-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathseq2seq_model.py
223 lines (194 loc) · 9.15 KB
/
seq2seq_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
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
import random
from modeling_unilm import *
from transformers import BertPreTrainedModel
from loss_function import LabelSmoothLoss
from utils import BeamHypotheses
class Seq2SeqModel(BertPreTrainedModel):
def __init__(self, config, tokenizer, output_max_length=60):
super(Seq2SeqModel, self).__init__(config)
self.tokenizer = tokenizer
self.bert = BertModel(config)
self.cls = BertOnlyMLMHead(config)
# self.loss_fct = nn.CrossEntropyLoss(reduction='none', ignore_index=0)
self.loss_fct = LabelSmoothLoss(smoothing=0.1)
self.output_max_length = output_max_length
def forward(self,
input_ids,
token_type_ids,
position_ids,
attention_mask,
answer_tag_ids=None):
outputs = self.forward_step(
# input_ids=input_ids,
input_ids=self.random_choice(input_ids, token_type_ids, attention_mask),
token_type_ids=token_type_ids,
position_ids=position_ids,
attention_mask=attention_mask,
answer_tag_ids=answer_tag_ids,
)
predictions = outputs[:, :-1, :].contiguous()
output_ids = input_ids[:, 1:].contiguous()
output_mask = token_type_ids[:, 1:].contiguous()
# loss = self.compute_loss_for_sentence(predictions, labels=output_ids, target_mask=output_mask)
loss = self.compute_loss(predictions, labels=output_ids, target_mask=output_mask)
return loss
def forward_step(self,
input_ids,
token_type_ids,
position_ids,
attention_mask,
answer_tag_ids=None):
# the step for one time forward
extend_attention_mask = self.create_attention_mask(
token_type_ids,
attention_mask,
)
output = self.bert(
input_ids=input_ids,
token_type_ids=token_type_ids,
position_ids=position_ids,
attention_mask=extend_attention_mask,
answer_tag_ids=answer_tag_ids,
)
sequence_output = output[0]
predictions = self.cls(sequence_output)
return predictions
def random_choice(self, input_ids, token_type_ids, attention_mask, type=1):
# type 0, no random choice
# type 1, target sentence random choice
# type 2, source and target all use random choice, in this type,
# you should use mlm task for source sentence
if type == 0:
return input_ids
# random choice
if random.random() < 0.5:
input_ids = self._random_choice(input_ids, target_mask=token_type_ids)
elif type == 2:
input_ids = self._random_choice(
input_ids=input_ids,
target_mask=(1 - token_type_ids) & attention_mask
)
return input_ids
def _random_choice(self, input_ids, target_mask):
batch_size, sequence_length = input_ids.shape
device = input_ids.device
vocab_size = self.bert.embeddings.word_embeddings.weight.shape[0]
rand_ids = torch.rand(batch_size, sequence_length) * vocab_size
rand_ids = rand_ids.to(input_ids.dtype).to(device) % vocab_size
rand_mask = torch.rand(batch_size, sequence_length).to(device) > 0.3
disturb_ids = torch.where(rand_mask, input_ids, rand_ids)
input_ids = torch.where(target_mask == 0, input_ids, disturb_ids)
return input_ids
def compute_loss(self, predictions, labels, target_mask):
# compute loss for batch
vocab_size = predictions.shape[-1]
predictions = predictions.view(-1, vocab_size)
labels = labels.view(-1)
target_mask = target_mask.view(-1).float()
loss = self.loss_fct(predictions, labels) * target_mask
return loss.sum() / target_mask.sum()
def compute_loss_for_sentence(self, predictions, labels, target_mask):
# compute loss for sentence, then compute the mean of loss
batch_size = predictions.shape[0]
loss = 0.0
for i in range(batch_size):
loss += self.compute_loss(predictions[i], labels[i], target_mask[i])
return loss / batch_size
def create_attention_mask(self, token_type_ids, attention_mask):
idxs = torch.cumsum(token_type_ids, dim=1)
mask = idxs[:, None, :] <= idxs[:, :, None]
mask = mask.to(token_type_ids.dtype)
attention_mask = attention_mask[:, None, :]
return mask & attention_mask
def beam_serach_predict(self,
input_ids,
token_type_ids,
position_ids,
attention_mask,
beam_size=5):
# too slow...
# predict sentence one by one.
batch_size, sequence_length = input_ids.shape
device = input_ids.device
output_ids = torch.empty(beam_size, 0, dtype=input_ids.dtype, device=device)
output_scores = torch.zeros(beam_size, dtype=torch.float, device=device)
output_scores[1:] = -1e9
generated_hyps = BeamHypotheses(num_beams=beam_size, max_length=self.output_max_length, length_penalty=1)
vocab_size = self.bert.embeddings.word_embeddings.weight.shape[0]
end_id = self.tokenizer.token_to_id(self.tokenizer._token_end)
# end_id = self.tokenizer.token_to_id('？')
for step in range(self.output_max_length):
if step == 0:
input_ids = input_ids.repeat(beam_size, 1)
token_type_ids = token_type_ids.repeat(beam_size, 1)
position_ids = position_ids.repeat(beam_size, 1)
attention_mask = attention_mask.repeat(beam_size, 1)
predictions = self.forward_step(
input_ids=input_ids,
token_type_ids=token_type_ids,
position_ids=position_ids,
attention_mask=attention_mask,
)
pred_scores = torch.log_softmax(predictions[:, -1, :], dim=-1)
next_scores = output_scores[:, None] + pred_scores
next_scores = next_scores.reshape(-1)
next_scores, next_ids = torch.topk(next_scores, beam_size, dim=-1, largest=True, sorted=True)
beam_ids = next_ids // vocab_size
token_ids = next_ids % vocab_size
for beam_id, token_id, next_score in zip(beam_ids, token_ids, next_scores):
if token_id == end_id:
generated_hyps.add(output_ids[beam_id].tolist(), next_score.item())
if generated_hyps.is_done(next_scores[0], step):
break
pred_ids = token_ids.reshape(beam_size, -1)
# reorder input and output
output_ids = output_ids[beam_ids, :]
output_scores = next_scores.view(-1)
input_ids = input_ids[beam_ids, :]
token_type_ids = token_type_ids[beam_ids, :]
position_ids = position_ids[beam_ids, :]
attention_mask = attention_mask[beam_ids, :]
# add new id in output ids
output_ids = torch.cat([output_ids, pred_ids], dim=-1)
input_ids = torch.cat([input_ids, pred_ids], dim=-1)
token_type_ids = torch.cat([token_type_ids, torch.ones_like(pred_ids)], dim=-1)
position_ids = torch.cat([position_ids, position_ids[:, -1:]+1], dim=-1)
attention_mask = torch.cat([attention_mask, torch.ones_like(pred_ids)], dim=-1)
sorted_hyps = sorted(generated_hyps.beams, key=lambda x: x[0])
try:
result = sorted_hyps.pop()[1]
except:
result = output_ids[0].tolist()
return [result]
def predict(self,
input_ids,
token_type_ids,
position_ids,
attention_mask,
answer_tag_ids=None,
):
# predict sentence one by one.
batch_size, sequence_length = input_ids.shape
device = input_ids.device
output_ids = torch.empty(batch_size, 0, dtype=input_ids.dtype, device=device)
end_id = self.tokenizer.token_to_id(self.tokenizer._token_end)
# end_id = self.tokenizer.token_to_id('？')
for step in range(self.output_max_length):
predictions = self.forward_step(
input_ids=input_ids,
token_type_ids=token_type_ids,
position_ids=position_ids,
attention_mask=attention_mask,
answer_tag_ids=answer_tag_ids
)
_, pred_ids = torch.max(predictions[:, -1, :], dim=-1, keepdim=True)
if pred_ids == end_id:
break
output_ids = torch.cat([output_ids, pred_ids], dim=-1)
input_ids = torch.cat([input_ids, pred_ids], dim=-1)
token_type_ids = torch.cat([token_type_ids, torch.ones_like(pred_ids)], dim=-1)
position_ids = torch.cat([position_ids, position_ids[:, -1:] + 1], dim=-1)
attention_mask = torch.cat([attention_mask, torch.ones_like(pred_ids)], dim=-1)
if answer_tag_ids is not None:
answer_tag_ids = torch.cat([answer_tag_ids, torch.ones_like(pred_ids)], dim=-1)
return output_ids.tolist()