-
Notifications
You must be signed in to change notification settings - Fork 2
/
search.py
205 lines (172 loc) · 7.35 KB
/
search.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
import torch
from util import use_cuda
beam_size = 5
MAX_DECODING_STEPS = 12
EPS = 1e-8
class Beam(object):
def __init__(self, word2idx, h_t, c_t):
self.word2idx = word2idx
# (beam_size, t) after t time steps
self.tokens = torch.LongTensor(beam_size, 1).fill_(self.word2idx['<SOS>'])
# (beam_size, 1), Initial score of beams = -30
self.scores = torch.FloatTensor(beam_size, 1).fill_(-30)
self.tokens = use_cuda(self.tokens)
self.scores = use_cuda(self.scores)
# At time step t = 0, all beams should extend from a single beam,
# so giving high initial score to 1st beam
self.scores[0][0] = 0
self.h_t = h_t.unsqueeze(0).repeat(beam_size, 1)
self.c_t = c_t.unsqueeze(0).repeat(beam_size, 1)
self.et_sum = None
self.dec_out = None
self.done = False
def get_current_state(self):
tokens = self.tokens[:,-1].clone()
for i in range(len(tokens)):
if tokens[i].item() >= len(self.word2idx):
tokens[i] = self.word2idx['<UNK>']
return tokens
def advance(self, prob_dist, h_t, c_t, et_sum, dec_out):
'''
:param prob_dist: (beam, n_extended_vocab)
:param h_t: (beam, hidden_dim)
:param c_t: (beam, hidden_dim)
:param et_sum: (beam, n_seq)
:param dec_out: (beam, t, n_hid)
'''
n_extended_vocab = prob_dist.size(1)
log_probs = torch.log(prob_dist + EPS)
# maintain for each beam overall score so far
scores = log_probs + self.scores
# extract the top beam candidates out of beam*n_extended_vocab candidates, (beam*n_extended_vocab, 1)
scores_t = scores.view(-1,1)
# will be sorted in descending order of scores
best_scores, best_scores_id = torch.topk(input=scores_t, k=beam_size, dim=0)
self.scores = best_scores
beams_order = best_scores_id.squeeze(1) // n_extended_vocab
best_words = best_scores_id % n_extended_vocab
self.h_t = h_t[beams_order]
self.c_t = c_t[beams_order]
if et_sum is not None:
self.et_sum = et_sum[beams_order]
if dec_out is not None:
self.dec_out = dec_out[beams_order]
self.tokens = self.tokens[beams_order]
self.tokens = torch.cat([self.tokens, best_words], dim=1)
#End condition is when top-of-beam is EOS.
if best_words[0][0] == self.word2idx['<EOS>']:
self.done = True
def get_best(self):
# Since beams are always in sorted (descending) order, 1st beam is the best beam
best_token = self.tokens[0].cpu().numpy().tolist()
try:
end_idx = best_token.index(self.word2idx['<EOS>'])
except ValueError:
end_idx = len(best_token)
best_token = best_token[1:end_idx]
return best_token
def get_all(self):
all_tokens = []
for i in range(len(self.tokens)):
all_tokens.append(self.tokens[i].cpu().numpy())
return all_tokens
def beam_search(h_t, c_t, enc_out, dec_tar, enc_padding_mask, enc_ext_vocab,
max_zeros_ext_vocab, model, word2idx, hidden_dim, evaluation='val'):
batch_size = len(h_t)
beam_idx = torch.LongTensor(list(range(batch_size)))
# For each example in batch, create Beam object
beams = [Beam(word2idx, h_t[i], c_t[i]) for i in range(batch_size)]
# Index of beams that are active, i.e: didn't generate [STOP] yet
n_rem = batch_size
et_sum = None
dec_out = None
# limit is set to max length steps for beam search decoding
max_dec_steps = dec_tar.shape[1] if evaluation == 'val' else MAX_DECODING_STEPS
for t in range(max_dec_steps):
inputs = torch.stack(
[beam.get_current_state() for beam in beams if beam.done == False]
).contiguous().view(-1)
dec_h = torch.stack(
[beam.h_t for beam in beams if beam.done == False]
).contiguous().view(-1, hidden_dim)
dec_c = torch.stack(
[beam.c_t for beam in beams if beam.done == False]
).contiguous().view(-1, hidden_dim)
if et_sum is not None:
# rem*beam_size, n_seq
et_sum = torch.stack(
[beam.et_sum for beam in beams if beam.done == False]
).contiguous().view(-1, enc_out.size(1))
if dec_out is not None:
# rem*beam_size, t-1, n_hid
dec_out = torch.stack(
[beam.dec_out for beam in beams if beam.done == False]
).contiguous().view(-1, t, hidden_dim)
# following steps make a bigger batch size by multiplying beam-width with (remaining) batch size
enc_out_beam = enc_out[beam_idx].view(n_rem, -1).repeat(1, beam_size).view(-1, enc_out.size(1), enc_out.size(2))
enc_pad_mask_beam = enc_padding_mask[beam_idx].repeat(1, beam_size).view(-1, enc_padding_mask.size(1))
max_zeros_ext_vocab_beam = None
if max_zeros_ext_vocab is not None:
max_zeros_ext_vocab_beam = max_zeros_ext_vocab[beam_idx].repeat(1, beam_size).view(-1, max_zeros_ext_vocab.size(1))
enc_ext_vocab_beam = enc_ext_vocab[beam_idx].repeat(1, beam_size).view(-1, enc_ext_vocab.size(1))
inputs = model.embedding_matrix(inputs)
p_y, dec_h, dec_c, et_sum, dec_out = model.decoder(t, dec_h, dec_c, enc_out_beam, dec_out,
et_sum, enc_pad_mask_beam, enc_ext_vocab_beam,
max_zeros_ext_vocab_beam, inputs)
# following steps separate the (remaining) batch size dimension from beam width dimension
p_y = p_y.view(n_rem, beam_size, -1)
dec_h = dec_h.view(n_rem, beam_size, -1)
dec_c = dec_c.view(n_rem, beam_size, -1)
if et_sum is not None:
# rem, beam_size, n_seq
et_sum = et_sum.view(n_rem, beam_size, -1)
if dec_out is not None:
# rem, beam_size, t, hidden_dim
dec_out = dec_out.view(n_rem, beam_size, -1, hidden_dim)
# For all the active beams, perform beam search
# indices of active beams after beam search
active = []
for i in range(n_rem):
# here beam means batch
b = beam_idx[i].item()
beam = beams[b]
if beam.done:
continue
et_sum_i = prev_s_i = None
if et_sum is not None:
et_sum_i = et_sum[i]
if dec_out is not None:
dec_out_i = dec_out[i]
beam.advance(p_y[i], dec_h[i], dec_c[i], et_sum_i, dec_out_i)
if beam.done == False:
active.append(b)
if len(active) == 0:
break
beam_idx = torch.LongTensor(active)
n_rem = len(beam_idx)
predicted_words = []
for beam in beams:
predicted_words.append(beam.get_best())
return predicted_words
def search(h_t, c_t, enc_out, enc_inp_len, dec_tar, enc_padding_mask, enc_ext_vocab,
max_zeros_ext_vocab, model, word2idx, greedy=True, evaluation='val'):
et_sum = None
dec_out = None
inputs = torch.zeros_like(enc_inp_len).fill_(word2idx['<SOS>'])
outputs = []
max_dec_steps = dec_tar.shape[1] if evaluation == 'val' else MAX_DECODING_STEPS
for t in range(max_dec_steps):
inputs = model.embedding_matrix(inputs)
p_y, h_t, c_t, et_sum, dec_out = model.decoder(t, h_t, c_t, enc_out, dec_out, et_sum,
enc_padding_mask, enc_ext_vocab,
max_zeros_ext_vocab, inputs)
if greedy:
_, inputs = torch.max(p_y, dim=1)
else:
inputs = torch.multinomial(p_y, 1).squeeze(1)
outputs.append(inputs)
is_oov = (inputs >= len(word2idx)).type(torch.LongTensor)
is_oov = use_cuda(is_oov)
inputs = is_oov * word2idx['<UNK>'] + (1 - is_oov) * inputs
outputs = torch.stack(outputs, dim=1).cpu().numpy()
return outputs