File size: 8,423 Bytes
9982ca9 1596979 9982ca9 1596979 9982ca9 1596979 9982ca9 1596979 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 |
from dataclasses import dataclass
from typing import Union, Type
import torch
from transformers.modeling_outputs import SequenceClassifierOutput
from transformers import (
PreTrainedModel,
PretrainedConfig,
WavLMConfig,
BertConfig,
WavLMModel,
BertModel,
Wav2Vec2Config,
Wav2Vec2Model
)
from transformers.models.wav2vec2.modeling_wav2vec2 import (
Wav2Vec2Encoder,
Wav2Vec2EncoderStableLayerNorm,
Wav2Vec2FeatureEncoder
)
from transformers.models.bert.modeling_bert import BertEncoder
class MultiModalConfig(PretrainedConfig):
"""Base class for multimodal configs"""
def __init__(self, **kwargs):
super().__init__(**kwargs)
class Wav2Vec2BertConfig(MultiModalConfig):
...
class BaseClassificationModel(PreTrainedModel):
config: Type[Union[PretrainedConfig, None]] = None
def compute_loss(self, logits, labels):
"""Compute loss
Args:
logits (torch.FloatTensor): logits
labels (torch.LongTensor): labels
Returns:
torch.FloatTensor: loss
Raises:
ValueError: Invalid number of labels
"""
if self.config.problem_type is None:
if self.num_labels == 1:
self.config.problem_type = "regression"
elif self.num_labels > 1:
self.config.problem_type = "single_label_classification"
else:
raise ValueError("Invalid number of labels: {}".format(self.num_labels))
if self.config.problem_type == "single_label_classification":
loss_fct = torch.nn.CrossEntropyLoss()
loss = loss_fct(logits.view(-1, self.num_labels), labels.view(-1))
elif self.config.problem_type == "multi_label_classification":
loss_fct = torch.nn.BCEWithLogitsLoss()
loss = loss_fct(logits.view(-1, self.num_labels), labels.view(-1, self.num_labels))
elif self.config.problem_type == "regression":
loss_fct = torch.nn.MSELoss()
loss = loss_fct(logits.view(-1), labels.view(-1))
else:
raise ValueError("Problem_type {} not supported".format(self.config.problem_type))
return loss
@staticmethod
def merged_strategy(
hidden_states,
mode="mean"
):
"""Merged strategy for pooling
Args:
hidden_states (torch.FloatTensor): hidden states
mode (str, optional): pooling mode. Defaults to "mean".
Returns:
torch.FloatTensor: pooled hidden states
"""
if mode == "mean":
outputs = torch.mean(hidden_states, dim=1)
elif mode == "sum":
outputs = torch.sum(hidden_states, dim=1)
elif mode == "max":
outputs = torch.max(hidden_states, dim=1)[0]
else:
raise Exception(
"The pooling method hasn't been defined! Your pooling mode must be one of these ['mean', 'sum', 'max']")
return outputs
class AudioTextModelForSequenceBaseClassification(BaseClassificationModel):
config_class = MultiModalConfig
def __init__(self, config):
"""
Args:
config (MultiModalConfig): config
Attributes:
config (MultiModalConfig): config
num_labels (int): number of labels
audio_config (Union[PretrainedConfig, None]): audio config
text_config (Union[PretrainedConfig, None]): text config
audio_model (Union[PreTrainedModel, None]): audio model
text_model (Union[PreTrainedModel, None]): text model
classifier (Union[torch.nn.Linear, None]): classifier
"""
super().__init__(config)
self.config = config
self.num_labels = self.config.num_labels
self.audio_config: Union[PretrainedConfig, None] = None
self.text_config: Union[PretrainedConfig, None] = None
self.audio_model: Union[PreTrainedModel, None] = None
self.text_model: Union[PreTrainedModel, None] = None
self.classifier: Union[torch.nn.Linear, None] = None
def forward(
self,
input_ids=None,
input_values=None,
text_attention_mask=None,
audio_attention_mask=None,
token_type_ids=None,
position_ids=None,
head_mask=None,
inputs_embeds=None,
labels=None,
output_attentions=None,
output_hidden_states=None,
return_dict=True,
):
"""Forward method for multimodal model for sequence classification task (e.g. text + audio)
Args:
input_ids (torch.LongTensor, optional): input ids. Defaults to None.
input_values (torch.FloatTensor, optional): input values. Defaults to None.
text_attention_mask (torch.LongTensor, optional): text attention mask. Defaults to None.
audio_attention_mask (torch.LongTensor, optional): audio attention mask. Defaults to None.
token_type_ids (torch.LongTensor, optional): token type ids. Defaults to None.
position_ids (torch.LongTensor, optional): position ids. Defaults to None.
head_mask (torch.FloatTensor, optional): head mask. Defaults to None.
inputs_embeds (torch.FloatTensor, optional): inputs embeds. Defaults to None.
labels (torch.LongTensor, optional): labels. Defaults to None.
output_attentions (bool, optional): output attentions. Defaults to None.
output_hidden_states (bool, optional): output hidden states. Defaults to None.
return_dict (bool, optional): return dict. Defaults to True.
Returns:
torch.FloatTensor: logits
"""
audio_output = self.audio_model(
input_values=input_values,
attention_mask=audio_attention_mask,
output_attentions=output_attentions,
output_hidden_states=output_hidden_states,
return_dict=return_dict
)
text_output = self.text_model(
input_ids=input_ids,
attention_mask=text_attention_mask,
token_type_ids=token_type_ids,
position_ids=position_ids,
head_mask=head_mask,
inputs_embeds=inputs_embeds,
output_attentions=output_attentions,
output_hidden_states=output_hidden_states,
return_dict=return_dict,
)
audio_mean = self.merged_strategy(audio_output.last_hidden_state, mode=self.config.pooling_mode)
pooled_output = torch.cat(
(audio_mean, text_output.pooler_output), dim=1
)
logits = self.classifier(pooled_output)
loss = None
if labels is not None:
loss = self.compute_loss(logits, labels)
return SequenceClassifierOutput(
loss=loss,
logits=logits
)
class Wav2Vec2BertForSequenceClassification(AudioTextModelForSequenceBaseClassification):
"""
Wav2Vec2BertForSequenceClassification is a model for sequence classification task
(e.g. sentiment analysis, text classification, etc.)
Args:
config (Wav2Vec2BertConfig): config
Attributes:
config (Wav2Vec2BertConfig): config
audio_config (Wav2Vec2Config): wav2vec2 config
text_config (BertConfig): bert config
audio_model (Wav2Vec2Model): wav2vec2 model
text_model (BertModel): bert model
classifier (torch.nn.Linear): classifier
"""
def __init__(self, config):
super().__init__(config)
self.supports_gradient_checkpointing = getattr(config, "gradient_checkpointing", True)
self.audio_config = Wav2Vec2Config.from_dict(self.config.Wav2Vec2Model)
self.text_config = BertConfig.from_dict(self.config.BertModel)
self.audio_model = Wav2Vec2Model(self.audio_config)
self.text_model = BertModel(self.text_config)
self.classifier = torch.nn.Linear(
self.audio_config.hidden_size + self.text_config.hidden_size, self.num_labels
)
self.init_weights()
@staticmethod
def _set_gradient_checkpointing(module, value=False):
if isinstance(module, (Wav2Vec2Encoder, Wav2Vec2EncoderStableLayerNorm, Wav2Vec2FeatureEncoder, BertEncoder)):
module.gradient_checkpointing = value |