-
Notifications
You must be signed in to change notification settings - Fork 10
/
Copy pathbert_model.py
180 lines (152 loc) · 8.58 KB
/
bert_model.py
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
"""
transformer based models
"""
import torch
from torch import nn
import torch.nn.functional as F
from transformers import BertPreTrainedModel, BertModel
from model.linear_crf_inferencer import LinearCRF
from typing import Tuple
import numpy as np
class BertCRF(BertPreTrainedModel):
def __init__(self, cfig):
super(BertCRF, self).__init__(cfig)
self.devices = cfig.device
self.num_labels = len(cfig.label2idx)
self.bert = BertModel(cfig)
self.dynamic_bert_layer_combine = False
self.R_Drop = False
if self.R_Drop:
self.R_Drop_K = cfig.R_Drop_K
self.dropout = nn.Dropout(cfig.hidden_dropout_prob)
# self.dropout = nn.Dropout(0.0)
self.classifier = nn.Linear(cfig.hidden_size, len(cfig.label2idx))
if self.dynamic_bert_layer_combine:
self.layer_classifier = nn.Linear(cfig.hidden_size, 1)
self.inferencer = LinearCRF(cfig)
self.init_weights()
def forward(self, input_ids, input_seq_lens=None, annotation_mask=None, labels=None,
attention_mask=None, token_type_ids=None,
position_ids=None, head_mask=None, add_crf=False):
if self.dynamic_bert_layer_combine:
outputs = self.bert(input_ids, attention_mask=attention_mask,
token_type_ids=token_type_ids, position_ids=position_ids, head_mask=head_mask, output_hidden_states=True)
# sequence_output = outputs[0]
# 动态融合bert各层hidden_state
all_encoder_layers = outputs[2][:-1]
layer_logits = []
for i, layer in enumerate(all_encoder_layers):
print("layer: ", layer)
layer_logits.append(self.layer_classifier(layer))
print("np.array(layer_logits).shape:", np.array(layer_logits).shape)
layer_logits = torch.cat((layer_logits), 2)
print("layer_logits.shape:", layer_logits.shape)
layer_dist = F.softmax(layer_logits, dim=2)
print("layer_dist.shape:", layer_dist.shape)
seq_out = torch.cat([torch.unsqueeze(x, dim=2) for x in all_encoder_layers], dim=2)
print("seq_out.shape:", seq_out.shape)
sequence_output = torch.matmul(torch.unsqueeze(layer_dist, dim=2), seq_out)
sequence_output = torch.squeeze(sequence_output, dim=2)
print("sequence_output.shape:", sequence_output.shape)
else:
outputs = self.bert(input_ids, attention_mask=attention_mask,
token_type_ids=token_type_ids, position_ids=position_ids, head_mask=head_mask,
output_hidden_states=False)
sequence_output = outputs[0]
sequence_output_1 = self.dropout(sequence_output) # (batch_size, seq_length, hidden_size)
logits = self.classifier(sequence_output_1) # (batch_size, seq_length, num_labels)
if labels is not None:
batch_size = input_ids.size(0)
sent_len = input_ids.size(1) # one batch max seq length
maskTemp = torch.arange(1, sent_len + 1, dtype=torch.long).view(1, sent_len).expand(batch_size, sent_len).to(self.device)
mask = torch.le(maskTemp, input_seq_lens.view(batch_size, 1).expand(batch_size, sent_len)).to(self.device)
unlabed_score, labeled_score = self.inferencer(logits, input_seq_lens, labels, attention_mask)
crf_loss = unlabed_score - labeled_score
if self.R_Drop:
outputs_2 = self.bert(input_ids, attention_mask=attention_mask,
token_type_ids=token_type_ids, position_ids=position_ids, head_mask=head_mask,
output_hidden_states=False)
sequence_output_2 = outputs_2[0]
sequence_output_2 = self.dropout(sequence_output_2)
logits_2 = self.classifier(sequence_output_2)
unlabed_score_2, labeled_score_2 = self.inferencer(logits_2, input_seq_lens, labels, attention_mask)
crf_loss_2 = unlabed_score_2 - labeled_score_2
crf_loss = 0.5 * (crf_loss + crf_loss_2)
# if self.R_Drop:
# sequence_output_2 = self.dropout(sequence_output)
# logits_2 = self.classifier(sequence_output_2)
pad_mask = torch.ones_like(attention_mask, device=self.devices).int() - attention_mask.int()
pad_mask = torch.unsqueeze(pad_mask, dim=-1)
pad_mask = pad_mask.gt(0)
kl_loss = self.compute_kl_loss(sequence_output_1, sequence_output_2, pad_mask)
# carefully choose hyper-parameters
loss = crf_loss + self.R_Drop_K * kl_loss
return loss
return crf_loss
else:
bestScores, decodeIdx = self.inferencer.decode(logits, input_seq_lens, annotation_mask)
return bestScores, decodeIdx
def compute_kl_loss(self, p, q, pad_mask=None):
p_loss = F.kl_div(F.log_softmax(p, dim=-1), F.softmax(q, dim=-1), reduction='none')
q_loss = F.kl_div(F.log_softmax(q, dim=-1), F.softmax(p, dim=-1), reduction='none')
# pad_mask is for seq-level tasks
if pad_mask is not None:
p_loss.masked_fill_(pad_mask, 0.)
q_loss.masked_fill_(pad_mask, 0.)
# You can choose whether to use function "sum" and "mean" depending on your task
p_loss = p_loss.sum()
q_loss = q_loss.sum()
loss = (p_loss + q_loss) / 2
return loss
# obsolete
def decode(self, input_ids, input_seq_lens=None, annotation_mask=None, attention_mask=None) -> Tuple[torch.Tensor, torch.Tensor]:
"""
Decode the batch input
:param batchInput:
:return:
"""
features = self.bert(input_ids, attention_mask=attention_mask, token_type_ids=None, position_ids=None, head_mask=None)
features = self.dropout(features) # (batch_size, seq_length, hidden_size)
logits = self.classifier(features) # (batch_size, seq_length, num_labels)
bestScores, decodeIdx = self.inferencer.decode(logits, input_seq_lens, annotation_mask)
return bestScores, decodeIdx
class BertCRF_pre(BertPreTrainedModel):
def __init__(self, cfig):
super(BertCRF_pre, self).__init__(cfig)
# self.device = cfig.device
self.num_labels = len(cfig.label2idx)
self.bert = BertModel(cfig)
self.dropout = nn.Dropout(cfig.hidden_dropout_prob)
self.classifier_pre = nn.Linear(cfig.hidden_size, len(cfig.label2idx))
self.inferencer_pre = LinearCRF(cfig)
self.init_weights()
def forward(self, input_ids, input_seq_lens=None, annotation_mask=None, labels=None,
attention_mask=None, token_type_ids=None,
position_ids=None, head_mask=None, add_crf=False):
outputs = self.bert(input_ids, attention_mask=attention_mask,
token_type_ids=token_type_ids, position_ids=position_ids, head_mask=head_mask)
sequence_output = outputs[0]
sequence_output = self.dropout(sequence_output) # (batch_size, seq_length, hidden_size)
logits = self.classifier_pre(sequence_output) # (batch_size, seq_length, num_labels)
if labels is not None:
batch_size = input_ids.size(0)
sent_len = input_ids.size(1) # one batch max seq length
maskTemp = torch.arange(1, sent_len + 1, dtype=torch.long).view(1, sent_len).expand(batch_size, sent_len).to(self.device)
mask = torch.le(maskTemp, input_seq_lens.view(batch_size, 1).expand(batch_size, sent_len)).to(self.device)
unlabed_score, labeled_score = self.inferencer_pre(logits, input_seq_lens, labels, attention_mask)
return unlabed_score - labeled_score
else:
bestScores, decodeIdx = self.inferencer_pre.decode(logits, input_seq_lens, annotation_mask)
return bestScores, decodeIdx
# obsolete
def decode(self, input_ids, input_seq_lens=None, annotation_mask=None, attention_mask=None) -> Tuple[torch.Tensor, torch.Tensor]:
"""
Decode the batch input
:param batchInput:
:return:
"""
features = self.bert(input_ids, attention_mask=attention_mask, token_type_ids=None, position_ids=None, head_mask=None)
features = self.dropout(features) # (batch_size, seq_length, hidden_size)
logits = self.classifier_pre(features) # (batch_size, seq_length, num_labels)
bestScores, decodeIdx = self.inferencer_pre.decode(logits, input_seq_lens, annotation_mask)
return bestScores, decodeIdx