/
iterative_refinement_generator.py
223 lines (193 loc) · 8.01 KB
/
iterative_refinement_generator.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
# Copyright (c) Facebook, Inc. and its affiliates.
#
# This source code is licensed under the MIT license found in the
# LICENSE file in the root directory of this source tree.
from collections import namedtuple
import torch
from fairseq import utils
DecoderOut = namedtuple('IterativeRefinementDecoderOut', [
'output_tokens',
'output_scores',
'attn',
'step',
'max_step',
])
class IterativeRefinementGenerator(object):
def __init__(
self,
tgt_dict,
models=None,
eos_penalty=0.0,
max_iter=10,
max_ratio=2,
decoding_format=None,
retain_dropout=False,
adaptive=True,
):
"""
Generates translations based on iterative refinement.
Args:
tgt_dict: target dictionary
eos_penalty: if > 0.0, it penalized early-stopping in decoding
max_iter: maximum number of refinement iterations
max_ratio: generate sequences of maximum length ax, where x is the source length
decoding_format: decoding mode in {'unigram', 'ensemble', 'vote', 'dp', 'bs'}
retain_dropout: retaining dropout in the inference
adaptive: decoding with early stop
"""
self.bos = tgt_dict.bos()
self.pad = tgt_dict.pad()
self.unk = tgt_dict.unk()
self.eos = tgt_dict.eos()
self.vocab_size = len(tgt_dict)
self.eos_penalty = eos_penalty
self.max_iter = max_iter
self.max_ratio = max_ratio
self.decoding_format = decoding_format
self.retain_dropout = retain_dropout
self.adaptive = adaptive
self.models = models
def generate_batched_itr(
self,
data_itr,
maxlen_a=None,
maxlen_b=None,
cuda=False,
timer=None,
prefix_size=0,
):
"""Iterate over a batched dataset and yield individual translations.
Args:
maxlen_a/b: generate sequences of maximum length ax + b,
where x is the source sentence length.
cuda: use GPU for generation
timer: StopwatchMeter for timing generations.
"""
for sample in data_itr:
if "net_input" not in sample:
continue
if timer is not None:
timer.start()
with torch.no_grad():
hypos = self.generate(
self.models,
sample,
prefix_tokens=sample["target"][:, :prefix_size]
if prefix_size > 0
else None,
)
if timer is not None:
timer.stop(sample["ntokens"])
for i, id in enumerate(sample["id"]):
# remove padding
src = utils.strip_pad(sample["net_input"]["src_tokens"][i, :], self.pad)
ref = utils.strip_pad(sample["target"][i, :], self.pad)
yield id, src, ref, hypos[i]
@torch.no_grad()
def generate(self, models, sample, prefix_tokens=None):
from fairseq.models.levenshtein_transformer import LevenshteinTransformerModel
from fairseq.models.nonautoregressive_ensembles import EnsembleLevT
if len(models) == 1:
# Keep this for other NAT models for which we have yet to implement ensemble wrappers. Later delete this.
model = models[0]
elif isinstance(models[0], LevenshteinTransformerModel):
model = EnsembleLevT(models)
else:
raise NotImplementedError
if not self.retain_dropout:
model.eval()
# TODO: better encoder inputs?
src_tokens = sample["net_input"]["src_tokens"]
src_lengths = sample["net_input"]["src_lengths"]
bsz, src_len = src_tokens.size()
sent_idxs = torch.arange(bsz)
# encoding
encoder_out = model.forward_encoder([src_tokens, src_lengths])
# initialize buffers (very model specific, with length prediction or not)
prev_decoder_out = model.initialize_output_tokens(encoder_out, src_tokens)
prev_output_tokens = prev_decoder_out.output_tokens.clone()
finalized = [[] for _ in range(bsz)]
def is_a_loop(x, y, s, a):
b, l_x, l_y = x.size(0), x.size(1), y.size(1)
if l_x > l_y:
y = torch.cat([y, x.new_zeros(b, l_x - l_y).fill_(self.pad)], 1)
s = torch.cat([s, s.new_zeros(b, l_x - l_y)], 1)
if a is not None:
a = torch.cat([a, a.new_zeros(b, l_x - l_y, a.size(2))], 1)
elif l_x < l_y:
x = torch.cat([x, y.new_zeros(b, l_y - l_x).fill_(self.pad)], 1)
return (x == y).all(1), y, s, a
def finalized_hypos(step, prev_out_token, prev_out_score, prev_out_attn):
cutoff = prev_out_token.ne(self.pad)
tokens = prev_out_token[cutoff]
scores = prev_out_score[cutoff]
if prev_out_attn is None:
hypo_attn, alignment = None, None
else:
hypo_attn = prev_out_attn[cutoff]
alignment = hypo_attn.max(dim=1)[1]
return {
"steps": step,
"tokens": tokens,
"positional_scores": scores,
"score": scores.mean(),
"hypo_attn": hypo_attn,
"alignment": alignment,
}
for step in range(self.max_iter + 1):
decoder_options = {
"eos_penalty": self.eos_penalty,
"max_ratio": self.max_ratio,
"decoding_format": self.decoding_format,
}
prev_decoder_out = prev_decoder_out._replace(
step=step,
max_step=self.max_iter + 1,
)
decoder_out = model.forward_decoder(
prev_decoder_out, encoder_out, **decoder_options
)
if self.adaptive:
# terminate if there is a loop
terminated, out_tokens, out_scores, out_attn = is_a_loop(
prev_output_tokens, decoder_out.output_tokens, decoder_out.output_scores, decoder_out.attn
)
decoder_out = decoder_out._replace(
output_tokens=out_tokens,
output_scores=out_scores,
attn=out_attn,
)
else:
terminated = decoder_out.output_tokens.new_zeros(decoder_out.output_tokens.size(0)).bool()
if step == self.max_iter: # reach last iteration, terminate
terminated.fill_(1)
# collect finalized sentences
finalized_idxs = sent_idxs[terminated]
finalized_tokens = decoder_out.output_tokens[terminated]
finalized_scores = decoder_out.output_scores[terminated]
finalized_attn = (
None if decoder_out.attn is None else decoder_out.attn[terminated]
)
for i in range(finalized_idxs.size(0)):
finalized[finalized_idxs[i]] = [
finalized_hypos(
step,
finalized_tokens[i],
finalized_scores[i],
None if finalized_attn is None else finalized_attn[i],
)
]
# check if all terminated
if terminated.sum() == terminated.size(0):
break
# for next step
not_terminated = ~terminated
prev_decoder_out = decoder_out._replace(
output_tokens=decoder_out.output_tokens[not_terminated],
output_scores=decoder_out.output_scores[not_terminated],
attn=decoder_out.attn[not_terminated] if decoder_out.attn is not None else None,
)
encoder_out = model.encoder.reorder_encoder_out(encoder_out, not_terminated.nonzero().squeeze())
sent_idxs = sent_idxs[not_terminated]
prev_output_tokens = prev_decoder_out.output_tokens.clone()
return finalized