Create modeling_whisper.py
Browse files- modeling_whisper.py +161 -0
modeling_whisper.py
ADDED
@@ -0,0 +1,161 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import math
|
2 |
+
from typing import Optional, Tuple, Union
|
3 |
+
|
4 |
+
import numpy as np
|
5 |
+
import torch
|
6 |
+
import torch.nn.functional as F
|
7 |
+
import torch.utils.checkpoint
|
8 |
+
from torch import nn
|
9 |
+
from torch.nn import CrossEntropyLoss
|
10 |
+
|
11 |
+
from transformers.modeling_outputs import (
|
12 |
+
SequenceClassifierOutput
|
13 |
+
)
|
14 |
+
from transformers.modeling_utils import PreTrainedModel
|
15 |
+
from transformers.utils import (
|
16 |
+
add_start_docstrings,
|
17 |
+
add_start_docstrings_to_model_forward,
|
18 |
+
logging,
|
19 |
+
replace_return_docstrings,
|
20 |
+
)
|
21 |
+
from transformers.models.whisper.configuration_whisper import WhisperConfig
|
22 |
+
from transformers.models.whisper.generation_whisper import WhisperGenerationMixin
|
23 |
+
from transformers.models.whisper.modeling_whisper import WhisperPreTrainedModel, WHISPER_ENCODER_INPUTS_DOCSTRING, _CONFIG_FOR_DOC, WhisperEncoder
|
24 |
+
|
25 |
+
class ViSpeechClassification(WhisperPreTrainedModel):
|
26 |
+
def __init__(self, config):
|
27 |
+
super().__init__(config)
|
28 |
+
|
29 |
+
self.encoder = WhisperEncoder(config)
|
30 |
+
num_layers = config.num_hidden_layers + 1 # transformer layers + input embeddings
|
31 |
+
self.projector = nn.Sequential(
|
32 |
+
nn.Linear(self.encoder.config.hidden_size, 1024),
|
33 |
+
nn.ReLU(),
|
34 |
+
nn.Dropout(config.dropout),
|
35 |
+
|
36 |
+
nn.Linear(1024, 512),
|
37 |
+
nn.ReLU(),
|
38 |
+
nn.Dropout(config.dropout),
|
39 |
+
|
40 |
+
nn.Linear(512, config.classifier_proj_size),
|
41 |
+
nn.ReLU(),
|
42 |
+
nn.Dropout(config.dropout)
|
43 |
+
)
|
44 |
+
|
45 |
+
self.classifier = nn.Linear(config.classifier_proj_size, config.num_labels)
|
46 |
+
self.config.use_weighted_layer_sum = False
|
47 |
+
|
48 |
+
# Initialize weights and apply final processing
|
49 |
+
self.post_init()
|
50 |
+
|
51 |
+
def freeze_encoder(self):
|
52 |
+
"""
|
53 |
+
Calling this function will disable the gradient computation for the Whisper encoder so that its parameters will
|
54 |
+
not be updated during training. Only the projection layers and classification head will be updated.
|
55 |
+
"""
|
56 |
+
self.encoder._freeze_parameters()
|
57 |
+
|
58 |
+
def get_input_embeddings(self) -> nn.Module:
|
59 |
+
return self.encoder.get_input_embeddings()
|
60 |
+
|
61 |
+
def set_input_embeddings(self, value: nn.Module):
|
62 |
+
self.encoder.set_input_embeddings(value)
|
63 |
+
|
64 |
+
@add_start_docstrings_to_model_forward(WHISPER_ENCODER_INPUTS_DOCSTRING)
|
65 |
+
@replace_return_docstrings(output_type=SequenceClassifierOutput, config_class=_CONFIG_FOR_DOC)
|
66 |
+
def forward(
|
67 |
+
self,
|
68 |
+
input_features: Optional[torch.LongTensor] = None,
|
69 |
+
head_mask: Optional[torch.Tensor] = None,
|
70 |
+
encoder_outputs: Optional[Tuple[Tuple[torch.FloatTensor]]] = None,
|
71 |
+
labels: Optional[torch.LongTensor] = None,
|
72 |
+
output_attentions: Optional[bool] = None,
|
73 |
+
output_hidden_states: Optional[bool] = None,
|
74 |
+
return_dict: Optional[bool] = None,
|
75 |
+
) -> Union[Tuple[torch.Tensor], SequenceClassifierOutput]:
|
76 |
+
r"""
|
77 |
+
labels (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
|
78 |
+
Labels for computing the sequence classification/regression loss. Indices should be in `[0, ...,
|
79 |
+
config.num_labels - 1]`. If `config.num_labels == 1` a regression loss is computed (Mean-Square loss), If
|
80 |
+
`config.num_labels > 1` a classification loss is computed (Cross-Entropy).
|
81 |
+
|
82 |
+
Returns:
|
83 |
+
|
84 |
+
Example:
|
85 |
+
|
86 |
+
```python
|
87 |
+
>>> import torch
|
88 |
+
>>> from transformers import AutoFeatureExtractor, WhisperForAudioClassification
|
89 |
+
>>> from datasets import load_dataset
|
90 |
+
|
91 |
+
>>> feature_extractor = AutoFeatureExtractor.from_pretrained("sanchit-gandhi/whisper-medium-fleurs-lang-id")
|
92 |
+
>>> model = WhisperForAudioClassification.from_pretrained("sanchit-gandhi/whisper-medium-fleurs-lang-id")
|
93 |
+
|
94 |
+
>>> ds = load_dataset("google/fleurs", "all", split="validation", streaming=True)
|
95 |
+
>>> sample = next(iter(ds))
|
96 |
+
|
97 |
+
>>> inputs = feature_extractor(
|
98 |
+
... sample["audio"]["array"], sampling_rate=sample["audio"]["sampling_rate"], return_tensors="pt"
|
99 |
+
... )
|
100 |
+
>>> input_features = inputs.input_features
|
101 |
+
|
102 |
+
>>> with torch.no_grad():
|
103 |
+
... logits = model(input_features).logits
|
104 |
+
|
105 |
+
>>> predicted_class_ids = torch.argmax(logits).item()
|
106 |
+
>>> predicted_label = model.config.id2label[predicted_class_ids]
|
107 |
+
>>> predicted_label
|
108 |
+
'Afrikaans'
|
109 |
+
```"""
|
110 |
+
|
111 |
+
output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
|
112 |
+
output_hidden_states = (
|
113 |
+
output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
|
114 |
+
)
|
115 |
+
if self.config.use_weighted_layer_sum:
|
116 |
+
output_hidden_states = True
|
117 |
+
elif output_hidden_states is None:
|
118 |
+
output_hidden_states = self.config.output_hidden_states
|
119 |
+
|
120 |
+
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
|
121 |
+
|
122 |
+
if encoder_outputs is None:
|
123 |
+
encoder_outputs = self.encoder(
|
124 |
+
input_features,
|
125 |
+
head_mask=head_mask,
|
126 |
+
output_attentions=output_attentions,
|
127 |
+
output_hidden_states=output_hidden_states,
|
128 |
+
return_dict=return_dict,
|
129 |
+
)
|
130 |
+
|
131 |
+
if self.config.use_weighted_layer_sum:
|
132 |
+
hidden_states = encoder_outputs[_HIDDEN_STATES_START_POSITION]
|
133 |
+
hidden_states = torch.stack(hidden_states, dim=1)
|
134 |
+
norm_weights = nn.functional.softmax(self.layer_weights, dim=-1)
|
135 |
+
hidden_states = (hidden_states * norm_weights.view(-1, 1, 1)).sum(dim=1)
|
136 |
+
else:
|
137 |
+
hidden_states = encoder_outputs[0]
|
138 |
+
|
139 |
+
hidden_states = self.projector(hidden_states)
|
140 |
+
pooled_output = hidden_states.mean(dim=1)
|
141 |
+
|
142 |
+
logits = self.classifier(pooled_output)
|
143 |
+
|
144 |
+
loss = None
|
145 |
+
|
146 |
+
if labels is not None:
|
147 |
+
loss_fct = CrossEntropyLoss()
|
148 |
+
# move labels to correct device to enable PP
|
149 |
+
labels = labels.to(logits.device)
|
150 |
+
loss = loss_fct(logits.view(-1, self.config.num_labels), labels.view(-1))
|
151 |
+
|
152 |
+
if not return_dict:
|
153 |
+
output = (logits,) + encoder_outputs[1:]
|
154 |
+
return ((loss,) + output) if loss is not None else output
|
155 |
+
|
156 |
+
return SequenceClassifierOutput(
|
157 |
+
loss=loss,
|
158 |
+
logits=logits,
|
159 |
+
hidden_states=encoder_outputs.hidden_states,
|
160 |
+
attentions=encoder_outputs.attentions,
|
161 |
+
)
|