/
gokc_model.py
323 lines (265 loc) · 14.2 KB
/
gokc_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
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
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
#!/usr/bin/env python
import torch
import torch.nn as nn
from torch.nn.utils import clip_grad_norm_
import torch.nn.functional as F
from base_model import BaseModel
from util.embedder import Embedder
from encoder import RNNEncoder
from decoder import Decoder
from util.criterions import NLLLoss, CopyGeneratorLoss
from util.misc import Pack
from evaluation.metrics import accuracy, perplexity
from util.attention import Attention
class GOKC(BaseModel):
def __init__(self, src_vocab_size, tgt_vocab_size, cue_vocab_size, goal_vocab_size, embed_size, hidden_size,
padding_idx=None, num_layers=1, bidirectional=True, attn_mode="mlp", with_bridge=False,
tie_embedding=False, dropout=0.0, use_gpu=False, use_bow=False, use_kd=False,
use_posterior=False, device=None, unk_idx=None, force_copy=True, stage=None):
super().__init__()
self.src_vocab_size = src_vocab_size
self.tgt_vocab_size = tgt_vocab_size
self.cue_vocab_size = cue_vocab_size
self.goal_vocab_size = goal_vocab_size
self.embed_size = embed_size
self.hidden_size = hidden_size
self.padding_idx = padding_idx
self.num_layers = num_layers
self.bidirectional = bidirectional
self.attn_mode = attn_mode
self.with_bridge = with_bridge
self.tie_embedding = tie_embedding
self.dropout = dropout
self.use_gpu = use_gpu
self.use_bow = use_bow
self.use_kd = use_kd
self.use_posterior = use_posterior
self.baseline = 0
self.device = device if device >= 0 else "cpu"
self.unk_idx = unk_idx
self.force_copy = force_copy
self.stage = stage
# the utterance embedding
enc_embedder = Embedder(num_embeddings=self.src_vocab_size,
embedding_dim=self.embed_size, padding_idx=self.padding_idx)
self.utt_encoder = RNNEncoder(input_size=self.embed_size, hidden_size=self.hidden_size,
embedder=enc_embedder, num_layers=self.num_layers,
bidirectional=self.bidirectional, dropout=self.dropout)
if self.with_bridge:
self.utt_bridge = nn.Sequential(nn.Linear(self.hidden_size, self.hidden_size), nn.Tanh())
self.goal_bridge = nn.Sequential(nn.Linear(self.hidden_size, self.hidden_size), nn.Tanh())
# self.prior_query_mlp = nn.Sequential(nn.Linear(self.hidden_size * 2, self.hidden_size), nn.Tanh())
self.fc1 = nn.Linear(self.hidden_size, self.hidden_size)
self.fc2 = nn.Linear(self.hidden_size, self.hidden_size)
self.fc3 = nn.Linear(self.hidden_size * 2, 1)
if self.tie_embedding:
# share the same embedding with utt encoder
assert self.src_vocab_size == self.tgt_vocab_size == self.cue_vocab_size == self.goal_vocab_size
self.dec_embedder = enc_embedder
knowledge_embedder = enc_embedder
goal_embedder = enc_embedder
else:
self.dec_embedder = Embedder(num_embeddings=self.tgt_vocab_size,
embedding_dim=self.embed_size,
padding_idx=self.padding_idx)
knowledge_embedder = Embedder(num_embeddings=self.cue_vocab_size,
embedding_dim=self.embed_size,
padding_idx=self.padding_idx)
goal_embedder = Embedder(num_embeddings=self.goal_vocab_size,
embedding_dim=self.embed_size,
padding_idx=self.padding_idx)
self.knowledge_encoder = RNNEncoder(input_size=self.embed_size,
hidden_size=self.hidden_size,
embedder=knowledge_embedder,
num_layers=self.num_layers,
bidirectional=self.bidirectional,
dropout=self.dropout)
self.goal_encoder = RNNEncoder(input_size=self.embed_size,
hidden_size=self.hidden_size,
embedder=goal_embedder,
num_layers=self.num_layers,
bidirectional=self.bidirectional,
dropout=self.dropout)
self.prior_attention = Attention(query_size=self.hidden_size,
memory_size=self.hidden_size,
hidden_size=self.hidden_size,
mode="dot",
device=self.device)
self.posterior_attention = Attention(query_size=self.hidden_size,
memory_size=self.hidden_size,
hidden_size=self.hidden_size,
mode="dot",
device=self.device)
self.decoder = Decoder(input_size=self.embed_size, hidden_size=self.hidden_size,
output_size=self.tgt_vocab_size, embedder=self.dec_embedder,
num_layers=self.num_layers, attn_mode=self.attn_mode,
memory_size=self.hidden_size, dropout=self.dropout,
device=self.device)
self.softmax = nn.Softmax(dim=-1)
self.sigmoid = nn.Sigmoid()
self.tanh = nn.Tanh()
if self.use_bow:
self.bow_output_layer = nn.Sequential(
nn.Linear(in_features=self.hidden_size, out_features=self.hidden_size),
nn.Tanh(),
nn.Linear(in_features=self.hidden_size, out_features=self.tgt_vocab_size),
nn.LogSoftmax(dim=-1))
if self.use_kd:
self.knowledge_dropout = nn.Dropout(self.dropout)
if self.padding_idx is not None:
self.weight = torch.ones(self.tgt_vocab_size)
self.weight[self.padding_idx] = 0
else:
self.weight = None
self.nll_loss = NLLLoss(weight=self.weight, ignore_index=self.padding_idx,
reduction='mean')
self.copy_gen_loss = CopyGeneratorLoss(vocab_size=self.tgt_vocab_size,
force_copy=self.force_copy,
unk_index=self.unk_idx,
ignore_index=self.padding_idx)
self.kl_loss = torch.nn.KLDivLoss(reduction="mean")
if self.use_gpu:
self.cuda()
self.weight = self.weight.cuda()
def encode(self, inputs, is_training=False):
"""
encode
"""
outputs = Pack()
# utt encoding info.
utt_inputs = _, utt_lengths = inputs.src[0][:, 1:-1], inputs.src[1] - 2
utt_enc_outputs, utt_enc_hidden = self.utt_encoder(utt_inputs)
if self.with_bridge:
utt_enc_hidden = self.utt_bridge(utt_enc_hidden)
# goal encoding info.
goal_inputs = _, goal_lengths = inputs.goal[0][:, 1:-1], inputs.goal[1] - 2
# goal_enc_hidden.size == [1, batch_size, hidden_size]
goal_enc_outputs, goal_enc_hidden = self.goal_encoder(goal_inputs)
if self.with_bridge:
goal_enc_hidden = self.goal_bridge(goal_enc_hidden)
# knowledge
batch_size, sent_num, sent = inputs.cue[0].size()
tmp_len = inputs.cue[1] # [batch_size, sent_num]
tmp_len[tmp_len > 0] -= 2
cue_inputs = inputs.cue[0].view(-1, sent)[:, 1:-1], tmp_len.view(-1)
# cue_enc_hidden.size() == [1, batch_size * sent_num, hidden_size]
# cue_enc_outputs.size() == [batch_size * sent_num, sent_len, hidden_size]
cue_enc_outputs, cue_enc_hidden = self.knowledge_encoder(cue_inputs)
# cue_enc_hidden[-1].size() == [batch_size * sent_num, hidden_size]
# [batch_size, sent_num, hidden_size]
cue_enc_outputs = cue_enc_outputs.view(batch_size, sent_num, cue_enc_outputs.size(-2), -1)
cue_outputs = cue_enc_hidden[-1].view(batch_size, sent_num, -1)
# Attention
p_U = self.tanh(self.fc1(utt_enc_hidden[-1].unsqueeze(0)))
p_G = self.tanh(self.fc2(goal_enc_hidden[-1].unsqueeze(0)))
k = self.sigmoid(self.fc3(torch.cat([p_U, p_G], dim=-1)))
prior_query = self.tanh(k * utt_enc_hidden + (1 - k) * goal_enc_hidden)
weighted_cue, cue_attn = self.prior_attention(query=prior_query[-1].unsqueeze(1),
memory=self.tanh(cue_outputs),
mask=inputs.cue[1].eq(0))
prior_attn = cue_attn.squeeze(1)
outputs.add(prior_attn=prior_attn)
posterior_attn = None
if self.use_posterior:
tgt_enc_inputs = inputs.tgt[0][:, 1:-1], inputs.tgt[1] - 2
_, tgt_enc_hidden = self.knowledge_encoder(tgt_enc_inputs)
posterior_weighted_cue, posterior_attn = self.posterior_attention(
query=tgt_enc_hidden[-1].unsqueeze(1),
memory=self.tanh(cue_outputs),
mask=inputs.cue[1].eq(0))
posterior_attn = posterior_attn.squeeze(1)
outputs.add(posterior_attn=posterior_attn)
knowledge = posterior_weighted_cue
if self.use_kd:
knowledge = self.knowledge_dropout(knowledge)
if self.use_bow:
bow_logits = self.bow_output_layer(knowledge)
outputs.add(bow_logits=bow_logits)
# Initialize the context vector of decoder
dec_init_context = torch.zeros(size=[batch_size, 1, self.hidden_size],
dtype=torch.float,
device=self.device)
dec_init_state = self.decoder.initialize_state(
utt_hidden=utt_enc_hidden,
utt_outputs=utt_enc_outputs if self.attn_mode else None,
utt_input_len=utt_lengths if self.attn_mode else None,
cue_hidden=cue_outputs.transpose(0, 1),
cue_outputs=cue_enc_outputs if self.attn_mode else None,
cue_input_len=tmp_len if self.attn_mode else None,
goal_hidden=goal_enc_hidden,
goal_outputs=goal_enc_outputs,
goal_input_len=goal_lengths,
pr_attn_dist=prior_attn,
po_attn_dist=posterior_attn,
dec_init_context=dec_init_context
)
return outputs, dec_init_state
def decode(self, input, state, oovs_max, src_extend_vocab, cue_extend_vocab, goal_extend_vocab):
output, dec_state = self.decoder.decode(input=input,
state=state,
oovs_max=oovs_max,
valid_src_extend_vocab=src_extend_vocab,
valid_cue_extend_vocab=cue_extend_vocab,
valid_goal_extend_vocab=goal_extend_vocab)
return output, dec_state
def forward(self, enc_inputs, dec_inputs, is_training=False):
outputs, dec_init_state = self.encode(
enc_inputs, is_training=is_training)
log_probs, _ = self.decoder(dec_inputs, dec_init_state, is_training=is_training)
outputs.add(logits=log_probs)
return outputs
def collect_metrics(self, outputs, oovs_target, no_extend_target, epoch=-1):
num_samples = no_extend_target.size(0)
metrics = Pack(num_samples=num_samples)
loss = 0
logits = outputs.logits # [batch_size, dec_seq_len, vocab_size]
nll_loss_ori = self.copy_gen_loss(scores=logits.transpose(1, 2).contiguous(),
align=oovs_target,
target=no_extend_target) # [batch_size, tgt_len]
nll_loss = torch.mean(torch.sum(nll_loss_ori, dim=-1))
num_words = no_extend_target.ne(self.padding_idx).sum() # .item()
ppl = nll_loss_ori.sum() / num_words
ppl = ppl.exp()
acc = accuracy(logits, no_extend_target, padding_idx=self.padding_idx)
metrics.add(nll=(nll_loss, num_words), acc=acc, ppl=ppl)
if self.use_posterior:
kl_loss = self.kl_loss(torch.log(outputs.prior_attn + 1e-20),
outputs.posterior_attn.detach())
metrics.add(kl=kl_loss)
if self.stage == 1:
loss += nll_loss
loss += kl_loss
if self.use_bow:
bow_logits = outputs.bow_logits # size = [batch_size, 1, vocab_size]
bow_logits = bow_logits.repeat(1, no_extend_target.size(-1), 1)
bow = self.nll_loss(bow_logits, no_extend_target)
loss += bow
metrics.add(bow=bow)
else:
loss += nll_loss
metrics.add(loss=loss)
return metrics
def iterate(self, inputs, optimizer=None, grad_clip=None, is_training=False, epoch=-1):
enc_inputs = inputs
dec_inputs = (inputs.tgt[0][:, :-1],
inputs.tgt[1] - 1,
inputs.src_extend_vocab,
inputs.cue_extend_vocab,
inputs.goal_extend_vocab,
inputs.merge_oovs_str)
outputs = self.forward(enc_inputs, dec_inputs, is_training=is_training)
oovs_target = inputs.tgt_oovs_vocab[0][:, 1:]
no_extend_target = inputs.tgt[0][:, 1:]
metrics = self.collect_metrics(outputs, oovs_target, no_extend_target, epoch=epoch)
loss = metrics.loss
if torch.isnan(loss):
raise ValueError("nan loss encountered")
if is_training:
assert optimizer is not None
optimizer.zero_grad()
loss.backward()
if grad_clip is not None and grad_clip > 0:
clip_grad_norm_(parameters=self.parameters(),
max_norm=grad_clip)
optimizer.step()
return metrics