model.py
8.77 KB
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
# Copyright (c) Microsoft Corporation.
# Licensed under the MIT license.
import torch
import torch.nn as nn
class Seq2Seq(nn.Module):
"""
Build Seqence-to-Sequence.
Parameters:
* `encoder`- encoder of seq2seq model. e.g. roberta
* `decoder`- decoder of seq2seq model. e.g. transformer
* `config`- configuration of encoder model.
* `beam_size`- beam size for beam search.
* `max_length`- max length of target for beam search.
* `sos_id`- start of symbol ids in target for beam search.
* `eos_id`- end of symbol ids in target for beam search.
"""
def __init__(self, encoder,decoder,config,beam_size=None,max_length=None,sos_id=None,eos_id=None):
super(Seq2Seq, self).__init__()
self.encoder = encoder
self.decoder=decoder
self.config=config
self.register_buffer("bias", torch.tril(torch.ones(2048, 2048)))
self.dense = nn.Linear(config.hidden_size, config.hidden_size)
self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
self.lsm = nn.LogSoftmax(dim=-1)
self.tie_weights()
self.beam_size=beam_size
self.max_length=max_length
self.sos_id=sos_id
self.eos_id=eos_id
def _tie_or_clone_weights(self, first_module, second_module):
""" Tie or clone module weights depending of weither we are using TorchScript or not
"""
if self.config.torchscript:
first_module.weight = nn.Parameter(second_module.weight.clone())
else:
first_module.weight = second_module.weight
def tie_weights(self):
""" Make sure we are sharing the input and output embeddings.
Export to TorchScript can't handle parameter sharing so we are cloning them instead.
"""
self._tie_or_clone_weights(self.lm_head,
self.encoder.embeddings.word_embeddings)
def forward(self, source_ids=None,source_mask=None,target_ids=None,target_mask=None,patch_ids=None,args=None):
outputs = self.encoder(source_ids, attention_mask=source_mask, patch_ids=patch_ids)
encoder_output = outputs[0].permute([1,0,2]).contiguous()
if target_ids is not None:
attn_mask=-1e4 *(1-self.bias[:target_ids.shape[1],:target_ids.shape[1]])
tgt_embeddings = self.encoder.embeddings(target_ids).permute([1,0,2]).contiguous()
out = self.decoder(tgt_embeddings,encoder_output,tgt_mask=attn_mask,memory_key_padding_mask=(1-source_mask).bool())
hidden_states = torch.tanh(self.dense(out)).permute([1,0,2]).contiguous()
lm_logits = self.lm_head(hidden_states)
# Shift so that tokens < n predict n
active_loss = target_mask[..., 1:].ne(0).view(-1) == 1
shift_logits = lm_logits[..., :-1, :].contiguous()
shift_labels = target_ids[..., 1:].contiguous()
# Flatten the tokens
loss_fct = nn.CrossEntropyLoss(ignore_index=-1)
loss = loss_fct(shift_logits.view(-1, shift_logits.size(-1))[active_loss],
shift_labels.view(-1)[active_loss])
outputs = loss,loss*active_loss.sum(),active_loss.sum()
return outputs
else:
#Predict
preds=[]
zero=torch.cuda.LongTensor(1).fill_(0)
for i in range(source_ids.shape[0]):
context=encoder_output[:,i:i+1]
context_mask=source_mask[i:i+1,:]
beam = Beam(self.beam_size,self.sos_id,self.eos_id)
input_ids=beam.getCurrentState()
context=context.repeat(1, self.beam_size,1)
context_mask=context_mask.repeat(self.beam_size,1)
for _ in range(self.max_length):
if beam.done():
break
attn_mask=-1e4 *(1-self.bias[:input_ids.shape[1],:input_ids.shape[1]])
tgt_embeddings = self.encoder.embeddings(input_ids).permute([1,0,2]).contiguous()
out = self.decoder(tgt_embeddings,context,tgt_mask=attn_mask,memory_key_padding_mask=(1-context_mask).bool())
out = torch.tanh(self.dense(out))
hidden_states=out.permute([1,0,2]).contiguous()[:,-1,:]
out = self.lsm(self.lm_head(hidden_states)).data
beam.advance(out)
input_ids.data.copy_(input_ids.data.index_select(0, beam.getCurrentOrigin()))
input_ids=torch.cat((input_ids,beam.getCurrentState()),-1)
hyp= beam.getHyp(beam.getFinal())
pred=beam.buildTargetTokens(hyp)[:self.beam_size]
pred=[torch.cat([x.view(-1) for x in p]+[zero]*(self.max_length-len(p))).view(1,-1) for p in pred]
preds.append(torch.cat(pred,0).unsqueeze(0))
preds=torch.cat(preds,0)
return preds
class Beam(object):
def __init__(self, size,sos,eos):
self.size = size
self.tt = torch.cuda
# The score for each translation on the beam.
self.scores = self.tt.FloatTensor(size).zero_()
# The backpointers at each time-step.
self.prevKs = []
# The outputs at each time-step.
self.nextYs = [self.tt.LongTensor(size)
.fill_(0)]
self.nextYs[0][0] = sos
# Has EOS topped the beam yet.
self._eos = eos
self.eosTop = False
# Time and k pair for finished.
self.finished = []
def getCurrentState(self):
"Get the outputs for the current timestep."
batch = self.tt.LongTensor(self.nextYs[-1]).view(-1, 1)
return batch
def getCurrentOrigin(self):
"Get the backpointers for the current timestep."
return self.prevKs[-1]
def advance(self, wordLk):
"""
Given prob over words for every last beam `wordLk` and attention
`attnOut`: Compute and update the beam search.
Parameters:
* `wordLk`- probs of advancing from the last step (K x words)
* `attnOut`- attention at the last step
Returns: True if beam search is complete.
"""
numWords = wordLk.size(1)
# Sum the previous scores.
if len(self.prevKs) > 0:
beamLk = wordLk + self.scores.unsqueeze(1).expand_as(wordLk)
# Don't let EOS have children.
for i in range(self.nextYs[-1].size(0)):
if self.nextYs[-1][i] == self._eos:
beamLk[i] = -1e20
else:
beamLk = wordLk[0]
flatBeamLk = beamLk.view(-1)
bestScores, bestScoresId = flatBeamLk.topk(self.size, 0, True, True)
self.scores = bestScores
# bestScoresId is flattened beam x word array, so calculate which
# word and beam each score came from
prevK = bestScoresId // numWords
self.prevKs.append(prevK)
self.nextYs.append((bestScoresId - prevK * numWords))
for i in range(self.nextYs[-1].size(0)):
if self.nextYs[-1][i] == self._eos:
s = self.scores[i]
self.finished.append((s, len(self.nextYs) - 1, i))
# End condition is when top-of-beam is EOS and no global score.
if self.nextYs[-1][0] == self._eos:
self.eosTop = True
def done(self):
return self.eosTop and len(self.finished) >=self.size
def getFinal(self):
if len(self.finished) == 0:
self.finished.append((self.scores[0], len(self.nextYs) - 1, 0))
self.finished.sort(key=lambda a: -a[0])
if len(self.finished) != self.size:
unfinished=[]
for i in range(self.nextYs[-1].size(0)):
if self.nextYs[-1][i] != self._eos:
s = self.scores[i]
unfinished.append((s, len(self.nextYs) - 1, i))
unfinished.sort(key=lambda a: -a[0])
self.finished+=unfinished[:self.size-len(self.finished)]
return self.finished[:self.size]
def getHyp(self, beam_res):
"""
Walk back to construct the full hypothesis.
"""
hyps=[]
for _,timestep, k in beam_res:
hyp = []
for j in range(len(self.prevKs[:timestep]) - 1, -1, -1):
hyp.append(self.nextYs[j+1][k])
k = self.prevKs[j][k]
hyps.append(hyp[::-1])
return hyps
def buildTargetTokens(self, preds):
sentence=[]
for pred in preds:
tokens = []
for tok in pred:
if tok==self._eos:
break
tokens.append(tok)
sentence.append(tokens)
return sentence