-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathmodel.py
193 lines (159 loc) · 9.5 KB
/
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
181
182
183
184
185
186
187
188
189
190
191
192
193
import torch
import label
import torch.nn as nn
from torchcrf import CRF
from transformers import AutoModel
from torch.nn.utils.rnn import pad_sequence
# from transformers.models.bert.modeling_bert import *
class MultiChannelAttention(nn.Module):
def __init__(self, ngram_size, hidden_size, cat_num):
super(MultiChannelAttention, self).__init__()
self.word_embedding = nn.Embedding(ngram_size, hidden_size, padding_idx=0)
self.channel_weight = nn.Embedding(cat_num, 1)
self.temper = hidden_size ** 0.5
def forward(self, word_seq, hidden_state, char_word_mask_matrix, channel_ids):
# word_seq: (batch_size, channel, word_seq_len)
# hidden_state: (batch_size, character_seq_len, hidden_size)
# mask_matrix: (batch_size, channel, character_seq_len, word_seq_len)
# embedding (batch_size, channel, word_seq_len, word_embedding_dim)
batch_size, character_seq_len, hidden_size = hidden_state.shape
channel = char_word_mask_matrix.shape[1]
word_seq_length = word_seq.shape[2]
embedding = self.word_embedding(word_seq)
tmp = embedding.permute(0, 1, 3, 2) #(batch_size, channel, hidden_size, word_seq_len)
tmp_hidden_state = torch.stack([hidden_state] * channel, 1) #(batch_size, channel, character_seq_len, hidden_size)
# u (batch_size, channel, character_seq_len, word_seq_len)
u = torch.matmul(tmp_hidden_state, tmp) / self.temper
# attention (batch_size, channel, character_seq_len, word_seq_len)
tmp_word_mask_metrix = torch.clamp(char_word_mask_matrix, 0, 1)
exp_u = torch.exp(u)
delta_exp_u = torch.mul(exp_u, tmp_word_mask_metrix)
sum_delta_exp_u = torch.stack([torch.sum(delta_exp_u, 3)] * delta_exp_u.shape[3], 3)
# denominator (batch_size, channel, character_seq_len, word_seq_len) 分母
attention = torch.div(delta_exp_u, sum_delta_exp_u + 1e-10)
attention = attention.view(batch_size * channel, character_seq_len, word_seq_length)
embedding = embedding.view(batch_size * channel, word_seq_length, hidden_size)
character_attention = torch.bmm(attention, embedding)
character_attention = character_attention.view(batch_size, channel, character_seq_len, hidden_size)
channel_w = self.channel_weight(channel_ids)
channel_w = nn.Softmax(dim=1)(channel_w)
channel_w = channel_w.view(batch_size, -1, 1, 1)
character_attention = torch.mul(character_attention, channel_w)
character_attention = character_attention.permute(0, 2, 1, 3)
#(batch_size, character_seq_len, channel, hidden_size)
character_attention = character_attention.flatten(start_dim=2)
return character_attention
class BertSegPos(nn.Module):
def __init__(self, config, gram2id):
super(BertSegPos, self).__init__( )
self.num_seglabels = label.num_seglabels
self.num_poslabels = label.num_poslabels
self.num_segposlabels = label.num_segposlabels
self.bert = AutoModel.from_pretrained(config.berta_model,add_pooling_layer=False)
#add_pooling_layer=False命令可以让模型没有池化层
self.dropout = nn.Dropout(config.hidden_dropout_prob)
self.gram2id = gram2id
# self.classifier_seg = nn.Linear(config.hidden_size, self.num_seglabels)
# self.classifier_pos = nn.Linear(config.hidden_size, self.num_poslabels)
# self.classifier_segpos = nn.Linear(config.hidden_size, self.num_segposlabels)
if config.use_attention:
self.multi_attention = MultiChannelAttention(len(self.gram2id), config.hidden_size, config.cat_num)
self.classifier_seg = nn.Linear(config.hidden_size * (1 + config.cat_num), self.num_seglabels)
self.classifier_pos = nn.Linear(config.hidden_size * (1 + config.cat_num), self.num_poslabels)
# self.classifier = nn.Linear(config.hidden_size * (1 + self.cat_num), self.num_labels, bias=False)
else:
self.multi_attention = None
self.classifier_seg = nn.Linear(config.hidden_size, self.num_seglabels)
self.classifier_pos = nn.Linear(config.hidden_size, self.num_poslabels)
self.classifier_segpos = nn.Linear(config.hidden_size, self.num_segposlabels)
if config.use_augment:
self.transform2noise_seg = nn.Linear(self.num_seglabels, self.num_seglabels, bias=False)
self.transform2noise_pos = nn.Linear(self.num_poslabels, self.num_poslabels, bias=False)
# self.softmax = nn.Softmax(dim=1)
torch.nn.init.eye_(self.transform2noise_seg.weight.data)
torch.nn.init.eye_(self.transform2noise_pos.weight.data)
# # self.tanh = nn.Tanh()
# self.flag_noise = True
# transform2noise_seg = torch.empty(size=(self.num_seglabels,self.num_seglabels), requires_grad=True)
# transform2noise_pos = torch.empty(size=(self.num_poslabels,self.num_poslabels), requires_grad=True)
# self.transform2noise_seg = nn.Parameter(transform2noise_seg)
# self.transform2noise_pos = nn.Parameter(transform2noise_pos)
# self.register_parameter('transform2noise_seg', self.transform2noise_seg)
# self.register_parameter('transform2noise_pos', self.transform2noise_pos)
# self.softmax = nn.Softmax(dim=0)
else:
self.transform2noise_seg = None
self.transform2noise_pos = None
self.crf_seg = CRF(self.num_seglabels, batch_first=True)
self.crf_pos = CRF(self.num_poslabels, batch_first=True)
# self.crf_segpos = CRF(self.num_segposlabels, batch_first=True)
# self.init_weights()
def forward(self, input_data, token_type_ids=None, attention_mask=None, seglabels=None,
poslabels=None, segposlabels=None, gram_list=None, matching_matrix=None, \
channel_ids=None, position_ids=None, inputs_embeds=None, head_mask=None, \
reduction='sum', mode='train', augment=False):
input_ids = input_data
outputs = self.bert(input_ids,
attention_mask=attention_mask,
token_type_ids=token_type_ids,
position_ids=position_ids,
head_mask=head_mask,
inputs_embeds=inputs_embeds)
sequence_output = outputs[0]
if self.multi_attention is not None:
attention_output = self.multi_attention(gram_list, sequence_output, matching_matrix, channel_ids)
sequence_output = torch.cat([sequence_output, attention_output], dim=2)
# 去除[CLS]、[SEP]标签等位置，获得与label对齐的pre_label表示
origin_sequence_output = [ layer[1:] for layer in sequence_output ]
# 将sequence_output的pred_label维度padding到最大长度
padded_sequence_output = pad_sequence(origin_sequence_output, batch_first=True, padding_value=0)
# dropout pred_label的一部分feature
padded_sequence_output = self.dropout(padded_sequence_output)
#得到判别值
logits_seg = self.classifier_seg(padded_sequence_output)
logits_pos = self.classifier_pos(padded_sequence_output)
# if self.transform2noise_seg and mode=='train' and augment:
# logits_seg = self.transform2noise_seg(logits_seg)
# logits_pos = self.transform2noise_pos(logits_pos)
# logits_seg = self.tanh(logits_seg_temp)
# logits_pos = self.tanh(logits_pos_temp)
# seg_noise_matrix = self.softmax(self.transform2noise_seg)
# pos_noise_matrix = self.softmax(self.transform2noise_pos)
# print(seg_noise_matrix)
# logits_seg = torch.matmul(logits_seg,seg_noise_matrix)
# logits_pos = torch.matmul(logits_pos,pos_noise_matrix)
# logits_segpos = self.classifier_segpos(padded_sequence_output)
outputs = ((logits_seg, logits_pos),)
# outputs = ((logits_segpos),)
if segposlabels is not None:
lossseg_mask = seglabels.gt(-1)
losspos_mask = poslabels.gt(-1)
# losssegpos_mask = segposlabels.gt(-1)
loss_seg = self.crf_seg
loss_pos = self.crf_pos
# loss_segpos = self.crf_segpos
# Only keep active parts of the loss
if lossseg_mask is not None:
activeseg_mask = lossseg_mask == 1
seg_loss = loss_seg(logits_seg, seglabels, activeseg_mask, reduction) * (-1)
else:
seg_loss = loss_seg(logits_seg, seglabels, reduction=reduction) * (-1)
if losspos_mask is not None:
activepos_mask = losspos_mask == 1
pos_loss = loss_pos(logits_pos, poslabels, activepos_mask, reduction) * (-1)
else:
pos_loss = loss_pos(logits_pos, poslabels, reduction=reduction) * (-1)
# if losssegpos_mask is not None:
# activesegpos_mask = losssegpos_mask == 1
# segpos_loss = loss_segpos(logits_segpos, segposlabels, activesegpos_mask) * (-1)
# else:
# segpos_loss = loss_segpos(logits_segpos, segposlabels) * (-1)
if reduction == 'sum':
outputs = (seg_loss+pos_loss,) + outputs
else:
outputs = (seg_loss, pos_loss,) + outputs
# if reduction == 'sum':
# outputs = (segpos_loss,) + outputs
# else:
# outputs = (segpos_loss,) + outputs
return outputs