/
Utils.py
208 lines (173 loc) · 6.42 KB
/
Utils.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
import torch
import numpy as np
import random
import time
import codecs
from Constants import *
from torch.distributions.categorical import *
import torch.nn.functional as F
from modules.Utils import *
def get_ms():
return time.time() * 1000
def init_seed(seed=None):
if seed is None:
seed = int(get_ms() // 1000)
np.random.seed(seed)
torch.manual_seed(seed)
random.seed(seed)
torch.cuda.manual_seed(seed)
torch.cuda.manual_seed_all(seed)
def importance_sampling(prob,topk):
m = Categorical(logits=prob)
indices = m.sample((topk,)).transpose(0,1) # batch, topk
values = prob.gather(1, indices)
return values, indices
def sequence_mask(lengths, max_len=None):
"""
Creates a boolean mask from sequence lengths.
"""
batch_size = lengths.numel()
max_len = max_len or lengths.max()
mask= (torch.arange(0, max_len)
.type_as(lengths)
.repeat(batch_size, 1)
.lt(lengths.unsqueeze(1)))
if torch.cuda.is_available():
mask=mask.cuda()
return mask
def start_end_mask(starts, ends, max_len):
batch_size=len(starts)
mask = torch.arange(1, max_len + 1)
if torch.cuda.is_available():
mask = mask.cuda()
mask = mask.unsqueeze(0).expand(batch_size, -1)
mask1 = mask >= starts.unsqueeze(1).expand_as(mask)
mask2 = mask <= ends.unsqueeze(1).expand_as(mask)
mask = (mask1 * mask2)
return mask
def decode_to_end(model, data, vocab2id, max_target_length=None, schedule_rate=1, softmax=False, encode_outputs=None, init_decoder_states=None, tgt=None):
# if tgt is None:
# tgt = data['output']
batch_size = len(data['id'])
if max_target_length is None:
max_target_length = tgt.size(1)
if encode_outputs is None:
encode_outputs = model.encode(data)
if init_decoder_states is None:
init_decoder_states = model.init_decoder_states(data, encode_outputs)
decoder_input = new_tensor([vocab2id[BOS_WORD]] * batch_size, requires_grad=False)
prob = torch.ones((batch_size,)) * schedule_rate
if torch.cuda.is_available():
prob=prob.cuda()
all_gen_outputs = list()
all_decode_outputs = [dict({'state': init_decoder_states})]
for t in range(max_target_length):
# decoder_outputs, decoder_states,...
decode_outputs = model.decode(
data, decoder_input, encode_outputs, all_decode_outputs[-1]
)
output = model.generate(data, encode_outputs, decode_outputs, softmax=softmax)
all_gen_outputs.append(output)
all_decode_outputs.append(decode_outputs)
if schedule_rate >=1:
decoder_input = tgt[:, t]
elif schedule_rate<=0:
probs, ids = model.to_word(data, output, 1)
decoder_input = model.generation_to_decoder_input(data, ids[:, 0])
else:
probs, ids = model.to_word(data, output, 1)
indices = model.generation_to_decoder_input(data, ids[:, 0])
draws = torch.bernoulli(prob).long()
decoder_input = tgt[:, t] * draws + indices * (1 - draws)
# all_gen_outputs = torch.cat(all_gen_outputs, dim=0).transpose(0, 1).contiguous()
return encode_outputs, init_decoder_states, all_decode_outputs, all_gen_outputs
def randomk(gen_output, k=5, PAD=None, BOS=None, UNK=None):
if PAD is not None:
gen_output[:, PAD] = -float('inf')
if BOS is not None:
gen_output[:, BOS] = -float('inf')
if UNK is not None:
gen_output[:, UNK] = -float('inf')
values, indices = importance_sampling(gen_output, k)
# words=[[tgt_id2vocab[id.item()] for id in one] for one in indices]
return values, indices
def topk(gen_output, k=5, PAD=None, BOS=None, UNK=None):
if PAD is not None:
gen_output[:, PAD] = 0
if BOS is not None:
gen_output[:, BOS] = 0
if UNK is not None:
gen_output[:, UNK] = 0
if k>1:
values, indices = torch.topk(gen_output, k, dim=1, largest=True,
sorted=True, out=None)
else:
values, indices = torch.max(gen_output, dim=1, keepdim=True)
return values, indices
def copy_topk(gen_output, vocab_map, vocab_overlap, k=5, PAD=None, BOS=None, UNK=None):
vocab=gen_output[:, :vocab_map.size(-1)]
dy_vocab=gen_output[:, vocab_map.size(-1):]
vocab=vocab+torch.bmm(dy_vocab.unsqueeze(1), vocab_map).squeeze(1)
dy_vocab=dy_vocab*vocab_overlap
gen_output=torch.cat([vocab, dy_vocab], dim=-1)
return topk(gen_output, k, PAD=PAD, BOS=BOS, UNK=UNK)
def remove_duplicate_once(sents, n=3):
changed=False
for b in range(len(sents)):
sent=sents[b]
if len(sent)<=n:
continue
for i in range(len(sent)-n):
index= len(sent) - i - n
if all(elem in sent[:index] for elem in sent[index:]):
sents[b]=sent[:index]
changed=True
break
return changed
def remove_duplicate(sents, n=3):
changed = remove_duplicate_once(sents, n)
while changed:
changed = remove_duplicate_once(sents, n)
def to_sentence(batch_indices, id2vocab):
batch_size=len(batch_indices)
summ=list()
for i in range(batch_size):
indexes=batch_indices[i]
text_summ2 = []
for index in indexes:
index = index.item()
w = id2vocab[index]
if w == BOS_WORD or w == PAD_WORD:
continue
if w == EOS_WORD:
break
text_summ2.append(w)
if len(text_summ2)==0:
text_summ2.append(UNK_WORD)
summ.append(text_summ2)
return summ
def to_copy_sentence(data, batch_indices, id2vocab, dyn_id2vocab_map):
ids=data['id']
batch_size=len(batch_indices)
summ=list()
for i in range(batch_size):
indexes=batch_indices[i]
text_summ2 = []
dyn_id2vocab=dyn_id2vocab_map[ids[i].item()]
for index in indexes:
index = index.item()
if index < len(id2vocab):
w = id2vocab[index]
elif index - len(id2vocab) in dyn_id2vocab:
w = dyn_id2vocab[index - len(id2vocab)]
else:
w = PAD_WORD
if w == BOS_WORD or w == PAD_WORD:
continue
if w == EOS_WORD:
break
text_summ2.append(w)
if len(text_summ2)==0:
text_summ2.append(UNK_WORD)
summ.append(text_summ2)
return summ