Spaces:
Sleeping
Sleeping
Upload 2 files
Browse files- model.py +54 -200
- wav2vec_aligen.py +12 -12
model.py
CHANGED
@@ -1,243 +1,97 @@
|
|
1 |
-
from transformers import
|
2 |
-
from transformers.modeling_outputs import
|
3 |
from typing import Optional, Tuple, Union
|
4 |
-
import
|
5 |
import torch
|
6 |
import torch.nn as nn
|
7 |
-
import math
|
8 |
|
9 |
-
|
10 |
-
|
11 |
-
|
12 |
-
|
13 |
-
def _no_grad_trunc_normal_(tensor, mean, std, a, b):
|
14 |
-
# Cut & paste from PyTorch official master until it's in a few official releases - RW
|
15 |
-
# Method based on https://people.sc.fsu.edu/~jburkardt/presentations/truncated_normal.pdf
|
16 |
-
def norm_cdf(x):
|
17 |
-
# Computes standard normal cumulative distribution function
|
18 |
-
return (1. + math.erf(x / math.sqrt(2.))) / 2.
|
19 |
-
|
20 |
-
|
21 |
-
with torch.no_grad():
|
22 |
-
# Values are generated by using a truncated uniform distribution and
|
23 |
-
# then using the inverse CDF for the normal distribution.
|
24 |
-
# Get upper and lower cdf values
|
25 |
-
l = norm_cdf((a - mean) / std)
|
26 |
-
u = norm_cdf((b - mean) / std)
|
27 |
-
|
28 |
-
# Uniformly fill tensor with values from [l, u], then translate to
|
29 |
-
# [2l-1, 2u-1].
|
30 |
-
tensor.uniform_(2 * l - 1, 2 * u - 1)
|
31 |
-
|
32 |
-
# Use inverse cdf transform for normal distribution to get truncated
|
33 |
-
# standard normal
|
34 |
-
tensor.erfinv_()
|
35 |
-
|
36 |
-
# Transform to proper mean, std
|
37 |
-
tensor.mul_(std * math.sqrt(2.))
|
38 |
-
tensor.add_(mean)
|
39 |
-
|
40 |
-
# Clamp to ensure it's in the proper range
|
41 |
-
tensor.clamp_(min=a, max=b)
|
42 |
-
return tensor
|
43 |
-
|
44 |
-
def trunc_normal_(tensor, mean=0., std=1., a=-2., b=2.):
|
45 |
-
return _no_grad_trunc_normal_(tensor, mean, std, a, b)
|
46 |
-
|
47 |
-
|
48 |
-
class Wav2Vec2ForWav2Vec2ForCTCAndUttranceRegression(Wav2Vec2PreTrainedModel):
|
49 |
-
def __init__(self, config, target_lang: Optional[str] = None):
|
50 |
super().__init__(config)
|
51 |
|
52 |
-
|
53 |
-
self.dropout = nn.Dropout(config.final_dropout)
|
54 |
-
|
55 |
-
self.target_lang = target_lang
|
56 |
-
|
57 |
-
if config.vocab_size is None:
|
58 |
raise ValueError(
|
59 |
-
|
60 |
-
"does not define the vocabulary size of the language model head. Please "
|
61 |
-
"instantiate the model as follows: `Wav2Vec2ForCTC.from_pretrained(..., vocab_size=vocab_size)`. "
|
62 |
-
"or define `vocab_size` of your model's configuration."
|
63 |
)
|
64 |
-
|
65 |
-
|
66 |
-
|
67 |
-
|
68 |
-
|
69 |
-
|
70 |
-
self.cls_token1 = nn.Parameter(torch.zeros(1, 1, config.hidden_size))
|
71 |
-
self.mlp_head_utt1 = nn.Sequential(nn.LayerNorm(config.hidden_size), nn.Linear(config.hidden_size, 1))
|
72 |
-
|
73 |
-
self.cls_token2 = nn.Parameter(torch.zeros(1, 1, config.hidden_size))
|
74 |
-
self.mlp_head_utt2 = nn.Sequential(nn.LayerNorm(config.hidden_size), nn.Linear(config.hidden_size, 1))
|
75 |
-
|
76 |
-
self.cls_token3 = nn.Parameter(torch.zeros(1, 1, config.hidden_size))
|
77 |
-
self.mlp_head_utt3 = nn.Sequential(nn.LayerNorm(config.hidden_size), nn.Linear(config.hidden_size, 1))
|
78 |
-
|
79 |
-
self.cls_token4 = nn.Parameter(torch.zeros(1, 1, config.hidden_size))
|
80 |
-
self.mlp_head_utt4 = nn.Sequential(nn.LayerNorm(config.hidden_size), nn.Linear(config.hidden_size, 1))
|
81 |
-
self.post_init()
|
82 |
-
# initialize the cls tokens
|
83 |
-
trunc_normal_(self.cls_token1, std=.092)
|
84 |
-
trunc_normal_(self.cls_token2, std=.01)
|
85 |
-
trunc_normal_(self.cls_token3, std=.052)
|
86 |
-
trunc_normal_(self.cls_token4, std=.02)
|
87 |
-
# Initialize weights and apply final processing
|
88 |
-
|
89 |
-
|
90 |
-
def tie_weights(self):
|
91 |
-
"""
|
92 |
-
This method overwrites [`~PreTrainedModel.tie_weights`] so that adapter weights can be correctly loaded when
|
93 |
-
passing `target_lang=...` to `from_pretrained(...)`.
|
94 |
-
|
95 |
-
This method is **not** supposed to be called by the user and is prone to be changed in the future.
|
96 |
-
"""
|
97 |
-
|
98 |
-
# Note that `tie_weights` is usually used to tie input and output embedding weights. The method is re-purposed to
|
99 |
-
# correctly load adapter layers for Wav2Vec2 so that we do not have to introduce a new API to
|
100 |
-
# [`PreTrainedModel`]. While slightly hacky, Wav2Vec2 never has to tie input and output embeddings, so that it is
|
101 |
-
# ok to repurpose this function here.
|
102 |
-
target_lang = self.target_lang
|
103 |
-
|
104 |
-
if target_lang is not None and getattr(self.config, "adapter_attn_dim", None) is None:
|
105 |
-
raise ValueError(f"Cannot pass `target_lang`: {target_lang} if `config.adapter_attn_dim` is not defined.")
|
106 |
-
elif target_lang is None and getattr(self.config, "adapter_attn_dim", None) is not None:
|
107 |
-
print("By default `target_lang` is set to 'eng'.")
|
108 |
-
elif target_lang is not None:
|
109 |
-
self.load_adapter(target_lang, force_load=True)
|
110 |
-
|
111 |
-
|
112 |
-
def freeze_feature_extractor(self):
|
113 |
-
"""
|
114 |
-
Calling this function will disable the gradient computation for the feature encoder so that its parameters will
|
115 |
-
not be updated during training.
|
116 |
-
"""
|
117 |
-
warnings.warn(
|
118 |
-
"The method `freeze_feature_extractor` is deprecated and will be removed in Transformers v5. "
|
119 |
-
"Please use the equivalent `freeze_feature_encoder` method instead.",
|
120 |
-
FutureWarning,
|
121 |
-
)
|
122 |
-
self.freeze_feature_encoder()
|
123 |
|
124 |
-
|
125 |
-
|
126 |
-
Calling this function will disable the gradient computation for the feature encoder so that its parameter will
|
127 |
-
not be updated during training.
|
128 |
-
"""
|
129 |
-
self.wav2vec2.feature_extractor._freeze_parameters()
|
130 |
|
131 |
def freeze_base_model(self):
|
132 |
"""
|
133 |
Calling this function will disable the gradient computation for the base model so that its parameters will not
|
134 |
be updated during training. Only the classification head will be updated.
|
135 |
"""
|
136 |
-
for param in self.
|
137 |
param.requires_grad = False
|
138 |
|
139 |
-
|
|
|
140 |
def forward(
|
141 |
self,
|
142 |
-
|
143 |
attention_mask: Optional[torch.Tensor] = None,
|
144 |
output_attentions: Optional[bool] = None,
|
145 |
output_hidden_states: Optional[bool] = None,
|
146 |
return_dict: Optional[bool] = None,
|
147 |
labels: Optional[torch.Tensor] = None,
|
148 |
-
) -> Union[Tuple,
|
149 |
r"""
|
150 |
-
labels (`torch.LongTensor` of shape `(batch_size,
|
151 |
-
Labels for
|
152 |
-
|
153 |
-
|
154 |
-
config.vocab_size - 1]`.
|
155 |
"""
|
156 |
|
157 |
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
|
158 |
-
|
159 |
-
|
160 |
-
extract_features = self.wav2vec2.feature_extractor(input_values)
|
161 |
-
extract_features = extract_features.transpose(1, 2)
|
162 |
-
|
163 |
-
if attention_mask is not None:
|
164 |
-
# compute reduced attention_mask corresponding to feature vectors
|
165 |
-
attention_mask = self.wav2vec2._get_feature_vector_attention_mask(
|
166 |
-
extract_features.shape[1], attention_mask, add_adapter=False
|
167 |
-
)
|
168 |
-
|
169 |
-
hidden_states, extract_features = self.wav2vec2.feature_projection(extract_features)
|
170 |
-
hidden_states = self.wav2vec2._mask_hidden_states(
|
171 |
-
hidden_states, mask_time_indices=None, attention_mask=attention_mask
|
172 |
-
)
|
173 |
|
174 |
-
|
175 |
-
|
176 |
-
cls_token3 = self.cls_token3.expand(B, -1, -1)
|
177 |
-
cls_token4 = self.cls_token4.expand(B, -1, -1)
|
178 |
-
hidden_states = torch.cat((cls_token1, cls_token2, cls_token3, cls_token4, hidden_states), dim=1) #cls_token4
|
179 |
-
# hidden_states = torch.cat((cls_token1, cls_token3, hidden_states), dim=1) #cls_token4
|
180 |
-
outputs = self.wav2vec2.encoder(
|
181 |
-
hidden_states,
|
182 |
attention_mask=attention_mask,
|
183 |
output_attentions=output_attentions,
|
184 |
output_hidden_states=output_hidden_states,
|
185 |
return_dict=return_dict,
|
186 |
)
|
187 |
-
hidden_states = outputs[0]
|
188 |
-
hidden_states = self.dropout(hidden_states)
|
189 |
|
190 |
-
|
191 |
-
|
192 |
-
|
193 |
-
|
194 |
-
|
195 |
-
|
196 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
197 |
|
198 |
loss = None
|
199 |
if labels is not None:
|
200 |
-
|
201 |
-
|
202 |
-
raise ValueError(f"Label values must be <= vocab_size: {self.config.vocab_size}")
|
203 |
-
|
204 |
-
# retrieve loss input_lengths from attention_mask
|
205 |
-
attention_mask = (
|
206 |
-
attention_mask if attention_mask is not None else torch.ones_like(input_values, dtype=torch.long)
|
207 |
-
)
|
208 |
-
input_lengths = self._get_feat_extract_output_lengths(attention_mask.sum(-1)).to(torch.long)
|
209 |
-
|
210 |
-
# assuming that padded tokens are filled with -100
|
211 |
-
# when not being attended to
|
212 |
-
labels_mask = labels >= 0
|
213 |
-
target_lengths = labels_mask.sum(-1)
|
214 |
-
flattened_targets = labels.masked_select(labels_mask)
|
215 |
-
|
216 |
-
log_probs = nn.functional.log_softmax(logits, dim=-1, dtype=torch.float32).transpose(0, 1)
|
217 |
-
|
218 |
-
with torch.backends.cudnn.flags(enabled=False):
|
219 |
-
# utterance level loss, also mse
|
220 |
-
utt_preds = torch.cat((u1, u2, u3, u4), dim=1)
|
221 |
-
# utt_preds = torch.cat((u1, u2), dim=1)
|
222 |
-
|
223 |
-
loss_utt = nn.functional.mse_loss(utt_preds ,utt_label)
|
224 |
-
|
225 |
-
|
226 |
-
loss_ph = nn.functional.ctc_loss(
|
227 |
-
log_probs,
|
228 |
-
flattened_targets,
|
229 |
-
input_lengths,
|
230 |
-
target_lengths,
|
231 |
-
blank=self.config.pad_token_id,
|
232 |
-
reduction=self.config.ctc_loss_reduction,
|
233 |
-
zero_infinity=self.config.ctc_zero_infinity,
|
234 |
-
)
|
235 |
-
loss = loss_utt + loss_ph
|
236 |
|
237 |
if not return_dict:
|
238 |
output = (logits,) + outputs[_HIDDEN_STATES_START_POSITION:]
|
239 |
return ((loss,) + output) if loss is not None else output
|
240 |
-
|
241 |
-
return
|
242 |
-
loss=loss,
|
|
|
|
|
|
|
243 |
)
|
|
|
1 |
+
from transformers import Wav2Vec2BertPreTrainedModel, Wav2Vec2BertModel
|
2 |
+
from transformers.modeling_outputs import SequenceClassifierOutput
|
3 |
from typing import Optional, Tuple, Union
|
4 |
+
from torch.nn import MSELoss
|
5 |
import torch
|
6 |
import torch.nn as nn
|
|
|
7 |
|
8 |
+
class Wav2Vec2BertForSequenceClassification(Wav2Vec2BertPreTrainedModel):
|
9 |
+
# Copied from transformers.models.wav2vec2.modeling_wav2vec2.Wav2Vec2ForSequenceClassification.__init__ with Wav2Vec2->Wav2Vec2Bert,wav2vec2->wav2vec2_bert
|
10 |
+
def __init__(self, config):
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
11 |
super().__init__(config)
|
12 |
|
13 |
+
if hasattr(config, "add_adapter") and config.add_adapter:
|
|
|
|
|
|
|
|
|
|
|
14 |
raise ValueError(
|
15 |
+
"Sequence classification does not support the use of Wav2Vec2Bert adapters (config.add_adapter=True)"
|
|
|
|
|
|
|
16 |
)
|
17 |
+
self.wav2vec2_bert = Wav2Vec2BertModel(config)
|
18 |
+
num_layers = config.num_hidden_layers + 1 # transformer layers + input embeddings
|
19 |
+
if config.use_weighted_layer_sum:
|
20 |
+
self.layer_weights = nn.Parameter(torch.ones(num_layers) / num_layers)
|
21 |
+
self.projector = nn.Linear(config.hidden_size, config.classifier_proj_size)
|
22 |
+
self.classifier = nn.Linear(config.classifier_proj_size, config.num_labels)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
23 |
|
24 |
+
# Initialize weights and apply final processing
|
25 |
+
self.post_init()
|
|
|
|
|
|
|
|
|
26 |
|
27 |
def freeze_base_model(self):
|
28 |
"""
|
29 |
Calling this function will disable the gradient computation for the base model so that its parameters will not
|
30 |
be updated during training. Only the classification head will be updated.
|
31 |
"""
|
32 |
+
for param in self.wav2vec2_bert.parameters():
|
33 |
param.requires_grad = False
|
34 |
|
35 |
+
|
36 |
+
# Copied from transformers.models.wav2vec2.modeling_wav2vec2.Wav2Vec2ForSequenceClassification.forward with Wav2Vec2->Wav2Vec2Bert,wav2vec2->wav2vec2_bert,WAV_2_VEC_2->WAV2VEC2_BERT, input_values->input_features
|
37 |
def forward(
|
38 |
self,
|
39 |
+
input_features: Optional[torch.Tensor],
|
40 |
attention_mask: Optional[torch.Tensor] = None,
|
41 |
output_attentions: Optional[bool] = None,
|
42 |
output_hidden_states: Optional[bool] = None,
|
43 |
return_dict: Optional[bool] = None,
|
44 |
labels: Optional[torch.Tensor] = None,
|
45 |
+
) -> Union[Tuple, SequenceClassifierOutput]:
|
46 |
r"""
|
47 |
+
labels (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
|
48 |
+
Labels for computing the sequence classification/regression loss. Indices should be in `[0, ...,
|
49 |
+
config.num_labels - 1]`. If `config.num_labels == 1` a regression loss is computed (Mean-Square loss), If
|
50 |
+
`config.num_labels > 1` a classification loss is computed (Cross-Entropy).
|
|
|
51 |
"""
|
52 |
|
53 |
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
|
54 |
+
output_hidden_states = True if self.config.use_weighted_layer_sum else output_hidden_states
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
55 |
|
56 |
+
outputs = self.wav2vec2_bert(
|
57 |
+
input_features,
|
|
|
|
|
|
|
|
|
|
|
|
|
58 |
attention_mask=attention_mask,
|
59 |
output_attentions=output_attentions,
|
60 |
output_hidden_states=output_hidden_states,
|
61 |
return_dict=return_dict,
|
62 |
)
|
|
|
|
|
63 |
|
64 |
+
if self.config.use_weighted_layer_sum:
|
65 |
+
hidden_states = outputs[_HIDDEN_STATES_START_POSITION]
|
66 |
+
hidden_states = torch.stack(hidden_states, dim=1)
|
67 |
+
norm_weights = nn.functional.softmax(self.layer_weights, dim=-1)
|
68 |
+
hidden_states = (hidden_states * norm_weights.view(-1, 1, 1)).sum(dim=1)
|
69 |
+
else:
|
70 |
+
hidden_states = outputs[0]
|
71 |
+
|
72 |
+
hidden_states = self.projector(hidden_states)
|
73 |
+
if attention_mask is None:
|
74 |
+
pooled_output = hidden_states.mean(dim=1)
|
75 |
+
else:
|
76 |
+
padding_mask = self._get_feature_vector_attention_mask(hidden_states.shape[1], attention_mask)
|
77 |
+
hidden_states[~padding_mask] = 0.0
|
78 |
+
pooled_output = hidden_states.sum(dim=1) / padding_mask.sum(dim=1).view(-1, 1)
|
79 |
+
|
80 |
+
logits = self.classifier(pooled_output)
|
81 |
+
logits = nn.functional.relu(logits)
|
82 |
|
83 |
loss = None
|
84 |
if labels is not None:
|
85 |
+
loss_fct = MSELoss()
|
86 |
+
loss = loss_fct(logits.view(-1, self.config.num_labels), labels.view(-1, self.config.num_labels))
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
87 |
|
88 |
if not return_dict:
|
89 |
output = (logits,) + outputs[_HIDDEN_STATES_START_POSITION:]
|
90 |
return ((loss,) + output) if loss is not None else output
|
91 |
+
|
92 |
+
return SequenceClassifierOutput(
|
93 |
+
loss=loss,
|
94 |
+
logits=logits,
|
95 |
+
hidden_states=outputs.hidden_states,
|
96 |
+
attentions=outputs.attentions,
|
97 |
)
|
wav2vec_aligen.py
CHANGED
@@ -1,8 +1,8 @@
|
|
1 |
import torch
|
2 |
import librosa
|
3 |
import os
|
4 |
-
from model import
|
5 |
-
from transformers import
|
6 |
from optimum.bettertransformer import BetterTransformer
|
7 |
|
8 |
device = 'cuda' if torch.cuda.is_available() else 'cpu'
|
@@ -12,21 +12,20 @@ os.environ['TRANSFORMERS_VERBOSITY'] = 'error'
|
|
12 |
torch.random.manual_seed(0);
|
13 |
# protobuf==3.20.0
|
14 |
|
15 |
-
model_name = "
|
16 |
-
processor =
|
17 |
-
model =
|
18 |
model = BetterTransformer.transform(model)
|
19 |
|
20 |
def load_audio(audio_path, processor):
|
21 |
audio, sr = librosa.load(audio_path, sr=16000)
|
22 |
|
23 |
-
input_values = processor(audio, sampling_rate=16000, return_tensors="pt").
|
24 |
return input_values
|
25 |
|
26 |
@torch.inference_mode()
|
27 |
def get_emissions(input_values, model):
|
28 |
results = model(input_values,).logits
|
29 |
-
results.pop('logits')
|
30 |
return results
|
31 |
|
32 |
|
@@ -34,10 +33,11 @@ def speaker_pronunciation_assesment(audio_path):
|
|
34 |
input_values = load_audio(audio_path, processor)
|
35 |
result_scores = get_emissions(input_values, model)
|
36 |
|
37 |
-
|
38 |
-
|
39 |
-
|
40 |
-
|
|
|
41 |
|
42 |
|
43 |
result = {'pronunciation_accuracy': pronunciation_score,
|
@@ -47,5 +47,5 @@ def speaker_pronunciation_assesment(audio_path):
|
|
47 |
return result
|
48 |
|
49 |
if __name__ == '__main__':
|
50 |
-
|
51 |
|
|
|
1 |
import torch
|
2 |
import librosa
|
3 |
import os
|
4 |
+
from model import Wav2Vec2BertForSequenceClassification
|
5 |
+
from transformers import AutoFeatureExtractor
|
6 |
from optimum.bettertransformer import BetterTransformer
|
7 |
|
8 |
device = 'cuda' if torch.cuda.is_available() else 'cpu'
|
|
|
12 |
torch.random.manual_seed(0);
|
13 |
# protobuf==3.20.0
|
14 |
|
15 |
+
model_name = "arslanarjumand/wav2vec-reptiles"
|
16 |
+
processor = AutoFeatureExtractor.from_pretrained(model_name)
|
17 |
+
model = Wav2Vec2BertForSequenceClassification.from_pretrained(model_name).to(device)
|
18 |
model = BetterTransformer.transform(model)
|
19 |
|
20 |
def load_audio(audio_path, processor):
|
21 |
audio, sr = librosa.load(audio_path, sr=16000)
|
22 |
|
23 |
+
input_values = processor(audio, sampling_rate=16000, return_tensors="pt").input_features
|
24 |
return input_values
|
25 |
|
26 |
@torch.inference_mode()
|
27 |
def get_emissions(input_values, model):
|
28 |
results = model(input_values,).logits
|
|
|
29 |
return results
|
30 |
|
31 |
|
|
|
33 |
input_values = load_audio(audio_path, processor)
|
34 |
result_scores = get_emissions(input_values, model)
|
35 |
|
36 |
+
pronunciation_score = round(result_scores[0].cpu().item())
|
37 |
+
fluency_score = round(result_scores[1].cpu().item())
|
38 |
+
total_score = round(result_scores[2].cpu().item())
|
39 |
+
content_scores = round(result_scores[3].cpu().item())
|
40 |
+
|
41 |
|
42 |
|
43 |
result = {'pronunciation_accuracy': pronunciation_score,
|
|
|
47 |
return result
|
48 |
|
49 |
if __name__ == '__main__':
|
50 |
+
pass
|
51 |
|