forked from SonjayaJetBrain/VNMT
-
Notifications
You must be signed in to change notification settings - Fork 0
/
sample.py
309 lines (267 loc) · 11.4 KB
/
sample.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
#!/usr/bin/env python
import argparse
import cPickle
import traceback
import logging
import time
import sys
import numpy
import experiments.nmt
from experiments.nmt import\
prototype_state,\
parse_input
from encdec import RNNEncoderDecoder
from experiments.nmt.numpy_compat import argpartition
logger = logging.getLogger(__name__)
class Timer(object):
def __init__(self):
self.total = 0
def start(self):
self.start_time = time.time()
def finish(self):
self.total += time.time() - self.start_time
class BeamSearch(object):
def __init__(self, enc_dec):
self.enc_dec = enc_dec
state = self.enc_dec.state
self.eos_id = state['null_sym_target']
self.unk_id = state['unk_sym_target']
def compile(self):
self.comp_repr = self.enc_dec.create_representation_computer()
self.comp_init_states = self.enc_dec.create_initializers()
self.comp_next_probs = self.enc_dec.create_next_probs_computer()
self.comp_next_states = self.enc_dec.create_next_states_computer()
def search(self, seq, n_samples, ignore_unk=False, minlen=1):
c = self.comp_repr(seq)[0]
states = map(lambda x : x[None, :], self.comp_init_states(c))
dim = states[0].shape[1]
num_levels = len(states)
fin_trans = []
fin_costs = []
trans = [[]]
costs = [0.0]
for k in range(3 * len(seq)):
if n_samples == 0:
break
# Compute probabilities of the next words for
# all the elements of the beam.
beam_size = len(trans)
last_words = (numpy.array(map(lambda t : t[-1], trans))
if k > 0
else numpy.zeros(beam_size, dtype="int64"))
log_probs = numpy.log(self.comp_next_probs(c, k, last_words, *states)[0])
# Adjust log probs according to search restrictions
if ignore_unk:
log_probs[:,self.unk_id] = -numpy.inf
# TODO: report me in the paper!!!
if k < minlen:
log_probs[:,self.eos_id] = -numpy.inf
# Find the best options by calling argpartition of flatten array
next_costs = numpy.array(costs)[:, None] - log_probs
flat_next_costs = next_costs.flatten()
best_costs_indices = argpartition(
flat_next_costs.flatten(),
n_samples)[:n_samples]
# Decypher flatten indices
voc_size = log_probs.shape[1]
trans_indices = best_costs_indices / voc_size
word_indices = best_costs_indices % voc_size
costs = flat_next_costs[best_costs_indices]
# Form a beam for the next iteration
new_trans = [[]] * n_samples
new_costs = numpy.zeros(n_samples)
new_states = [numpy.zeros((n_samples, dim), dtype="float32") for level
in range(num_levels)]
inputs = numpy.zeros(n_samples, dtype="int64")
for i, (orig_idx, next_word, next_cost) in enumerate(
zip(trans_indices, word_indices, costs)):
new_trans[i] = trans[orig_idx] + [next_word]
new_costs[i] = next_cost
for level in range(num_levels):
new_states[level][i] = states[level][orig_idx]
inputs[i] = next_word
new_states = self.comp_next_states(c, k, inputs, *new_states)
# Filter the sequences that end with end-of-sequence character
trans = []
costs = []
indices = []
for i in range(n_samples):
if new_trans[i][-1] != self.enc_dec.state['null_sym_target']:
trans.append(new_trans[i])
costs.append(new_costs[i])
indices.append(i)
else:
n_samples -= 1
fin_trans.append(new_trans[i])
fin_costs.append(new_costs[i])
states = map(lambda x : x[indices], new_states)
# Dirty tricks to obtain any translation
if not len(fin_trans):
if ignore_unk:
logger.warning("Did not manage without UNK")
return self.search(seq, n_samples, False, minlen)
elif n_samples < self.enc_dec.state['maxsamp']:
logger.warning("Still no translations: try beam size {}".format(n_samples * 2))
return self.search(seq, n_samples * 2, False, minlen)
else:
# changed by BiaoZhang for selflogging
'''
logger.error("Translation failed")
'''
logger.warning("Translation failed")
fin_trans = numpy.array([[self.unk_id, self.eos_id]])
fin_costs = numpy.array([-numpy.inf])
else:
fin_trans = numpy.array(fin_trans)[numpy.argsort(fin_costs)]
fin_costs = numpy.array(sorted(fin_costs))
return fin_trans, fin_costs
def indices_to_words(i2w, seq):
sen = []
for k in xrange(len(seq)):
if i2w[seq[k]] == '<eol>':
break
sen.append(i2w[seq[k]])
return sen
def sample(lm_model, seq, n_samples,
sampler=None, beam_search=None,
ignore_unk=False, normalize=False,
alpha=1, verbose=False):
if beam_search:
sentences = []
trans, costs = beam_search.search(seq, n_samples,
ignore_unk=ignore_unk, minlen=len(seq) / 2)
if normalize:
counts = [len(s) for s in trans]
costs = [co / cn for co, cn in zip(costs, counts)]
for i in range(len(trans)):
sen = indices_to_words(lm_model.word_indxs, trans[i])
sentences.append(" ".join(sen))
for i in range(len(costs)):
if verbose:
print "{}: {}".format(costs[i], sentences[i])
return sentences, costs, trans
elif sampler:
sentences = []
all_probs = []
costs = []
values, cond_probs = sampler(n_samples, 3 * (len(seq) - 1), alpha, seq)
for sidx in xrange(n_samples):
sen = []
for k in xrange(values.shape[0]):
if lm_model.word_indxs[values[k, sidx]] == '<eol>':
break
sen.append(lm_model.word_indxs[values[k, sidx]])
sentences.append(" ".join(sen))
probs = cond_probs[:, sidx]
probs = numpy.array(cond_probs[:len(sen) + 1, sidx])
all_probs.append(numpy.exp(-probs))
costs.append(-numpy.sum(probs))
if normalize:
counts = [len(s.strip().split(" ")) for s in sentences]
costs = [co / cn for co, cn in zip(costs, counts)]
sprobs = numpy.argsort(costs)
if verbose:
for pidx in sprobs:
print "{}: {} {} {}".format(pidx, -costs[pidx], all_probs[pidx], sentences[pidx])
print
return sentences, costs, None
else:
raise Exception("I don't know what to do")
def parse_args():
parser = argparse.ArgumentParser(
"Sample (of find with beam-serch) translations from a translation model")
parser.add_argument("--state",
required=True, help="State to use")
parser.add_argument("--beam-search",
action="store_true", help="Beam size, turns on beam-search")
parser.add_argument("--beam-size",
type=int, help="Beam size")
parser.add_argument("--ignore-unk",
default=False, action="store_true",
help="Ignore unknown words")
parser.add_argument("--source",
help="File of source sentences")
parser.add_argument("--trans",
help="File to save translations in")
parser.add_argument("--normalize",
action="store_true", default=False,
help="Normalize log-prob with the word count")
parser.add_argument("--verbose",
action="store_true", default=False,
help="Be verbose")
parser.add_argument("model_path",
help="Path to the model")
parser.add_argument("changes",
nargs="?", default="",
help="Changes to state")
return parser.parse_args()
def main():
args = parse_args()
state = prototype_state()
with open(args.state) as src:
state.update(cPickle.load(src))
state.update(eval("dict({})".format(args.changes)))
logging.basicConfig(level=getattr(logging, state['level']), format="%(asctime)s: %(name)s: %(levelname)s: %(message)s")
rng = numpy.random.RandomState(state['seed'])
enc_dec = RNNEncoderDecoder(state, rng, skip_init=True)
enc_dec.build()
lm_model = enc_dec.create_lm_model()
lm_model.load(args.model_path)
indx_word = cPickle.load(open(state['word_indx'],'rb'))
sampler = None
beam_search = None
if args.beam_search:
beam_search = BeamSearch(enc_dec)
beam_search.compile()
else:
sampler = enc_dec.create_sampler(many_samples=True)
idict_src = cPickle.load(open(state['indx_word'],'r'))
if args.source and args.trans:
# Actually only beam search is currently supported here
assert beam_search
assert args.beam_size
fsrc = open(args.source, 'r')
ftrans = open(args.trans, 'w')
start_time = time.time()
n_samples = args.beam_size
total_cost = 0.0
logging.debug("Beam size: {}".format(n_samples))
for i, line in enumerate(fsrc):
seqin = line.strip()
seq, parsed_in = parse_input(state, indx_word, seqin, idx2word=idict_src)
if args.verbose:
print "Parsed Input:", parsed_in
trans, costs, _ = sample(lm_model, seq, n_samples, sampler=sampler,
beam_search=beam_search, ignore_unk=args.ignore_unk, normalize=args.normalize)
best = numpy.argmin(costs)
print >>ftrans, trans[best]
if args.verbose:
print "Translation:", trans[best]
total_cost += costs[best]
if (i + 1) % 100 == 0:
ftrans.flush()
logger.debug("Current speed is {} per sentence".
format((time.time() - start_time) / (i + 1)))
print "Total cost of the translations: {}".format(total_cost)
fsrc.close()
ftrans.close()
else:
while True:
try:
seqin = raw_input('Input Sequence: ')
n_samples = int(raw_input('How many samples? '))
alpha = None
if not args.beam_search:
alpha = float(raw_input('Inverse Temperature? '))
seq,parsed_in = parse_input(state, indx_word, seqin, idx2word=idict_src)
print "Parsed Input:", parsed_in
except Exception:
print "Exception while parsing your input:"
traceback.print_exc()
continue
sample(lm_model, seq, n_samples, sampler=sampler,
beam_search=beam_search,
ignore_unk=args.ignore_unk, normalize=args.normalize,
alpha=alpha, verbose=True)
if __name__ == "__main__":
main()