Upload model
Browse files- config.json +4 -2
- model.safetensors +3 -0
- modeling_hubert_spkreg.py +613 -0
config.json
CHANGED
@@ -3,11 +3,12 @@
|
|
3 |
"activation_dropout": 0.1,
|
4 |
"apply_spec_augment": true,
|
5 |
"architectures": [
|
6 |
-
"
|
7 |
],
|
8 |
"attention_dropout": 0.1,
|
9 |
"auto_map": {
|
10 |
-
"AutoConfig": "configuration_hubert_spkreg.HubertSpkRegConfig"
|
|
|
11 |
},
|
12 |
"bos_token_id": 1,
|
13 |
"classifier_proj_size": 256,
|
@@ -78,6 +79,7 @@
|
|
78 |
"reduction": "mean",
|
79 |
"scale": 30.0,
|
80 |
"tokenizer_class": "Wav2Vec2CTCTokenizer",
|
|
|
81 |
"transformers_version": "4.46.2",
|
82 |
"use_weighted_layer_sum": false,
|
83 |
"vocab_size": 32
|
|
|
3 |
"activation_dropout": 0.1,
|
4 |
"apply_spec_augment": true,
|
5 |
"architectures": [
|
6 |
+
"HubertSpkRegModel"
|
7 |
],
|
8 |
"attention_dropout": 0.1,
|
9 |
"auto_map": {
|
10 |
+
"AutoConfig": "configuration_hubert_spkreg.HubertSpkRegConfig",
|
11 |
+
"AutoModel": "modeling_hubert_spkreg.HubertSpkRegModel"
|
12 |
},
|
13 |
"bos_token_id": 1,
|
14 |
"classifier_proj_size": 256,
|
|
|
79 |
"reduction": "mean",
|
80 |
"scale": 30.0,
|
81 |
"tokenizer_class": "Wav2Vec2CTCTokenizer",
|
82 |
+
"torch_dtype": "float32",
|
83 |
"transformers_version": "4.46.2",
|
84 |
"use_weighted_layer_sum": false,
|
85 |
"vocab_size": 32
|
model.safetensors
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:411e8e7f967ba2a68bc6fba072e6374effc390225c7fdb75b8731edd95717e15
|
3 |
+
size 377510584
|
modeling_hubert_spkreg.py
ADDED
@@ -0,0 +1,613 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import math
|
2 |
+
import warnings
|
3 |
+
from typing import Union, Tuple, Optional
|
4 |
+
|
5 |
+
import numpy as np
|
6 |
+
|
7 |
+
import torch
|
8 |
+
import torch.nn as nn
|
9 |
+
import torch.nn.functional as F
|
10 |
+
|
11 |
+
from transformers.modeling_utils import PreTrainedModel
|
12 |
+
from transformers.modeling_outputs import SequenceClassifierOutput, BaseModelOutput
|
13 |
+
from transformers.integrations.deepspeed import is_deepspeed_zero3_enabled
|
14 |
+
from transformers.integrations.fsdp import is_fsdp_managed_module
|
15 |
+
from transformers.models.hubert.modeling_hubert import (
|
16 |
+
HubertFeatureEncoder,
|
17 |
+
HubertFeatureProjection,
|
18 |
+
HubertEncoderStableLayerNorm,
|
19 |
+
HubertEncoder,
|
20 |
+
_HIDDEN_STATES_START_POSITION
|
21 |
+
)
|
22 |
+
|
23 |
+
from .configuration_hubert_spkreg import HubertSpkRegConfig
|
24 |
+
|
25 |
+
|
26 |
+
class HubertSpkRegPreTrainedModel(PreTrainedModel):
|
27 |
+
"""
|
28 |
+
An abstract class to handle weights initialization and a simple interface for downloading and loading pretrained
|
29 |
+
models.
|
30 |
+
"""
|
31 |
+
|
32 |
+
config_class = HubertSpkRegConfig
|
33 |
+
base_model_prefix = "hubert"
|
34 |
+
main_input_name = "input_values"
|
35 |
+
supports_gradient_checkpointing = True
|
36 |
+
_supports_flash_attn_2 = True
|
37 |
+
_supports_sdpa = True
|
38 |
+
|
39 |
+
def _init_weights(self, module):
|
40 |
+
"""Initialize the weights"""
|
41 |
+
if isinstance(module, nn.Linear):
|
42 |
+
# Slightly different from the TF version which uses truncated_normal for initialization
|
43 |
+
# cf https://github.com/pytorch/pytorch/pull/5617
|
44 |
+
module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)
|
45 |
+
elif isinstance(module, (nn.LayerNorm, nn.GroupNorm)):
|
46 |
+
module.bias.data.zero_()
|
47 |
+
module.weight.data.fill_(1.0)
|
48 |
+
elif isinstance(module, nn.Conv1d):
|
49 |
+
if is_deepspeed_zero3_enabled():
|
50 |
+
import deepspeed
|
51 |
+
|
52 |
+
if hasattr(module, "weight_v") and hasattr(module, "weight_g"):
|
53 |
+
with deepspeed.zero.GatheredParameters([module.weight_v, module.weight_g], modifier_rank=0):
|
54 |
+
nn.init.kaiming_normal_(module.weight.data)
|
55 |
+
else:
|
56 |
+
with deepspeed.zero.GatheredParameters(module.weight, modifier_rank=0):
|
57 |
+
nn.init.kaiming_normal_(module.weight.data)
|
58 |
+
else:
|
59 |
+
nn.init.kaiming_normal_(module.weight.data)
|
60 |
+
|
61 |
+
if isinstance(module, (nn.Linear, nn.Conv1d)) and module.bias is not None:
|
62 |
+
module.bias.data.zero_()
|
63 |
+
|
64 |
+
def _get_feat_extract_output_lengths(self, input_lengths: Union[torch.LongTensor, int]):
|
65 |
+
"""
|
66 |
+
Computes the output length of the convolutional layers
|
67 |
+
"""
|
68 |
+
|
69 |
+
def _conv_out_length(input_length, kernel_size, stride):
|
70 |
+
# 1D convolutional layer output length formula taken
|
71 |
+
# from https://pytorch.org/docs/stable/generated/torch.nn.Conv1d.html
|
72 |
+
return torch.div(input_length - kernel_size, stride, rounding_mode="floor") + 1
|
73 |
+
|
74 |
+
for kernel_size, stride in zip(self.config.conv_kernel, self.config.conv_stride):
|
75 |
+
input_lengths = _conv_out_length(input_lengths, kernel_size, stride)
|
76 |
+
|
77 |
+
return input_lengths
|
78 |
+
|
79 |
+
def _get_feature_vector_attention_mask(self, feature_vector_length: int, attention_mask: torch.LongTensor):
|
80 |
+
output_lengths = self._get_feat_extract_output_lengths(attention_mask.sum(-1)).to(torch.long)
|
81 |
+
batch_size = attention_mask.shape[0]
|
82 |
+
|
83 |
+
attention_mask = torch.zeros(
|
84 |
+
(batch_size, feature_vector_length), dtype=attention_mask.dtype, device=attention_mask.device
|
85 |
+
)
|
86 |
+
# these two operations makes sure that all values before the output lengths idxs are attended to
|
87 |
+
attention_mask[(torch.arange(attention_mask.shape[0], device=attention_mask.device), output_lengths - 1)] = 1
|
88 |
+
attention_mask = attention_mask.flip([-1]).cumsum(-1).flip([-1]).bool()
|
89 |
+
return attention_mask
|
90 |
+
|
91 |
+
|
92 |
+
# Copied from transformers.models.wav2vec2.modeling_wav2vec2._compute_mask_indices
|
93 |
+
def _compute_mask_indices(
|
94 |
+
shape: Tuple[int, int],
|
95 |
+
mask_prob: float,
|
96 |
+
mask_length: int,
|
97 |
+
attention_mask: Optional[torch.LongTensor] = None,
|
98 |
+
min_masks: int = 0,
|
99 |
+
) -> np.ndarray:
|
100 |
+
"""
|
101 |
+
Computes random mask spans for a given shape. Used to implement [SpecAugment: A Simple Data Augmentation Method for
|
102 |
+
ASR](https://arxiv.org/abs/1904.08779). Note that this method is not optimized to run on TPU and should be run on
|
103 |
+
CPU as part of the preprocessing during training.
|
104 |
+
|
105 |
+
Args:
|
106 |
+
shape: The shape for which to compute masks. This should be of a tuple of size 2 where
|
107 |
+
the first element is the batch size and the second element is the length of the axis to span.
|
108 |
+
mask_prob: The percentage of the whole axis (between 0 and 1) which will be masked. The number of
|
109 |
+
independently generated mask spans of length `mask_length` is computed by
|
110 |
+
`mask_prob*shape[1]/mask_length`. Note that due to overlaps, `mask_prob` is an upper bound and the
|
111 |
+
actual percentage will be smaller.
|
112 |
+
mask_length: size of the mask
|
113 |
+
min_masks: minimum number of masked spans
|
114 |
+
attention_mask: A (right-padded) attention mask which independently shortens the feature axis of
|
115 |
+
each batch dimension.
|
116 |
+
"""
|
117 |
+
batch_size, sequence_length = shape
|
118 |
+
|
119 |
+
if mask_length < 1:
|
120 |
+
raise ValueError("`mask_length` has to be bigger than 0.")
|
121 |
+
|
122 |
+
if mask_length > sequence_length:
|
123 |
+
raise ValueError(
|
124 |
+
f"`mask_length` has to be smaller than `sequence_length`, but got `mask_length`: {mask_length}"
|
125 |
+
f" and `sequence_length`: {sequence_length}`"
|
126 |
+
)
|
127 |
+
|
128 |
+
# epsilon is used for probabilistic rounding
|
129 |
+
epsilon = np.random.rand(1).item()
|
130 |
+
|
131 |
+
def compute_num_masked_span(input_length):
|
132 |
+
"""Given input length, compute how many spans should be masked"""
|
133 |
+
num_masked_span = int(mask_prob * input_length / mask_length + epsilon)
|
134 |
+
num_masked_span = max(num_masked_span, min_masks)
|
135 |
+
|
136 |
+
# make sure num masked span <= sequence_length
|
137 |
+
if num_masked_span * mask_length > sequence_length:
|
138 |
+
num_masked_span = sequence_length // mask_length
|
139 |
+
|
140 |
+
# make sure num_masked span is also <= input_length - (mask_length - 1)
|
141 |
+
if input_length - (mask_length - 1) < num_masked_span:
|
142 |
+
num_masked_span = max(input_length - (mask_length - 1), 0)
|
143 |
+
|
144 |
+
return num_masked_span
|
145 |
+
|
146 |
+
# compute number of masked spans in batch
|
147 |
+
input_lengths = (
|
148 |
+
attention_mask.sum(-1).detach().tolist()
|
149 |
+
if attention_mask is not None
|
150 |
+
else [sequence_length for _ in range(batch_size)]
|
151 |
+
)
|
152 |
+
|
153 |
+
# SpecAugment mask to fill
|
154 |
+
spec_aug_mask = np.zeros((batch_size, sequence_length), dtype=bool)
|
155 |
+
spec_aug_mask_idxs = []
|
156 |
+
|
157 |
+
max_num_masked_span = compute_num_masked_span(sequence_length)
|
158 |
+
|
159 |
+
if max_num_masked_span == 0:
|
160 |
+
return spec_aug_mask
|
161 |
+
|
162 |
+
for input_length in input_lengths:
|
163 |
+
# compute num of masked spans for this input
|
164 |
+
num_masked_span = compute_num_masked_span(input_length)
|
165 |
+
|
166 |
+
# get random indices to mask
|
167 |
+
spec_aug_mask_idx = np.random.choice(
|
168 |
+
np.arange(input_length - (mask_length - 1)), num_masked_span, replace=False
|
169 |
+
)
|
170 |
+
|
171 |
+
# pick first sampled index that will serve as a dummy index to pad vector
|
172 |
+
# to ensure same dimension for all batches due to probabilistic rounding
|
173 |
+
# Picking first sample just pads those vectors twice.
|
174 |
+
if len(spec_aug_mask_idx) == 0:
|
175 |
+
# this case can only happen if `input_length` is strictly smaller then
|
176 |
+
# `sequence_length` in which case the last token has to be a padding
|
177 |
+
# token which we can use as a dummy mask id
|
178 |
+
dummy_mask_idx = sequence_length - 1
|
179 |
+
else:
|
180 |
+
dummy_mask_idx = spec_aug_mask_idx[0]
|
181 |
+
|
182 |
+
spec_aug_mask_idx = np.concatenate(
|
183 |
+
[spec_aug_mask_idx, np.ones(max_num_masked_span - num_masked_span, dtype=np.int32) * dummy_mask_idx]
|
184 |
+
)
|
185 |
+
spec_aug_mask_idxs.append(spec_aug_mask_idx)
|
186 |
+
|
187 |
+
spec_aug_mask_idxs = np.array(spec_aug_mask_idxs)
|
188 |
+
|
189 |
+
# expand masked indices to masked spans
|
190 |
+
spec_aug_mask_idxs = np.broadcast_to(
|
191 |
+
spec_aug_mask_idxs[:, :, None], (batch_size, max_num_masked_span, mask_length)
|
192 |
+
)
|
193 |
+
spec_aug_mask_idxs = spec_aug_mask_idxs.reshape(batch_size, max_num_masked_span * mask_length)
|
194 |
+
|
195 |
+
# add offset to the starting indexes so that indexes now create a span
|
196 |
+
offsets = np.arange(mask_length)[None, None, :]
|
197 |
+
offsets = np.broadcast_to(offsets, (batch_size, max_num_masked_span, mask_length)).reshape(
|
198 |
+
batch_size, max_num_masked_span * mask_length
|
199 |
+
)
|
200 |
+
spec_aug_mask_idxs = spec_aug_mask_idxs + offsets
|
201 |
+
|
202 |
+
# ensure that we cannot have indices larger than sequence_length
|
203 |
+
if spec_aug_mask_idxs.max() > sequence_length - 1:
|
204 |
+
spec_aug_mask_idxs[spec_aug_mask_idxs > sequence_length - 1] = sequence_length - 1
|
205 |
+
|
206 |
+
# scatter indices to mask
|
207 |
+
np.put_along_axis(spec_aug_mask, spec_aug_mask_idxs, 1, -1)
|
208 |
+
|
209 |
+
return spec_aug_mask
|
210 |
+
|
211 |
+
|
212 |
+
class HubertSpkRegModel(HubertSpkRegPreTrainedModel):
|
213 |
+
|
214 |
+
def __init__(self, config: HubertSpkRegConfig):
|
215 |
+
super().__init__(config)
|
216 |
+
self.config = config
|
217 |
+
self.feature_extractor = HubertFeatureEncoder(config)
|
218 |
+
self.feature_projection = HubertFeatureProjection(config)
|
219 |
+
|
220 |
+
if config.mask_time_prob > 0.0 or config.mask_feature_prob > 0.0:
|
221 |
+
self.masked_spec_embed = nn.Parameter(torch.Tensor(config.hidden_size).uniform_())
|
222 |
+
|
223 |
+
if config.do_stable_layer_norm:
|
224 |
+
self.encoder = HubertEncoderStableLayerNorm(config)
|
225 |
+
else:
|
226 |
+
self.encoder = HubertEncoder(config)
|
227 |
+
|
228 |
+
# Initialize weights and apply final processing
|
229 |
+
self.post_init()
|
230 |
+
|
231 |
+
# Copied from transformers.models.wav2vec2.modeling_wav2vec2.Wav2Vec2Model._mask_hidden_states
|
232 |
+
def _mask_hidden_states(
|
233 |
+
self,
|
234 |
+
hidden_states: torch.FloatTensor,
|
235 |
+
mask_time_indices: Optional[torch.FloatTensor] = None,
|
236 |
+
attention_mask: Optional[torch.LongTensor] = None,
|
237 |
+
):
|
238 |
+
"""
|
239 |
+
Masks extracted features along time axis and/or along feature axis according to
|
240 |
+
[SpecAugment](https://arxiv.org/abs/1904.08779).
|
241 |
+
"""
|
242 |
+
|
243 |
+
# `config.apply_spec_augment` can set masking to False
|
244 |
+
if not getattr(self.config, "apply_spec_augment", True):
|
245 |
+
return hidden_states
|
246 |
+
|
247 |
+
# generate indices & apply SpecAugment along time axis
|
248 |
+
batch_size, sequence_length, hidden_size = hidden_states.size()
|
249 |
+
|
250 |
+
if mask_time_indices is not None:
|
251 |
+
# apply SpecAugment along time axis with given mask_time_indices
|
252 |
+
hidden_states[mask_time_indices] = self.masked_spec_embed.to(hidden_states.dtype)
|
253 |
+
elif self.config.mask_time_prob > 0 and self.training:
|
254 |
+
mask_time_indices = _compute_mask_indices(
|
255 |
+
(batch_size, sequence_length),
|
256 |
+
mask_prob=self.config.mask_time_prob,
|
257 |
+
mask_length=self.config.mask_time_length,
|
258 |
+
attention_mask=attention_mask,
|
259 |
+
min_masks=self.config.mask_time_min_masks,
|
260 |
+
)
|
261 |
+
mask_time_indices = torch.tensor(mask_time_indices, device=hidden_states.device, dtype=torch.bool)
|
262 |
+
hidden_states[mask_time_indices] = self.masked_spec_embed.to(hidden_states.dtype)
|
263 |
+
|
264 |
+
if self.config.mask_feature_prob > 0 and self.training:
|
265 |
+
# generate indices & apply SpecAugment along feature axis
|
266 |
+
mask_feature_indices = _compute_mask_indices(
|
267 |
+
(batch_size, hidden_size),
|
268 |
+
mask_prob=self.config.mask_feature_prob,
|
269 |
+
mask_length=self.config.mask_feature_length,
|
270 |
+
min_masks=self.config.mask_feature_min_masks,
|
271 |
+
)
|
272 |
+
mask_feature_indices = torch.tensor(mask_feature_indices, device=hidden_states.device, dtype=torch.bool)
|
273 |
+
mask_feature_indices = mask_feature_indices[:, None].expand(-1, sequence_length, -1)
|
274 |
+
hidden_states[mask_feature_indices] = 0
|
275 |
+
|
276 |
+
return hidden_states
|
277 |
+
|
278 |
+
def forward(
|
279 |
+
self,
|
280 |
+
input_values: Optional[torch.Tensor],
|
281 |
+
attention_mask: Optional[torch.Tensor] = None,
|
282 |
+
mask_time_indices: Optional[torch.FloatTensor] = None,
|
283 |
+
output_attentions: Optional[bool] = None,
|
284 |
+
output_hidden_states: Optional[bool] = None,
|
285 |
+
return_dict: Optional[bool] = None,
|
286 |
+
) -> Union[Tuple, BaseModelOutput]:
|
287 |
+
"""
|
288 |
+
|
289 |
+
Returns:
|
290 |
+
|
291 |
+
Example:
|
292 |
+
|
293 |
+
```python
|
294 |
+
>>> from transformers import AutoProcessor, HubertModel
|
295 |
+
>>> from datasets import load_dataset
|
296 |
+
>>> import soundfile as sf
|
297 |
+
|
298 |
+
>>> processor = AutoProcessor.from_pretrained("facebook/hubert-large-ls960-ft")
|
299 |
+
>>> model = HubertModel.from_pretrained("facebook/hubert-large-ls960-ft")
|
300 |
+
|
301 |
+
|
302 |
+
>>> def map_to_array(batch):
|
303 |
+
... speech, _ = sf.read(batch["file"])
|
304 |
+
... batch["speech"] = speech
|
305 |
+
... return batch
|
306 |
+
|
307 |
+
|
308 |
+
>>> ds = load_dataset("hf-internal-testing/librispeech_asr_dummy", "clean", split="validation")
|
309 |
+
>>> ds = ds.map(map_to_array)
|
310 |
+
|
311 |
+
>>> input_values = processor(ds["speech"][0], return_tensors="pt").input_values # Batch size 1
|
312 |
+
>>> hidden_states = model(input_values).last_hidden_state
|
313 |
+
```"""
|
314 |
+
output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
|
315 |
+
output_hidden_states = (
|
316 |
+
output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
|
317 |
+
)
|
318 |
+
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
|
319 |
+
|
320 |
+
extract_features = self.feature_extractor(input_values)
|
321 |
+
extract_features = extract_features.transpose(1, 2)
|
322 |
+
|
323 |
+
if attention_mask is not None:
|
324 |
+
# compute reduced attention_mask corresponding to feature vectors
|
325 |
+
attention_mask = self._get_feature_vector_attention_mask(extract_features.shape[1], attention_mask)
|
326 |
+
|
327 |
+
hidden_states = self.feature_projection(extract_features)
|
328 |
+
hidden_states = self._mask_hidden_states(hidden_states, mask_time_indices=mask_time_indices)
|
329 |
+
|
330 |
+
encoder_outputs = self.encoder(
|
331 |
+
hidden_states,
|
332 |
+
attention_mask=attention_mask,
|
333 |
+
output_attentions=output_attentions,
|
334 |
+
output_hidden_states=output_hidden_states,
|
335 |
+
return_dict=return_dict,
|
336 |
+
)
|
337 |
+
|
338 |
+
hidden_states = encoder_outputs[0]
|
339 |
+
|
340 |
+
if not return_dict:
|
341 |
+
return (hidden_states,) + encoder_outputs[1:]
|
342 |
+
|
343 |
+
return BaseModelOutput(
|
344 |
+
last_hidden_state=hidden_states,
|
345 |
+
hidden_states=encoder_outputs.hidden_states,
|
346 |
+
attentions=encoder_outputs.attentions,
|
347 |
+
)
|
348 |
+
|
349 |
+
|
350 |
+
class AngularLinear(nn.Module):
|
351 |
+
|
352 |
+
def __init__(self, in_features: int, out_features: int):
|
353 |
+
super(AngularLinear, self).__init__()
|
354 |
+
self.in_features = in_features
|
355 |
+
self.out_features = out_features
|
356 |
+
self.weight = torch.nn.Parameter(
|
357 |
+
torch.FloatTensor(out_features, in_features), requires_grad=True
|
358 |
+
)
|
359 |
+
nn.init.xavier_normal_(self.weight, gain=1)
|
360 |
+
|
361 |
+
def forward(
|
362 |
+
self,
|
363 |
+
inputs: torch.Tensor,
|
364 |
+
):
|
365 |
+
# Calculation of cos(theta)
|
366 |
+
cosine = F.linear(F.normalize(inputs), F.normalize(self.weight))
|
367 |
+
return cosine
|
368 |
+
|
369 |
+
def extra_repr(self) -> str:
|
370 |
+
return 'in_features={}, out_features={}'.format(
|
371 |
+
self.in_features, self.out_features
|
372 |
+
)
|
373 |
+
|
374 |
+
|
375 |
+
class AMSoftmaxLoss(nn.Module):
|
376 |
+
"""Additive Margin Softmax (CosFace).
|
377 |
+
|
378 |
+
Paper: Wang, Feng, et al. "Additive margin softmax for face verification."
|
379 |
+
IEEE Signal Processing Letters 25.7 (2018): 926-930.
|
380 |
+
"""
|
381 |
+
def __init__(
|
382 |
+
self,
|
383 |
+
scale: float = 30.0,
|
384 |
+
margin: float = 0.35,
|
385 |
+
label_smoothing: float = 0.0,
|
386 |
+
reduction: str = "mean"
|
387 |
+
):
|
388 |
+
"""
|
389 |
+
Args:
|
390 |
+
num_classes: Number of classes (output dimension)
|
391 |
+
scale: Scaling factor for logits (default: 30.0)
|
392 |
+
margin: Angular margin (default: 0.35)
|
393 |
+
"""
|
394 |
+
super(AMSoftmaxLoss, self).__init__()
|
395 |
+
self.scale = scale
|
396 |
+
self.margin = margin
|
397 |
+
self.label_smoothing = label_smoothing
|
398 |
+
self.reduction = reduction
|
399 |
+
|
400 |
+
def forward(
|
401 |
+
self,
|
402 |
+
inputs: torch.Tensor,
|
403 |
+
targets: torch.Tensor,
|
404 |
+
):
|
405 |
+
"""
|
406 |
+
Args:
|
407 |
+
inputs: Input features of shape (batch_size, num_labels)
|
408 |
+
targets: Ground truth labels of shape (batch_size)
|
409 |
+
label_smoothing: Label smoothing factor (default: 0.0)
|
410 |
+
reduction: Reduction method (default: "mean")
|
411 |
+
Returns:
|
412 |
+
Loss value
|
413 |
+
"""
|
414 |
+
_, num_labels = inputs.shape
|
415 |
+
# `inputs` are the outputs from AngularLinear()
|
416 |
+
cos_theta = torch.clamp(inputs, -1.0 + 1e-7, 1.0 - 1e-7)
|
417 |
+
psi = cos_theta - self.margin
|
418 |
+
one_hot = nn.functional.one_hot(targets, num_labels)
|
419 |
+
outputs = self.scale * torch.where(one_hot.bool(), psi, cos_theta)
|
420 |
+
loss = F.cross_entropy(
|
421 |
+
outputs, targets, label_smoothing=self.label_smoothing, reduction=self.reduction
|
422 |
+
)
|
423 |
+
return loss
|
424 |
+
|
425 |
+
|
426 |
+
class AAMSoftmaxLoss(nn.Module):
|
427 |
+
"""Additive Angular Margin Softmax (ArcFace).
|
428 |
+
|
429 |
+
Paper: Deng, Jiankang, et al. "Arcface: Additive angular margin loss for deep face recognition."
|
430 |
+
Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 2019.
|
431 |
+
"""
|
432 |
+
def __init__(
|
433 |
+
self,
|
434 |
+
scale: float = 30.0,
|
435 |
+
margin: float = 0.35,
|
436 |
+
easy_margin: bool = False,
|
437 |
+
label_smoothing: float = 0.0,
|
438 |
+
reduction: str = "mean"
|
439 |
+
):
|
440 |
+
"""
|
441 |
+
Args:
|
442 |
+
num_classes: Number of classes (output dimension)
|
443 |
+
scale: Scaling factor for logits (default: 30.0)
|
444 |
+
margin: Angular margin (default: 0.35)
|
445 |
+
easy_margin: Use the easy margin loss (default: False)
|
446 |
+
"""
|
447 |
+
super(AAMSoftmaxLoss, self).__init__()
|
448 |
+
self.scale = scale
|
449 |
+
self.margin = margin
|
450 |
+
self.easy_margin = easy_margin
|
451 |
+
self.label_smoothing = label_smoothing
|
452 |
+
self.reduction = reduction
|
453 |
+
|
454 |
+
def forward(
|
455 |
+
self,
|
456 |
+
inputs: torch.Tensor,
|
457 |
+
targets: torch.Tensor,
|
458 |
+
):
|
459 |
+
"""
|
460 |
+
Args:
|
461 |
+
inputs: Input features of shape (batch_size, num_labels)
|
462 |
+
targets: Ground truth labels of shape (batch_size)
|
463 |
+
Returns:
|
464 |
+
Loss value
|
465 |
+
"""
|
466 |
+
_, num_labels = inputs.shape
|
467 |
+
# `inputs` are the outputs from AngularLinear()
|
468 |
+
cos_theta = torch.clamp(inputs, -1.0 + 1e-7, 1.0 - 1e-7)
|
469 |
+
theta = torch.acos(cos_theta)
|
470 |
+
psi = torch.cos(theta + self.margin)
|
471 |
+
one_hot = nn.functional.one_hot(targets, num_labels)
|
472 |
+
outputs = self.scale * torch.where(one_hot.bool(), psi, cos_theta)
|
473 |
+
loss = F.cross_entropy(
|
474 |
+
outputs, targets, label_smoothing=self.label_smoothing, reduction=self.reduction
|
475 |
+
)
|
476 |
+
return loss
|
477 |
+
|
478 |
+
|
479 |
+
class HubertSpkRegForSequenceClassification(HubertSpkRegPreTrainedModel):
|
480 |
+
def __init__(self, config):
|
481 |
+
super().__init__(config)
|
482 |
+
|
483 |
+
if hasattr(config, "add_adapter") and config.add_adapter:
|
484 |
+
raise ValueError(
|
485 |
+
"Sequence classification does not support the use of Hubert adapters (config.add_adapter=True)"
|
486 |
+
)
|
487 |
+
self.hubert = HubertSpkRegModel(config)
|
488 |
+
num_layers = config.num_hidden_layers + 1 # transformer layers + input embeddings
|
489 |
+
if config.use_weighted_layer_sum:
|
490 |
+
self.layer_weights = nn.Parameter(torch.ones(num_layers) / num_layers)
|
491 |
+
self.projector = nn.Linear(config.hidden_size, config.classifier_proj_size)
|
492 |
+
|
493 |
+
if self.config.loss_fct == 'cross_entropy':
|
494 |
+
self.classifier = nn.Linear(config.classifier_proj_size, config.num_labels)
|
495 |
+
elif self.config.loss_fct == 'additive_margin':
|
496 |
+
self.classifier = AngularLinear(config.classifier_proj_size, config.num_labels)
|
497 |
+
elif self.config.loss_fct == 'additive_angular_margin':
|
498 |
+
self.classifier = AngularLinear(config.classifier_proj_size, config.num_labels)
|
499 |
+
else:
|
500 |
+
raise ValueError(f"Unsupported loss function: {self.config.loss_fct}")
|
501 |
+
|
502 |
+
# Initialize weights and apply final processing
|
503 |
+
self.post_init()
|
504 |
+
|
505 |
+
def freeze_feature_extractor(self):
|
506 |
+
"""
|
507 |
+
Calling this function will disable the gradient computation for the feature encoder so that its parameters will
|
508 |
+
not be updated during training.
|
509 |
+
"""
|
510 |
+
warnings.warn(
|
511 |
+
"The method `freeze_feature_extractor` is deprecated and will be removed in Transformers v5. "
|
512 |
+
"Please use the equivalent `freeze_feature_encoder` method instead.",
|
513 |
+
FutureWarning,
|
514 |
+
)
|
515 |
+
self.freeze_feature_encoder()
|
516 |
+
|
517 |
+
def freeze_feature_encoder(self):
|
518 |
+
"""
|
519 |
+
Calling this function will disable the gradient computation for the feature encoder so that its parameter will
|
520 |
+
not be updated during training.
|
521 |
+
"""
|
522 |
+
self.hubert.feature_extractor._freeze_parameters()
|
523 |
+
|
524 |
+
def freeze_base_model(self):
|
525 |
+
"""
|
526 |
+
Calling this function will disable the gradient computation for the base model so that its parameters will not
|
527 |
+
be updated during training. Only the classification head will be updated.
|
528 |
+
"""
|
529 |
+
for param in self.hubert.parameters():
|
530 |
+
param.requires_grad = False
|
531 |
+
|
532 |
+
def forward(
|
533 |
+
self,
|
534 |
+
input_values: Optional[torch.Tensor],
|
535 |
+
attention_mask: Optional[torch.Tensor] = None,
|
536 |
+
output_attentions: Optional[bool] = None,
|
537 |
+
output_hidden_states: Optional[bool] = None,
|
538 |
+
return_dict: Optional[bool] = None,
|
539 |
+
labels: Optional[torch.Tensor] = None,
|
540 |
+
) -> Union[Tuple, SequenceClassifierOutput]:
|
541 |
+
r"""
|
542 |
+
labels (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
|
543 |
+
Labels for computing the sequence classification/regression loss. Indices should be in `[0, ...,
|
544 |
+
config.num_labels - 1]`. If `config.num_labels == 1` a regression loss is computed (Mean-Square loss), If
|
545 |
+
`config.num_labels > 1` a classification loss is computed (Cross-Entropy).
|
546 |
+
"""
|
547 |
+
|
548 |
+
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
|
549 |
+
output_hidden_states = True if self.config.use_weighted_layer_sum else output_hidden_states
|
550 |
+
|
551 |
+
outputs = self.hubert(
|
552 |
+
input_values,
|
553 |
+
attention_mask=attention_mask,
|
554 |
+
output_attentions=output_attentions,
|
555 |
+
output_hidden_states=output_hidden_states,
|
556 |
+
return_dict=return_dict,
|
557 |
+
)
|
558 |
+
|
559 |
+
if self.config.use_weighted_layer_sum:
|
560 |
+
hidden_states = outputs[_HIDDEN_STATES_START_POSITION]
|
561 |
+
hidden_states = torch.stack(hidden_states, dim=1)
|
562 |
+
norm_weights = nn.functional.softmax(self.layer_weights, dim=-1)
|
563 |
+
hidden_states = (hidden_states * norm_weights.view(-1, 1, 1)).sum(dim=1)
|
564 |
+
else:
|
565 |
+
hidden_states = outputs[0]
|
566 |
+
|
567 |
+
hidden_states = self.projector(hidden_states)
|
568 |
+
if attention_mask is None:
|
569 |
+
pooled_output = hidden_states.mean(dim=1)
|
570 |
+
else:
|
571 |
+
padding_mask = self._get_feature_vector_attention_mask(hidden_states.shape[1], attention_mask)
|
572 |
+
hidden_states[~padding_mask] = 0.0
|
573 |
+
pooled_output = hidden_states.sum(dim=1) / padding_mask.sum(dim=1).view(-1, 1)
|
574 |
+
|
575 |
+
logits = self.classifier(pooled_output)
|
576 |
+
|
577 |
+
loss = None
|
578 |
+
if labels is not None:
|
579 |
+
if self.config.loss_fct == 'cross_entropy':
|
580 |
+
loss_fct = nn.CrossEntropyLoss(
|
581 |
+
label_smoothing=self.config.label_smoothing,
|
582 |
+
reduction=self.config.reduction
|
583 |
+
)
|
584 |
+
elif self.config.loss_fct == 'additive_margin':
|
585 |
+
loss_fct = AMSoftmaxLoss(
|
586 |
+
scale=self.config.scale,
|
587 |
+
margin=self.config.margin,
|
588 |
+
label_smoothing=self.config.label_smoothing,
|
589 |
+
reduction=self.config.reduction
|
590 |
+
)
|
591 |
+
elif self.config.loss_fct == 'additive_angular_margin':
|
592 |
+
loss_fct = AAMSoftmaxLoss(
|
593 |
+
scale=self.config.scale,
|
594 |
+
margin=self.config.margin,
|
595 |
+
easy_margin=self.config.easy_margin,
|
596 |
+
label_smoothing=self.config.label_smoothing,
|
597 |
+
reduction=self.config.reduction
|
598 |
+
)
|
599 |
+
loss = loss_fct(
|
600 |
+
logits.view(-1, self.config.num_labels),
|
601 |
+
labels.view(-1),
|
602 |
+
)
|
603 |
+
|
604 |
+
if not return_dict:
|
605 |
+
output = (logits,) + outputs[_HIDDEN_STATES_START_POSITION:]
|
606 |
+
return ((loss,) + output) if loss is not None else output
|
607 |
+
|
608 |
+
return SequenceClassifierOutput(
|
609 |
+
loss=loss,
|
610 |
+
logits=logits,
|
611 |
+
hidden_states=outputs.hidden_states,
|
612 |
+
attentions=outputs.attentions,
|
613 |
+
)
|