Create wav2vec2speechclassification.py
Browse files- wav2vec2speechclassification.py +127 -0
wav2vec2speechclassification.py
ADDED
@@ -0,0 +1,127 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from dataclasses import dataclass
|
2 |
+
from typing import Optional, Tuple
|
3 |
+
import torch
|
4 |
+
from transformers.file_utils import ModelOutput
|
5 |
+
|
6 |
+
|
7 |
+
@dataclass
|
8 |
+
class SpeechClassifierOutput(ModelOutput):
|
9 |
+
loss: Optional[torch.FloatTensor] = None
|
10 |
+
logits: torch.FloatTensor = None
|
11 |
+
hidden_states: Optional[Tuple[torch.FloatTensor]] = None
|
12 |
+
attentions: Optional[Tuple[torch.FloatTensor]] = None
|
13 |
+
|
14 |
+
|
15 |
+
import torch
|
16 |
+
import torch.nn as nn
|
17 |
+
from torch.nn import BCEWithLogitsLoss, CrossEntropyLoss, MSELoss
|
18 |
+
|
19 |
+
from transformers.models.wav2vec2.modeling_wav2vec2 import (
|
20 |
+
Wav2Vec2PreTrainedModel,
|
21 |
+
Wav2Vec2Model
|
22 |
+
)
|
23 |
+
|
24 |
+
|
25 |
+
class Wav2Vec2ClassificationHead(nn.Module):
|
26 |
+
"""Head for wav2vec classification task."""
|
27 |
+
|
28 |
+
def __init__(self, config):
|
29 |
+
super().__init__()
|
30 |
+
self.dense = nn.Linear(config.hidden_size, config.hidden_size)
|
31 |
+
self.dropout = nn.Dropout(config.final_dropout)
|
32 |
+
self.out_proj = nn.Linear(config.hidden_size, config.num_labels)
|
33 |
+
|
34 |
+
def forward(self, features, **kwargs):
|
35 |
+
x = features
|
36 |
+
x = self.dropout(x)
|
37 |
+
x = self.dense(x)
|
38 |
+
x = torch.tanh(x)
|
39 |
+
x = self.dropout(x)
|
40 |
+
x = self.out_proj(x)
|
41 |
+
return x
|
42 |
+
|
43 |
+
|
44 |
+
class Wav2Vec2ForSpeechClassification(Wav2Vec2PreTrainedModel):
|
45 |
+
def __init__(self, config):
|
46 |
+
super().__init__(config)
|
47 |
+
self.num_labels = config.num_labels
|
48 |
+
self.pooling_mode = config.pooling_mode
|
49 |
+
self.config = config
|
50 |
+
|
51 |
+
self.wav2vec2 = Wav2Vec2Model(config)
|
52 |
+
self.classifier = Wav2Vec2ClassificationHead(config)
|
53 |
+
|
54 |
+
self.init_weights()
|
55 |
+
|
56 |
+
def freeze_feature_extractor(self):
|
57 |
+
self.wav2vec2.feature_extractor._freeze_parameters()
|
58 |
+
|
59 |
+
def merged_strategy(
|
60 |
+
self,
|
61 |
+
hidden_states,
|
62 |
+
mode="mean"
|
63 |
+
):
|
64 |
+
if mode == "mean":
|
65 |
+
outputs = torch.mean(hidden_states, dim=1)
|
66 |
+
elif mode == "sum":
|
67 |
+
outputs = torch.sum(hidden_states, dim=1)
|
68 |
+
elif mode == "max":
|
69 |
+
outputs = torch.max(hidden_states, dim=1)[0]
|
70 |
+
else:
|
71 |
+
raise Exception(
|
72 |
+
"The pooling method hasn't been defined! Your pooling mode must be one of these ['mean', 'sum', 'max']")
|
73 |
+
|
74 |
+
return outputs
|
75 |
+
|
76 |
+
def forward(
|
77 |
+
self,
|
78 |
+
input_values,
|
79 |
+
attention_mask=None,
|
80 |
+
output_attentions=None,
|
81 |
+
output_hidden_states=None,
|
82 |
+
return_dict=None,
|
83 |
+
labels=None,
|
84 |
+
):
|
85 |
+
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
|
86 |
+
outputs = self.wav2vec2(
|
87 |
+
input_values,
|
88 |
+
attention_mask=attention_mask,
|
89 |
+
output_attentions=output_attentions,
|
90 |
+
output_hidden_states=output_hidden_states,
|
91 |
+
return_dict=return_dict,
|
92 |
+
)
|
93 |
+
hidden_states = outputs[0]
|
94 |
+
hidden_states = self.merged_strategy(hidden_states, mode=self.pooling_mode)
|
95 |
+
logits = self.classifier(hidden_states)
|
96 |
+
|
97 |
+
loss = None
|
98 |
+
if labels is not None:
|
99 |
+
if self.config.problem_type is None:
|
100 |
+
if self.num_labels == 1:
|
101 |
+
self.config.problem_type = "regression"
|
102 |
+
elif self.num_labels > 1 and (labels.dtype == torch.long or labels.dtype == torch.int):
|
103 |
+
self.config.problem_type = "single_label_classification"
|
104 |
+
else:
|
105 |
+
self.config.problem_type = "multi_label_classification"
|
106 |
+
|
107 |
+
if self.config.problem_type == "regression":
|
108 |
+
loss_fct = MSELoss()
|
109 |
+
loss = loss_fct(logits.view(-1, self.num_labels), labels)
|
110 |
+
elif self.config.problem_type == "single_label_classification":
|
111 |
+
loss_fct = CrossEntropyLoss()
|
112 |
+
loss = loss_fct(logits.view(-1, self.num_labels), labels.view(-1))
|
113 |
+
elif self.config.problem_type == "multi_label_classification":
|
114 |
+
loss_fct = BCEWithLogitsLoss()
|
115 |
+
loss = loss_fct(logits, labels)
|
116 |
+
|
117 |
+
if not return_dict:
|
118 |
+
output = (logits,) + outputs[2:]
|
119 |
+
return ((loss,) + output) if loss is not None else output
|
120 |
+
|
121 |
+
return SpeechClassifierOutput(
|
122 |
+
loss=loss,
|
123 |
+
logits=logits,
|
124 |
+
hidden_states=outputs.hidden_states,
|
125 |
+
attentions=outputs.attentions,
|
126 |
+
)
|
127 |
+
|