Delete src/miocodec/module/ssl_extractor.py with huggingface_hub
Browse files
src/miocodec/module/ssl_extractor.py
DELETED
|
@@ -1,106 +0,0 @@
|
|
| 1 |
-
import torch
|
| 2 |
-
import torch.nn as nn
|
| 3 |
-
import torchaudio
|
| 4 |
-
import torchaudio.pipelines as pipelines
|
| 5 |
-
from torchaudio.models.wav2vec2 import Wav2Vec2Model
|
| 6 |
-
from torchaudio.models.wav2vec2.components import ConvLayerBlock
|
| 7 |
-
|
| 8 |
-
from ..util import get_logger
|
| 9 |
-
|
| 10 |
-
logger = get_logger()
|
| 11 |
-
|
| 12 |
-
|
| 13 |
-
# Map of friendly names to torchaudio pipeline bundles
|
| 14 |
-
MODEL_REGISTRY = {
|
| 15 |
-
"wav2vec2_base": pipelines.WAV2VEC2_BASE,
|
| 16 |
-
"wav2vec2_large": pipelines.WAV2VEC2_LARGE,
|
| 17 |
-
"wav2vec2_large_lv60k": pipelines.WAV2VEC2_LARGE_LV60K,
|
| 18 |
-
"hubert_base": pipelines.HUBERT_BASE,
|
| 19 |
-
"hubert_large": pipelines.HUBERT_LARGE,
|
| 20 |
-
"hubert_xlarge": pipelines.HUBERT_XLARGE,
|
| 21 |
-
"wavlm_base": pipelines.WAVLM_BASE,
|
| 22 |
-
"wavlm_base_plus": pipelines.WAVLM_BASE_PLUS,
|
| 23 |
-
"wavlm_large": pipelines.WAVLM_LARGE,
|
| 24 |
-
}
|
| 25 |
-
|
| 26 |
-
|
| 27 |
-
class SSLFeatureExtractor(nn.Module):
|
| 28 |
-
def __init__(self, model_name: str = "wavlm_base_plus", output_layer: int | None = None, sample_rate: int = 16000):
|
| 29 |
-
"""
|
| 30 |
-
Args:
|
| 31 |
-
model_name: Name of the SSL model to use
|
| 32 |
-
output_layer: Which layer's features to extract (None for last layer), 1-based indexing
|
| 33 |
-
sample_rate: Sample rate of input audio
|
| 34 |
-
"""
|
| 35 |
-
super().__init__()
|
| 36 |
-
self.output_layer = output_layer if output_layer is not None else -1
|
| 37 |
-
|
| 38 |
-
if model_name not in MODEL_REGISTRY:
|
| 39 |
-
raise ValueError(f"Unknown model: {model_name}. Available models: {list(MODEL_REGISTRY.keys())}")
|
| 40 |
-
bundle = MODEL_REGISTRY[model_name]
|
| 41 |
-
self.model: Wav2Vec2Model = bundle.get_model()
|
| 42 |
-
self.model.eval()
|
| 43 |
-
self.feature_dim: int = bundle._params["encoder_embed_dim"]
|
| 44 |
-
|
| 45 |
-
self.ssl_sample_rate = bundle.sample_rate
|
| 46 |
-
# Create resampler if needed
|
| 47 |
-
if sample_rate != self.ssl_sample_rate:
|
| 48 |
-
logger.debug(f"Resampling from {sample_rate} to {self.ssl_sample_rate} required by {model_name}.")
|
| 49 |
-
self.resampler = torchaudio.transforms.Resample(orig_freq=sample_rate, new_freq=self.ssl_sample_rate)
|
| 50 |
-
else:
|
| 51 |
-
self.resampler = None
|
| 52 |
-
|
| 53 |
-
@property
|
| 54 |
-
def hop_size(self) -> int:
|
| 55 |
-
"""Get the hop size of the model's convolutional layers."""
|
| 56 |
-
hop_size = 1
|
| 57 |
-
for _, stride in self.conv_config:
|
| 58 |
-
hop_size *= stride
|
| 59 |
-
return hop_size
|
| 60 |
-
|
| 61 |
-
@property
|
| 62 |
-
def conv_config(self) -> list[tuple[int, int]]:
|
| 63 |
-
"""Get the configuration of the convolutional layers in the model."""
|
| 64 |
-
conv_layers = []
|
| 65 |
-
for layer in self.model.feature_extractor.conv_layers:
|
| 66 |
-
layer: ConvLayerBlock
|
| 67 |
-
conv_layers.append((layer.kernel_size, layer.stride))
|
| 68 |
-
return conv_layers
|
| 69 |
-
|
| 70 |
-
def get_minimum_input_length(self, desired_output_length: int) -> int:
|
| 71 |
-
"""Calculate the minimum input length required to produce a given output length."""
|
| 72 |
-
length = desired_output_length
|
| 73 |
-
for kernel_size, stride in reversed(self.conv_config):
|
| 74 |
-
length = (length - 1) * stride + kernel_size
|
| 75 |
-
return length
|
| 76 |
-
|
| 77 |
-
@torch.no_grad()
|
| 78 |
-
def forward(
|
| 79 |
-
self,
|
| 80 |
-
waveform: torch.Tensor,
|
| 81 |
-
lengths: torch.Tensor | None = None,
|
| 82 |
-
num_layers: int | None = None,
|
| 83 |
-
return_lengths: bool = False,
|
| 84 |
-
) -> list[torch.Tensor]:
|
| 85 |
-
"""
|
| 86 |
-
Args:
|
| 87 |
-
waveform: (batch_size, num_samples)
|
| 88 |
-
lengths: Optional tensor of sequence lengths for each batch item (used for attention masking)
|
| 89 |
-
|
| 90 |
-
Returns:
|
| 91 |
-
features: List of feature tensors for each layer (batch_size, frame, dim)
|
| 92 |
-
lengths: Sequence lengths for each batch item
|
| 93 |
-
"""
|
| 94 |
-
if waveform.dim() == 1:
|
| 95 |
-
waveform = waveform.unsqueeze(0)
|
| 96 |
-
# Resample if needed
|
| 97 |
-
if self.resampler is not None:
|
| 98 |
-
waveform = self.resampler(waveform)
|
| 99 |
-
|
| 100 |
-
features, feature_lengths = self.model.extract_features(
|
| 101 |
-
waveform, lengths, num_layers=num_layers or self.output_layer
|
| 102 |
-
)
|
| 103 |
-
|
| 104 |
-
if return_lengths:
|
| 105 |
-
return features, feature_lengths
|
| 106 |
-
return features
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|