File size: 13,007 Bytes
c14c4fe | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 | from typing import List
from typing import Optional
from typing import Sequence
from typing import Tuple
from typing import Union
import logging
import torch
import torch.nn as nn
from funasr.models.scama.chunk_utilis import overlap_chunk
import numpy as np
from funasr.models.transformer.utils.nets_utils import make_pad_mask
from funasr.models.sond.attention import MultiHeadSelfAttention
from funasr.models.transformer.embedding import SinusoidalPositionEncoder
from funasr.models.transformer.layer_norm import LayerNorm
from funasr.models.transformer.utils.multi_layer_conv import Conv1dLinear
from funasr.models.transformer.utils.multi_layer_conv import MultiLayeredConv1d
from funasr.models.transformer.positionwise_feed_forward import (
PositionwiseFeedForward, # noqa: H301
)
from funasr.models.transformer.utils.repeat import repeat
from funasr.models.transformer.utils.subsampling import Conv2dSubsampling
from funasr.models.transformer.utils.subsampling import Conv2dSubsampling2
from funasr.models.transformer.utils.subsampling import Conv2dSubsampling6
from funasr.models.transformer.utils.subsampling import Conv2dSubsampling8
from funasr.models.transformer.utils.subsampling import TooShortUttError
from funasr.models.transformer.utils.subsampling import check_short_utt
from funasr.models.ctc import CTC
from funasr.models.encoder.abs_encoder import AbsEncoder
class EncoderLayer(nn.Module):
def __init__(
self,
in_size,
size,
self_attn,
feed_forward,
dropout_rate,
normalize_before=True,
concat_after=False,
stochastic_depth_rate=0.0,
):
"""Construct an EncoderLayer object."""
super(EncoderLayer, self).__init__()
self.self_attn = self_attn
self.feed_forward = feed_forward
self.norm1 = LayerNorm(in_size)
self.norm2 = LayerNorm(size)
self.dropout = nn.Dropout(dropout_rate)
self.in_size = in_size
self.size = size
self.normalize_before = normalize_before
self.concat_after = concat_after
if self.concat_after:
self.concat_linear = nn.Linear(size + size, size)
self.stochastic_depth_rate = stochastic_depth_rate
self.dropout_rate = dropout_rate
def forward(self, x, mask, cache=None, mask_att_chunk_encoder=None):
"""Compute encoded features.
Args:
x_input (torch.Tensor): Input tensor (#batch, time, size).
mask (torch.Tensor): Mask tensor for the input (#batch, time).
cache (torch.Tensor): Cache tensor of the input (#batch, time - 1, size).
Returns:
torch.Tensor: Output tensor (#batch, time, size).
torch.Tensor: Mask tensor (#batch, time).
"""
skip_layer = False
# with stochastic depth, residual connection `x + f(x)` becomes
# `x <- x + 1 / (1 - p) * f(x)` at training time.
stoch_layer_coeff = 1.0
if self.training and self.stochastic_depth_rate > 0:
skip_layer = torch.rand(1).item() < self.stochastic_depth_rate
stoch_layer_coeff = 1.0 / (1 - self.stochastic_depth_rate)
if skip_layer:
if cache is not None:
x = torch.cat([cache, x], dim=1)
return x, mask
residual = x
if self.normalize_before:
x = self.norm1(x)
if self.concat_after:
x_concat = torch.cat(
(x, self.self_attn(x, mask, mask_att_chunk_encoder=mask_att_chunk_encoder)), dim=-1
)
if self.in_size == self.size:
x = residual + stoch_layer_coeff * self.concat_linear(x_concat)
else:
x = stoch_layer_coeff * self.concat_linear(x_concat)
else:
if self.in_size == self.size:
x = residual + stoch_layer_coeff * self.dropout(
self.self_attn(x, mask, mask_att_chunk_encoder=mask_att_chunk_encoder)
)
else:
x = stoch_layer_coeff * self.dropout(
self.self_attn(x, mask, mask_att_chunk_encoder=mask_att_chunk_encoder)
)
if not self.normalize_before:
x = self.norm1(x)
residual = x
if self.normalize_before:
x = self.norm2(x)
x = residual + stoch_layer_coeff * self.dropout(self.feed_forward(x))
if not self.normalize_before:
x = self.norm2(x)
return x, mask, cache, mask_att_chunk_encoder
class SelfAttentionEncoder(AbsEncoder):
"""
Author: Speech Lab of DAMO Academy, Alibaba Group
Self attention encoder in OpenNMT framework
"""
def __init__(
self,
input_size: int,
output_size: int = 256,
attention_heads: int = 4,
linear_units: int = 2048,
num_blocks: int = 6,
dropout_rate: float = 0.1,
positional_dropout_rate: float = 0.1,
attention_dropout_rate: float = 0.0,
input_layer: Optional[str] = "conv2d",
pos_enc_class=SinusoidalPositionEncoder,
normalize_before: bool = True,
concat_after: bool = False,
positionwise_layer_type: str = "linear",
positionwise_conv_kernel_size: int = 1,
padding_idx: int = -1,
interctc_layer_idx: List[int] = [],
interctc_use_conditioning: bool = False,
tf2torch_tensor_name_prefix_torch: str = "encoder",
tf2torch_tensor_name_prefix_tf: str = "seq2seq/encoder",
out_units=None,
):
super().__init__()
self._output_size = output_size
if input_layer == "linear":
self.embed = torch.nn.Sequential(
torch.nn.Linear(input_size, output_size),
torch.nn.LayerNorm(output_size),
torch.nn.Dropout(dropout_rate),
torch.nn.ReLU(),
pos_enc_class(output_size, positional_dropout_rate),
)
elif input_layer == "conv2d":
self.embed = Conv2dSubsampling(input_size, output_size, dropout_rate)
elif input_layer == "conv2d2":
self.embed = Conv2dSubsampling2(input_size, output_size, dropout_rate)
elif input_layer == "conv2d6":
self.embed = Conv2dSubsampling6(input_size, output_size, dropout_rate)
elif input_layer == "conv2d8":
self.embed = Conv2dSubsampling8(input_size, output_size, dropout_rate)
elif input_layer == "embed":
self.embed = torch.nn.Sequential(
torch.nn.Embedding(input_size, output_size, padding_idx=padding_idx),
SinusoidalPositionEncoder(),
)
elif input_layer is None:
if input_size == output_size:
self.embed = None
else:
self.embed = torch.nn.Linear(input_size, output_size)
elif input_layer == "pe":
self.embed = SinusoidalPositionEncoder()
elif input_layer == "null":
self.embed = None
else:
raise ValueError("unknown input_layer: " + input_layer)
self.normalize_before = normalize_before
if positionwise_layer_type == "linear":
positionwise_layer = PositionwiseFeedForward
positionwise_layer_args = (
output_size,
linear_units,
dropout_rate,
)
elif positionwise_layer_type == "conv1d":
positionwise_layer = MultiLayeredConv1d
positionwise_layer_args = (
output_size,
linear_units,
positionwise_conv_kernel_size,
dropout_rate,
)
elif positionwise_layer_type == "conv1d-linear":
positionwise_layer = Conv1dLinear
positionwise_layer_args = (
output_size,
linear_units,
positionwise_conv_kernel_size,
dropout_rate,
)
else:
raise NotImplementedError("Support only linear or conv1d.")
self.encoders = repeat(
num_blocks,
lambda lnum: (
EncoderLayer(
output_size,
output_size,
MultiHeadSelfAttention(
attention_heads,
output_size,
output_size,
attention_dropout_rate,
),
positionwise_layer(*positionwise_layer_args),
dropout_rate,
normalize_before,
concat_after,
)
if lnum > 0
else EncoderLayer(
input_size,
output_size,
MultiHeadSelfAttention(
attention_heads,
input_size if input_layer == "pe" or input_layer == "null" else output_size,
output_size,
attention_dropout_rate,
),
positionwise_layer(*positionwise_layer_args),
dropout_rate,
normalize_before,
concat_after,
)
),
)
if self.normalize_before:
self.after_norm = LayerNorm(output_size)
self.interctc_layer_idx = interctc_layer_idx
if len(interctc_layer_idx) > 0:
assert 0 < min(interctc_layer_idx) and max(interctc_layer_idx) < num_blocks
self.interctc_use_conditioning = interctc_use_conditioning
self.conditioning_layer = None
self.dropout = nn.Dropout(dropout_rate)
self.tf2torch_tensor_name_prefix_torch = tf2torch_tensor_name_prefix_torch
self.tf2torch_tensor_name_prefix_tf = tf2torch_tensor_name_prefix_tf
self.out_units = out_units
if out_units is not None:
self.output_linear = nn.Linear(output_size, out_units)
def output_size(self) -> int:
return self._output_size
def forward(
self,
xs_pad: torch.Tensor,
ilens: torch.Tensor,
prev_states: torch.Tensor = None,
ctc: CTC = None,
) -> Tuple[torch.Tensor, torch.Tensor, Optional[torch.Tensor]]:
"""Embed positions in tensor.
Args:
xs_pad: input tensor (B, L, D)
ilens: input length (B)
prev_states: Not to be used now.
Returns:
position embedded tensor and mask
"""
masks = (~make_pad_mask(ilens)[:, None, :]).to(xs_pad.device)
xs_pad = xs_pad * self.output_size() ** 0.5
if self.embed is None:
xs_pad = xs_pad
elif (
isinstance(self.embed, Conv2dSubsampling)
or isinstance(self.embed, Conv2dSubsampling2)
or isinstance(self.embed, Conv2dSubsampling6)
or isinstance(self.embed, Conv2dSubsampling8)
):
short_status, limit_size = check_short_utt(self.embed, xs_pad.size(1))
if short_status:
raise TooShortUttError(
f"has {xs_pad.size(1)} frames and is too short for subsampling "
+ f"(it needs more than {limit_size} frames), return empty results",
xs_pad.size(1),
limit_size,
)
xs_pad, masks = self.embed(xs_pad, masks)
else:
xs_pad = self.embed(xs_pad)
xs_pad = self.dropout(xs_pad)
# encoder_outs = self.encoders0(xs_pad, masks)
# xs_pad, masks = encoder_outs[0], encoder_outs[1]
intermediate_outs = []
if len(self.interctc_layer_idx) == 0:
encoder_outs = self.encoders(xs_pad, masks)
xs_pad, masks = encoder_outs[0], encoder_outs[1]
else:
for layer_idx, encoder_layer in enumerate(self.encoders):
encoder_outs = encoder_layer(xs_pad, masks)
xs_pad, masks = encoder_outs[0], encoder_outs[1]
if layer_idx + 1 in self.interctc_layer_idx:
encoder_out = xs_pad
# intermediate outputs are also normalized
if self.normalize_before:
encoder_out = self.after_norm(encoder_out)
intermediate_outs.append((layer_idx + 1, encoder_out))
if self.interctc_use_conditioning:
ctc_out = ctc.softmax(encoder_out)
xs_pad = xs_pad + self.conditioning_layer(ctc_out)
if self.normalize_before:
xs_pad = self.after_norm(xs_pad)
if self.out_units is not None:
xs_pad = self.output_linear(xs_pad)
olens = masks.squeeze(1).sum(1)
if len(intermediate_outs) > 0:
return (xs_pad, intermediate_outs), olens, None
return xs_pad, olens, None
|