-
Notifications
You must be signed in to change notification settings - Fork 2
/
Copy pathmodelling.py
240 lines (178 loc) · 9.42 KB
/
modelling.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
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
231
232
233
234
235
236
237
238
239
240
import torch
from transformers import BertPreTrainedModel
from transformers import AutoModelForTokenClassification, TrainingArguments, Trainer, AutoConfig, AutoModel
from torch import nn
class Simple_BERT(BertPreTrainedModel):
def __init__(self, config, actual_model_checkpoint):
super(Simple_BERT, self).__init__(config)
self.bert = AutoModel.from_pretrained(actual_model_checkpoint)
self.classifier = nn.Linear(768, 5)
self.softmax = nn.Softmax(dim=2)
self.criterion = nn.CrossEntropyLoss(reduction='none')
self.token_dropout = nn.Dropout(0.1)
def forward(self, input_ids, attn_mask, labels, lambda_max_loss = 0, lambda_mask_loss = 0):
lhs = self.bert(input_ids, attn_mask).last_hidden_state
logits = self.classifier(self.token_dropout(lhs))
#ypreds = torch.argmax(self.softmax(logits), dim=2)
active_loss = attn_mask.view(-1) == 1
active_logits = logits.view(-1, 5)[active_loss]
active_labels = labels.view(-1)[active_loss]
loss = self.criterion(active_logits, active_labels)
loss_crossEntropy = torch.mean(loss)
if lambda_max_loss == 0.0 and lambda_mask_loss == 0.0:
ypreds = torch.argmax(self.softmax(logits), dim=2)
return ypreds, loss_crossEntropy
batch_size = input_ids.shape[0]
max_seq_len = input_ids.shape[1]
active_loss = active_loss.view(batch_size, max_seq_len)
active_max = []
active_mask = []
start_id = 0
for i in range(batch_size):
sent_len = torch.sum(active_loss[i])
# mask-loss
if lambda_mask_loss != 0.0:
active_mask.append((input_ids[i] == 103)[: sent_len]) # id of [MASK] is 103, according to the bertTokenizer
# max-loss
if lambda_max_loss != 0.0:
end_id = start_id + sent_len
active_max.append(torch.max(loss[start_id: end_id]))
start_id = end_id
if lambda_max_loss != 0:
loss_max = torch.mean(torch.stack(active_max))
else:
loss_max = 0.0
if lambda_mask_loss != 0:
active_mask = torch.cat(active_mask)
if sum(active_mask) != 0:
loss_mask = torch.sum(loss[active_mask]) / sum(active_mask)
else:
loss_mask = 0.0
ypreds = torch.argmax(self.softmax(logits), dim=2)
#loss = self.criterion(logits.view(-1, 5), labels.view(-1))
return ypreds, loss_crossEntropy + lambda_max_loss * loss_max + lambda_mask_loss * loss_mask
#loss = self.criterion(logits.view(-1, 5), labels.view(-1))
#return ypreds, loss
class Transform_CharacterBERT(BertPreTrainedModel):
def __init__(self, config, actual_model_checkpoint, char_vocab, embed_matrix, tokenizer, max_word_len, cnn_size):
super(Transform_CharacterBERT, self).__init__(config)
self.bert = AutoModel.from_pretrained(actual_model_checkpoint)
self.classifier = nn.Linear(868, 5)
self.softmax = nn.Softmax(dim=2)
self.criterion = nn.CrossEntropyLoss(reduction='none')
self.char_embeddings = CharEmbeddings(len(char_vocab), 300, embed_matrix,max_word_len, cnn_size, 0.1)
self.token_dropout = nn.Dropout(0.1)
self.tokenizer = tokenizer
self.max_word_len = max_word_len
self.conv_filter_size = cnn_size
self.char_vocab = char_vocab
def get_char_seq(self, words, max_len):
char_seq = list()
for i in range(0, self.max_word_len + self.conv_filter_size - 1): #### CLS
char_seq.append(self.char_vocab['<PAD>'])
for i in range(0, self.conv_filter_size - 1): #### Extra Padding
char_seq.append(self.char_vocab['<PAD>'])
for word in words:
for c in word[0:min(len(word), self.max_word_len)]:
if c in self.char_vocab:
char_seq.append(self.char_vocab[c])
else:
char_seq.append(self.char_vocab['<UNK>'])
pad_len = self.max_word_len - len(word)
for i in range(0, pad_len):
char_seq.append(self.char_vocab['<PAD>'])
for i in range(0, self.conv_filter_size - 1):
char_seq.append(self.char_vocab['<PAD>'])
pad_len = max_len - len(words) - 1
for i in range(0, pad_len):
for i in range(0, self.max_word_len + self.conv_filter_size - 1):
char_seq.append(self.char_vocab['<PAD>'])
return char_seq
def get_all_char_seq_tensor(self, input_ids):
all_char_seqs = []
max_len = input_ids.shape[1]
for i in range(len(input_ids)):
i_id = input_ids[i]
all_toks = self.tokenizer.convert_ids_to_tokens(i_id, skip_special_tokens = True)
char_seq = self.get_char_seq(all_toks, max_len)
all_char_seqs.append(char_seq)
return torch.tensor(all_char_seqs)
def forward(self, input_ids, attn_mask, labels, lambda_max_loss = 0, lambda_mask_loss = 0):
lhs = self.bert(input_ids, attn_mask).last_hidden_state
#print('LHS', lhs.device)
all_char_seqs = self.get_all_char_seq_tensor(input_ids)
all_char_seqs = all_char_seqs.to('cuda')
#print('allchar', all_char_seqs.device)
char_token_embs = self.char_embeddings(all_char_seqs)
char_token_embs = char_token_embs.to('cuda')
#print('chartok', char_token_embs.device)
lhs_plus_char = torch.cat([lhs, char_token_embs], dim = -1)
# print("lhs shape", lhs.shape)
logits = self.classifier(self.token_dropout(lhs_plus_char))
active_loss = attn_mask.view(-1) == 1
active_logits = logits.view(-1, 5)[active_loss]
active_labels = labels.view(-1)[active_loss]
loss = self.criterion(active_logits, active_labels)
loss_crossEntropy = torch.mean(loss)
if lambda_max_loss == 0.0 and lambda_mask_loss == 0.0:
ypreds = torch.argmax(self.softmax(logits), dim=2)
return ypreds, loss_crossEntropy
batch_size = input_ids.shape[0]
max_seq_len = input_ids.shape[1]
active_loss = active_loss.view(batch_size, max_seq_len)
active_max = []
active_mask = []
start_id = 0
for i in range(batch_size):
sent_len = torch.sum(active_loss[i])
# mask-loss
if lambda_mask_loss != 0.0:
active_mask.append((input_ids[i] == 103)[: sent_len]) # id of [MASK] is 103, according to the bertTokenizer
# max-loss
if lambda_max_loss != 0.0:
end_id = start_id + sent_len
active_max.append(torch.max(loss[start_id: end_id]))
start_id = end_id
if lambda_max_loss != 0:
loss_max = torch.mean(torch.stack(active_max))
else:
loss_max = 0.0
if lambda_mask_loss != 0:
active_mask = torch.cat(active_mask)
if sum(active_mask) != 0:
loss_mask = torch.sum(loss[active_mask]) / sum(active_mask)
else:
loss_mask = 0.0
ypreds = torch.argmax(self.softmax(logits), dim=2)
#loss = self.criterion(logits.view(-1, 5), labels.view(-1))
return ypreds, loss_crossEntropy + lambda_max_loss * loss_max + lambda_mask_loss * loss_mask
# lhs = self.bert(input_ids, attn_mask).last_hidden_state
# #print('LHS', lhs.device)
# all_char_seqs = self.get_all_char_seq_tensor(input_ids)
# all_char_seqs = all_char_seqs.to('cuda')
# #print('allchar', all_char_seqs.device)
# char_token_embs = self.char_embeddings(all_char_seqs)
# char_token_embs = char_token_embs.to('cuda')
# #print('chartok', char_token_embs.device)
# lhs_plus_char = torch.cat([lhs, char_token_embs], dim = -1)
# # print("lhs shape", lhs.shape)
# logits = self.classifier(self.token_dropout(lhs_plus_char))
# ypreds = torch.argmax(self.softmax(logits), dim=2)
# loss = self.criterion(logits.view(-1, 5), labels.view(-1))
# return ypreds, loss
class CharEmbeddings(nn.Module):
def __init__(self, vocab_size, embed_dim, embed_matrix, max_word_len, conv_filter_size, drop_out_rate):
super(CharEmbeddings, self).__init__()
self.embeddings = nn.Embedding(vocab_size, embed_dim, padding_idx=0) ### B * Pad_length * 300
self.embeddings.weight.data.copy_((embed_matrix))
self.embeddings.weight.requires_grad = True
self.conv1d = nn.Conv1d(embed_dim, 100, conv_filter_size)
self.max_pool = nn.MaxPool1d(max_word_len + conv_filter_size - 1, max_word_len + conv_filter_size - 1)
self.dropout = nn.Dropout(drop_out_rate)
def forward(self, char_seq):
char_embeds = self.embeddings(char_seq) ### B * Pad_length * 300
char_embeds = self.dropout(char_embeds) ##same
char_embeds = char_embeds.permute(0, 2, 1) ##B*300_pad_lenght
char_feature = torch.tanh(self.max_pool(self.conv1d(char_embeds))) ##B*100*num_of_tokens
char_feature = char_feature.permute(0, 2, 1) ##B * num_of_tokens * 100
return char_feature