, :] + return output + + # reshape weight + weight = self._get_linearized_weight() + kw = self.kernel_size[0] + + bsz = input.size(0) # input: bsz x len x dim + if kw > 1: + input = input.data + input_buffer = self._get_input_buffer(incremental_state) + if input_buffer is None: + input_buffer = input.new(bsz, kw, input.size(2)).zero_() + self._set_input_buffer(incremental_state, input_buffer) + else: + # shift buffer + input_buffer[:, :-1, :] = input_buffer[:, 1:, :].clone() + # append next input + input_buffer[:, -1, :] = input[:, -1, :] + input = input_buffer + with torch.no_grad(): + output = F.linear(input.view(bsz, -1), weight, self.bias) + return output.view(bsz, 1, -1) + + @torch.jit.unused + def reorder_incremental_state( + self, + incremental_state: Optional[Dict[str, Dict[str, Optional[Tensor]]]], + new_order, + ): + input_buffer = self._get_input_buffer(incremental_state) + if input_buffer is not None: + input_buffer = input_buffer.index_select(0, new_order) + self._set_input_buffer(incremental_state, input_buffer) + + @torch.jit.unused + def _get_input_buffer( + self, incremental_state: Optional[Dict[str, Dict[str, Optional[Tensor]]]] + ): + return utils.get_incremental_state(self, incremental_state, "input_buffer") + + @torch.jit.unused + def _set_input_buffer( + self, + incremental_state: Optional[Dict[str, Dict[str, Optional[Tensor]]]], + new_buffer, + ): + return utils.set_incremental_state( + self, incremental_state, "input_buffer", new_buffer + ) + + @torch.jit.unused + def _get_linearized_weight(self): + if self._linearized_weight is None: + kw = self.kernel_size[0] + weight = self.weight.transpose(2, 1).transpose(1, 0).contiguous() + assert weight.size() == (self.out_channels, kw, self.in_channels) + return weight.view(self.out_channels, -1) + return self._linearized_weight + + @torch.jit.unused + def _clear_linearized_weight(self, *args): + self._linearized_weight = None diff --git a/data/fairseq/fairseq/modules/location_attention.py b/data/fairseq/fairseq/modules/location_attention.py new file mode 100644 index 0000000000000000000000000000000000000000..dbbbfb9f2d4b64a2ccc0281cc1679d789ef8374f --- /dev/null +++ b/data/fairseq/fairseq/modules/location_attention.py @@ -0,0 +1,83 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import torch.nn as nn +import torch +import torch.nn.functional as F + + +class LocationAttention(nn.Module): + """ + Attention-Based Models for Speech Recognition + https://arxiv.org/pdf/1506.07503.pdf + + :param int encoder_dim: # projection-units of encoder + :param int decoder_dim: # units of decoder + :param int attn_dim: attention dimension + :param int conv_dim: # channels of attention convolution + :param int conv_kernel_size: filter size of attention convolution + """ + + def __init__( + self, + attn_dim, + encoder_dim, + decoder_dim, + attn_state_kernel_size, + conv_dim, + conv_kernel_size, + scaling=2.0, + ): + super(LocationAttention, self).__init__() + self.attn_dim = attn_dim + self.decoder_dim = decoder_dim + self.scaling = scaling + self.proj_enc = nn.Linear(encoder_dim, attn_dim) + self.proj_dec = nn.Linear(decoder_dim, attn_dim, bias=False) + self.proj_attn = nn.Linear(conv_dim, attn_dim, bias=False) + self.conv = nn.Conv1d( + attn_state_kernel_size, + conv_dim, + 2 * conv_kernel_size + 1, + padding=conv_kernel_size, + bias=False, + ) + self.proj_out = nn.Sequential(nn.Tanh(), nn.Linear(attn_dim, 1)) + + self.proj_enc_out = None # cache + + def clear_cache(self): + self.proj_enc_out = None + + def forward(self, encoder_out, encoder_padding_mask, decoder_h, attn_state): + """ + :param torch.Tensor encoder_out: padded encoder hidden state B x T x D + :param torch.Tensor encoder_padding_mask: encoder padding mask + :param torch.Tensor decoder_h: decoder hidden state B x D + :param torch.Tensor attn_prev: previous attention weight B x K x T + :return: attention weighted encoder state (B, D) + :rtype: torch.Tensor + :return: previous attention weights (B x T) + :rtype: torch.Tensor + """ + bsz, seq_len, _ = encoder_out.size() + if self.proj_enc_out is None: + self.proj_enc_out = self.proj_enc(encoder_out) + + # B x K x T -> B x C x T + attn = self.conv(attn_state) + # B x C x T -> B x T x C -> B x T x D + attn = self.proj_attn(attn.transpose(1, 2)) + + if decoder_h is None: + decoder_h = encoder_out.new_zeros(bsz, self.decoder_dim) + dec_h = self.proj_dec(decoder_h).view(bsz, 1, self.attn_dim) + + out = self.proj_out(attn + self.proj_enc_out + dec_h).squeeze(2) + out.masked_fill_(encoder_padding_mask, -float("inf")) + + w = F.softmax(self.scaling * out, dim=1) + c = torch.sum(encoder_out * w.view(bsz, seq_len, 1), dim=1) + return c, w diff --git a/data/fairseq/fairseq/modules/lstm_cell_with_zoneout.py b/data/fairseq/fairseq/modules/lstm_cell_with_zoneout.py new file mode 100644 index 0000000000000000000000000000000000000000..273308951f7b3d1924675fde4a62359602a11bca --- /dev/null +++ b/data/fairseq/fairseq/modules/lstm_cell_with_zoneout.py @@ -0,0 +1,37 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import torch.nn as nn + + +class LSTMCellWithZoneOut(nn.Module): + """ + Zoneout: Regularizing RNNs by Randomly Preserving Hidden Activations + https://arxiv.org/abs/1606.01305 + """ + + def __init__( + self, prob: float, input_size: int, hidden_size: int, bias: bool = True + ): + super(LSTMCellWithZoneOut, self).__init__() + self.lstm_cell = nn.LSTMCell(input_size, hidden_size, bias=bias) + self.prob = prob + if prob > 1.0 or prob < 0.0: + raise ValueError( + "zoneout probability must be in the range from " "0.0 to 1.0." + ) + + def zoneout(self, h, next_h, prob): + if isinstance(h, tuple): + return tuple([self.zoneout(h[i], next_h[i], prob) for i in range(len(h))]) + + if self.training: + mask = h.new_zeros(*h.size()).bernoulli_(prob) + return mask * h + (1 - mask) * next_h + + return prob * h + (1 - prob) * next_h + + def forward(self, x, h): + return self.zoneout(h, self.lstm_cell(x, h), self.prob) diff --git a/data/fairseq/fairseq/modules/multihead_attention.py b/data/fairseq/fairseq/modules/multihead_attention.py new file mode 100644 index 0000000000000000000000000000000000000000..262132dfe7575d99705048dd6c6092505eb6898e --- /dev/null +++ b/data/fairseq/fairseq/modules/multihead_attention.py @@ -0,0 +1,910 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import math +from typing import Dict, List, Optional, Tuple + +import torch +import torch.nn.functional as F +from torch import Tensor, nn +from torch.nn import Parameter + +try: + from xformers.components.attention import build_attention + from xformers.components.attention.utils import maybe_merge_masks + + _xformers_available = True +except ImportError: + _xformers_available = False + +from fairseq import utils +from fairseq.modules.fairseq_dropout import FairseqDropout +from fairseq.modules.quant_noise import quant_noise +from fairseq.models.fairseq_incremental_decoder import FairseqIncrementalDecoder + + +# TODO: move this into xformers? +# TODO: uint8 input type should just output a bool +def _mask_for_xformers(mask: Tensor, to_dtype: Optional[torch.dtype] = None): + """ + call to pytorch multihead accepts three mask types: + - ByteTensor where non-zero means to mask + - FloatTensor which is an additive mask + - BoolTensor where True means to mask + xFormers currently accepts boolean and additive maks. For boolean masks + the values have opposite meaning. For a BoolTensor True mean to keep the value. + """ + float_types = [torch.float, torch.float16] + # If an input mask is a float it is an additive mask. Otherwise it is either uint8 or bool. + additive = mask.dtype in float_types + # If to_dype is not specified, keep same dtype as mask. + to_dtype = mask.dtype if to_dtype is None else to_dtype + to_additive = to_dtype in float_types + + if additive: + if to_additive: + return mask.to(to_dtype) + mask = mask < 0 + + if to_additive: + # return additive mask + new_mask = torch.zeros_like(mask, dtype=to_dtype) + new_mask = new_mask.masked_fill_(mask, -float("inf")) + return new_mask + + # In xFormers True is value to keep rather than value to mask + mask = ~mask.to(torch.bool) + mask = mask.to(to_dtype) + return mask + + +class MultiheadAttention(FairseqIncrementalDecoder): + """Multi-headed attention. + + See "Attention Is All You Need" for more details. + """ + + def __init__( + self, + embed_dim, + num_heads, + kdim=None, + vdim=None, + dropout=0.0, + bias=True, + add_bias_kv=False, + add_zero_attn=False, + self_attention=False, + encoder_decoder_attention=False, + dictionary=None, + q_noise=0.0, + qn_block_size=8, + # TODO: pass in config rather than string. + # config defined in xformers.components.attention.AttentionConfig + xformers_att_config: Optional[str] = None, + xformers_blocksparse_layout: Optional[ + torch.Tensor + ] = None, # This should be part of the config + xformers_blocksparse_blocksize: Optional[ + int + ] = 16, # This should be part of the config + ): + super().__init__(dictionary) + + xformers_att_config = utils.eval_str_dict(xformers_att_config) + self.use_xformers = xformers_att_config is not None + if self.use_xformers and not _xformers_available: + raise ImportError("\n\n Please install xFormers.") + self.embed_dim = embed_dim + self.kdim = kdim if kdim is not None else embed_dim + self.vdim = vdim if vdim is not None else embed_dim + self.qkv_same_dim = self.kdim == embed_dim and self.vdim == embed_dim + + self.num_heads = num_heads + self.dropout_module = FairseqDropout( + dropout, module_name=self.__class__.__name__ + ) + + self.head_dim = embed_dim // num_heads + assert ( + self.head_dim * num_heads == self.embed_dim + ), "embed_dim must be divisible by num_heads" + self.scaling = self.head_dim**-0.5 + + self.self_attention = self_attention + self.encoder_decoder_attention = encoder_decoder_attention + + assert not self.self_attention or self.qkv_same_dim, ( + "Self-attention requires query, key and " "value to be of the same size" + ) + + self.k_proj = quant_noise( + nn.Linear(self.kdim, embed_dim, bias=bias), q_noise, qn_block_size + ) + self.v_proj = quant_noise( + nn.Linear(self.vdim, embed_dim, bias=bias), q_noise, qn_block_size + ) + self.q_proj = quant_noise( + nn.Linear(embed_dim, embed_dim, bias=bias), q_noise, qn_block_size + ) + + self.out_proj = quant_noise( + nn.Linear(embed_dim, embed_dim, bias=bias), q_noise, qn_block_size + ) + + if add_bias_kv: + self.bias_k = Parameter(torch.Tensor(1, 1, embed_dim)) + self.bias_v = Parameter(torch.Tensor(1, 1, embed_dim)) + else: + self.bias_k = self.bias_v = None + + self.add_zero_attn = add_zero_attn + self.beam_size = 1 + self.reset_parameters() + + if self.use_xformers: + xformers_att_config["dropout"] = xformers_att_config.get("dropout", dropout) + xformers_att_config["num_heads"] = xformers_att_config.get( + "num_heads", num_heads + ) + + if xformers_blocksparse_layout is not None: + # Could be part of a single config passed only once + xformers_att_config["block_size"] = xformers_blocksparse_blocksize + xformers_att_config["layout"] = xformers_blocksparse_layout + xformers_att_config["name"] = "blocksparse" + + self.attention = build_attention(xformers_att_config) + + self.onnx_trace = False + self.skip_embed_dim_check = False + self.init_incremental_state() + + def prepare_for_onnx_export_(self): + self.onnx_trace = True + + def reset_parameters(self): + if self.qkv_same_dim: + # Empirically observed the convergence to be much better with + # the scaled initialization + nn.init.xavier_uniform_(self.k_proj.weight, gain=1 / math.sqrt(2)) + nn.init.xavier_uniform_(self.v_proj.weight, gain=1 / math.sqrt(2)) + nn.init.xavier_uniform_(self.q_proj.weight, gain=1 / math.sqrt(2)) + else: + nn.init.xavier_uniform_(self.k_proj.weight) + nn.init.xavier_uniform_(self.v_proj.weight) + nn.init.xavier_uniform_(self.q_proj.weight) + + nn.init.xavier_uniform_(self.out_proj.weight) + if self.out_proj.bias is not None: + nn.init.constant_(self.out_proj.bias, 0.0) + if self.bias_k is not None: + nn.init.xavier_normal_(self.bias_k) + if self.bias_v is not None: + nn.init.xavier_normal_(self.bias_v) + + def _get_reserve_head_index(self, num_heads_to_keep: int): + k_proj_heads_norm = [] + q_proj_heads_norm = [] + v_proj_heads_norm = [] + + for i in range(self.num_heads): + start_idx = i * self.head_dim + end_idx = (i + 1) * self.head_dim + k_proj_heads_norm.append( + torch.sum( + torch.abs( + self.k_proj.weight[ + start_idx:end_idx, + ] + ) + ).tolist() + + torch.sum(torch.abs(self.k_proj.bias[start_idx:end_idx])).tolist() + ) + q_proj_heads_norm.append( + torch.sum( + torch.abs( + self.q_proj.weight[ + start_idx:end_idx, + ] + ) + ).tolist() + + torch.sum(torch.abs(self.q_proj.bias[start_idx:end_idx])).tolist() + ) + v_proj_heads_norm.append( + torch.sum( + torch.abs( + self.v_proj.weight[ + start_idx:end_idx, + ] + ) + ).tolist() + + torch.sum(torch.abs(self.v_proj.bias[start_idx:end_idx])).tolist() + ) + + heads_norm = [] + for i in range(self.num_heads): + heads_norm.append( + k_proj_heads_norm[i] + q_proj_heads_norm[i] + v_proj_heads_norm[i] + ) + + sorted_head_index = sorted( + range(self.num_heads), key=lambda k: heads_norm[k], reverse=True + ) + reserve_head_index = [] + for i in range(num_heads_to_keep): + start = sorted_head_index[i] * self.head_dim + end = (sorted_head_index[i] + 1) * self.head_dim + reserve_head_index.append((start, end)) + return reserve_head_index + + def _adaptive_prune_heads(self, reserve_head_index: List[Tuple[int, int]]): + new_q_weight = [] + new_q_bias = [] + new_k_weight = [] + new_k_bias = [] + new_v_weight = [] + new_v_bias = [] + new_out_proj_weight = [] + + for ele in reserve_head_index: + start_idx, end_idx = ele + new_q_weight.append( + self.q_proj.weight[ + start_idx:end_idx, + ] + ) + new_q_bias.append(self.q_proj.bias[start_idx:end_idx]) + + new_k_weight.append( + self.k_proj.weight[ + start_idx:end_idx, + ] + ) + + new_k_bias.append(self.k_proj.bias[start_idx:end_idx]) + + new_v_weight.append( + self.v_proj.weight[ + start_idx:end_idx, + ] + ) + new_v_bias.append(self.v_proj.bias[start_idx:end_idx]) + + new_out_proj_weight.append(self.out_proj.weight[:, start_idx:end_idx]) + + new_q_weight = torch.cat(new_q_weight).detach() + new_k_weight = torch.cat(new_k_weight).detach() + new_v_weight = torch.cat(new_v_weight).detach() + new_out_proj_weight = torch.cat(new_out_proj_weight, dim=-1).detach() + new_q_weight.requires_grad = True + new_k_weight.requires_grad = True + new_v_weight.requires_grad = True + new_out_proj_weight.requires_grad = True + + new_q_bias = torch.cat(new_q_bias).detach() + new_q_bias.requires_grad = True + + new_k_bias = torch.cat(new_k_bias).detach() + new_k_bias.requires_grad = True + + new_v_bias = torch.cat(new_v_bias).detach() + new_v_bias.requires_grad = True + + self.q_proj.weight = torch.nn.Parameter(new_q_weight) + self.q_proj.bias = torch.nn.Parameter(new_q_bias) + + self.k_proj.weight = torch.nn.Parameter(new_k_weight) + self.k_proj.bias = torch.nn.Parameter(new_k_bias) + + self.v_proj.weight = torch.nn.Parameter(new_v_weight) + self.v_proj.bias = torch.nn.Parameter(new_v_bias) + + self.out_proj.weight = torch.nn.Parameter(new_out_proj_weight) + + self.num_heads = len(reserve_head_index) + self.embed_dim = self.head_dim * self.num_heads + self.q_proj.out_features = self.embed_dim + self.k_proj.out_features = self.embed_dim + self.v_proj.out_features = self.embed_dim + + def _set_skip_embed_dim_check(self): + self.skip_embed_dim_check = True + + def _pad_masks( + self, + key_padding_mask: Optional[Tensor], + attn_mask: Optional[Tensor], + ) -> Tuple[Optional[Tensor], Optional[Tensor]]: + if attn_mask is not None: + shape = attn_mask.size()[:-1] + torch.Size([1]) + attn_mask = torch.cat([attn_mask, attn_mask.new_zeros(shape)], dim=-1) + if key_padding_mask is not None: + shape = key_padding_mask.size()[:-1] + torch.Size([1]) + key_padding_mask = torch.cat( + [ + key_padding_mask, + key_padding_mask.new_zeros(shape), + ], + dim=-1, + ) + return key_padding_mask, attn_mask + + def _add_bias( + self, + k: Tensor, + v: Tensor, + key_padding_mask: Optional[Tensor], + attn_mask: Optional[Tensor], + bsz: int, + ) -> Tuple[Tensor, Tensor, Optional[Tensor], Optional[Tensor]]: + assert self.bias_k is not None + assert self.bias_v is not None + k = torch.cat([k, self.bias_k.repeat(1, bsz, 1)]) + v = torch.cat([v, self.bias_v.repeat(1, bsz, 1)]) + key_padding_mask, attn_mask = self._pad_masks( + key_padding_mask=key_padding_mask, attn_mask=attn_mask + ) + return k, v, key_padding_mask, attn_mask + + def _append_zero_attn( + self, + k: Tensor, + v: Tensor, + key_padding_mask: Optional[Tensor], + attn_mask: Optional[Tensor], + ) -> Tuple[Tensor, Tensor, Optional[Tensor], Optional[Tensor]]: + zero_attn_shape = k.size()[:-2] + torch.Size([1]) + k.size()[-1:] + k = torch.cat( + [k, torch.zeros(zero_attn_shape, dtype=k.dtype, device=k.device)], dim=-2 + ) + v = torch.cat( + [v, torch.zeros(zero_attn_shape, dtype=v.dtype, device=v.device)], dim=-2 + ) + key_padding_mask, attn_mask = self._pad_masks( + key_padding_mask=key_padding_mask, attn_mask=attn_mask + ) + return k, v, key_padding_mask, attn_mask + + def _xformers_attn_forward( + self, + query, + key: Optional[Tensor], + value: Optional[Tensor], + key_padding_mask: Optional[Tensor] = None, + need_weights: bool = True, + attn_mask: Optional[Tensor] = None, + ) -> Tuple[Tensor, Optional[Tensor]]: + + tgt_len, bsz, embed_dim = query.size() + + if key_padding_mask is not None: + assert key_padding_mask.size(0) == bsz + assert key_padding_mask.size(1) == tgt_len + + if self.self_attention: + key = query + value = query + elif self.encoder_decoder_attention: + value = key + + q = self.q_proj(query) + k = self.k_proj(key) + v = self.v_proj(value) + + if self.bias_k is not None: + assert self.bias_v is not None + k, v, attn_mask, key_padding_mask = self._add_bias( + k, v, attn_mask, key_padding_mask, bsz + ) + + def fold_heads(x): + return ( + x.contiguous() + .view(-1, bsz * self.num_heads, self.head_dim) + .transpose(0, 1) + ) + + def split_heads(x): + return ( + x.contiguous() + .view(-1, bsz, self.num_heads, self.head_dim) + .transpose(0, 1) + .transpose(1, 2) + ) + + massage = split_heads if self.attention.requires_head_dimension else fold_heads + q = massage(q) + if k is not None: + k = massage(k) + if v is not None: + v = massage(v) + + if self.add_zero_attn: + k, v, key_padding_mask, attn_mask = self._append_zero_attn( + k=k, v=v, key_padding_mask=key_padding_mask, attn_mask=attn_mask + ) + + kwargs = {} + + if attn_mask is not None and self.attention.supports_attention_mask: + attn_mask = _mask_for_xformers(attn_mask, to_dtype=q.dtype) + kwargs["att_mask"] = attn_mask + + if key_padding_mask is not None: + key_padding_mask = _mask_for_xformers(key_padding_mask, to_dtype=torch.bool) + if not self.attention.requires_separate_masks: + attn_mask = maybe_merge_masks( + attn_mask, + key_padding_mask, + batch_size=bsz, + src_len=k.size(-2), + tgt_len=q.size(-2), + num_heads=self.num_heads, + ) + key_padding_mask = None + kwargs["att_mask"] = attn_mask + if self.attention.supports_key_padding_mask: + kwargs["key_padding_mask"] = key_padding_mask + + y = self.attention(q, k, v, **kwargs) + + y = ( + y.view(bsz, self.num_heads, tgt_len, self.head_dim) + .transpose(1, 2) + .flatten(start_dim=2, end_dim=3) + .transpose(0, 1) + ) + assert list(y.size()) == [tgt_len, bsz, embed_dim] + + # Dropout not needed because already applied in attention. + # It is applied to the attention weights before matmul with v. + y = self.out_proj(y) + + # TODO: support returning attention weights if needed. + return y, None + + def forward( + self, + query: Tensor, + key: Optional[Tensor], + value: Optional[Tensor], + key_padding_mask: Optional[Tensor] = None, + incremental_state: Optional[Dict[str, Dict[str, Optional[Tensor]]]] = None, + need_weights: bool = True, + static_kv: bool = False, + attn_mask: Optional[Tensor] = None, + before_softmax: bool = False, + need_head_weights: bool = False, + ) -> Tuple[Tensor, Optional[Tensor]]: + """Input shape: Time x Batch x Channel + + Args: + key_padding_mask (ByteTensor, optional): mask to exclude + keys that are pads, of shape `(batch, src_len)`, where + padding elements are indicated by 1s. + need_weights (bool, optional): return the attention weights, + averaged over heads (default: False). + attn_mask (ByteTensor, optional): typically used to + implement causal attention, where the mask prevents the + attention from looking forward in time (default: None). + before_softmax (bool, optional): return the raw attention + weights and values before the attention softmax. + need_head_weights (bool, optional): return the attention + weights for each head. Implies *need_weights*. Default: + return the average attention weights over all heads. + """ + if need_head_weights: + need_weights = True + + is_tpu = query.device.type == "xla" + + tgt_len, bsz, embed_dim = query.size() + src_len = tgt_len + if not self.skip_embed_dim_check: + assert ( + embed_dim == self.embed_dim + ), f"query dim {embed_dim} != {self.embed_dim}" + assert list(query.size()) == [tgt_len, bsz, embed_dim] + if key is not None: + src_len, key_bsz, _ = key.size() + if not torch.jit.is_scripting(): + assert value is not None + assert src_len, key_bsz == value.shape[:2] + + if ( + not self.onnx_trace + and not is_tpu # don't use PyTorch version on TPUs + and incremental_state is None + and not static_kv + # A workaround for quantization to work. Otherwise JIT compilation + # treats bias in linear module as method. + and not torch.jit.is_scripting() + # The Multihead attention implemented in pytorch forces strong dimension check + # for input embedding dimention and K,Q,V projection dimension. + # Since pruning will break the dimension check and it is not easy to modify the pytorch API, + # it is preferred to bypass the pytorch MHA when we need to skip embed_dim_check + and not self.skip_embed_dim_check + ): + assert key is not None and value is not None + + if self.use_xformers: + return self._xformers_attn_forward( + query, key, value, key_padding_mask, need_weights, attn_mask + ) + + else: + return F.multi_head_attention_forward( + query, + key, + value, + self.embed_dim, + self.num_heads, + torch.empty([0]), + torch.cat((self.q_proj.bias, self.k_proj.bias, self.v_proj.bias)), + self.bias_k, + self.bias_v, + self.add_zero_attn, + self.dropout_module.p, + self.out_proj.weight, + self.out_proj.bias, + self.training or self.dropout_module.apply_during_inference, + key_padding_mask.bool() if key_padding_mask is not None else None, + need_weights, + attn_mask, + use_separate_proj_weight=True, + q_proj_weight=self.q_proj.weight, + k_proj_weight=self.k_proj.weight, + v_proj_weight=self.v_proj.weight, + ) + + if incremental_state is not None: + saved_state = self._get_input_buffer(incremental_state) + if saved_state is not None and "prev_key" in saved_state: + # previous time steps are cached - no need to recompute + # key and value if they are static + if static_kv: + assert self.encoder_decoder_attention and not self.self_attention + key = value = None + else: + saved_state = None + + if self.self_attention: + q = self.q_proj(query) + k = self.k_proj(query) + v = self.v_proj(query) + elif self.encoder_decoder_attention: + # encoder-decoder attention + q = self.q_proj(query) + if key is None: + assert value is None + k = v = None + else: + if self.beam_size > 1 and bsz == key.size(1): + # key is [T, bsz*beam_size, C], reduce to [T, bsz, C] + key = key.view(key.size(0), -1, self.beam_size, key.size(2))[ + :, :, 0, : + ] + if key_padding_mask is not None: + key_padding_mask = key_padding_mask.view( + -1, self.beam_size, key_padding_mask.size(1) + )[:, 0, :] + k = self.k_proj(key) + v = self.v_proj(key) + + else: + assert key is not None and value is not None + q = self.q_proj(query) + k = self.k_proj(key) + v = self.v_proj(value) + q *= self.scaling + + if self.bias_k is not None: + assert self.bias_v is not None + k, v, attn_mask, key_padding_mask = self._add_bias( + k, v, attn_mask, key_padding_mask, bsz + ) + + q = ( + q.contiguous() + .view(tgt_len, bsz * self.num_heads, self.head_dim) + .transpose(0, 1) + ) + kv_bsz = bsz # need default value for scripting + if k is not None: + kv_bsz = k.size(1) + k = ( + k.contiguous() + .view(-1, kv_bsz * self.num_heads, self.head_dim) + .transpose(0, 1) + ) + if v is not None: + v = ( + v.contiguous() + .view(-1, kv_bsz * self.num_heads, self.head_dim) + .transpose(0, 1) + ) + + if saved_state is not None: + # saved states are stored with shape (bsz, num_heads, seq_len, head_dim) + if "prev_key" in saved_state: + _prev_key = saved_state["prev_key"] + assert _prev_key is not None + kv_bsz = _prev_key.size(0) + prev_key = _prev_key.view(kv_bsz * self.num_heads, -1, self.head_dim) + if static_kv: + k = prev_key + else: + assert k is not None + k = torch.cat([prev_key, k], dim=1) + src_len = k.size(1) + if "prev_value" in saved_state: + _prev_value = saved_state["prev_value"] + assert _prev_value is not None + assert kv_bsz == _prev_value.size(0) + prev_value = _prev_value.view( + kv_bsz * self.num_heads, -1, self.head_dim + ) + if static_kv: + v = prev_value + else: + assert v is not None + v = torch.cat([prev_value, v], dim=1) + prev_key_padding_mask: Optional[Tensor] = None + if "prev_key_padding_mask" in saved_state: + prev_key_padding_mask = saved_state["prev_key_padding_mask"] + assert k is not None and v is not None + key_padding_mask = MultiheadAttention._append_prev_key_padding_mask( + key_padding_mask=key_padding_mask, + prev_key_padding_mask=prev_key_padding_mask, + batch_size=kv_bsz, + src_len=k.size(1), + static_kv=static_kv, + ) + + saved_state["prev_key"] = k.view(kv_bsz, self.num_heads, -1, self.head_dim) + saved_state["prev_value"] = v.view( + kv_bsz, self.num_heads, -1, self.head_dim + ) + saved_state["prev_key_padding_mask"] = key_padding_mask + # In this branch incremental_state is never None + assert incremental_state is not None + incremental_state = self._set_input_buffer(incremental_state, saved_state) + assert k is not None + assert k.size(1) == src_len + + # This is part of a workaround to get around fork/join parallelism + # not supporting Optional types. + if key_padding_mask is not None and key_padding_mask.dim() == 0: + key_padding_mask = None + + if key_padding_mask is not None: + assert key_padding_mask.size(0) == kv_bsz + assert key_padding_mask.size(1) == src_len + + if self.add_zero_attn: + assert v is not None + src_len += 1 + k, v, key_padding_mask, attn_mask = self._append_zero_attn( + k=k, v=v, key_padding_mask=key_padding_mask, attn_mask=attn_mask + ) + + if self.encoder_decoder_attention and bsz != kv_bsz: + attn_weights = torch.einsum( + "bxhtd,bhsd->bxhts", + q.view((kv_bsz, -1, self.num_heads) + q.size()[1:]), + k.view((kv_bsz, self.num_heads) + k.size()[1:]), + ) + attn_weights = attn_weights.reshape((-1,) + attn_weights.size()[-2:]) + else: + attn_weights = torch.bmm(q, k.transpose(1, 2)) + attn_weights = self.apply_sparse_mask(attn_weights, tgt_len, src_len, bsz) + + assert list(attn_weights.size()) == [bsz * self.num_heads, tgt_len, src_len] + + if attn_mask is not None: + attn_mask = attn_mask.unsqueeze(0) + if self.onnx_trace: + attn_mask = attn_mask.repeat(attn_weights.size(0), 1, 1) + attn_weights += attn_mask + + if key_padding_mask is not None: + # don't attend to padding symbols + attn_weights = attn_weights.view(bsz, self.num_heads, tgt_len, src_len) + if not is_tpu: + attn_weights = attn_weights.view( + kv_bsz, -1, self.num_heads, tgt_len, src_len + ) + attn_weights = attn_weights.masked_fill( + key_padding_mask.unsqueeze(1) + .unsqueeze(2) + .unsqueeze(3) + .to(torch.bool), + float("-inf"), + ) + else: + attn_weights = attn_weights.transpose(0, 2) + attn_weights = attn_weights.masked_fill(key_padding_mask, float("-inf")) + attn_weights = attn_weights.transpose(0, 2) + attn_weights = attn_weights.view(bsz * self.num_heads, tgt_len, src_len) + + if before_softmax: + return attn_weights, v + + attn_weights_float = utils.softmax( + attn_weights, dim=-1, onnx_trace=self.onnx_trace + ) + attn_weights = attn_weights_float.type_as(attn_weights) + attn_probs = self.dropout_module(attn_weights) + + assert v is not None + attn: Optional[Tensor] = None + if self.encoder_decoder_attention and bsz != kv_bsz: + attn = torch.einsum( + "bxhts,bhsd->bxhtd", + attn_probs.view( + ( + kv_bsz, + -1, + self.num_heads, + ) + + attn_probs.size()[1:] + ), + v.view( + ( + kv_bsz, + self.num_heads, + ) + + v.size()[1:] + ), + ) + attn = attn.reshape((-1,) + attn.size()[-2:]) + else: + attn = torch.bmm(attn_probs, v) + assert list(attn.size()) == [bsz * self.num_heads, tgt_len, self.head_dim] + if self.onnx_trace and attn.size(1) == 1: + # when ONNX tracing a single decoder step (sequence length == 1) + # the transpose is a no-op copy before view, thus unnecessary + attn = attn.contiguous().view(tgt_len, bsz, self.embed_dim) + else: + attn = attn.transpose(0, 1).contiguous().view(tgt_len, bsz, self.embed_dim) + attn = self.out_proj(attn) + attn_weights: Optional[Tensor] = None + if need_weights: + attn_weights = attn_weights_float.view( + bsz, self.num_heads, tgt_len, src_len + ).transpose(1, 0) + if not need_head_weights: + # average attention weights over heads + attn_weights = attn_weights.mean(dim=0) + + return attn, attn_weights + + @staticmethod + def _append_prev_key_padding_mask( + key_padding_mask: Optional[Tensor], + prev_key_padding_mask: Optional[Tensor], + batch_size: int, + src_len: int, + static_kv: bool, + ) -> Optional[Tensor]: + # saved key padding masks have shape (bsz, seq_len) + if prev_key_padding_mask is not None and static_kv: + new_key_padding_mask = prev_key_padding_mask + elif prev_key_padding_mask is not None and key_padding_mask is not None: + new_key_padding_mask = torch.cat( + [prev_key_padding_mask.float(), key_padding_mask.float()], dim=1 + ) + # During incremental decoding, as the padding token enters and + # leaves the frame, there will be a time when prev or current + # is None + elif prev_key_padding_mask is not None: + if src_len > prev_key_padding_mask.size(1): + filler = torch.zeros( + (batch_size, src_len - prev_key_padding_mask.size(1)), + device=prev_key_padding_mask.device, + ) + new_key_padding_mask = torch.cat( + [prev_key_padding_mask.float(), filler.float()], dim=1 + ) + else: + new_key_padding_mask = prev_key_padding_mask.float() + elif key_padding_mask is not None: + if src_len > key_padding_mask.size(1): + filler = torch.zeros( + (batch_size, src_len - key_padding_mask.size(1)), + device=key_padding_mask.device, + ) + new_key_padding_mask = torch.cat( + [filler.float(), key_padding_mask.float()], dim=1 + ) + else: + new_key_padding_mask = key_padding_mask.float() + else: + new_key_padding_mask = prev_key_padding_mask + return new_key_padding_mask + + @torch.jit.export + def reorder_incremental_state( + self, + incremental_state: Optional[Dict[str, Dict[str, Optional[Tensor]]]], + new_order: Tensor, + ): + """Reorder buffered internal state (for incremental generation).""" + input_buffer = self._get_input_buffer(incremental_state) + if input_buffer is not None: + for k in input_buffer.keys(): + input_buffer_k = input_buffer[k] + if input_buffer_k is not None: + if self.encoder_decoder_attention: + if input_buffer_k.size(0) * self.beam_size == new_order.size(0): + return incremental_state + elif self.beam_size > 1: + input_buffer[k] = input_buffer_k.index_select( + 0, + new_order.reshape(-1, self.beam_size)[:, 0] + // self.beam_size, + ) + else: + input_buffer[k] = input_buffer_k.index_select(0, new_order) + else: + input_buffer[k] = input_buffer_k.index_select(0, new_order) + incremental_state = self._set_input_buffer(incremental_state, input_buffer) + return incremental_state + + def set_beam_size(self, beam_size): + """Used for effiecient beamable enc-dec attention""" + self.beam_size = beam_size + + def _get_input_buffer( + self, incremental_state: Optional[Dict[str, Dict[str, Optional[Tensor]]]] + ) -> Dict[str, Optional[Tensor]]: + result = self.get_incremental_state(incremental_state, "attn_state") + if result is not None: + return result + else: + empty_result: Dict[str, Optional[Tensor]] = {} + return empty_result + + def _set_input_buffer( + self, + incremental_state: Optional[Dict[str, Dict[str, Optional[Tensor]]]], + buffer: Dict[str, Optional[Tensor]], + ): + return self.set_incremental_state(incremental_state, "attn_state", buffer) + + def apply_sparse_mask(self, attn_weights, tgt_len: int, src_len: int, bsz: int): + return attn_weights + + def upgrade_state_dict_named(self, state_dict, name): + prefix = name + "." if name != "" else "" + items_to_add = {} + keys_to_remove = [] + for k in state_dict.keys(): + if k.endswith(prefix + "in_proj_weight"): + # in_proj_weight used to be q + k + v with same dimensions + dim = int(state_dict[k].shape[0] / 3) + items_to_add[prefix + "q_proj.weight"] = state_dict[k][:dim] + items_to_add[prefix + "k_proj.weight"] = state_dict[k][dim : 2 * dim] + items_to_add[prefix + "v_proj.weight"] = state_dict[k][2 * dim :] + + keys_to_remove.append(k) + + k_bias = prefix + "in_proj_bias" + if k_bias in state_dict.keys(): + dim = int(state_dict[k].shape[0] / 3) + items_to_add[prefix + "q_proj.bias"] = state_dict[k_bias][:dim] + items_to_add[prefix + "k_proj.bias"] = state_dict[k_bias][ + dim : 2 * dim + ] + items_to_add[prefix + "v_proj.bias"] = state_dict[k_bias][2 * dim :] + + keys_to_remove.append(prefix + "in_proj_bias") + + for k in keys_to_remove: + del state_dict[k] + + for key, value in items_to_add.items(): + state_dict[key] = value diff --git a/data/fairseq/fairseq/modules/positional_embedding.py b/data/fairseq/fairseq/modules/positional_embedding.py new file mode 100644 index 0000000000000000000000000000000000000000..fbc13d80acc3982a7e9c39943598813975a25fc9 --- /dev/null +++ b/data/fairseq/fairseq/modules/positional_embedding.py @@ -0,0 +1,37 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import torch.nn as nn + +from .learned_positional_embedding import LearnedPositionalEmbedding +from .sinusoidal_positional_embedding import SinusoidalPositionalEmbedding + + +def PositionalEmbedding( + num_embeddings: int, + embedding_dim: int, + padding_idx: int, + learned: bool = False, + auto_expand: bool = True, +): + if learned: + # if padding_idx is specified then offset the embedding ids by + # this index and adjust num_embeddings appropriately + # TODO: The right place for this offset would be inside + # LearnedPositionalEmbedding. Move this there for a cleaner implementation. + if padding_idx is not None: + num_embeddings = num_embeddings + padding_idx + 1 + m = LearnedPositionalEmbedding(num_embeddings, embedding_dim, padding_idx) + nn.init.normal_(m.weight, mean=0, std=embedding_dim**-0.5) + if padding_idx is not None: + nn.init.constant_(m.weight[padding_idx], 0) + else: + m = SinusoidalPositionalEmbedding( + embedding_dim, + padding_idx, + init_size=num_embeddings + padding_idx + 1, + auto_expand=auto_expand, + ) + return m diff --git a/data/fairseq/fairseq/modules/positional_encoding.py b/data/fairseq/fairseq/modules/positional_encoding.py new file mode 100644 index 0000000000000000000000000000000000000000..67f635353918d7b485da3c31ec9a836dedec6b7b --- /dev/null +++ b/data/fairseq/fairseq/modules/positional_encoding.py @@ -0,0 +1,129 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import torch.nn as nn +import math +import torch + + +class PositionalEncoding(nn.Module): + """Positional encoding. + + Args: + d_model: Embedding dimension. + dropout_rate: Dropout rate. + max_len: Maximum input length. + reverse: Whether to reverse the input position. + """ + + def __init__(self, d_model, dropout_rate, max_len=5000, reverse=False): + """Construct an PositionalEncoding object.""" + super(PositionalEncoding, self).__init__() + self.d_model = d_model + self.reverse = reverse + self.xscale = math.sqrt(self.d_model) + self.dropout = nn.Dropout(p=dropout_rate) + self.pe = None + self.extend_pe(torch.tensor(0.0).expand(1, max_len)) + + def extend_pe(self, x): + """Reset the positional encodings.""" + if self.pe is not None: + if self.pe.size(1) >= x.size(1): + if self.pe.dtype != x.dtype or self.pe.device != x.device: + self.pe = self.pe.to(dtype=x.dtype, device=x.device) + return + pe = torch.zeros(x.size(1), self.d_model) + if self.reverse: + position = torch.arange( + x.size(1) - 1, -1, -1.0, dtype=torch.float32 + ).unsqueeze(1) + else: + position = torch.arange(0, x.size(1), dtype=torch.float32).unsqueeze(1) + div_term = torch.exp( + torch.arange(0, self.d_model, 2, dtype=torch.float32) + * -(math.log(10000.0) / self.d_model) + ) + pe[:, 0::2] = torch.sin(position * div_term) + pe[:, 1::2] = torch.cos(position * div_term) + pe = pe.unsqueeze(0) + self.pe = pe.to(device=x.device, dtype=x.dtype) + + def forward(self, x: torch.Tensor): + """Add positional encoding. + Args: + x (torch.Tensor): Input tensor B X T X C + Returns: + torch.Tensor: Encoded tensor B X T X C + """ + self.extend_pe(x) + x = x * self.xscale + self.pe[:, : x.size(1)] + return self.dropout(x) + + +class RelPositionalEncoding(nn.Module): + """Relative positional encoding module (new implementation). + + Args: + d_model: Embedding dimension. + dropout_rate: Dropout rate. + max_len: Maximum input length. + """ + + def __init__(self, max_len, d_model): + """Construct an PositionalEncoding object.""" + super(RelPositionalEncoding, self).__init__() + self.d_model = d_model + self.pe = None + self.extend_pe(torch.tensor(0.0).expand(1, max_len)) + + def extend_pe(self, x): + """Reset the positional encodings.""" + if self.pe is not None: + # self.pe contains both positive and negative parts + # the length of self.pe is 2 * input_len - 1 + if self.pe.size(1) >= x.size(1) * 2 - 1: + if self.pe.dtype != x.dtype or self.pe.device != x.device: + self.pe = self.pe.to(dtype=x.dtype, device=x.device) + return + # Suppose `i` means to the position of query vecotr and `j` means the + # position of key vector. We use position relative positions when keys + # are to the left (i>j) and negative relative positions otherwise (i 0: + # given an empty cluster, find most populated cluster and split it into two + k = random.choice(list(empty_clusters)) + m = counts.most_common(1)[0][0] + e = torch.randn_like(self.centroids[m]) * self.eps + self.centroids[k] = self.centroids[m].clone() + self.centroids[k] += e + self.centroids[m] -= e + + # recompute assignments + distances = self.compute_distances() # (n_centroids x out_features) + self.assignments = torch.argmin(distances, dim=0) # (out_features) + + # check for empty clusters + counts = Counter(map(lambda x: x.item(), self.assignments)) + empty_clusters = set(range(self.n_centroids)) - set(counts.keys()) + + # increment tentatives + if tentatives == self.max_tentatives: + logging.info( + f"Could not resolve all empty clusters, {len(empty_clusters)} remaining" + ) + raise EmptyClusterResolveError + tentatives += 1 + + return n_empty_clusters + + def compute_distances(self): + """ + For every centroid m, computes + + ||M - m[None, :]||_2 + + Remarks: + - We rely on PyTorch's broadcasting to speed up computations + and reduce the memory overhead + - Without chunking, the sizes in the broadcasting are modified as: + (n_centroids x n_samples x out_features) -> (n_centroids x out_features) + - The broadcasting computation is automatically chunked so that + the tensors fit into the memory of the GPU + """ + + nb_centroids_chunks = 1 + + while True: + try: + return torch.cat( + [ + (self.W[None, :, :] - centroids_c[:, :, None]).norm(p=2, dim=1) + for centroids_c in self.centroids.chunk( + nb_centroids_chunks, dim=0 + ) + ], + dim=0, + ) + except RuntimeError: + nb_centroids_chunks *= 2 + + def assign(self): + """ + Assigns each column of W to its closest centroid, thus essentially + performing the E-step in train(). + + Remarks: + - The function must be called after train() or after loading + centroids using self.load(), otherwise it will return empty tensors + """ + + distances = self.compute_distances() # (n_centroids x out_features) + self.assignments = torch.argmin(distances, dim=0) # (out_features) + + def save(self, path, layer): + """ + Saves centroids and assignments. + + Args: + - path: folder used to save centroids and assignments + """ + + torch.save(self.centroids, os.path.join(path, "{}_centroids.pth".format(layer))) + torch.save( + self.assignments, os.path.join(path, "{}_assignments.pth".format(layer)) + ) + torch.save(self.objective, os.path.join(path, "{}_objective.pth".format(layer))) + + def load(self, path, layer): + """ + Loads centroids and assignments from a given path + + Args: + - path: folder use to load centroids and assignments + """ + + self.centroids = torch.load( + os.path.join(path, "{}_centroids.pth".format(layer)) + ) + self.assignments = torch.load( + os.path.join(path, "{}_assignments.pth".format(layer)) + ) + self.objective = torch.load( + os.path.join(path, "{}_objective.pth".format(layer)) + ) + + +class EmptyClusterResolveError(Exception): + pass diff --git a/data/fairseq/fairseq/modules/quantization/pq/modules/__init__.py b/data/fairseq/fairseq/modules/quantization/pq/modules/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..b67c8e8ad691aa01e9e10e904d69d94595387668 --- /dev/null +++ b/data/fairseq/fairseq/modules/quantization/pq/modules/__init__.py @@ -0,0 +1,8 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from .qconv import PQConv2d # NOQA +from .qemb import PQEmbedding # NOQA +from .qlinear import PQLinear # NOQA diff --git a/data/fairseq/fairseq/modules/quantization/pq/modules/__pycache__/__init__.cpython-38.pyc b/data/fairseq/fairseq/modules/quantization/pq/modules/__pycache__/__init__.cpython-38.pyc new file mode 100644 index 0000000000000000000000000000000000000000..228087e99705064a57a8c6964721e600282bb1d4 Binary files /dev/null and b/data/fairseq/fairseq/modules/quantization/pq/modules/__pycache__/__init__.cpython-38.pyc differ diff --git a/data/fairseq/fairseq/modules/quantization/pq/modules/__pycache__/qconv.cpython-38.pyc b/data/fairseq/fairseq/modules/quantization/pq/modules/__pycache__/qconv.cpython-38.pyc new file mode 100644 index 0000000000000000000000000000000000000000..b9c906ea1833281878d10b3fae8e31416d144520 Binary files /dev/null and b/data/fairseq/fairseq/modules/quantization/pq/modules/__pycache__/qconv.cpython-38.pyc differ diff --git a/data/fairseq/fairseq/modules/quantization/pq/modules/__pycache__/qemb.cpython-38.pyc b/data/fairseq/fairseq/modules/quantization/pq/modules/__pycache__/qemb.cpython-38.pyc new file mode 100644 index 0000000000000000000000000000000000000000..51077e255906d918a1db28378bdd734f66ff0411 Binary files /dev/null and b/data/fairseq/fairseq/modules/quantization/pq/modules/__pycache__/qemb.cpython-38.pyc differ diff --git a/data/fairseq/fairseq/modules/quantization/pq/modules/__pycache__/qlinear.cpython-38.pyc b/data/fairseq/fairseq/modules/quantization/pq/modules/__pycache__/qlinear.cpython-38.pyc new file mode 100644 index 0000000000000000000000000000000000000000..440b4a9a17015fa8c62ffb80cf7042894df9ea0a Binary files /dev/null and b/data/fairseq/fairseq/modules/quantization/pq/modules/__pycache__/qlinear.cpython-38.pyc differ diff --git a/data/fairseq/fairseq/modules/quantization/pq/modules/qconv.py b/data/fairseq/fairseq/modules/quantization/pq/modules/qconv.py new file mode 100644 index 0000000000000000000000000000000000000000..d15ec192e8cda6265a198e583a9bf7fb194dd129 --- /dev/null +++ b/data/fairseq/fairseq/modules/quantization/pq/modules/qconv.py @@ -0,0 +1,115 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import numpy as np +import torch +import torch.nn as nn +import torch.nn.functional as F +from torch.nn.modules.utils import _pair + + +class PQConv2d(nn.Module): + """ + Quantized counterpart of nn.Conv2d module. Stores the centroid, the assignments + and the non-quantized biases. The full weight is re-instantiated at each forward + pass and autograd automatically computes the gradients with respect to the + centroids. + + Args: + - centroids: centroids of size n_centroids x block_size + - assignments: assignments of the centroids to the subvectors + of size self.out_channels x n_blocks + - bias: the non-quantized bias, must be either torch.Tensor or None + + Remarks: + - We refer the reader to the official documentation of the nn.Conv2d module + for the other arguments and the behavior of the module. + - Performance tests on GPU show that this implementation is 10% slower than + the non-quantized nn.Conv2d module for a standard training loop. + - During the backward, the gradients are averaged by cluster and not summed. + This explains the hook registered to the centroids. + """ + + def __init__( + self, + centroids, + assignments, + bias, + in_channels, + out_channels, + kernel_size, + stride=1, + padding=0, + dilation=1, + groups=1, + padding_mode="zeros", + ): + super(PQConv2d, self).__init__() + self.block_size = centroids.size(1) + self.n_centroids = centroids.size(0) + self.in_channels = in_channels + self.out_channels = out_channels + self.kernel_size = _pair(kernel_size) + self.stride = _pair(stride) + self.padding = _pair(padding) + self.dilation = _pair(dilation) + self.groups = groups + self.padding_mode = padding_mode + # check compatibility + if in_channels // groups * np.prod(self.kernel_size) % self.block_size != 0: + raise ValueError("Wrong PQ sizes") + if len(assignments) % out_channels != 0: + raise ValueError("Wrong PQ sizes") + if in_channels % groups != 0: + raise ValueError("in_channels must be divisible by groups") + if out_channels % groups != 0: + raise ValueError("out_channels must be divisible by groups") + # define parameters + self.centroids = nn.Parameter(centroids, requires_grad=True) + self.register_buffer("assignments", assignments) + self.register_buffer("counts", torch.bincount(assignments).type_as(centroids)) + if bias is not None: + self.bias = nn.Parameter(bias) + else: + self.register_parameter("bias", None) + # register hook for averaging gradients per centroids instead of summing + self.centroids.register_hook(lambda x: x / self.counts[:, None]) + + @property + def weight(self): + return ( + self.centroids[self.assignments] + .reshape(-1, self.out_channels, self.block_size) + .permute(1, 0, 2) + .reshape( + self.out_channels, self.in_channels // self.groups, *self.kernel_size + ) + ) + + def forward(self, x): + return F.conv2d( + x, + self.weight, + self.bias, + self.stride, + self.padding, + self.dilation, + self.groups, + ) + + def extra_repr(self): + s = "{in_channels}, {out_channels}, kernel_size={kernel_size}, stride={stride}" + if self.padding != (0,) * len(self.padding): + s += ", padding={padding}" + if self.dilation != (1,) * len(self.dilation): + s += ", dilation={dilation}" + if self.groups != 1: + s += ", groups={groups}" + if self.bias is None: + s += ", bias=False" + if self.padding_mode != "zeros": + s += ", padding_mode={padding_mode}" + s += ", n_centroids={n_centroids}, block_size={block_size}" + return s.format(**self.__dict__) diff --git a/data/fairseq/fairseq/modules/quantization/pq/modules/qemb.py b/data/fairseq/fairseq/modules/quantization/pq/modules/qemb.py new file mode 100644 index 0000000000000000000000000000000000000000..3a74ad3c4c7c9d3203d26e7885864ba578951bfe --- /dev/null +++ b/data/fairseq/fairseq/modules/quantization/pq/modules/qemb.py @@ -0,0 +1,107 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import torch +import torch.nn as nn +import torch.nn.functional as F + + +class PQEmbedding(nn.Module): + """ + Quantized counterpart of nn.Embedding module. Stores the centroids and + the assignments. The full weight is re-instantiated at each forward + pass. + + Args: + - centroids: centroids of size n_centroids x block_size + - assignments: assignments of the centroids to the subvectors + of size self.out_features x n_blocks + - bias: the non-quantized bias + + Remarks: + - We refer the reader to the official documentation of the nn.Embedding module + for the other arguments and the behavior of the module + - Performance tests on GPU show that this implementation is 10% slower than + the non-quantized nn.Embedding module for a standard training loop. + """ + + def __init__( + self, + centroids, + assignments, + num_embeddings, + embedding_dim, + padding_idx=None, + max_norm=None, + norm_type=2.0, + scale_grad_by_freq=False, + sparse=False, + _weight=None, + ): + super(PQEmbedding, self).__init__() + self.block_size = centroids.size(1) + self.n_centroids = centroids.size(0) + self.num_embeddings = num_embeddings + self.embedding_dim = embedding_dim + if padding_idx is not None: + if padding_idx > 0: + assert ( + padding_idx < self.num_embeddings + ), "Padding_idx must be within num_embeddings" + elif padding_idx < 0: + assert ( + padding_idx >= -self.num_embeddings + ), "Padding_idx must be within num_embeddings" + padding_idx = self.num_embeddings + padding_idx + self.padding_idx = padding_idx + self.max_norm = max_norm + self.norm_type = norm_type + self.scale_grad_by_freq = scale_grad_by_freq + self.sparse = sparse + # check compatibility + if self.embedding_dim % self.block_size != 0: + raise ValueError("Wrong PQ sizes") + if len(assignments) % self.num_embeddings != 0: + raise ValueError("Wrong PQ sizes") + # define parameters + self.centroids = nn.Parameter(centroids, requires_grad=True) + self.register_buffer("assignments", assignments) + self.register_buffer("counts", torch.bincount(assignments).type_as(centroids)) + + @property + def weight(self): + return ( + self.centroids[self.assignments] + .reshape(-1, self.num_embeddings, self.block_size) + .permute(1, 0, 2) + .flatten(1, 2) + ) + + def forward(self, input): + return F.embedding( + input, + self.weight, + self.padding_idx, + self.max_norm, + self.norm_type, + self.scale_grad_by_freq, + self.sparse, + ) + + def extra_repr(self): + s = "{num_embeddings}, {embedding_dim}" + if self.padding_idx is not None: + s += ", padding_idx={padding_idx}" + if self.max_norm is not None: + s += ", max_norm={max_norm}" + if self.norm_type != 2: + s += ", norm_type={norm_type}" + if self.scale_grad_by_freq is not False: + s += ", scale_grad_by_freq={scale_grad_by_freq}" + if self.sparse is not False: + s += ", sparse=True" + s += ", n_centroids={n_centroids}, block_size={block_size}" + + return s.format(**self.__dict__) diff --git a/data/fairseq/fairseq/modules/quantization/pq/modules/qlinear.py b/data/fairseq/fairseq/modules/quantization/pq/modules/qlinear.py new file mode 100644 index 0000000000000000000000000000000000000000..9bdd25a8685bb7c7b32e1f02372aaeb26d8ba53a --- /dev/null +++ b/data/fairseq/fairseq/modules/quantization/pq/modules/qlinear.py @@ -0,0 +1,71 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import torch +import torch.nn as nn +import torch.nn.functional as F + + +class PQLinear(nn.Module): + """ + Quantized counterpart of nn.Linear module. Stores the centroid, the assignments + and the non-quantized biases. The full weight is re-instantiated at each forward + pass. + + Args: + - centroids: centroids of size n_centroids x block_size + - assignments: assignments of the centroids to the subvectors + of size self.out_features x n_blocks + - bias: the non-quantized bias + + Remarks: + - We refer the reader to the official documentation of the nn.Linear module + for the other arguments and the behavior of the module + - Performance tests on GPU show that this implementation is 15% slower than + the non-quantized nn.Linear module for a standard training loop. + """ + + def __init__(self, centroids, assignments, bias, in_features, out_features): + super(PQLinear, self).__init__() + self.block_size = centroids.size(1) + self.n_centroids = centroids.size(0) + self.in_features = in_features + self.out_features = out_features + # check compatibility + if self.in_features % self.block_size != 0: + raise ValueError("Wrong PQ sizes") + if len(assignments) % self.out_features != 0: + raise ValueError("Wrong PQ sizes") + # define parameters + self.centroids = nn.Parameter(centroids, requires_grad=True) + self.register_buffer("assignments", assignments) + self.register_buffer("counts", torch.bincount(assignments).type_as(centroids)) + if bias is not None: + self.bias = nn.Parameter(bias) + else: + self.register_parameter("bias", None) + + @property + def weight(self): + return ( + self.centroids[self.assignments] + .reshape(-1, self.out_features, self.block_size) + .permute(1, 0, 2) + .flatten(1, 2) + ) + + def forward(self, x): + return F.linear( + x, + self.weight, + self.bias, + ) + + def extra_repr(self): + return f"in_features={self.in_features},\ + out_features={self.out_features},\ + n_centroids={self.n_centroids},\ + block_size={self.block_size},\ + bias={self.bias is not None}" diff --git a/data/fairseq/fairseq/modules/quantization/pq/pq.py b/data/fairseq/fairseq/modules/quantization/pq/pq.py new file mode 100644 index 0000000000000000000000000000000000000000..eddc2eb34602403f10979f54cd23a45bc2f104d5 --- /dev/null +++ b/data/fairseq/fairseq/modules/quantization/pq/pq.py @@ -0,0 +1,128 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from .em import EM, EmptyClusterResolveError + + +class PQ(EM): + """ + Quantizes the layer weights W with the standard Product Quantization + technique. This learns a codebook of codewords or centroids of size + block_size from W. For further reference on using PQ to quantize + neural networks, see "And the Bit Goes Down: Revisiting the Quantization + of Neural Networks", Stock et al., ICLR 2020. + + PQ is performed in two steps: + (1) The matrix W (weights or fully-connected or convolutional layer) + is reshaped to (block_size, -1). + - If W is fully-connected (2D), its columns are split into + blocks of size block_size. + - If W is convolutional (4D), its filters are split along the + spatial dimension. + (2) We apply the standard EM/k-means algorithm to the resulting reshaped matrix. + + Args: + - W: weight matrix to quantize of size (in_features x out_features) + - block_size: size of the blocks (subvectors) + - n_centroids: number of centroids + - n_iter: number of k-means iterations + - eps: for cluster reassignment when an empty cluster is found + - max_tentatives for cluster reassignment when an empty cluster is found + - verbose: print information after each iteration + + Remarks: + - block_size be compatible with the shape of W + """ + + def __init__( + self, + W, + block_size, + n_centroids=256, + n_iter=20, + eps=1e-6, + max_tentatives=30, + verbose=True, + ): + self.block_size = block_size + W_reshaped = self._reshape(W) + super(PQ, self).__init__( + W_reshaped, + n_centroids=n_centroids, + n_iter=n_iter, + eps=eps, + max_tentatives=max_tentatives, + verbose=verbose, + ) + + def _reshape(self, W): + """ + Reshapes the matrix W as expained in step (1). + """ + + # fully connected: by convention the weight has size out_features x in_features + if len(W.size()) == 2: + self.out_features, self.in_features = W.size() + assert ( + self.in_features % self.block_size == 0 + ), "Linear: n_blocks must be a multiple of in_features" + return ( + W.reshape(self.out_features, -1, self.block_size) + .permute(2, 1, 0) + .flatten(1, 2) + ) + + # convolutional: we reshape along the spatial dimension + elif len(W.size()) == 4: + self.out_channels, self.in_channels, self.k_h, self.k_w = W.size() + assert ( + self.in_channels * self.k_h * self.k_w + ) % self.block_size == 0, ( + "Conv2d: n_blocks must be a multiple of in_channels * k_h * k_w" + ) + return ( + W.reshape(self.out_channels, -1, self.block_size) + .permute(2, 1, 0) + .flatten(1, 2) + ) + # not implemented + else: + raise NotImplementedError(W.size()) + + def encode(self): + """ + Performs self.n_iter EM steps. + """ + + self.initialize_centroids() + for i in range(self.n_iter): + try: + self.step(i) + except EmptyClusterResolveError: + break + + def decode(self): + """ + Returns the encoded full weight matrix. Must be called after + the encode function. + """ + + # fully connected case + if "k_h" not in self.__dict__: + return ( + self.centroids[self.assignments] + .reshape(-1, self.out_features, self.block_size) + .permute(1, 0, 2) + .flatten(1, 2) + ) + + # convolutional case + else: + return ( + self.centroids[self.assignments] + .reshape(-1, self.out_channels, self.block_size) + .permute(1, 0, 2) + .reshape(self.out_channels, self.in_channels, self.k_h, self.k_w) + ) diff --git a/data/fairseq/fairseq/modules/quantization/pq/utils.py b/data/fairseq/fairseq/modules/quantization/pq/utils.py new file mode 100644 index 0000000000000000000000000000000000000000..eceeef8ba3eb5657374256caf371080ee7d97aa4 --- /dev/null +++ b/data/fairseq/fairseq/modules/quantization/pq/utils.py @@ -0,0 +1,376 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import logging +import re +from operator import attrgetter, itemgetter +import torch +import numpy as np +import torch.distributed as dist +import torch.nn as nn + +from .modules import PQConv2d, PQEmbedding, PQLinear +from .pq import PQ + + +def quantize_model_( + model, + size_tracker, + layers_to_quantize, + block_sizes_config, + n_centroids_config, + step=0, + n_iter=15, + eps=1e-6, + max_tentatives=100, + remove_weights=False, + verbose=True, + state_dict=None, +): + """ + Quantize a model in-place by stages. All the targeted + layers are replaced by their quantized counterpart, + and the model is ready for the finetuning of the + centroids in a standard training loop (no modifications + required). Note that we do not quantize biases. + + Args: + - model: a nn.Module + - size_tracker: useful for tracking quatization statistics + - layers_to_quantize: a list containing regexps for + filtering the layers to quantize at each stage according + to their name (as in model.named_parameters()) + - block_sizes_config: dict like + { + 'Conv2d': ('kernel_size', {'(3, 3)': 9, '(1, 1)': 4}), + 'Linear': ('in_features', {'*': 8}) + } + For instance, all conv2d layers with kernel size 3x3 have + a block size of 9 and all Linear layers are quantized with + a block size of 8, irrespective of their size. + - n_centroids_config: dict like + { + 'Conv2d': ('kernel_size', {'*': 256}), + 'Linear': ('in_features', {'*': 256}) + } + For instance, all conv2d layers are quantized with 256 centroids + - step: the layers to quantize inplace corresponding + to layers_to_quantize[step] + """ + + quantized_layers = get_layers( + model, layers_to_quantize[step], remove_weights=remove_weights + ) + + for layer in quantized_layers: + + # book-keeping + is_master_process = (not dist.is_initialized()) or ( + dist.is_initialized() and dist.get_rank() == 0 + ) + verbose = verbose and is_master_process + + # get block size and centroids + module = attrgetter(layer)(model) + block_size = get_param(module, layer, block_sizes_config) + n_centroids = get_param(module, layer, n_centroids_config) + if verbose: + logging.info( + f"Quantizing layer {layer} with block size {block_size} and {n_centroids} centroids" + ) + + # quantize layer + weight = module.weight.data.clone() + is_bias = "bias" in [x[0] for x in module.named_parameters()] + bias = module.bias.data.clone() if is_bias else None + quantizer = PQ( + weight, + block_size, + n_centroids=n_centroids, + n_iter=n_iter, + eps=eps, + max_tentatives=max_tentatives, + verbose=verbose, + ) + + # quantization performed on all GPUs with same seed + quantizer.encode() + centroids = quantizer.centroids.contiguous() + assignments = quantizer.assignments.contiguous() + + # If n_iter = 0 and state_dict is provided, then + # we initialize random assignments and centroids to + # random values of the appropriate dimensions + # because the quantized model parameters will + # overwritten by the state_dict later on. + if n_iter == 0 and state_dict: + # Initialize random centroids of the correct size + centroids = torch.rand(centroids.size()) + centroids.cuda() + # Get counts and assignment keys from layer in loaded checkpoint. + counts_key = layer + "." + "counts" + assignment_key = layer + "." + "assignments" + # Get number of different bins to include. + counts = list(state_dict[counts_key].shape)[0] + print(layer) + print(state_dict[counts_key]) + print(counts) + # Initialize random assignments of the correct size + # with an appropriate number of bins. + num_assignments = list(state_dict[assignment_key].shape)[0] + num_extra = num_assignments - counts + print(num_assignments) + print(num_extra) + assignments_bins = torch.arange(counts) + assignments_rand = torch.randint(0, counts - 1, (num_extra,)) + assignments = torch.cat((assignments_bins, assignments_rand), 0) + # assignments = assignments.type(torch.IntTensor) + assignments.cuda() + print("assignments") + print(assignments) + + # broadcast results to make sure weights are up-to-date + if dist.is_initialized(): + dist.broadcast(centroids, 0) + dist.broadcast(assignments, 0) + + # instantiate the quantized counterpart + if isinstance(module, nn.Linear): + out_features, in_features = map( + lambda k: module.__dict__[k], ["out_features", "in_features"] + ) + quantized_module = PQLinear( + centroids, assignments, bias, in_features, out_features + ) + elif isinstance(module, nn.Embedding): + num_embeddings, embedding_dim = map( + lambda k: module.__dict__[k], ["num_embeddings", "embedding_dim"] + ) + quantized_module = PQEmbedding( + centroids, assignments, num_embeddings, embedding_dim + ) + elif isinstance(module, nn.Conv2d): + out_channels, in_channels, kernel_size = map( + lambda k: module.__dict__[k], + ["out_channels", "in_channels", "kernel_size"], + ) + stride, padding, dilation, groups, padding_mode = map( + lambda k: module.__dict__[k], + ["stride", "padding", "dilation", "groups", "padding_mode"], + ) + + quantized_module = PQConv2d( + centroids, + assignments, + bias, + in_channels, + out_channels, + kernel_size, + stride=stride, + padding=padding, + dilation=dilation, + groups=groups, + padding_mode=padding_mode, + ) + else: + raise ValueError(f"Module {module} not yet supported for quantization") + + # replace layer by its quantized counterpart + attrsetter(layer)(model, quantized_module) + + # update statistics + size_tracker.update(weight, block_size, n_centroids) + + # return name of quantized layers + return quantized_layers + + +def get_layers(model, filter_regexp, remove_weights=False): + """ + Filters out the layers according to a regexp. Note that + we omit biases. + + Args: + - model: a nn.Module + - filter_regexp: a regexp to filter the layers to keep + according to their name in model.named_parameters(). + For instance, the regexp: + + down_layers\\.[123456]\\.(conv[12]|identity\\.conv)) + + is keeping blocks down_layers from 1 to 6, and inside + each block is keeping conv1, conv2 and identity.conv. + + Remarks: + - We add (module\\.)? at the beginning of the regexp to + account for the possible use of nn.parallel.DataParallel + """ + + # get all parameter names + all_layers = map(itemgetter(0), model.named_parameters()) + + # remove biases + all_layers = filter(lambda x: "bias" not in x, all_layers) + + # remove .weight in all other names (or .weight_orig is spectral norm) + all_layers = map(lambda x: x.replace(".weight_orig", ""), all_layers) + # remove weights indicates whether the weights extension should be removed, in addition to + # weight_orig and weight extension on names + if remove_weights: + all_layers = map(lambda x: x.replace(".weights", ""), all_layers) + all_layers = map(lambda x: x.replace(".weight", ""), all_layers) + + # return filtered layers + filter_regexp = "(module\\.)?" + "(" + filter_regexp + ")" + r = re.compile(filter_regexp) + + return list(filter(r.match, all_layers)) + + +def get_param(module, layer_name, param_config): + """ + Given a quantization configuration, get the right parameter + for the module to be quantized. + + Args: + - module: a nn.Module + - layer_name: the name of the layer + - param_config: a dict like + { + 'Conv2d': ('kernel_size', {'(3, 3)': 9, '(1, 1)': 4}), + 'Linear': ('in_features', {'*': 8}) + } + For instance, all conv2d layers with kernel size 3x3 have + a block size of 9 and all Linear layers are quantized with + a block size of 8, irrespective of their size. + + Remarks: + - if 'fuzzy_name' is passed as a parameter, layers whose layer_name + include 'fuzzy_name' will be assigned the given parameter. + In the following example, conv.expand layers will have a block + size of 9 while conv.reduce will have a block size of 4 and all + other layers will have a block size of 2. + { + 'Conv2d': ('fuzzy_name', {'expand': 9, 'reduce': 4, '*': 2}), + 'Linear': ('fuzzy_name', {'classifier': 8, 'projection': 4}) + } + + """ + + layer_type = module.__class__.__name__ + + if layer_type not in param_config: + raise KeyError(f"Layer type {layer_type} not in config for layer {module}") + + feature, params = param_config[module.__class__.__name__] + + if feature != "fuzzy_name": + feature_value = str(getattr(module, feature)) + if feature_value not in params: + if "*" in params: + feature_value = "*" + else: + raise KeyError( + f"{feature}={feature_value} not in config for layer {module}" + ) + else: + feature_values = [name for name in params if name in layer_name] + if len(feature_values) == 0: + if "*" in params: + feature_value = "*" + else: + raise KeyError(f"name={layer_name} not in config for {module}") + else: + feature_value = feature_values[0] + + return params[feature_value] + + +class SizeTracker(object): + """ + Class to keep track of the compressed network size with iPQ. + + Args: + - model: a nn.Module + + Remarks: + - The compressed size is the sum of three components + for each layer in the network: + (1) Storing the centroids given by iPQ in fp16 + (2) Storing the assignments of the blocks in int8 + (3) Storing all non-compressed elements such as biases + - This cost in only valid if we use 256 centroids (then + indexing can indeed by done with int8). + """ + + def __init__(self, model): + self.model = model + self.size_non_compressed_model = self.compute_size() + self.size_non_quantized = self.size_non_compressed_model + self.size_index = 0 + self.size_centroids = 0 + self.n_quantized_layers = 0 + + def compute_size(self): + """ + Computes the size of the model (in MB). + """ + + res = 0 + for _, p in self.model.named_parameters(): + res += p.numel() + return res * 4 / 1024 / 1024 + + def update(self, W, block_size, n_centroids): + """ + Updates the running statistics when quantizing a new layer. + """ + + # bits per weights + bits_per_weight = np.log2(n_centroids) / block_size + self.n_quantized_layers += 1 + + # size of indexing the subvectors of size block_size (in MB) + size_index_layer = bits_per_weight * W.numel() / 8 / 1024 / 1024 + self.size_index += size_index_layer + + # size of the centroids stored in float16 (in MB) + size_centroids_layer = n_centroids * block_size * 2 / 1024 / 1024 + self.size_centroids += size_centroids_layer + + # size of non-compressed layers, e.g. LayerNorms or biases (in MB) + size_uncompressed_layer = W.numel() * 4 / 1024 / 1024 + self.size_non_quantized -= size_uncompressed_layer + + def __repr__(self): + size_compressed = ( + self.size_index + self.size_centroids + self.size_non_quantized + ) + compression_ratio = self.size_non_compressed_model / size_compressed # NOQA + return ( + f"Non-compressed model size: {self.size_non_compressed_model:.2f} MB. " + f"After quantizing {self.n_quantized_layers} layers, size " + f"(indexing + centroids + other): {self.size_index:.2f} MB + " + f"{self.size_centroids:.2f} MB + {self.size_non_quantized:.2f} MB = " + f"{size_compressed:.2f} MB, compression ratio: {compression_ratio:.2f}x" + ) + + +def attrsetter(*items): + def resolve_attr(obj, attr): + attrs = attr.split(".") + head = attrs[:-1] + tail = attrs[-1] + + for name in head: + obj = getattr(obj, name) + return obj, tail + + def g(obj, val): + for attr in items: + resolved_obj, resolved_attr = resolve_attr(obj, attr) + setattr(resolved_obj, resolved_attr, val) + + return g diff --git a/data/fairseq/fairseq/modules/quantization/quantization_options.py b/data/fairseq/fairseq/modules/quantization/quantization_options.py new file mode 100644 index 0000000000000000000000000000000000000000..b46d682c0edaeaaf2a230e51d50da2a32d4bda98 --- /dev/null +++ b/data/fairseq/fairseq/modules/quantization/quantization_options.py @@ -0,0 +1,44 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + + +def parse_config_yaml(yaml_data): + # Initialize to default options. + quantization_options = { + "n_centroids": { + "Linear": ["in_features", {"*": 256}], + "Embedding": ["embedding_dim", {"*": 256}], + }, + "block_sizes": { + "Linear": ["fuzzy_name", {"fc": 8, "attn": 4, "emb": 4}], + "Embedding": ["fuzzy_name", {"emb": 8}], + }, + "layers_to_quantize": [ + "decoder\\.layers\\.\\d+\\.fc[12]", + "decoder\\.embed_tokens\\.embeddings\\.[012]\\.[01]", + "decoder\\.layers\\.\\d+\\.self_attn\\.(k_proj|v_proj|q_proj|out_proj)", + ], + } + + if "n_centroids" in yaml_data: + quantization_options["n_centroids"] = { + layer: convert_yaml_to_tuple(layer_data) + for layer, layer_data in yaml_data["n_centroids"].items() + } + if "block_sizes" in yaml_data: + quantization_options["block_sizes"] = { + layer: convert_yaml_to_tuple(layer_data) + for layer, layer_data in yaml_data["block_sizes"].items() + } + if "layers_to_quantize" in yaml_data: + quantization_options["layers_to_quantize"] = yaml_data["layers_to_quantize"] + + return quantization_options + + +def convert_yaml_to_tuple(yaml_dictionary): + """Converts a yaml dictionary with two keys: `key` and `value` into a two + argument tuple of those values.""" + return (yaml_dictionary["key"], yaml_dictionary["value"]) diff --git a/data/fairseq/fairseq/modules/quantization/scalar/__init__.py b/data/fairseq/fairseq/modules/quantization/scalar/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..143834f3d036780eb6844c82f0c6f2d10cfe2f61 --- /dev/null +++ b/data/fairseq/fairseq/modules/quantization/scalar/__init__.py @@ -0,0 +1,6 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from .utils import quantize_model_ # NOQA diff --git a/data/fairseq/fairseq/modules/quantization/scalar/__pycache__/__init__.cpython-38.pyc b/data/fairseq/fairseq/modules/quantization/scalar/__pycache__/__init__.cpython-38.pyc new file mode 100644 index 0000000000000000000000000000000000000000..45331f0f0e6c2f2bfc06aeff7bc82f275732b363 Binary files /dev/null and b/data/fairseq/fairseq/modules/quantization/scalar/__pycache__/__init__.cpython-38.pyc differ diff --git a/data/fairseq/fairseq/modules/quantization/scalar/__pycache__/ops.cpython-38.pyc b/data/fairseq/fairseq/modules/quantization/scalar/__pycache__/ops.cpython-38.pyc new file mode 100644 index 0000000000000000000000000000000000000000..a9ffaf3cb5f4ae4a61232a78b0808b6c9de3d506 Binary files /dev/null and b/data/fairseq/fairseq/modules/quantization/scalar/__pycache__/ops.cpython-38.pyc differ diff --git a/data/fairseq/fairseq/modules/quantization/scalar/__pycache__/utils.cpython-38.pyc b/data/fairseq/fairseq/modules/quantization/scalar/__pycache__/utils.cpython-38.pyc new file mode 100644 index 0000000000000000000000000000000000000000..8e5744ac2693d64848bba437b437312a60c2efc8 Binary files /dev/null and b/data/fairseq/fairseq/modules/quantization/scalar/__pycache__/utils.cpython-38.pyc differ diff --git a/data/fairseq/fairseq/modules/quantization/scalar/modules/__init__.py b/data/fairseq/fairseq/modules/quantization/scalar/modules/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..8031d9cdb23f2bc72596f8bc9cfa4965f96e3e6c --- /dev/null +++ b/data/fairseq/fairseq/modules/quantization/scalar/modules/__init__.py @@ -0,0 +1,9 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from .qact import ActivationQuantizer # NOQA +from .qconv import IntConv2d # NOQA +from .qemb import IntEmbedding # NOQA +from .qlinear import IntLinear # NOQA diff --git a/data/fairseq/fairseq/modules/quantization/scalar/modules/__pycache__/__init__.cpython-38.pyc b/data/fairseq/fairseq/modules/quantization/scalar/modules/__pycache__/__init__.cpython-38.pyc new file mode 100644 index 0000000000000000000000000000000000000000..13ffdb32c40ad7c0f75d376ddfc531b1ee6b9052 Binary files /dev/null and b/data/fairseq/fairseq/modules/quantization/scalar/modules/__pycache__/__init__.cpython-38.pyc differ diff --git a/data/fairseq/fairseq/modules/quantization/scalar/modules/__pycache__/qact.cpython-38.pyc b/data/fairseq/fairseq/modules/quantization/scalar/modules/__pycache__/qact.cpython-38.pyc new file mode 100644 index 0000000000000000000000000000000000000000..1c3e29c9c1025e7cb7fdf43f0b16c7b5e419d7c5 Binary files /dev/null and b/data/fairseq/fairseq/modules/quantization/scalar/modules/__pycache__/qact.cpython-38.pyc differ diff --git a/data/fairseq/fairseq/modules/quantization/scalar/modules/__pycache__/qconv.cpython-38.pyc b/data/fairseq/fairseq/modules/quantization/scalar/modules/__pycache__/qconv.cpython-38.pyc new file mode 100644 index 0000000000000000000000000000000000000000..fdc5a401a411e8328cc540c54abe256624a10404 Binary files /dev/null and b/data/fairseq/fairseq/modules/quantization/scalar/modules/__pycache__/qconv.cpython-38.pyc differ diff --git a/data/fairseq/fairseq/modules/quantization/scalar/modules/__pycache__/qemb.cpython-38.pyc b/data/fairseq/fairseq/modules/quantization/scalar/modules/__pycache__/qemb.cpython-38.pyc new file mode 100644 index 0000000000000000000000000000000000000000..2d3b378d88b7bd0ef8bf41cae3170acbf8304e54 Binary files /dev/null and b/data/fairseq/fairseq/modules/quantization/scalar/modules/__pycache__/qemb.cpython-38.pyc differ diff --git a/data/fairseq/fairseq/modules/quantization/scalar/modules/__pycache__/qlinear.cpython-38.pyc b/data/fairseq/fairseq/modules/quantization/scalar/modules/__pycache__/qlinear.cpython-38.pyc new file mode 100644 index 0000000000000000000000000000000000000000..abf0e02681217a38a80aa7c905807c23a7d53834 Binary files /dev/null and b/data/fairseq/fairseq/modules/quantization/scalar/modules/__pycache__/qlinear.cpython-38.pyc differ diff --git a/data/fairseq/fairseq/modules/quantization/scalar/modules/qact.py b/data/fairseq/fairseq/modules/quantization/scalar/modules/qact.py new file mode 100644 index 0000000000000000000000000000000000000000..b362c30dc7982e3717d1b764ae737ec1b24ea78e --- /dev/null +++ b/data/fairseq/fairseq/modules/quantization/scalar/modules/qact.py @@ -0,0 +1,88 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import torch + +from ..ops import emulate_int + + +class ActivationQuantizer: + """ + Fake scalar quantization of the activations using a forward hook. + + Args: + - module. a nn.Module for which we quantize the *post-activations* + - p: proportion of activations to quantize, set by default to 1 + - update_step: to recompute quantization parameters + - bits: number of bits for quantization + - method: choose among {"tensor", "histogram", "channel"} + - clamp_threshold: to prevent gradients overflow + + Remarks: + - Parameters scale and zero_point are recomputed every update_step + forward pass to reduce the overhead + - For the list of quantization methods and number of bits, see ops.py + - To remove the hook from the module, simply call self.handle.remove() + - At test time, the activations are fully quantized + - We use the straight-through estimator so that the gradients + back-propagate nicely in the network, this is implemented with + the detach() trick + - The activations are hard-clamped in [-clamp_threshold, clamp_threshold] + to prevent overflow during the backward pass + """ + + def __init__( + self, + module, + p=1, + update_step=1000, + bits=8, + method="histogram", + clamp_threshold=5, + ): + self.module = module + self.p = p + self.update_step = update_step + self.counter = 0 + self.bits = bits + self.method = method + self.clamp_threshold = clamp_threshold + self.handle = None + self.register_hook() + + def register_hook(self): + # forward hook + def quantize_hook(module, x, y): + + # update parameters every 1000 iterations + if self.counter % self.update_step == 0: + self.scale = None + self.zero_point = None + self.counter += 1 + + # train with QuantNoise and evaluate the fully quantized network + p = self.p if self.module.training else 1 + + # quantize activations + y_q, self.scale, self.zero_point = emulate_int( + y.detach(), + bits=self.bits, + method=self.method, + scale=self.scale, + zero_point=self.zero_point, + ) + + # mask to apply noise + mask = torch.zeros_like(y) + mask.bernoulli_(1 - p) + noise = (y_q - y).masked_fill(mask.bool(), 0) + + # using straight-through estimator (STE) + clamp_low = -self.scale * self.zero_point + clamp_high = self.scale * (2**self.bits - 1 - self.zero_point) + return torch.clamp(y, clamp_low.item(), clamp_high.item()) + noise.detach() + + # register hook + self.handle = self.module.register_forward_hook(quantize_hook) diff --git a/data/fairseq/fairseq/modules/quantization/scalar/modules/qconv.py b/data/fairseq/fairseq/modules/quantization/scalar/modules/qconv.py new file mode 100644 index 0000000000000000000000000000000000000000..29744744ecffaec6187888e343f6d6031660292f --- /dev/null +++ b/data/fairseq/fairseq/modules/quantization/scalar/modules/qconv.py @@ -0,0 +1,149 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import torch +import torch.nn.functional as F +from torch.nn.modules.conv import _ConvNd +from torch.nn.modules.utils import _pair + +from ..ops import emulate_int + + +class IntConv2d(_ConvNd): + """ + Quantized counterpart of the nn.Conv2d module that applies QuantNoise during training. + + Args: + - standard nn.Conv2d parameters + - p: amount of noise to inject (0 = no quantization, 1 = quantize all the weights) + - bits: number of bits + - method: choose among {"tensor", "histogram", "channel"} + - update_step: recompute scale and zero_point every update_steps iterations + + Remarks: + - We use the straight-thgourh estimator so that the gradients + back-propagate nicely in the network, this is implemented with + the detach() trick + - Parameters scale and zero_point are recomputed every update_step + forward pass to reduce the overhead + - At test time, the weights are fully quantized + """ + + def __init__( + self, + in_channels, + out_channels, + kernel_size, + stride=1, + padding=0, + dilation=1, + groups=1, + bias=True, + padding_mode="zeros", + p=0, + bits=8, + method="histogram", + update_step=1000, + ): + kernel_size = _pair(kernel_size) + stride = _pair(stride) + padding = _pair(padding) + dilation = _pair(dilation) + super(IntConv2d, self).__init__( + in_channels, + out_channels, + kernel_size, + stride, + padding, + dilation, + False, + _pair(0), + groups, + bias, + padding_mode, + ) + + # quantization parameters + self.p = p + self.bits = bits + self.method = method + self.update_step = update_step + self.counter = 0 + + def _conv_forward(self, input, weight): + if self.padding_mode != "zeros": + return F.conv2d( + F.pad(input, self._padding_repeated_twice, mode=self.padding_mode), + weight, + self.bias, + self.stride, + _pair(0), + self.dilation, + self.groups, + ) + return F.conv2d( + input, + weight, + self.bias, + self.stride, + self.padding, + self.dilation, + self.groups, + ) + + def forward(self, input): + # train with QuantNoise and evaluate the fully quantized network + p = self.p if self.training else 1 + + # update parameters every 100 iterations + if self.counter % self.update_step == 0: + self.scale = None + self.zero_point = None + self.counter += 1 + + # quantize weight + weight_quantized, self.scale, self.zero_point = emulate_int( + self.weight.detach(), + bits=self.bits, + method=self.method, + scale=self.scale, + zero_point=self.zero_point, + ) + + # mask to apply noise + mask = torch.zeros_like(self.weight) + mask.bernoulli_(1 - p) + noise = (weight_quantized - self.weight).masked_fill(mask.bool(), 0) + + # using straight-through estimator (STE) + clamp_low = -self.scale * self.zero_point + clamp_high = self.scale * (2**self.bits - 1 - self.zero_point) + weight = ( + torch.clamp(self.weight, clamp_low.item(), clamp_high.item()) + + noise.detach() + ) + + # return output + output = self._conv_forward(input, weight) + return output + + def extra_repr(self): + return ( + "in_channels={}, out_channels={}, kernel_size={}, stride={}, " + "padding={}, dilation={}, groups={}, bias={}, quant_noise={}, " + "bits={}, method={}".format( + self.in_channels, + self.out_channels, + self.kernel_size, + self.stride, + self.padding, + self.dilation, + self.groups, + self.bias is not None, + self.p, + self.bits, + self.method, + ) + ) diff --git a/data/fairseq/fairseq/modules/quantization/scalar/modules/qemb.py b/data/fairseq/fairseq/modules/quantization/scalar/modules/qemb.py new file mode 100644 index 0000000000000000000000000000000000000000..3b293ac31e9ab683c47e05b763c077ec181e64f7 --- /dev/null +++ b/data/fairseq/fairseq/modules/quantization/scalar/modules/qemb.py @@ -0,0 +1,147 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import torch +import torch.nn as nn +import torch.nn.functional as F + +from ..ops import emulate_int + + +class IntEmbedding(nn.Module): + """ + Quantized counterpart of the nn.Embedding module that applies QuantNoise during training. + + Args: + - num_embeddings: number of tokens + - embedding_dim: embedding dimension + - p: amount of noise to inject (0 = no quantization, 1 = quantize all the weights) + - bits: number of bits + - method: choose among {"tensor", "histogram", "channel"} + - update_step: recompute scale and zero_point every update_steps iterations + + Remarks: + - We use the straight-through estimator so that the gradients + back-propagate nicely in the network, this is implemented with + the detach() trick + - Parameters scale and zero_point are recomputed every update_step + forward pass to reduce the overhead + - At test time, the weights are fully quantized + """ + + def __init__( + self, + num_embeddings, + embedding_dim, + padding_idx=None, + max_norm=None, + norm_type=2.0, + scale_grad_by_freq=False, + sparse=False, + _weight=None, + p=0, + update_step=1000, + bits=8, + method="histogram", + ): + super(IntEmbedding, self).__init__() + self.num_embeddings = num_embeddings + self.embedding_dim = embedding_dim + if padding_idx is not None: + if padding_idx > 0: + assert ( + padding_idx < self.num_embeddings + ), "Padding_idx must be within num_embeddings" + elif padding_idx < 0: + assert ( + padding_idx >= -self.num_embeddings + ), "Padding_idx must be within num_embeddings" + padding_idx = self.num_embeddings + padding_idx + self.padding_idx = padding_idx + self.max_norm = max_norm + self.norm_type = norm_type + self.scale_grad_by_freq = scale_grad_by_freq + if _weight is None: + self.weight = nn.Parameter(torch.Tensor(num_embeddings, embedding_dim)) + self.reset_parameters() + else: + assert list(_weight.shape) == [ + num_embeddings, + embedding_dim, + ], "Shape of weight does not match num_embeddings and embedding_dim" + self.weight = nn.Parameter(_weight) + self.sparse = sparse + + # quantization parameters + self.p = p + self.bits = bits + self.method = method + self.update_step = update_step + self.counter = 0 + + def reset_parameters(self): + nn.init.normal_(self.weight) + if self.padding_idx is not None: + with torch.no_grad(): + self.weight[self.padding_idx].fill_(0) + + def forward(self, input): + # train with QuantNoise and evaluate the fully quantized network + p = self.p if self.training else 1 + + # update parameters every 1000 iterations + if self.counter % self.update_step == 0: + self.scale = None + self.zero_point = None + self.counter += 1 + + # quantize weight + weight_quantized, self.scale, self.zero_point = emulate_int( + self.weight.detach(), + bits=self.bits, + method=self.method, + scale=self.scale, + zero_point=self.zero_point, + ) + + # mask to apply noise + mask = torch.zeros_like(self.weight) + mask.bernoulli_(1 - p) + noise = (weight_quantized - self.weight).masked_fill(mask.bool(), 0) + + # using straight-through estimator (STE) + clamp_low = -self.scale * self.zero_point + clamp_high = self.scale * (2**self.bits - 1 - self.zero_point) + weight = ( + torch.clamp(self.weight, clamp_low.item(), clamp_high.item()) + + noise.detach() + ) + + # return output + output = F.embedding( + input, + weight, + self.padding_idx, + self.max_norm, + self.norm_type, + self.scale_grad_by_freq, + self.sparse, + ) + return output + + def extra_repr(self): + s = "{num_embeddings}, {embedding_dim}" + if self.padding_idx is not None: + s += ", padding_idx={padding_idx}" + if self.max_norm is not None: + s += ", max_norm={max_norm}" + if self.norm_type != 2: + s += ", norm_type={norm_type}" + if self.scale_grad_by_freq is not False: + s += ", scale_grad_by_freq={scale_grad_by_freq}" + if self.sparse is not False: + s += ", sparse=True" + s += "quant_noise={p}, bits={bits}, method={method}" + return s.format(**self.__dict__) diff --git a/data/fairseq/fairseq/modules/quantization/scalar/modules/qlinear.py b/data/fairseq/fairseq/modules/quantization/scalar/modules/qlinear.py new file mode 100644 index 0000000000000000000000000000000000000000..78606a25b98b69b42eef72742037616b21116133 --- /dev/null +++ b/data/fairseq/fairseq/modules/quantization/scalar/modules/qlinear.py @@ -0,0 +1,113 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import torch +import torch.nn as nn +import torch.nn.functional as F + +from ..ops import emulate_int + + +class IntLinear(nn.Module): + """ + Quantized counterpart of the nn.Linear module that applies QuantNoise during training. + + Args: + - in_features: input features + - out_features: output features + - bias: bias or not + - p: amount of noise to inject (0 = no quantization, 1 = quantize all the weights) + - bits: number of bits + - method: choose among {"tensor", "histogram", "channel"} + - update_step: recompute scale and zero_point every update_steps iterations + + Remarks: + - We use the straight-through estimator so that the gradients + back-propagate nicely in the network, this is implemented with + the detach() trick. + - Parameters scale and zero_point are recomputed every update_step + forward pass to reduce the overhead + - At test time, the weights are fully quantized + """ + + def __init__( + self, + in_features, + out_features, + bias=True, + p=0, + update_step=3000, + bits=8, + method="histogram", + ): + super(IntLinear, self).__init__() + self.in_features = int(in_features) + self.out_features = int(out_features) + self.weight = torch.nn.Parameter(torch.Tensor(out_features, in_features)) + self.chosen_bias = bias + if self.chosen_bias: + self.bias = torch.nn.Parameter(torch.Tensor(out_features)) + else: + self.register_parameter("bias", None) + self.reset_parameters() + + # quantization parameters + self.p = p + self.bits = bits + self.method = method + self.update_step = update_step + self.counter = 0 + + def reset_parameters(self): + nn.init.xavier_uniform_(self.weight) + if self.chosen_bias: + nn.init.constant_(self.bias, 0.0) + return + + def forward(self, input): + # train with QuantNoise and evaluate the fully quantized network + p = self.p if self.training else 1 + + # update parameters every 100 iterations + if self.counter % self.update_step == 0: + self.scale = None + self.zero_point = None + self.counter += 1 + + # quantize weight + weight_quantized, self.scale, self.zero_point = emulate_int( + self.weight.detach(), + bits=self.bits, + method=self.method, + scale=self.scale, + zero_point=self.zero_point, + ) + + # mask to apply noise + mask = torch.zeros_like(self.weight) + mask.bernoulli_(1 - p) + noise = (weight_quantized - self.weight).masked_fill(mask.bool(), 0) + + # using straight-through estimator (STE) + clamp_low = -self.scale * self.zero_point + clamp_high = self.scale * (2**self.bits - 1 - self.zero_point) + weight = ( + torch.clamp(self.weight, clamp_low.item(), clamp_high.item()) + + noise.detach() + ) + + # return output + output = F.linear(input, weight, self.bias) + return output + + def extra_repr(self): + return "in_features={}, out_features={}, bias={}, quant_noise={}, bits={}, method={}".format( + self.in_features, + self.out_features, + self.bias is not None, + self.p, + self.bits, + self.method, + ) diff --git a/data/fairseq/fairseq/modules/quantization/scalar/ops.py b/data/fairseq/fairseq/modules/quantization/scalar/ops.py new file mode 100644 index 0000000000000000000000000000000000000000..e0f9a0c1f83112e554cbbea37aafa7eea27900db --- /dev/null +++ b/data/fairseq/fairseq/modules/quantization/scalar/ops.py @@ -0,0 +1,59 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import torch + +try: + import torch.ao.quantization as quantization +except ImportError: + import torch.quantization as quantization + + +def emulate_int(w, bits, method, scale=None, zero_point=None): + q = globals()[f"emulate_int8_{method}"] + return q(w, scale=scale, zero_point=zero_point, bits=bits) + + +def quantize(w, scale, zero_point, bits=8): + # In the default behavior, max_val = 255. + max_val = 2**bits - 1 + return ( + torch.clamp(torch.round(w / scale + zero_point), 0, max_val) - zero_point + ) * scale + + +def emulate_int8_histogram(w, scale=None, zero_point=None, bits=8): + if scale is None: + obs = quantization.observer.HistogramObserver() + obs.to(device=w.device) + _ = obs(w.float()) + scale, zero_point = obs.calculate_qparams() + scale = scale.cuda().type_as(w) + zero_point = zero_point.cuda().type_as(w) + return quantize(w, scale, zero_point, bits=bits), scale, zero_point + + +def emulate_int8_channel(w, scale=None, zero_point=None, bits=8): + if scale is None: + obs = quantization.observer.PerChannelMinMaxObserver( + ch_axis=-1, qscheme=torch.per_channel_symmetric + ) + obs.to(device=w.device) + _ = obs(w) + scale, zero_point, ch_axis = obs.get_qparams() + scale = scale.cuda().type_as(w) + zero_point = zero_point.cuda().type_as(w) + return quantize(w, scale, zero_point, bits=bits), scale, zero_point + + +def emulate_int8_tensor(w, scale=None, zero_point=None, bits=8): + if scale is None: + obs = quantization.observer.MinMaxObserver() + obs.to(device=w.device) + _ = obs(w) + scale, zero_point = obs.calculate_qparams() + scale = scale.cuda().type_as(w) + zero_point = zero_point.cuda().type_as(w) + return quantize(w, scale, zero_point, bits=bits), scale, zero_point diff --git a/data/fairseq/fairseq/modules/quantization/scalar/utils.py b/data/fairseq/fairseq/modules/quantization/scalar/utils.py new file mode 100644 index 0000000000000000000000000000000000000000..d4b1cc255b931098bd63ef277dd4a609106326d5 --- /dev/null +++ b/data/fairseq/fairseq/modules/quantization/scalar/utils.py @@ -0,0 +1,80 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import logging +from operator import attrgetter + +import torch.distributed as dist +import torch.nn as nn + +from ..pq.utils import attrsetter, get_layers +from .modules import ActivationQuantizer, IntConv2d, IntEmbedding, IntLinear + + +MAPPING = {nn.Linear: IntLinear, nn.Embedding: IntEmbedding, nn.Conv2d: IntConv2d} + + +def quantize_model_( + model, p=0.2, bits=8, update_step=3000, method="histogram", remove_weights=False +): + """ + Replaces all modules with their scalar quantized counterpart and + registers hooks to quantize the post-ativations of those modules. + + Args: + - model: a nn.Module + - p: amount of noise (0 for no noise, 1 to quantize all the weights/activations) + - bits: number of bits + - update_step: update quantization parameters every update_step steps + """ + # quantize all layers + # remove weights indicates whether the weights extension should be removed, in addition to + # weight_orig and weight extension on names + quantized_layers = get_layers(model, "(.*?)", remove_weights=remove_weights) + + for layer in quantized_layers: + + # book-keeping + is_master_process = (not dist.is_initialized()) or ( + dist.is_initialized() and dist.get_rank() == 0 + ) + + # recover module + module = attrgetter(layer)(model) + if is_master_process: + logging.info( + f"Quantizing layer {layer} with bits={bits} and QuantNoise={p}" + ) + + # quantization params + q_params = { + "p": p, + "update_step": update_step, + "bits": bits, + "method": method, + "counter": 0, + } + + # instantiate the quantized counterpart + if isinstance(module, tuple(MAPPING.keys())): + QuantizedModule = MAPPING[module.__class__] + quantized_module = QuantizedModule.__new__(QuantizedModule) + params = module.__dict__ + params.update(q_params) + quantized_module.__dict__.update(params) + + else: + if is_master_process: + logging.info(f"Module {module} not yet supported for quantization") + continue + + # activation quantization + a_q = ActivationQuantizer(quantized_module, p=0, bits=bits, method=method) + + # replace layer by its quantized counterpart + attrsetter(layer)(model, quantized_module) + + # return name of quantized layers + return quantized_layers diff --git a/data/fairseq/fairseq/modules/rotary_positional_embedding.py b/data/fairseq/fairseq/modules/rotary_positional_embedding.py new file mode 100644 index 0000000000000000000000000000000000000000..b74028b0117ae94ff8567f45f37e67ae1079ee72 --- /dev/null +++ b/data/fairseq/fairseq/modules/rotary_positional_embedding.py @@ -0,0 +1,50 @@ +import torch + + +class RotaryPositionalEmbedding(torch.nn.Module): + def __init__(self, dim, base=10000, precision=torch.half): + """Rotary positional embedding + Reference : https://blog.eleuther.ai/rotary-embeddings/ + Paper: https://arxiv.org/pdf/2104.09864.pdf + Args: + dim: Dimension of embedding + base: Base value for exponential + precision: precision to use for numerical values + """ + super().__init__() + inv_freq = 1.0 / (base ** (torch.arange(0, dim, 2).float() / dim)) + self.register_buffer("inv_freq", inv_freq) + self.seq_len_cached = 0 + self.cos_cached = torch.empty(self.seq_len_cached, 1, 1, dim) + self.sin_cached = torch.empty(self.seq_len_cached, 1, 1, dim) + self.precision = precision + + def forward(self, x, seq_len: int = 0): + """ + Args: + x: Input x with T X B X C + seq_len: Sequence length of input x + """ + if seq_len > self.seq_len_cached: + self.seq_len_cached = seq_len + t = torch.arange(seq_len, device=x.device).type_as(self.inv_freq) + freqs = torch.einsum("i,j->ij", t, self.inv_freq) + emb = torch.cat((freqs, freqs), dim=-1).to(x.device) + self.cos_cached = emb.cos().view(emb.size(0), 1, 1, emb.size(1)) + self.sin_cached = emb.sin().view(emb.size(0), 1, 1, emb.size(1)) + return self.cos_cached, self.sin_cached + +# rotary pos emb helpers: +def rotate_half(x): + x1, x2 = x[..., : x.shape[-1] // 2], x[..., x.shape[-1] // 2 :] + return torch.cat( + (-x2, x1), dim=x1.ndim - 1 + ) # dim=-1 triggers a bug in earlier torch versions + + +def apply_rotary_pos_emb(q, k, cos, sin, offset: int = 0): + cos, sin = ( + cos[offset : q.shape[0] + offset, ...], + sin[offset : q.shape[0] + offset, ...], + ) + return (q * cos) + (rotate_half(q) * sin), (k * cos) + (rotate_half(k) * sin) diff --git a/data/fairseq/fairseq/modules/same_pad.py b/data/fairseq/fairseq/modules/same_pad.py new file mode 100644 index 0000000000000000000000000000000000000000..a3ce4131c686fddeb86b4392b0c662c8b8100ff3 --- /dev/null +++ b/data/fairseq/fairseq/modules/same_pad.py @@ -0,0 +1,33 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + + +from torch import nn + + +class SamePad(nn.Module): + def __init__(self, kernel_size, causal=False): + super().__init__() + if causal: + self.remove = kernel_size - 1 + else: + self.remove = 1 if kernel_size % 2 == 0 else 0 + + def forward(self, x): + if self.remove > 0: + x = x[:, :, : -self.remove] + return x + + +class SamePad2d(nn.Module): + def __init__(self, kernel_size): + super().__init__() + self.remove = 1 if kernel_size % 2 == 0 else 0 + + def forward(self, x): + assert len(x.size()) == 4 + if self.remove > 0: + x = x[:, :, : -self.remove, : -self.remove] + return x diff --git a/data/fairseq/fairseq/modules/scalar_bias.py b/data/fairseq/fairseq/modules/scalar_bias.py new file mode 100644 index 0000000000000000000000000000000000000000..c96247c75914fabb8a2b7ff731bb82b588f72690 --- /dev/null +++ b/data/fairseq/fairseq/modules/scalar_bias.py @@ -0,0 +1,31 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. +# + +import torch + + +class ScalarBias(torch.autograd.Function): + """ + Adds a vector of scalars, used in self-attention mechanism to allow + the model to optionally attend to this vector instead of the past + """ + + @staticmethod + def forward(ctx, input, dim, bias_init): + size = list(input.size()) + size[dim] += 1 + output = input.new(*size).fill_(bias_init) + output.narrow(dim, 1, size[dim] - 1).copy_(input) + ctx.dim = dim + return output + + @staticmethod + def backward(ctx, grad): + return grad.narrow(ctx.dim, 1, grad.size(ctx.dim) - 1), None, None + + +def scalar_bias(input, dim, bias_init=0): + return ScalarBias.apply(input, dim, bias_init) diff --git a/data/fairseq/fairseq/modules/sinusoidal_positional_embedding.py b/data/fairseq/fairseq/modules/sinusoidal_positional_embedding.py new file mode 100644 index 0000000000000000000000000000000000000000..dd93ddc397935027257515403dc316bd98d02bfa --- /dev/null +++ b/data/fairseq/fairseq/modules/sinusoidal_positional_embedding.py @@ -0,0 +1,122 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import math +from typing import Any, Optional + +import torch +import torch.onnx.operators +from fairseq import utils +from torch import nn, Tensor + + +class SinusoidalPositionalEmbedding(nn.Module): + """This module produces sinusoidal positional embeddings of any length. + + Padding symbols are ignored. + """ + + def __init__(self, embedding_dim, padding_idx, init_size=1024, auto_expand=True): + super().__init__() + self.embedding_dim = embedding_dim + self.padding_idx = padding_idx if padding_idx is not None else 0 + self.register_buffer( + "weights", + SinusoidalPositionalEmbedding.get_embedding( + init_size, embedding_dim, padding_idx + ), + persistent=False, + ) + self.max_positions = int(1e5) + self.auto_expand = auto_expand + self.onnx_trace = False + + def prepare_for_onnx_export_(self): + self.onnx_trace = True + + def _load_from_state_dict(self, state_dict, prefix, *args, **kwargs): + # Ignore some deprecated keys that were used in older versions + deprecated_keys = ["weights", "_float_tensor"] + for key in deprecated_keys: + if prefix + key in state_dict: + del state_dict[prefix + key] + super()._load_from_state_dict(state_dict, prefix, *args, **kwargs) + + @staticmethod + def get_embedding( + num_embeddings: int, embedding_dim: int, padding_idx: Optional[int] = None + ): + """Build sinusoidal embeddings. + + This matches the implementation in tensor2tensor, but differs slightly + from the description in Section 3.5 of "Attention Is All You Need". + """ + half_dim = embedding_dim // 2 + emb = math.log(10000) / (half_dim - 1) + emb = torch.exp(torch.arange(half_dim, dtype=torch.float) * -emb) + emb = torch.arange(num_embeddings, dtype=torch.float).unsqueeze( + 1 + ) * emb.unsqueeze(0) + emb = torch.cat([torch.sin(emb), torch.cos(emb)], dim=1).view( + num_embeddings, -1 + ) + if embedding_dim % 2 == 1: + # zero pad + emb = torch.cat([emb, torch.zeros(num_embeddings, 1)], dim=1) + if padding_idx is not None: + emb[padding_idx, :] = 0 + return emb + + def forward( + self, + input, + incremental_state: Optional[Any] = None, + timestep: Optional[Tensor] = None, + positions: Optional[Any] = None, + ): + """Input is expected to be of size [bsz x seqlen].""" + bspair = torch.onnx.operators.shape_as_tensor(input) + bsz, seq_len = bspair[0], bspair[1] + max_pos = self.padding_idx + 1 + seq_len + weights = self.weights + + if max_pos > self.weights.size(0): + # If the input is longer than the number of pre-computed embeddings, + # compute the extra embeddings on the fly. + # Only store the expanded embeddings if auto_expand=True. + # In multithreading environments, mutating the weights of a module + # may cause trouble. Set auto_expand=False if this happens. + weights = SinusoidalPositionalEmbedding.get_embedding( + max_pos, self.embedding_dim, self.padding_idx + ).to(self.weights) + if self.auto_expand: + self.weights = weights + + if incremental_state is not None: + # positions is the same for every token when decoding a single step + pos = timestep.view(-1)[0] + 1 if timestep is not None else seq_len + if self.onnx_trace: + return ( + weights.index_select(index=self.padding_idx + pos, dim=0) + .unsqueeze(1) + .repeat(bsz, 1, 1) + ) + return weights[self.padding_idx + pos, :].expand(bsz, 1, -1) + + positions = utils.make_positions( + input, self.padding_idx, onnx_trace=self.onnx_trace + ) + if self.onnx_trace: + flat_embeddings = weights.detach().index_select(0, positions.view(-1)) + embedding_shape = torch.cat( + (bsz.view(1), seq_len.view(1), torch.tensor([-1], dtype=torch.long)) + ) + embeddings = torch.onnx.operators.reshape_from_tensor_shape( + flat_embeddings, embedding_shape + ) + return embeddings + return ( + weights.index_select(0, positions.view(-1)).view(bsz, seq_len, -1).detach() + ) diff --git a/data/fairseq/fairseq/modules/sparse_multihead_attention.py b/data/fairseq/fairseq/modules/sparse_multihead_attention.py new file mode 100644 index 0000000000000000000000000000000000000000..3cbd9d6785886e319aab0601517e27df733b6f97 --- /dev/null +++ b/data/fairseq/fairseq/modules/sparse_multihead_attention.py @@ -0,0 +1,140 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import math + +import torch + +from .multihead_attention import MultiheadAttention + + +class SparseMultiheadAttention(MultiheadAttention): + """Sparse Multi-Headed Attention. + + "Generating Long Sequences with Sparse Transformers". Implements + fixed factorized self attention, where l=stride and c=expressivity. + A(1) includes all words in the stride window and A(2) takes a summary of c + words from the end of each stride window. + If is_bidirectional=False, we do not include any words past the current word, + as in the paper. + """ + + def __init__( + self, + embed_dim, + num_heads, + kdim=None, + vdim=None, + dropout=0.0, + bias=True, + add_bias_kv=False, + add_zero_attn=False, + self_attention=False, + encoder_decoder_attention=False, + stride=32, + expressivity=8, + is_bidirectional=True, + ): + + super().__init__( + embed_dim, + num_heads, + kdim, + vdim, + dropout, + bias, + add_bias_kv, + add_zero_attn, + self_attention, + encoder_decoder_attention, + ) + + self.is_bidirectional = is_bidirectional + self.stride = stride + self.expressivity = expressivity + assert self.stride > 0 and self.stride >= self.expressivity + + # Used for Ai(2) calculations - beginning of [l-c, l] range + def compute_checkpoint(self, word_index): + if word_index % self.stride == 0 and word_index != 0: + checkpoint_index = word_index - self.expressivity + else: + checkpoint_index = ( + math.floor(word_index / self.stride) * self.stride + + self.stride + - self.expressivity + ) + return checkpoint_index + + # Computes Ai(2) + def compute_subset_summaries(self, absolute_max): + checkpoint_index = self.compute_checkpoint(0) + subset_two = set() + while checkpoint_index <= absolute_max - 1: + summary = set( + range( + checkpoint_index, + min(checkpoint_index + self.expressivity + 1, absolute_max), + ) + ) + subset_two = subset_two.union(summary) + checkpoint_index = self.compute_checkpoint(checkpoint_index + self.stride) + return subset_two + + # Sparse Transformer Fixed Attention Pattern: https://arxiv.org/pdf/1904.10509.pdf + def compute_fixed_attention_subset(self, word_index, tgt_len): + # +1s account for range function; [min, max) -> [min, max] + if not self.is_bidirectional: + absolute_max = word_index + 1 + else: + absolute_max = tgt_len + + # Subset 1 - whole window + rounded_index = ( + math.floor((word_index + self.stride) / self.stride) * self.stride + ) + if word_index % self.stride == 0 and word_index != 0: + subset_one = set( + range(word_index - self.stride, min(absolute_max, word_index + 1)) + ) + else: + subset_one = set( + range( + max(0, rounded_index - self.stride), + min(absolute_max, rounded_index + 1), + ) + ) + + # Subset 2 - summary per window + # If bidirectional, subset 2 is the same for every index + subset_two = set() + if not self.is_bidirectional: + subset_two = self.compute_subset_summaries(absolute_max) + + return subset_one.union(subset_two) + + # Compute sparse mask - if bidirectional, can pre-compute and store + def buffered_sparse_mask(self, tensor, tgt_len, src_len): + assert tgt_len > self.stride + sparse_mask = torch.empty((tgt_len, src_len)).float().fill_(float("-inf")) + + # If bidirectional, subset 2 is the same for every index + subset_summaries = set() + if self.is_bidirectional: + subset_summaries = self.compute_subset_summaries(tgt_len) + + for i in range(tgt_len): + fixed_attention_subset = self.compute_fixed_attention_subset(i, tgt_len) + fixed_attention_subset = fixed_attention_subset.union(subset_summaries) + included_word_indices = torch.LongTensor(list(fixed_attention_subset)) + sparse_mask[i].index_fill_(0, included_word_indices, 0) + return sparse_mask.type_as(tensor) + + def apply_sparse_mask(self, attn_weights, tgt_len, src_len, bsz): + sparse_mask = self.buffered_sparse_mask(attn_weights, tgt_len, src_len) + sparse_mask = sparse_mask.unsqueeze(0).expand( + bsz * self.num_heads, tgt_len, src_len + ) + attn_weights += sparse_mask diff --git a/data/fairseq/fairseq/modules/sparse_transformer_sentence_encoder.py b/data/fairseq/fairseq/modules/sparse_transformer_sentence_encoder.py new file mode 100644 index 0000000000000000000000000000000000000000..f41ec09327fe80b50d20674e7482794ce45c531c --- /dev/null +++ b/data/fairseq/fairseq/modules/sparse_transformer_sentence_encoder.py @@ -0,0 +1,96 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import torch.nn as nn +from fairseq.modules import TransformerSentenceEncoder +from fairseq.modules.sparse_transformer_sentence_encoder_layer import ( + SparseTransformerSentenceEncoderLayer, +) + + +class SparseTransformerSentenceEncoder(TransformerSentenceEncoder): + """ + Sparse implementation of the TransformerSentenceEncoder + - see SparseMultiheadAttention + """ + + def __init__( + self, + padding_idx: int, + vocab_size: int, + num_encoder_layers: int = 6, + embedding_dim: int = 768, + ffn_embedding_dim: int = 3072, + num_attention_heads: int = 8, + dropout: float = 0.1, + attention_dropout: float = 0.1, + activation_dropout: float = 0.1, + max_seq_len: int = 256, + num_segments: int = 2, + use_position_embeddings: bool = True, + offset_positions_by_padding: bool = True, + encoder_normalize_before: bool = False, + apply_bert_init: bool = False, + activation_fn: str = "relu", + learned_pos_embedding: bool = True, + embed_scale: float = None, + freeze_embeddings: bool = False, + n_trans_layers_to_freeze: int = 0, + export: bool = False, + is_bidirectional: bool = True, + stride: int = 32, + expressivity: int = 8, + ) -> None: + + super().__init__( + padding_idx, + vocab_size, + num_encoder_layers, + embedding_dim, + ffn_embedding_dim, + num_attention_heads, + dropout, + attention_dropout, + activation_dropout, + max_seq_len, + num_segments, + use_position_embeddings, + offset_positions_by_padding, + encoder_normalize_before, + apply_bert_init, + activation_fn, + learned_pos_embedding, + embed_scale, + freeze_embeddings, + n_trans_layers_to_freeze, + export, + ) + + self.layers = nn.ModuleList( + [ + SparseTransformerSentenceEncoderLayer( + embedding_dim=self.embedding_dim, + ffn_embedding_dim=ffn_embedding_dim, + num_attention_heads=num_attention_heads, + dropout=dropout, + attention_dropout=attention_dropout, + activation_dropout=activation_dropout, + activation_fn=activation_fn, + export=export, + is_bidirectional=is_bidirectional, + stride=stride, + expressivity=expressivity, + ) + for _ in range(num_encoder_layers) + ] + ) + + def freeze_module_params(m): + if m is not None: + for p in m.parameters(): + p.requires_grad = False + + for layer in range(n_trans_layers_to_freeze): + freeze_module_params(self.layers[layer]) diff --git a/data/fairseq/fairseq/modules/sparse_transformer_sentence_encoder_layer.py b/data/fairseq/fairseq/modules/sparse_transformer_sentence_encoder_layer.py new file mode 100644 index 0000000000000000000000000000000000000000..d95da59c2471bfa858fd627605196d7f41f9ec12 --- /dev/null +++ b/data/fairseq/fairseq/modules/sparse_transformer_sentence_encoder_layer.py @@ -0,0 +1,51 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from fairseq.modules import TransformerSentenceEncoderLayer +from fairseq.modules.sparse_multihead_attention import SparseMultiheadAttention + + +class SparseTransformerSentenceEncoderLayer(TransformerSentenceEncoderLayer): + """ + Implements a Sprase Transformer Encoder Layer (see SparseMultiheadAttention) + """ + + def __init__( + self, + embedding_dim: int = 768, + ffn_embedding_dim: int = 3072, + num_attention_heads: int = 8, + dropout: float = 0.1, + attention_dropout: float = 0.1, + activation_dropout: float = 0.1, + activation_fn: str = "relu", + export: bool = False, + is_bidirectional: bool = True, + stride: int = 32, + expressivity: int = 8, + ) -> None: + + super().__init__( + embedding_dim, + ffn_embedding_dim, + num_attention_heads, + dropout, + attention_dropout, + activation_dropout, + activation_fn, + export, + ) + + self.self_attn = SparseMultiheadAttention( + self.embedding_dim, + num_attention_heads, + dropout=attention_dropout, + add_bias_kv=False, + add_zero_attn=False, + self_attention=True, + is_bidirectional=is_bidirectional, + stride=stride, + expressivity=expressivity, + ) diff --git a/data/fairseq/fairseq/modules/transformer_layer.py b/data/fairseq/fairseq/modules/transformer_layer.py new file mode 100644 index 0000000000000000000000000000000000000000..19e035dec53fcfc2ed832d9349c98cdd28824e94 --- /dev/null +++ b/data/fairseq/fairseq/modules/transformer_layer.py @@ -0,0 +1,562 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from typing import Dict, List, Optional + +import torch +import torch.nn as nn +from torch import Tensor + +from fairseq import utils +from fairseq.models.transformer import TransformerConfig +from fairseq.modules import LayerNorm, MultiheadAttention +from fairseq.modules.fairseq_dropout import FairseqDropout +from fairseq.modules.quant_noise import quant_noise + + +class TransformerEncoderLayerBase(nn.Module): + """Encoder layer block. + + In the original paper each operation (multi-head attention or FFN) is + postprocessed with: `dropout -> add residual -> layernorm`. In the + tensor2tensor code they suggest that learning is more robust when + preprocessing each layer with layernorm and postprocessing with: + `dropout -> add residual`. We default to the approach in the paper, but the + tensor2tensor approach can be enabled by setting + *cfg.encoder.normalize_before* to ``True``. + + Args: + cfg (argparse.Namespace): parsed command-line arguments + """ + + def __init__(self, cfg, return_fc=False): + super().__init__() + self.cfg = cfg + self.return_fc = return_fc + self.embed_dim = cfg.encoder.embed_dim + self.quant_noise = cfg.quant_noise.pq + self.quant_noise_block_size = cfg.quant_noise.pq_block_size + self.self_attn = self.build_self_attention(self.embed_dim, cfg) + self.self_attn_layer_norm = LayerNorm(self.embed_dim, export=cfg.export) + self.dropout_module = FairseqDropout( + cfg.dropout, module_name=self.__class__.__name__ + ) + self.activation_fn = utils.get_activation_fn(activation=cfg.activation_fn) + activation_dropout_p = cfg.activation_dropout + if activation_dropout_p == 0: + # for backwards compatibility with models that use cfg.relu_dropout + activation_dropout_p = cfg.relu_dropout or 0 + self.activation_dropout_module = FairseqDropout( + float(activation_dropout_p), module_name=self.__class__.__name__ + ) + self.normalize_before = cfg.encoder.normalize_before + self.fc1 = self.build_fc1( + self.embed_dim, + cfg.encoder.ffn_embed_dim, + self.quant_noise, + self.quant_noise_block_size, + ) + self.fc2 = self.build_fc2( + cfg.encoder.ffn_embed_dim, + self.embed_dim, + self.quant_noise, + self.quant_noise_block_size, + ) + + self.final_layer_norm = LayerNorm(self.embed_dim, export=cfg.export) + + def build_fc1(self, input_dim, output_dim, q_noise, qn_block_size): + return quant_noise( + nn.Linear(input_dim, output_dim), p=q_noise, block_size=qn_block_size + ) + + def build_fc2(self, input_dim, output_dim, q_noise, qn_block_size): + return quant_noise( + nn.Linear(input_dim, output_dim), p=q_noise, block_size=qn_block_size + ) + + def _get_fc_rank(self, remove_num: int) -> List[int]: + f1_filter_param = [] + for i in range(self.fc1.out_features): + f1_filter_param.append( + torch.sum(torch.abs(self.fc1.weight[i])) + + torch.sum(torch.abs(self.fc2.weight[:, i])) + + torch.abs(self.fc1.bias[i]) + ) + return sorted( + range(len(f1_filter_param)), key=lambda k: f1_filter_param[k], reverse=False + )[0:remove_num] + + def _prune_fc_layer(self, remove_index: List[int]): + new_fc1_weight = [] + new_fc1_bias = [] + for i in range(self.fc1.out_features): + if i not in remove_index: + new_fc1_weight.append(self.fc1.weight[i]) + new_fc1_bias.append(self.fc1.bias[i]) + + new_fc1_weight = torch.stack(new_fc1_weight).detach() + new_fc1_weight.requires_grad = True + + new_fc1_bias = torch.stack(new_fc1_bias).detach() + new_fc1_bias.requires_grad = True + + self.fc1 = quant_noise( + nn.Linear(self.fc1.in_features, self.fc1.out_features - len(remove_index)), + p=self.quant_noise, + block_size=self.quant_noise_block_size, + ) + self.fc1.weight = torch.nn.Parameter(new_fc1_weight) + self.fc1.bias = torch.nn.Parameter(new_fc1_bias) + + new_fc2_weight = [] + new_fc2_bias = [] + for i in range(self.fc2.in_features): + if i not in remove_index: + new_fc2_weight.append(self.fc2.weight[:, i]) + new_fc2_bias = self.fc2.bias.detach() + + new_fc2_weight = torch.stack(new_fc2_weight, dim=-1).detach() + new_fc2_weight.requires_grad = True + + new_fc2_bias = self.fc2.bias.detach() + new_fc2_bias.requires_grad = True + + self.fc2 = quant_noise( + nn.Linear(self.fc2.in_features - len(remove_index), self.fc2.out_features), + p=self.quant_noise, + block_size=self.quant_noise_block_size, + ) + self.fc2.weight = torch.nn.Parameter(new_fc2_weight) + self.fc2.bias = torch.nn.Parameter(new_fc2_bias) + + def build_self_attention(self, embed_dim, cfg): + return MultiheadAttention( + embed_dim, + cfg.encoder.attention_heads, + dropout=cfg.attention_dropout, + self_attention=True, + q_noise=self.quant_noise, + qn_block_size=self.quant_noise_block_size, + xformers_att_config=cfg.encoder.xformers_att_config, + ) + + def residual_connection(self, x, residual): + return residual + x + + def upgrade_state_dict_named(self, state_dict, name): + """ + Rename layer norm states from `...layer_norms.0.weight` to + `...self_attn_layer_norm.weight` and `...layer_norms.1.weight` to + `...final_layer_norm.weight` + """ + layer_norm_map = {"0": "self_attn_layer_norm", "1": "final_layer_norm"} + for old, new in layer_norm_map.items(): + for m in ("weight", "bias"): + k = "{}.layer_norms.{}.{}".format(name, old, m) + if k in state_dict: + state_dict["{}.{}.{}".format(name, new, m)] = state_dict[k] + del state_dict[k] + + def forward( + self, + x, + encoder_padding_mask: Optional[Tensor], + attn_mask: Optional[Tensor] = None, + ): + """ + Args: + x (Tensor): input to the layer of shape `(seq_len, batch, embed_dim)` + encoder_padding_mask (ByteTensor): binary ByteTensor of shape + `(batch, seq_len)` where padding elements are indicated by ``1``. + attn_mask (ByteTensor): binary tensor of shape `(tgt_len, src_len)`, + where `tgt_len` is the length of output and `src_len` is the + length of input, though here both are equal to `seq_len`. + `attn_mask[tgt_i, src_j] = 1` means that when calculating the + embedding for `tgt_i`, we exclude (mask out) `src_j`. This is + useful for strided self-attention. + + Returns: + encoded output of shape `(seq_len, batch, embed_dim)` + """ + # anything in original attn_mask = 1, becomes -1e8 + # anything in original attn_mask = 0, becomes 0 + # Note that we cannot use -inf here, because at some edge cases, + # the attention weight (before softmax) for some padded element in query + # will become -inf, which results in NaN in model parameters + if attn_mask is not None: + attn_mask = attn_mask.masked_fill( + attn_mask.to(torch.bool), -1e8 if x.dtype == torch.float32 else -1e4 + ) + + residual = x + if self.normalize_before: + x = self.self_attn_layer_norm(x) + x, _ = self.self_attn( + query=x, + key=x, + value=x, + key_padding_mask=encoder_padding_mask, + need_weights=False, + attn_mask=attn_mask, + ) + x = self.dropout_module(x) + x = self.residual_connection(x, residual) + if not self.normalize_before: + x = self.self_attn_layer_norm(x) + + residual = x + if self.normalize_before: + x = self.final_layer_norm(x) + x = self.activation_fn(self.fc1(x)) + x = self.activation_dropout_module(x) + x = self.fc2(x) + + fc_result = x + + x = self.dropout_module(x) + x = self.residual_connection(x, residual) + if not self.normalize_before: + x = self.final_layer_norm(x) + + if self.return_fc and not torch.jit.is_scripting(): + return x, fc_result + return x + + +# backward compatible with the legacy argparse format +class TransformerEncoderLayer(TransformerEncoderLayerBase): + def __init__(self, args): + super().__init__(TransformerConfig.from_namespace(args)) + self.args = args + + def build_self_attention(self, embed_dim, args): + return super().build_self_attention( + embed_dim, TransformerConfig.from_namespace(args) + ) + + +class TransformerDecoderLayerBase(nn.Module): + """Decoder layer block. + + In the original paper each operation (multi-head attention, encoder + attention or FFN) is postprocessed with: `dropout -> add residual -> + layernorm`. In the tensor2tensor code they suggest that learning is more + robust when preprocessing each layer with layernorm and postprocessing with: + `dropout -> add residual`. We default to the approach in the paper, but the + tensor2tensor approach can be enabled by setting + *cfg.decoder.normalize_before* to ``True``. + + Args: + args (argparse.Namespace): parsed command-line arguments + no_encoder_attn (bool, optional): whether to attend to encoder outputs + (default: False). + """ + + def __init__( + self, cfg, no_encoder_attn=False, add_bias_kv=False, add_zero_attn=False + ): + super().__init__() + self.embed_dim = cfg.decoder.embed_dim + self.dropout_module = FairseqDropout( + cfg.dropout, module_name=self.__class__.__name__ + ) + self.quant_noise = cfg.quant_noise.pq + self.quant_noise_block_size = cfg.quant_noise.pq_block_size + + self.cross_self_attention = cfg.cross_self_attention + + self.self_attn = self.build_self_attention( + self.embed_dim, + cfg, + add_bias_kv=add_bias_kv, + add_zero_attn=add_zero_attn, + ) + self.attn_ln = ( + LayerNorm(self.embed_dim) + if utils.safe_getattr(cfg, "scale_attn", False) + else None + ) + self.nh = self.self_attn.num_heads + self.head_dim = self.self_attn.head_dim + scale_heads = utils.safe_getattr(cfg, "scale_heads", False) + self.c_attn = ( + nn.Parameter(torch.ones((self.nh,)), requires_grad=True) + if scale_heads + else None + ) + + self.activation_fn = utils.get_activation_fn(activation=cfg.activation_fn) + activation_dropout_p = cfg.activation_dropout + if activation_dropout_p == 0: + # for backwards compatibility with models that use cfg.relu_dropout + activation_dropout_p = cfg.relu_dropout or 0 + self.activation_dropout_module = FairseqDropout( + float(activation_dropout_p), module_name=self.__class__.__name__ + ) + self.normalize_before = cfg.decoder.normalize_before + + self.self_attn_layer_norm = LayerNorm(self.embed_dim, export=cfg.export) + + if no_encoder_attn: + self.encoder_attn = None + self.encoder_attn_layer_norm = None + else: + self.encoder_attn = self.build_encoder_attention(self.embed_dim, cfg) + self.encoder_attn_layer_norm = LayerNorm(self.embed_dim, export=cfg.export) + + self.ffn_layernorm = ( + LayerNorm(cfg.decoder.ffn_embed_dim) + if utils.safe_getattr(cfg, "scale_fc", False) + else None + ) + self.w_resid = ( + nn.Parameter( + torch.ones( + self.embed_dim, + ), + requires_grad=True, + ) + if utils.safe_getattr(cfg, "scale_resids", False) + else None + ) + + self.fc1 = self.build_fc1( + self.embed_dim, + cfg.decoder.ffn_embed_dim, + self.quant_noise, + self.quant_noise_block_size, + ) + self.fc2 = self.build_fc2( + cfg.decoder.ffn_embed_dim, + self.embed_dim, + self.quant_noise, + self.quant_noise_block_size, + ) + + self.final_layer_norm = LayerNorm(self.embed_dim, export=cfg.export) + self.need_attn = True + + self.onnx_trace = False + + def build_fc1(self, input_dim, output_dim, q_noise, qn_block_size): + return quant_noise(nn.Linear(input_dim, output_dim), q_noise, qn_block_size) + + def build_fc2(self, input_dim, output_dim, q_noise, qn_block_size): + return quant_noise(nn.Linear(input_dim, output_dim), q_noise, qn_block_size) + + def build_self_attention( + self, embed_dim, cfg, add_bias_kv=False, add_zero_attn=False + ): + return MultiheadAttention( + embed_dim, + cfg.decoder.attention_heads, + dropout=cfg.attention_dropout, + add_bias_kv=add_bias_kv, + add_zero_attn=add_zero_attn, + self_attention=not cfg.cross_self_attention, + q_noise=self.quant_noise, + qn_block_size=self.quant_noise_block_size, + xformers_att_config=cfg.decoder.xformers_att_config, + ) + + def build_encoder_attention(self, embed_dim, cfg): + return MultiheadAttention( + embed_dim, + cfg.decoder.attention_heads, + kdim=cfg.encoder.embed_dim, + vdim=cfg.encoder.embed_dim, + dropout=cfg.attention_dropout, + encoder_decoder_attention=True, + q_noise=self.quant_noise, + qn_block_size=self.quant_noise_block_size, + xformers_att_config=cfg.encoder.xformers_att_config, + ) + + def prepare_for_onnx_export_(self): + self.onnx_trace = True + + def residual_connection(self, x, residual): + return residual + x + + def forward( + self, + x, + encoder_out: Optional[torch.Tensor] = None, + encoder_padding_mask: Optional[torch.Tensor] = None, + incremental_state: Optional[Dict[str, Dict[str, Optional[Tensor]]]] = None, + prev_self_attn_state: Optional[List[torch.Tensor]] = None, + prev_attn_state: Optional[List[torch.Tensor]] = None, + self_attn_mask: Optional[torch.Tensor] = None, + self_attn_padding_mask: Optional[torch.Tensor] = None, + need_attn: bool = False, + need_head_weights: bool = False, + ): + """ + Args: + x (Tensor): input to the layer of shape `(seq_len, batch, embed_dim)` + encoder_padding_mask (ByteTensor, optional): binary + ByteTensor of shape `(batch, src_len)` where padding + elements are indicated by ``1``. + need_attn (bool, optional): return attention weights + need_head_weights (bool, optional): return attention weights + for each head (default: return average over heads). + + Returns: + encoded output of shape `(seq_len, batch, embed_dim)` + """ + if need_head_weights: + need_attn = True + + residual = x + if self.normalize_before: + x = self.self_attn_layer_norm(x) + if prev_self_attn_state is not None: + prev_key, prev_value = prev_self_attn_state[:2] + saved_state: Dict[str, Optional[Tensor]] = { + "prev_key": prev_key, + "prev_value": prev_value, + } + if len(prev_self_attn_state) >= 3: + saved_state["prev_key_padding_mask"] = prev_self_attn_state[2] + assert incremental_state is not None + self.self_attn._set_input_buffer(incremental_state, saved_state) + _self_attn_input_buffer = self.self_attn._get_input_buffer(incremental_state) + if self.cross_self_attention and not ( + incremental_state is not None + and _self_attn_input_buffer is not None + and "prev_key" in _self_attn_input_buffer + ): + if self_attn_mask is not None: + assert encoder_out is not None + self_attn_mask = torch.cat( + (x.new_zeros(x.size(0), encoder_out.size(0)), self_attn_mask), dim=1 + ) + if self_attn_padding_mask is not None: + if encoder_padding_mask is None: + assert encoder_out is not None + encoder_padding_mask = self_attn_padding_mask.new_zeros( + encoder_out.size(1), encoder_out.size(0) + ) + self_attn_padding_mask = torch.cat( + (encoder_padding_mask, self_attn_padding_mask), dim=1 + ) + assert encoder_out is not None + y = torch.cat((encoder_out, x), dim=0) + else: + y = x + + x, attn = self.self_attn( + query=x, + key=y, + value=y, + key_padding_mask=self_attn_padding_mask, + incremental_state=incremental_state, + need_weights=False, + attn_mask=self_attn_mask, + ) + if self.c_attn is not None: + tgt_len, bsz = x.size(0), x.size(1) + x = x.view(tgt_len, bsz, self.nh, self.head_dim) + x = torch.einsum("tbhd,h->tbhd", x, self.c_attn) + x = x.reshape(tgt_len, bsz, self.embed_dim) + if self.attn_ln is not None: + x = self.attn_ln(x) + x = self.dropout_module(x) + x = self.residual_connection(x, residual) + if not self.normalize_before: + x = self.self_attn_layer_norm(x) + + if self.encoder_attn is not None and encoder_out is not None: + residual = x + if self.normalize_before: + x = self.encoder_attn_layer_norm(x) + if prev_attn_state is not None: + prev_key, prev_value = prev_attn_state[:2] + saved_state: Dict[str, Optional[Tensor]] = { + "prev_key": prev_key, + "prev_value": prev_value, + } + if len(prev_attn_state) >= 3: + saved_state["prev_key_padding_mask"] = prev_attn_state[2] + assert incremental_state is not None + self.encoder_attn._set_input_buffer(incremental_state, saved_state) + + x, attn = self.encoder_attn( + query=x, + key=encoder_out, + value=encoder_out, + key_padding_mask=encoder_padding_mask, + incremental_state=incremental_state, + static_kv=True, + need_weights=need_attn or (not self.training and self.need_attn), + need_head_weights=need_head_weights, + ) + x = self.dropout_module(x) + x = self.residual_connection(x, residual) + if not self.normalize_before: + x = self.encoder_attn_layer_norm(x) + + residual = x + if self.normalize_before: + x = self.final_layer_norm(x) + + x = self.activation_fn(self.fc1(x)) + x = self.activation_dropout_module(x) + if self.ffn_layernorm is not None: + x = self.ffn_layernorm(x) + x = self.fc2(x) + x = self.dropout_module(x) + if self.w_resid is not None: + residual = torch.mul(self.w_resid, residual) + x = self.residual_connection(x, residual) + if not self.normalize_before: + x = self.final_layer_norm(x) + if self.onnx_trace and incremental_state is not None: + saved_state = self.self_attn._get_input_buffer(incremental_state) + assert saved_state is not None + if self_attn_padding_mask is not None: + self_attn_state = [ + saved_state["prev_key"], + saved_state["prev_value"], + saved_state["prev_key_padding_mask"], + ] + else: + self_attn_state = [saved_state["prev_key"], saved_state["prev_value"]] + return x, attn, self_attn_state + return x, attn, None + + def make_generation_fast_(self, need_attn: bool = False, **kwargs): + self.need_attn = need_attn + + +# backward compatible with the legacy argparse format +class TransformerDecoderLayer(TransformerDecoderLayerBase): + def __init__( + self, args, no_encoder_attn=False, add_bias_kv=False, add_zero_attn=False + ): + super().__init__( + TransformerConfig.from_namespace(args), + no_encoder_attn=no_encoder_attn, + add_bias_kv=add_bias_kv, + add_zero_attn=add_zero_attn, + ) + self.args = args + + def build_self_attention( + self, embed_dim, args, add_bias_kv=False, add_zero_attn=False + ): + return super().build_self_attention( + embed_dim, + TransformerConfig.from_namespace(args), + add_bias_kv=add_bias_kv, + add_zero_attn=add_zero_attn, + ) + + def build_encoder_attention(self, embed_dim, args): + return super().build_encoder_attention( + embed_dim, + TransformerConfig.from_namespace(args), + ) diff --git a/data/fairseq/fairseq/modules/transformer_layer_aug.py b/data/fairseq/fairseq/modules/transformer_layer_aug.py new file mode 100644 index 0000000000000000000000000000000000000000..7eb816978a0c44b9be94221d8a886aa96511a67f --- /dev/null +++ b/data/fairseq/fairseq/modules/transformer_layer_aug.py @@ -0,0 +1,315 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from typing import Dict, List, Optional + +import torch +from numpy.random import uniform +from torch import Tensor + +from fairseq.modules import LayerNorm +from fairseq.modules.transformer_layer import TransformerDecoderLayerBase + + +class AugTransformerDecoderLayerBase(TransformerDecoderLayerBase): + """Decoder layer block augmented with an additional cross-attention. + + This decoder block is processed with the sequence of the following sub-modules. + self-attention -> cross-attention (first) -> cross-attention (second) -> FFN + + Args: + cfg (argparse.Namespace): parsed command-line arguments + encoder_attn_merge_type (str, optional): the way to combine outputs from + two cross-attention modules. If "sequential" is set, two cross-attention + modules are stacked sequentially. If "parallel" is set, they are processed + in parallel and combined before feeding it to FFN (default: sequential). + dropnet_ratio (float, optional): a probability to drop each cross-attention + module during training (default: 0.0). + """ + + def __init__( + self, + cfg, + add_bias_kv=False, + add_zero_attn=False, + encoder_attn_merge_type="sequential", + dropnet_ratio=0.0, + ): + super().__init__( + cfg, + no_encoder_attn=False, + add_bias_kv=add_bias_kv, + add_zero_attn=False, + ) + self.encoder_attn = self.build_encoder_attention(self.embed_dim, cfg) + self.encoder_attn_layer_norm = LayerNorm(self.embed_dim, export=cfg.export) + self.encoder_attn2 = self.build_encoder_attention(self.embed_dim, cfg) + if encoder_attn_merge_type == "sequential": + self.encoder_attn_layer_norm2 = LayerNorm(self.embed_dim, export=cfg.export) + else: + self.encoder_attn_layer_norm2 = None + + self.encoder_attn_merge_type = encoder_attn_merge_type + self.dropnet_ratio = dropnet_ratio + + def forward( + self, + x, + encoder_out: Optional[torch.Tensor] = None, + encoder_padding_mask: Optional[torch.Tensor] = None, + encoder_out_aug: Optional[torch.Tensor] = None, + encoder_padding_mask2: Optional[torch.Tensor] = None, + incremental_state: Optional[Dict[str, Dict[str, Optional[Tensor]]]] = None, + prev_self_attn_state: Optional[List[torch.Tensor]] = None, + prev_attn_state: Optional[List[torch.Tensor]] = None, + self_attn_mask: Optional[torch.Tensor] = None, + self_attn_padding_mask: Optional[torch.Tensor] = None, + need_attn: bool = False, + need_head_weights: bool = False, + ): + """ + Args: + x (Tensor): input to the layer of shape `(seq_len, batch, embed_dim)` + encoder_padding_mask (ByteTensor, optional): binary + ByteTensor of shape `(batch, src_len)` where padding + elements are indicated by ``1``. + need_attn (bool, optional): return attention weights + need_head_weights (bool, optional): return attention weights + for each head (default: return average over heads). + + Returns: + encoded output of shape `(seq_len, batch, embed_dim)` + """ + if need_head_weights: + need_attn = True + + residual = x + if self.normalize_before: + x = self.self_attn_layer_norm(x) + if prev_self_attn_state is not None: + prev_key, prev_value = prev_self_attn_state[:2] + saved_state: Dict[str, Optional[Tensor]] = { + "prev_key": prev_key, + "prev_value": prev_value, + } + if len(prev_self_attn_state) >= 3: + saved_state["prev_key_padding_mask"] = prev_self_attn_state[2] + assert incremental_state is not None + self.self_attn._set_input_buffer(incremental_state, saved_state) + _self_attn_input_buffer = self.self_attn._get_input_buffer(incremental_state) + if self.cross_self_attention and not ( + incremental_state is not None + and _self_attn_input_buffer is not None + and "prev_key" in _self_attn_input_buffer + ): + if self_attn_mask is not None: + assert encoder_out is not None + self_attn_mask = torch.cat( + (x.new_zeros(x.size(0), encoder_out.size(0)), self_attn_mask), dim=1 + ) + if self_attn_padding_mask is not None: + if encoder_padding_mask is None: + assert encoder_out is not None + encoder_padding_mask = self_attn_padding_mask.new_zeros( + encoder_out.size(1), encoder_out.size(0) + ) + self_attn_padding_mask = torch.cat( + (encoder_padding_mask, self_attn_padding_mask), dim=1 + ) + assert encoder_out is not None + y = torch.cat((encoder_out, x), dim=0) + else: + y = x + + x, attn = self.self_attn( + query=x, + key=y, + value=y, + key_padding_mask=self_attn_padding_mask, + incremental_state=incremental_state, + need_weights=False, + attn_mask=self_attn_mask, + ) + if self.c_attn is not None: + tgt_len, bsz = x.size(0), x.size(1) + x = x.view(tgt_len, bsz, self.nh, self.head_dim) + x = torch.einsum("tbhd,h->tbhd", x, self.c_attn) + x = x.reshape(tgt_len, bsz, self.embed_dim) + if self.attn_ln is not None: + x = self.attn_ln(x) + x = self.dropout_module(x) + x = self.residual_connection(x, residual) + if not self.normalize_before: + x = self.self_attn_layer_norm(x) + + assert encoder_out is not None + assert encoder_out_aug is not None + + if self.encoder_attn_merge_type == "sequential": + ratios = self.get_dropnet_ratio() + + # first encoder attention + if ratios[0] > 0: + residual = x + if self.normalize_before: + x = self.encoder_attn_layer_norm(x) + if prev_attn_state is not None: + prev_key, prev_value = prev_attn_state[:2] + saved_state: Dict[str, Optional[Tensor]] = { + "prev_key": prev_key, + "prev_value": prev_value, + } + if len(prev_attn_state) >= 3: + saved_state["prev_key_padding_mask"] = prev_attn_state[2] + assert incremental_state is not None + self.encoder_attn._set_input_buffer(incremental_state, saved_state) + + x, attn = self.encoder_attn( + query=x, + key=encoder_out, + value=encoder_out, + key_padding_mask=encoder_padding_mask, + incremental_state=incremental_state, + static_kv=True, + need_weights=need_attn or (not self.training and self.need_attn), + need_head_weights=need_head_weights, + ) + x = self.dropout_module(x) + x = self.residual_connection(x, residual) + if not self.normalize_before: + x = self.encoder_attn_layer_norm(x) + x = ratios[0] * x + + # second encoder attention + if ratios[1] > 0: + residual = x + if self.normalize_before: + x = self.encoder_attn_layer_norm2(x) + if prev_attn_state is not None: + prev_key, prev_value = prev_attn_state[:2] + saved_state: Dict[str, Optional[Tensor]] = { + "prev_key": prev_key, + "prev_value": prev_value, + } + if len(prev_attn_state) >= 3: + saved_state["prev_key_padding_mask"] = prev_attn_state[2] + assert incremental_state is not None + self.encoder_attn2._set_input_buffer(incremental_state, saved_state) + + x, attn2 = self.encoder_attn2( + query=x, + key=encoder_out_aug, + value=encoder_out_aug, + key_padding_mask=encoder_padding_mask2, + incremental_state=incremental_state, + static_kv=True, + need_weights=need_attn or (not self.training and self.need_attn), + need_head_weights=need_head_weights, + ) + x = self.dropout_module(x) + x = self.residual_connection(x, residual) + if not self.normalize_before: + x = self.encoder_attn_layer_norm2(x) + x = ratios[1] * x + + elif self.encoder_attn_merge_type == "parallel": + residual = x + if self.normalize_before: + x = self.encoder_attn_layer_norm(x) + if prev_attn_state is not None: + prev_key, prev_value = prev_attn_state[:2] + saved_state: Dict[str, Optional[Tensor]] = { + "prev_key": prev_key, + "prev_value": prev_value, + } + if len(prev_attn_state) >= 3: + saved_state["prev_key_padding_mask"] = prev_attn_state[2] + assert incremental_state is not None + self.encoder_attn._set_input_buffer(incremental_state, saved_state) + + x1, attn = self.encoder_attn( + query=x, + key=encoder_out, + value=encoder_out, + key_padding_mask=encoder_padding_mask, + incremental_state=incremental_state, + static_kv=True, + need_weights=need_attn or (not self.training and self.need_attn), + need_head_weights=need_head_weights, + ) + x2, attn2 = self.encoder_attn2( + query=x, + key=encoder_out_aug, + value=encoder_out_aug, + key_padding_mask=encoder_padding_mask2, + incremental_state=incremental_state, + static_kv=True, + need_weights=need_attn or (not self.training and self.need_attn), + need_head_weights=need_head_weights, + ) + x1 = self.dropout_module(x1) + x2 = self.dropout_module(x2) + ratios = self.get_dropnet_ratio() + x = ratios[0] * x1 + ratios[1] * x2 + x = self.residual_connection(x, residual) + if not self.normalize_before: + x = self.encoder_attn_layer_norm(x) + + else: + raise NotImplementedError(self.encoder_attn_merge_type) + + residual = x + if self.normalize_before: + x = self.final_layer_norm(x) + + x = self.activation_fn(self.fc1(x)) + x = self.activation_dropout_module(x) + if self.ffn_layernorm is not None: + x = self.ffn_layernorm(x) + x = self.fc2(x) + x = self.dropout_module(x) + if self.w_resid is not None: + residual = torch.mul(self.w_resid, residual) + x = self.residual_connection(x, residual) + if not self.normalize_before: + x = self.final_layer_norm(x) + if self.onnx_trace and incremental_state is not None: + saved_state = self.self_attn._get_input_buffer(incremental_state) + assert saved_state is not None + if self_attn_padding_mask is not None: + self_attn_state = [ + saved_state["prev_key"], + saved_state["prev_value"], + saved_state["prev_key_padding_mask"], + ] + else: + self_attn_state = [saved_state["prev_key"], saved_state["prev_value"]] + return x, attn, attn2, self_attn_state + return x, attn, attn2, None + + def get_dropnet_ratio(self): + if self.encoder_attn_merge_type == "sequential": + if self.dropnet_ratio > 0: + frand = float(uniform(0, 1)) + if frand < self.dropnet_ratio and self.training: + return [2, 0] + elif frand > 1 - self.dropnet_ratio and self.training: + return [0, 2] + else: + return [1, 1] + else: + return [1, 1] + + elif self.encoder_attn_merge_type == "parallel": + if self.dropnet_ratio > 0: + frand = float(uniform(0, 1)) + if frand < self.dropnet_ratio and self.training: + return [1, 0] + elif frand > 1 - self.dropnet_ratio and self.training: + return [0, 1] + else: + return [0.5, 0.5] + else: + return [0.5, 0.5] diff --git a/data/fairseq/fairseq/modules/transformer_sentence_encoder.py b/data/fairseq/fairseq/modules/transformer_sentence_encoder.py new file mode 100644 index 0000000000000000000000000000000000000000..5d2db91ad7b902386efa843f5ade0c0295758fa1 --- /dev/null +++ b/data/fairseq/fairseq/modules/transformer_sentence_encoder.py @@ -0,0 +1,291 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from typing import Optional, Tuple + +import torch +import torch.nn as nn +from fairseq.modules import ( + FairseqDropout, + LayerDropModuleList, + LayerNorm, + MultiheadAttention, + PositionalEmbedding, + TransformerSentenceEncoderLayer, +) +from fairseq.modules.quant_noise import quant_noise as apply_quant_noise_ + + +def init_bert_params(module): + """ + Initialize the weights specific to the BERT Model. + This overrides the default initializations depending on the specified arguments. + 1. If normal_init_linear_weights is set then weights of linear + layer will be initialized using the normal distribution and + bais will be set to the specified value. + 2. If normal_init_embed_weights is set then weights of embedding + layer will be initialized using the normal distribution. + 3. If normal_init_proj_weights is set then weights of + in_project_weight for MultiHeadAttention initialized using + the normal distribution (to be validated). + """ + + def normal_(data): + # with FSDP, module params will be on CUDA, so we cast them back to CPU + # so that the RNG is consistent with and without FSDP + data.copy_(data.cpu().normal_(mean=0.0, std=0.02).to(data.device)) + + if isinstance(module, nn.Linear): + normal_(module.weight.data) + if module.bias is not None: + module.bias.data.zero_() + if isinstance(module, nn.Embedding): + normal_(module.weight.data) + if module.padding_idx is not None: + module.weight.data[module.padding_idx].zero_() + if isinstance(module, MultiheadAttention): + normal_(module.q_proj.weight.data) + normal_(module.k_proj.weight.data) + normal_(module.v_proj.weight.data) + + +class TransformerSentenceEncoder(nn.Module): + """ + Implementation for a Bi-directional Transformer based Sentence Encoder used + in BERT/XLM style pre-trained models. + + This first computes the token embedding using the token embedding matrix, + position embeddings (if specified) and segment embeddings + (if specified). After applying the specified number of + TransformerEncoderLayers, it outputs all the internal states of the + encoder as well as the final representation associated with the first + token (usually CLS token). + + Input: + - tokens: B x T matrix representing sentences + - segment_labels: B x T matrix representing segment label for tokens + + Output: + - a tuple of the following: + - a list of internal model states used to compute the + predictions where each tensor has shape T x B x C + - sentence representation associated with first input token + in format B x C. + """ + + def __init__( + self, + padding_idx: int, + vocab_size: int, + num_encoder_layers: int = 6, + embedding_dim: int = 768, + ffn_embedding_dim: int = 3072, + num_attention_heads: int = 8, + dropout: float = 0.1, + attention_dropout: float = 0.1, + activation_dropout: float = 0.1, + layerdrop: float = 0.0, + max_seq_len: int = 256, + num_segments: int = 2, + use_position_embeddings: bool = True, + offset_positions_by_padding: bool = True, + encoder_normalize_before: bool = False, + apply_bert_init: bool = False, + activation_fn: str = "relu", + learned_pos_embedding: bool = True, + embed_scale: float = None, + freeze_embeddings: bool = False, + n_trans_layers_to_freeze: int = 0, + export: bool = False, + traceable: bool = False, + q_noise: float = 0.0, + qn_block_size: int = 8, + ) -> None: + + super().__init__() + self.padding_idx = padding_idx + self.vocab_size = vocab_size + self.dropout_module = FairseqDropout( + dropout, module_name=self.__class__.__name__ + ) + self.layerdrop = layerdrop + self.max_seq_len = max_seq_len + self.embedding_dim = embedding_dim + self.num_segments = num_segments + self.use_position_embeddings = use_position_embeddings + self.apply_bert_init = apply_bert_init + self.learned_pos_embedding = learned_pos_embedding + self.traceable = traceable + + self.embed_tokens = self.build_embedding( + self.vocab_size, self.embedding_dim, self.padding_idx + ) + self.embed_scale = embed_scale + + if q_noise > 0: + self.quant_noise = apply_quant_noise_( + nn.Linear(self.embedding_dim, self.embedding_dim, bias=False), + q_noise, + qn_block_size, + ) + else: + self.quant_noise = None + + self.segment_embeddings = ( + nn.Embedding(self.num_segments, self.embedding_dim, padding_idx=None) + if self.num_segments > 0 + else None + ) + + self.embed_positions = ( + PositionalEmbedding( + self.max_seq_len, + self.embedding_dim, + padding_idx=(self.padding_idx if offset_positions_by_padding else None), + learned=self.learned_pos_embedding, + ) + if self.use_position_embeddings + else None + ) + + if encoder_normalize_before: + self.emb_layer_norm = LayerNorm(self.embedding_dim, export=export) + else: + self.emb_layer_norm = None + + if self.layerdrop > 0.0: + self.layers = LayerDropModuleList(p=self.layerdrop) + else: + self.layers = nn.ModuleList([]) + self.layers.extend( + [ + self.build_transformer_sentence_encoder_layer( + embedding_dim=self.embedding_dim, + ffn_embedding_dim=ffn_embedding_dim, + num_attention_heads=num_attention_heads, + dropout=self.dropout_module.p, + attention_dropout=attention_dropout, + activation_dropout=activation_dropout, + activation_fn=activation_fn, + export=export, + q_noise=q_noise, + qn_block_size=qn_block_size, + ) + for _ in range(num_encoder_layers) + ] + ) + + # Apply initialization of model params after building the model + if self.apply_bert_init: + self.apply(init_bert_params) + + def freeze_module_params(m): + if m is not None: + for p in m.parameters(): + p.requires_grad = False + + if freeze_embeddings: + freeze_module_params(self.embed_tokens) + freeze_module_params(self.segment_embeddings) + freeze_module_params(self.embed_positions) + freeze_module_params(self.emb_layer_norm) + + for layer in range(n_trans_layers_to_freeze): + freeze_module_params(self.layers[layer]) + + def build_embedding(self, vocab_size, embedding_dim, padding_idx): + return nn.Embedding(vocab_size, embedding_dim, padding_idx) + + def build_transformer_sentence_encoder_layer( + self, + embedding_dim, + ffn_embedding_dim, + num_attention_heads, + dropout, + attention_dropout, + activation_dropout, + activation_fn, + export, + q_noise, + qn_block_size, + ): + return TransformerSentenceEncoderLayer( + embedding_dim=embedding_dim, + ffn_embedding_dim=ffn_embedding_dim, + num_attention_heads=num_attention_heads, + dropout=dropout, + attention_dropout=attention_dropout, + activation_dropout=activation_dropout, + activation_fn=activation_fn, + export=export, + q_noise=q_noise, + qn_block_size=qn_block_size, + ) + + def forward( + self, + tokens: torch.Tensor, + segment_labels: torch.Tensor = None, + last_state_only: bool = False, + positions: Optional[torch.Tensor] = None, + token_embeddings: Optional[torch.Tensor] = None, + attn_mask: Optional[torch.Tensor] = None, + ) -> Tuple[torch.Tensor, torch.Tensor]: + is_tpu = tokens.device.type == "xla" + + # compute padding mask. This is needed for multi-head attention + padding_mask = tokens.eq(self.padding_idx) + if not self.traceable and not is_tpu and not padding_mask.any(): + padding_mask = None + + if token_embeddings is not None: + x = token_embeddings + else: + x = self.embed_tokens(tokens) + + if self.embed_scale is not None: + x = x * self.embed_scale + + if self.embed_positions is not None: + x = x + self.embed_positions(tokens, positions=positions) + + if self.segment_embeddings is not None and segment_labels is not None: + x = x + self.segment_embeddings(segment_labels) + + if self.quant_noise is not None: + x = self.quant_noise(x) + + if self.emb_layer_norm is not None: + x = self.emb_layer_norm(x) + + x = self.dropout_module(x) + + # account for padding while computing the representation + if padding_mask is not None: + x = x * (1 - padding_mask.unsqueeze(-1).type_as(x)) + + # B x T x C -> T x B x C + x = x.transpose(0, 1) + + inner_states = [] + if not last_state_only: + inner_states.append(x) + + for layer in self.layers: + x, _ = layer( + x, self_attn_padding_mask=padding_mask, self_attn_mask=attn_mask + ) + if not last_state_only: + inner_states.append(x) + + sentence_rep = x[0, :, :] + + if last_state_only: + inner_states = [x] + + if self.traceable: + return torch.stack(inner_states), sentence_rep + else: + return inner_states, sentence_rep diff --git a/data/fairseq/fairseq/modules/transformer_sentence_encoder_layer.py b/data/fairseq/fairseq/modules/transformer_sentence_encoder_layer.py new file mode 100644 index 0000000000000000000000000000000000000000..f869c4b2f8fb15f96a292e39bd293df7898a4fce --- /dev/null +++ b/data/fairseq/fairseq/modules/transformer_sentence_encoder_layer.py @@ -0,0 +1,139 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from typing import Callable, Optional + +import torch +import torch.nn as nn +from fairseq import utils +from fairseq.modules import LayerNorm, MultiheadAttention +from fairseq.modules.fairseq_dropout import FairseqDropout +from fairseq.modules.quant_noise import quant_noise + + +class TransformerSentenceEncoderLayer(nn.Module): + """ + Implements a Transformer Encoder Layer used in BERT/XLM style pre-trained + models. + """ + + def __init__( + self, + embedding_dim: int = 768, + ffn_embedding_dim: int = 3072, + num_attention_heads: int = 8, + dropout: float = 0.1, + attention_dropout: float = 0.1, + activation_dropout: float = 0.1, + activation_fn: str = "relu", + export: bool = False, + q_noise: float = 0.0, + qn_block_size: int = 8, + init_fn: Callable = None, + ) -> None: + super().__init__() + + if init_fn is not None: + init_fn() + + # Initialize parameters + self.embedding_dim = embedding_dim + self.num_attention_heads = num_attention_heads + self.attention_dropout = attention_dropout + self.q_noise = q_noise + self.qn_block_size = qn_block_size + + self.dropout_module = FairseqDropout( + dropout, module_name=self.__class__.__name__ + ) + self.activation_dropout_module = FairseqDropout( + activation_dropout, module_name=self.__class__.__name__ + ) + + # Initialize blocks + self.activation_fn = utils.get_activation_fn(activation_fn) + self.self_attn = self.build_self_attention( + self.embedding_dim, + num_attention_heads, + dropout=attention_dropout, + self_attention=True, + q_noise=q_noise, + qn_block_size=qn_block_size, + ) + + # layer norm associated with the self attention layer + self.self_attn_layer_norm = LayerNorm(self.embedding_dim, export=export) + + self.fc1 = self.build_fc1( + self.embedding_dim, + ffn_embedding_dim, + q_noise=q_noise, + qn_block_size=qn_block_size, + ) + self.fc2 = self.build_fc2( + ffn_embedding_dim, + self.embedding_dim, + q_noise=q_noise, + qn_block_size=qn_block_size, + ) + + # layer norm associated with the position wise feed-forward NN + self.final_layer_norm = LayerNorm(self.embedding_dim, export=export) + + def build_fc1(self, input_dim, output_dim, q_noise, qn_block_size): + return quant_noise(nn.Linear(input_dim, output_dim), q_noise, qn_block_size) + + def build_fc2(self, input_dim, output_dim, q_noise, qn_block_size): + return quant_noise(nn.Linear(input_dim, output_dim), q_noise, qn_block_size) + + def build_self_attention( + self, + embed_dim, + num_attention_heads, + dropout, + self_attention, + q_noise, + qn_block_size, + ): + return MultiheadAttention( + embed_dim, + num_attention_heads, + dropout=dropout, + self_attention=True, + q_noise=q_noise, + qn_block_size=qn_block_size, + ) + + def forward( + self, + x: torch.Tensor, + self_attn_mask: Optional[torch.Tensor] = None, + self_attn_padding_mask: Optional[torch.Tensor] = None, + ): + """ + LayerNorm is applied either before or after the self-attention/ffn + modules similar to the original Transformer implementation. + """ + residual = x + x, attn = self.self_attn( + query=x, + key=x, + value=x, + key_padding_mask=self_attn_padding_mask, + need_weights=False, + attn_mask=self_attn_mask, + ) + x = self.dropout_module(x) + x = residual + x + x = self.self_attn_layer_norm(x) + + residual = x + x = self.activation_fn(self.fc1(x)) + x = self.activation_dropout_module(x) + x = self.fc2(x) + x = self.dropout_module(x) + x = residual + x + x = self.final_layer_norm(x) + return x, attn diff --git a/data/fairseq/fairseq/modules/transpose_last.py b/data/fairseq/fairseq/modules/transpose_last.py new file mode 100644 index 0000000000000000000000000000000000000000..d7cca9a4bbdb3f455217380f96a2f2d77eae8630 --- /dev/null +++ b/data/fairseq/fairseq/modules/transpose_last.py @@ -0,0 +1,21 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. +""" +transpose last 2 dimensions of the input +""" + +import torch.nn as nn + + +class TransposeLast(nn.Module): + def __init__(self, deconstruct_idx=None, tranpose_dim=-2): + super().__init__() + self.deconstruct_idx = deconstruct_idx + self.tranpose_dim = tranpose_dim + + def forward(self, x): + if self.deconstruct_idx is not None: + x = x[self.deconstruct_idx] + return x.transpose(self.tranpose_dim, -1) diff --git a/data/fairseq/fairseq/modules/unfold.py b/data/fairseq/fairseq/modules/unfold.py new file mode 100644 index 0000000000000000000000000000000000000000..bbaafbd6bfe1d206348a3673da8cae8ef9daeca4 --- /dev/null +++ b/data/fairseq/fairseq/modules/unfold.py @@ -0,0 +1,19 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import torch.nn.functional as F + + +def unfold1d(x, kernel_size: int, padding_l: int, pad_value: float = 0): + """unfold T x B x C to T x B x C x K""" + if kernel_size > 1: + T, B, C = x.size() + x = F.pad( + x, (0, 0, 0, 0, padding_l, kernel_size - 1 - padding_l), value=pad_value + ) + x = x.as_strided((T, B, C, kernel_size), (B * C, C, 1, B * C)) + else: + x = x.unsqueeze(3) + return x diff --git a/data/fairseq/fairseq/modules/vggblock.py b/data/fairseq/fairseq/modules/vggblock.py new file mode 100644 index 0000000000000000000000000000000000000000..ee5ee19a34816c7350c21fba7c4907fec8ca7a61 --- /dev/null +++ b/data/fairseq/fairseq/modules/vggblock.py @@ -0,0 +1,116 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from __future__ import absolute_import, division, print_function, unicode_literals + +from collections.abc import Iterable +from itertools import repeat + +import torch +import torch.nn as nn + + +def _pair(v): + if isinstance(v, Iterable): + assert len(v) == 2, "len(v) != 2" + return v + return tuple(repeat(v, 2)) + + +def infer_conv_output_dim(conv_op, input_dim, sample_inchannel): + sample_seq_len = 200 + sample_bsz = 10 + x = torch.randn(sample_bsz, sample_inchannel, sample_seq_len, input_dim) + # N x C x H x W + # N: sample_bsz, C: sample_inchannel, H: sample_seq_len, W: input_dim + x = conv_op(x) + # N x C x H x W + x = x.transpose(1, 2) + # N x H x C x W + bsz, seq = x.size()[:2] + per_channel_dim = x.size()[3] + # bsz: N, seq: H, CxW the rest + return x.contiguous().view(bsz, seq, -1).size(-1), per_channel_dim + + +class VGGBlock(torch.nn.Module): + """ + VGG motibated cnn module https://arxiv.org/pdf/1409.1556.pdf + + Args: + in_channels: (int) number of input channels (typically 1) + out_channels: (int) number of output channels + conv_kernel_size: convolution channels + pooling_kernel_size: the size of the pooling window to take a max over + num_conv_layers: (int) number of convolution layers + input_dim: (int) input dimension + conv_stride: the stride of the convolving kernel. + Can be a single number or a tuple (sH, sW) Default: 1 + padding: implicit paddings on both sides of the input. + Can be a single number or a tuple (padH, padW). Default: None + layer_norm: (bool) if layer norm is going to be applied. Default: False + + Shape: + Input: BxCxTxfeat, i.e. (batch_size, input_size, timesteps, features) + Output: BxCxTxfeat, i.e. (batch_size, input_size, timesteps, features) + """ + + def __init__( + self, + in_channels, + out_channels, + conv_kernel_size, + pooling_kernel_size, + num_conv_layers, + input_dim, + conv_stride=1, + padding=None, + layer_norm=False, + ): + assert ( + input_dim is not None + ), "Need input_dim for LayerNorm and infer_conv_output_dim" + super(VGGBlock, self).__init__() + self.in_channels = in_channels + self.out_channels = out_channels + self.conv_kernel_size = _pair(conv_kernel_size) + self.pooling_kernel_size = _pair(pooling_kernel_size) + self.num_conv_layers = num_conv_layers + self.padding = ( + tuple(e // 2 for e in self.conv_kernel_size) + if padding is None + else _pair(padding) + ) + self.conv_stride = _pair(conv_stride) + + self.layers = nn.ModuleList() + for layer in range(num_conv_layers): + conv_op = nn.Conv2d( + in_channels if layer == 0 else out_channels, + out_channels, + self.conv_kernel_size, + stride=self.conv_stride, + padding=self.padding, + ) + self.layers.append(conv_op) + if layer_norm: + conv_output_dim, per_channel_dim = infer_conv_output_dim( + conv_op, input_dim, in_channels if layer == 0 else out_channels + ) + self.layers.append(nn.LayerNorm(per_channel_dim)) + input_dim = per_channel_dim + self.layers.append(nn.ReLU()) + + if self.pooling_kernel_size is not None: + pool_op = nn.MaxPool2d(kernel_size=self.pooling_kernel_size, ceil_mode=True) + self.layers.append(pool_op) + self.total_output_dim, self.output_dim = infer_conv_output_dim( + pool_op, input_dim, out_channels + ) + + def forward(self, x): + for i, _ in enumerate(self.layers): + x = self.layers[i](x) + return x diff --git a/data/fairseq/fairseq/scoring/__init__.py b/data/fairseq/fairseq/scoring/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..58f2f563e493327394dff1265030d18f0814b5a2 --- /dev/null +++ b/data/fairseq/fairseq/scoring/__init__.py @@ -0,0 +1,55 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + + +import importlib +import os +from abc import ABC, abstractmethod + +from fairseq import registry +from omegaconf import DictConfig + + +class BaseScorer(ABC): + def __init__(self, cfg): + self.cfg = cfg + self.ref = [] + self.pred = [] + + def add_string(self, ref, pred): + self.ref.append(ref) + self.pred.append(pred) + + @abstractmethod + def score(self) -> float: + pass + + @abstractmethod + def result_string(self) -> str: + pass + + +_build_scorer, register_scorer, SCORER_REGISTRY, _ = registry.setup_registry( + "--scoring", default="bleu" +) + + +def build_scorer(choice, tgt_dict): + _choice = choice._name if isinstance(choice, DictConfig) else choice + + if _choice == "bleu": + from fairseq.scoring import bleu + + return bleu.Scorer( + bleu.BleuConfig(pad=tgt_dict.pad(), eos=tgt_dict.eos(), unk=tgt_dict.unk()) + ) + return _build_scorer(choice) + + +# automatically import any Python files in the current directory +for file in sorted(os.listdir(os.path.dirname(__file__))): + if file.endswith(".py") and not file.startswith("_"): + module = file[: file.find(".py")] + importlib.import_module("fairseq.scoring." + module) diff --git a/data/fairseq/fairseq/scoring/__pycache__/__init__.cpython-38.pyc b/data/fairseq/fairseq/scoring/__pycache__/__init__.cpython-38.pyc new file mode 100644 index 0000000000000000000000000000000000000000..5b7ae1ac5d25fa81f14c73726662defb3fcca6b4 Binary files /dev/null and b/data/fairseq/fairseq/scoring/__pycache__/__init__.cpython-38.pyc differ diff --git a/data/fairseq/fairseq/scoring/__pycache__/bertscore.cpython-38.pyc b/data/fairseq/fairseq/scoring/__pycache__/bertscore.cpython-38.pyc new file mode 100644 index 0000000000000000000000000000000000000000..e1dea994eaf44750c6f027629e5167c83eb63648 Binary files /dev/null and b/data/fairseq/fairseq/scoring/__pycache__/bertscore.cpython-38.pyc differ diff --git a/data/fairseq/fairseq/scoring/__pycache__/bleu.cpython-38.pyc b/data/fairseq/fairseq/scoring/__pycache__/bleu.cpython-38.pyc new file mode 100644 index 0000000000000000000000000000000000000000..04e84207810c5e1f5dffe3356672e2d2275afe6c Binary files /dev/null and b/data/fairseq/fairseq/scoring/__pycache__/bleu.cpython-38.pyc differ diff --git a/data/fairseq/fairseq/scoring/__pycache__/chrf.cpython-38.pyc b/data/fairseq/fairseq/scoring/__pycache__/chrf.cpython-38.pyc new file mode 100644 index 0000000000000000000000000000000000000000..c79eb44ff1c318ff41f3c0faeaf3c7969dd831bc Binary files /dev/null and b/data/fairseq/fairseq/scoring/__pycache__/chrf.cpython-38.pyc differ diff --git a/data/fairseq/fairseq/scoring/__pycache__/meteor.cpython-38.pyc b/data/fairseq/fairseq/scoring/__pycache__/meteor.cpython-38.pyc new file mode 100644 index 0000000000000000000000000000000000000000..2ef900e79ee6e37c81a980abaa76ccda6abe8d22 Binary files /dev/null and b/data/fairseq/fairseq/scoring/__pycache__/meteor.cpython-38.pyc differ diff --git a/data/fairseq/fairseq/scoring/__pycache__/tokenizer.cpython-38.pyc b/data/fairseq/fairseq/scoring/__pycache__/tokenizer.cpython-38.pyc new file mode 100644 index 0000000000000000000000000000000000000000..b185c71859cd677efd6802e5d57a2c2075c473a2 Binary files /dev/null and b/data/fairseq/fairseq/scoring/__pycache__/tokenizer.cpython-38.pyc differ diff --git a/data/fairseq/fairseq/scoring/__pycache__/wer.cpython-38.pyc b/data/fairseq/fairseq/scoring/__pycache__/wer.cpython-38.pyc new file mode 100644 index 0000000000000000000000000000000000000000..5f58be6af2d5838681ef891b5b49f5d393b10431 Binary files /dev/null and b/data/fairseq/fairseq/scoring/__pycache__/wer.cpython-38.pyc differ diff --git a/data/fairseq/fairseq/scoring/bertscore.py b/data/fairseq/fairseq/scoring/bertscore.py new file mode 100644 index 0000000000000000000000000000000000000000..6d5a8450d34a7f76ba08748f61f5d84fcf5570ee --- /dev/null +++ b/data/fairseq/fairseq/scoring/bertscore.py @@ -0,0 +1,44 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from dataclasses import dataclass, field + +import numpy as np + +from fairseq.dataclass import FairseqDataclass +from fairseq.scoring import BaseScorer, register_scorer + + +@dataclass +class BertScoreScorerConfig(FairseqDataclass): + bert_score_lang: str = field(default="en", metadata={"help": "BERTScore language"}) + + +@register_scorer("bert_score", dataclass=BertScoreScorerConfig) +class BertScoreScorer(BaseScorer): + def __init__(self, cfg): + super(BertScoreScorer, self).__init__(cfg) + try: + import bert_score as _bert_score + except ImportError: + raise ImportError("Please install BERTScore: pip install bert-score") + + self.cfg = cfg + self._bert_score = _bert_score + self.scores = None + + def add_string(self, ref, pred): + self.ref.append(ref) + self.pred.append(pred) + + def score(self, order=4): + _, _, self.scores = self._bert_score.score( + self.pred, self.ref, lang=self.cfg.bert_score_lang + ) + self.scores = self.scores.numpy() + return np.mean(self.scores) + + def result_string(self, order=4): + return f"BERTScore: {self.score():.4f}" diff --git a/data/fairseq/fairseq/scoring/bleu.py b/data/fairseq/fairseq/scoring/bleu.py new file mode 100644 index 0000000000000000000000000000000000000000..e55bd2f393e279d318e08d680c41cc20e8d49931 --- /dev/null +++ b/data/fairseq/fairseq/scoring/bleu.py @@ -0,0 +1,168 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import ctypes +import math +import sys +from dataclasses import dataclass, field + +import torch +from fairseq.dataclass import FairseqDataclass +from fairseq.scoring import BaseScorer, register_scorer +from fairseq.scoring.tokenizer import EvaluationTokenizer + + +class BleuStat(ctypes.Structure): + _fields_ = [ + ("reflen", ctypes.c_size_t), + ("predlen", ctypes.c_size_t), + ("match1", ctypes.c_size_t), + ("count1", ctypes.c_size_t), + ("match2", ctypes.c_size_t), + ("count2", ctypes.c_size_t), + ("match3", ctypes.c_size_t), + ("count3", ctypes.c_size_t), + ("match4", ctypes.c_size_t), + ("count4", ctypes.c_size_t), + ] + + +@dataclass +class SacrebleuConfig(FairseqDataclass): + sacrebleu_tokenizer: EvaluationTokenizer.ALL_TOKENIZER_TYPES = field( + default="13a", metadata={"help": "tokenizer"} + ) + sacrebleu_lowercase: bool = field( + default=False, metadata={"help": "apply lowercasing"} + ) + sacrebleu_char_level: bool = field( + default=False, metadata={"help": "evaluate at character level"} + ) + + +@register_scorer("sacrebleu", dataclass=SacrebleuConfig) +class SacrebleuScorer(BaseScorer): + def __init__(self, cfg): + super(SacrebleuScorer, self).__init__(cfg) + import sacrebleu + + self.sacrebleu = sacrebleu + self.tokenizer = EvaluationTokenizer( + tokenizer_type=cfg.sacrebleu_tokenizer, + lowercase=cfg.sacrebleu_lowercase, + character_tokenization=cfg.sacrebleu_char_level, + ) + + def add_string(self, ref, pred): + self.ref.append(self.tokenizer.tokenize(ref)) + self.pred.append(self.tokenizer.tokenize(pred)) + + def _score(self, order=4): + if order != 4: + raise NotImplementedError + # tokenization and lowercasing are performed by self.tokenizer instead. + return self.sacrebleu.corpus_bleu(self.pred, [self.ref], tokenize="none") + + def score(self, order=4): + return self._score(order).score + + def result_string(self, order=4): + return self._score(order).format() + + +@dataclass +class BleuConfig(FairseqDataclass): + pad: int = field(default=1, metadata={"help": "padding index"}) + eos: int = field(default=2, metadata={"help": "eos index"}) + unk: int = field(default=3, metadata={"help": "unk index"}) + + +@register_scorer("bleu", dataclass=BleuConfig) +class Scorer(object): + def __init__(self, cfg): + self.stat = BleuStat() + self.pad = cfg.pad + self.eos = cfg.eos + self.unk = cfg.unk + + try: + from fairseq import libbleu + except ImportError as e: + sys.stderr.write( + "ERROR: missing libbleu.so. run `pip install --editable .`\n" + ) + raise e + + self.C = ctypes.cdll.LoadLibrary(libbleu.__file__) + + self.reset() + + def reset(self, one_init=False): + if one_init: + self.C.bleu_one_init(ctypes.byref(self.stat)) + else: + self.C.bleu_zero_init(ctypes.byref(self.stat)) + + def add(self, ref, pred): + if not isinstance(ref, torch.IntTensor): + raise TypeError("ref must be a torch.IntTensor (got {})".format(type(ref))) + if not isinstance(pred, torch.IntTensor): + raise TypeError("pred must be a torch.IntTensor(got {})".format(type(pred))) + + # don't match unknown words + rref = ref.clone() + assert not rref.lt(0).any() + rref[rref.eq(self.unk)] = -999 + + rref = rref.contiguous().view(-1) + pred = pred.contiguous().view(-1) + + self.C.bleu_add( + ctypes.byref(self.stat), + ctypes.c_size_t(rref.size(0)), + ctypes.c_void_p(rref.data_ptr()), + ctypes.c_size_t(pred.size(0)), + ctypes.c_void_p(pred.data_ptr()), + ctypes.c_int(self.pad), + ctypes.c_int(self.eos), + ) + + def score(self, order=4): + psum = sum( + math.log(p) if p > 0 else float("-Inf") for p in self.precision()[:order] + ) + return self.brevity() * math.exp(psum / order) * 100 + + def precision(self): + def ratio(a, b): + return a / b if b > 0 else 0 + + return [ + ratio(self.stat.match1, self.stat.count1), + ratio(self.stat.match2, self.stat.count2), + ratio(self.stat.match3, self.stat.count3), + ratio(self.stat.match4, self.stat.count4), + ] + + def brevity(self): + r = self.stat.reflen / self.stat.predlen + return min(1, math.exp(1 - r)) + + def result_string(self, order=4): + assert order <= 4, "BLEU scores for order > 4 aren't supported" + fmt = "BLEU{} = {:2.2f}, {:2.1f}" + for _ in range(1, order): + fmt += "/{:2.1f}" + fmt += " (BP={:.3f}, ratio={:.3f}, syslen={}, reflen={})" + bleup = [p * 100 for p in self.precision()[:order]] + return fmt.format( + order, + self.score(order=order), + *bleup, + self.brevity(), + self.stat.predlen / self.stat.reflen, + self.stat.predlen, + self.stat.reflen + ) diff --git a/data/fairseq/fairseq/scoring/chrf.py b/data/fairseq/fairseq/scoring/chrf.py new file mode 100644 index 0000000000000000000000000000000000000000..5df5a1c011243fe2e836c38a5f8459aeb824f0e7 --- /dev/null +++ b/data/fairseq/fairseq/scoring/chrf.py @@ -0,0 +1,36 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + + +from dataclasses import dataclass + +from fairseq.dataclass import FairseqDataclass +from fairseq.scoring import BaseScorer, register_scorer + + +@dataclass +class ChrFScorerConfig(FairseqDataclass): + pass + + +@register_scorer("chrf", dataclass=ChrFScorerConfig) +class ChrFScorer(BaseScorer): + def __init__(self, args): + super(ChrFScorer, self).__init__(args) + import sacrebleu + + self.sacrebleu = sacrebleu + + def add_string(self, ref, pred): + self.ref.append(ref) + self.pred.append(pred) + + def score(self, order=4): + return self.result_string(order).score + + def result_string(self, order=4): + if order != 4: + raise NotImplementedError + return self.sacrebleu.corpus_chrf(self.pred, [self.ref]).format() diff --git a/data/fairseq/fairseq/scoring/meteor.py b/data/fairseq/fairseq/scoring/meteor.py new file mode 100644 index 0000000000000000000000000000000000000000..32719956fe32cc25ab071aa7ff6abf84c6533fe8 --- /dev/null +++ b/data/fairseq/fairseq/scoring/meteor.py @@ -0,0 +1,42 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import numpy as np +from dataclasses import dataclass + +from fairseq.dataclass import FairseqDataclass +from fairseq.scoring import BaseScorer, register_scorer + + +@dataclass +class MeteorScorerConfig(FairseqDataclass): + pass + + +@register_scorer("meteor", dataclass=MeteorScorerConfig) +class MeteorScorer(BaseScorer): + def __init__(self, args): + super(MeteorScorer, self).__init__(args) + try: + import nltk + except ImportError: + raise ImportError("Please install nltk to use METEOR scorer") + + self.nltk = nltk + self.scores = [] + + def add_string(self, ref, pred): + self.ref.append(ref) + self.pred.append(pred) + + def score(self, order=4): + self.scores = [ + self.nltk.translate.meteor_score.single_meteor_score(r, p) + for r, p in zip(self.ref, self.pred) + ] + return np.mean(self.scores) + + def result_string(self, order=4): + return f"METEOR: {self.score():.4f}" diff --git a/data/fairseq/fairseq/scoring/tokenizer.py b/data/fairseq/fairseq/scoring/tokenizer.py new file mode 100644 index 0000000000000000000000000000000000000000..b0cedd5099b9ddc7278205e1a4f2aa18359a14bf --- /dev/null +++ b/data/fairseq/fairseq/scoring/tokenizer.py @@ -0,0 +1,80 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import unicodedata + +import sacrebleu as sb + +from fairseq.dataclass import ChoiceEnum + +SACREBLEU_V2_ABOVE = int(sb.__version__[0]) >= 2 + + +class EvaluationTokenizer(object): + """A generic evaluation-time tokenizer, which leverages built-in tokenizers + in sacreBLEU (https://github.com/mjpost/sacrebleu). It additionally provides + lowercasing, punctuation removal and character tokenization, which are + applied after sacreBLEU tokenization. + + Args: + tokenizer_type (str): the type of sacreBLEU tokenizer to apply. + lowercase (bool): lowercase the text. + punctuation_removal (bool): remove punctuation (based on unicode + category) from text. + character_tokenization (bool): tokenize the text to characters. + """ + + SPACE = chr(32) + SPACE_ESCAPE = chr(9601) + _ALL_TOKENIZER_TYPES = ( + sb.BLEU.TOKENIZERS + if SACREBLEU_V2_ABOVE + else ["none", "13a", "intl", "zh", "ja-mecab"] + ) + ALL_TOKENIZER_TYPES = ChoiceEnum(_ALL_TOKENIZER_TYPES) + + def __init__( + self, + tokenizer_type: str = "13a", + lowercase: bool = False, + punctuation_removal: bool = False, + character_tokenization: bool = False, + ): + + assert ( + tokenizer_type in self._ALL_TOKENIZER_TYPES + ), f"{tokenizer_type}, {self._ALL_TOKENIZER_TYPES}" + self.lowercase = lowercase + self.punctuation_removal = punctuation_removal + self.character_tokenization = character_tokenization + if SACREBLEU_V2_ABOVE: + self.tokenizer = sb.BLEU(tokenize=str(tokenizer_type)).tokenizer + else: + self.tokenizer = sb.tokenizers.TOKENIZERS[tokenizer_type]() + + @classmethod + def remove_punctuation(cls, sent: str): + """Remove punctuation based on Unicode category.""" + return cls.SPACE.join( + t + for t in sent.split(cls.SPACE) + if not all(unicodedata.category(c)[0] == "P" for c in t) + ) + + def tokenize(self, sent: str): + tokenized = self.tokenizer(sent) + + if self.punctuation_removal: + tokenized = self.remove_punctuation(tokenized) + + if self.character_tokenization: + tokenized = self.SPACE.join( + list(tokenized.replace(self.SPACE, self.SPACE_ESCAPE)) + ) + + if self.lowercase: + tokenized = tokenized.lower() + + return tokenized diff --git a/data/fairseq/fairseq/scoring/wer.py b/data/fairseq/fairseq/scoring/wer.py new file mode 100644 index 0000000000000000000000000000000000000000..633dc47c247691c4c9e36cbdbab7d7cb74b38452 --- /dev/null +++ b/data/fairseq/fairseq/scoring/wer.py @@ -0,0 +1,58 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +from dataclasses import dataclass, field + +from fairseq.dataclass import FairseqDataclass +from fairseq.scoring import BaseScorer, register_scorer +from fairseq.scoring.tokenizer import EvaluationTokenizer + + +@dataclass +class WerScorerConfig(FairseqDataclass): + wer_tokenizer: EvaluationTokenizer.ALL_TOKENIZER_TYPES = field( + default="none", metadata={"help": "sacreBLEU tokenizer to use for evaluation"} + ) + wer_remove_punct: bool = field( + default=False, metadata={"help": "remove punctuation"} + ) + wer_char_level: bool = field( + default=False, metadata={"help": "evaluate at character level"} + ) + wer_lowercase: bool = field(default=False, metadata={"help": "lowercasing"}) + + +@register_scorer("wer", dataclass=WerScorerConfig) +class WerScorer(BaseScorer): + def __init__(self, cfg): + super().__init__(cfg) + self.reset() + try: + import editdistance as ed + except ImportError: + raise ImportError("Please install editdistance to use WER scorer") + self.ed = ed + self.tokenizer = EvaluationTokenizer( + tokenizer_type=self.cfg.wer_tokenizer, + lowercase=self.cfg.wer_lowercase, + punctuation_removal=self.cfg.wer_remove_punct, + character_tokenization=self.cfg.wer_char_level, + ) + + def reset(self): + self.distance = 0 + self.ref_length = 0 + + def add_string(self, ref, pred): + ref_items = self.tokenizer.tokenize(ref).split() + pred_items = self.tokenizer.tokenize(pred).split() + self.distance += self.ed.eval(ref_items, pred_items) + self.ref_length += len(ref_items) + + def result_string(self): + return f"WER: {self.score():.2f}" + + def score(self): + return 100.0 * self.distance / self.ref_length if self.ref_length > 0 else 0