-
Notifications
You must be signed in to change notification settings - Fork 5
/
Copy pathtrainer.py
290 lines (245 loc) · 12.2 KB
/
trainer.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
# coding: utf-8
from __future__ import absolute_import, print_function, unicode_literals, division
import dynet
import random, util, sys, math, time
import argparse
import numpy as np
import sequence2sequence as s2s
import embeddings
np.random.seed(442551802)
random.seed(442551802)
# input args
parser = argparse.ArgumentParser()
## dummy argument for dynet
parser.add_argument("--dynet-mem")
## locations of data
# a single file or directory for each argument
parser.add_argument("--train_src")
parser.add_argument("--train_tgt")
parser.add_argument("--valid_src")
parser.add_argument("--valid_tgt")
parser.add_argument("--test_src")
parser.add_argument("--test_tgt")
parser.add_argument("--reader_mode", default="parallel") # format of corpus (parallel..etc)
## alternatively, load one dataset and split it
parser.add_argument("--percent_valid", default=3000, type=int)
## vocab parameters
parser.add_argument('--rebuild_vocab', action='store_true')
parser.add_argument('--unk_thresh', default=2, type=int)
## rnn parameters
parser.add_argument("--layers", default=1, type=int)
parser.add_argument("--input_dim", default=512, type=int)
parser.add_argument("--hidden_dim", default=512, type=int)
parser.add_argument("--attention_dim", default=256, type=int)
parser.add_argument("--rnn", default="lstm")
parser.add_argument("--trainer", default="adam") # from (simple_sgd, momentum_sgd, adadelta, adagrad, adam)
## word embedding specific parameters
# loss function to be used among/or combination of: phonological/morphological, bilingual, regularization, external knowledge
parser.add_argument("--loss_function", default="cross_entropy")
parser.add_argument("--write_embeddings", action='store_true') # Write embeddings to file
parser.add_argument("--extract_lvs", type=str)
## experiment parameters
parser.add_argument("--epochs", default=10, type=int)
parser.add_argument("--dropout", default=0.5, type=float)
parser.add_argument("--learning_rate", default=0.001, type=float)
parser.add_argument("--log_train_every_n", default=5000, type=int)
parser.add_argument("--log_valid_every_n", default=40000, type=int)
parser.add_argument("--log_output")
## choose what model to use
parser.add_argument("--model_type", default="basic")
parser.add_argument("--load")
parser.add_argument("--save")
parser.add_argument("--eval", action='store_true')
parser.add_argument("--eval_all", action='store_true')
parser.add_argument("--test", action='store_true')
parser.add_argument("--results_filename", default='results')
## model-specific parameters
parser.add_argument("--beam_size", default=3, type=int)
parser.add_argument("--minibatch_size", default=64, type=int)
args = parser.parse_args()
print("ARGS:", args)
if args.rnn == "lstm": args.rnn = dynet.LSTMBuilder
elif args.rnn == "gru": args.rnn = dynet.GRUBuilder
else: args.rnn = dynet.SimpleRNNBuilder
BEGIN_TOKEN = '<s>'
END_TOKEN = '<e>'
# define model and obtain vocabulary
# (reload vocab files if saved model or create new vocab files if new model)
model = dynet.Model()
if not args.trainer or args.trainer=="simple_sgd":
trainer = dynet.SimpleSGDTrainer(model)
elif args.trainer == "momentum_sgd":
trainer = dynet.MomentumSGDTrainer(model)
elif args.trainer == "adadelta":
trainer = dynet.AdadeltaTrainer(model)
elif args.trainer == "adagrad":
trainer = dynet.AdagradTrainer(model)
elif args.trainer == "adam":
trainer = dynet.AdamTrainer(model)
else:
raise Exception("Trainer not recognized! Please use one of {simple_sgd, momentum_sgd, adadelta, adagrad, adam}")
# Set sparse updates for efficiency
trainer.set_sparse_updates(True)
# Load train/valid corpus
print("Loading corpus...")
train_data_src = list(util.get_reader(args.reader_mode)(args.train_src, mode=args.reader_mode, begin=BEGIN_TOKEN, end=END_TOKEN))
train_data_tgt = list(util.get_reader(args.reader_mode)(args.train_tgt, mode=args.reader_mode, begin=BEGIN_TOKEN, end=END_TOKEN))
if args.valid_src and args.valid_tgt:
valid_data_src = list(util.get_reader(args.reader_mode)(args.valid_src, mode=args.reader_mode, begin=BEGIN_TOKEN, end=END_TOKEN))
valid_data_tgt = list(util.get_reader(args.reader_mode)(args.valid_tgt, mode=args.reader_mode, begin=BEGIN_TOKEN, end=END_TOKEN))
else:
if args.percent_valid > 1: cutoff = args.percent_valid
else: cutoff = int(len(train_data_src)*(args.percent_valid))
valid_data_src = train_data_src[-cutoff:]
valid_data_tgt = train_data_tgt[-cutoff:]
train_data_src = train_data_src[:-cutoff]
train_data_tgt = train_data_tgt[:-cutoff]
print("Train set of size", len(train_data_src), "/ Validation set of size", len(valid_data_src))
assert len(train_data_src)==len(train_data_tgt), "Number of source and target sentences in training set not equal!"
assert len(valid_data_src)==len(valid_data_tgt), "Number of source and target sentences in validation set not equal!"
print("done.")
## Sort train/valid set before minibatching
train_data_src, train_data_tgt = util.sortbylength(train_data_src, train_data_tgt, 80)
valid_data_src, valid_data_tgt = util.sortbylength(valid_data_src, valid_data_tgt, 80)
# Initialize model
S2SModel = s2s.get_s2s(args.model_type)
if args.load:
print("Loading existing model...")
s2s = S2SModel.load(model, train_data_src, train_data_tgt, args.load)
src_vocab = s2s.src_vocab
tgt_vocab = s2s.tgt_vocab
else:
print("New model. Getting vocabulary from training set...")
src_reader = util.get_reader(args.reader_mode)(args.train_src, mode=args.reader_mode, begin=BEGIN_TOKEN, end=END_TOKEN)
src_vocab = util.Vocab.load_from_corpus(src_reader, remake=args.rebuild_vocab, src_or_tgt="src")
src_vocab.START_TOK = src_vocab[BEGIN_TOKEN]
src_vocab.END_TOK = src_vocab[END_TOKEN]
src_vocab.add_unk(args.unk_thresh)
tgt_reader = util.get_reader(args.reader_mode)(args.train_tgt, mode=args.reader_mode, end=END_TOKEN)
tgt_vocab = util.Vocab.load_from_corpus(tgt_reader, remake=args.rebuild_vocab, src_or_tgt="tgt")
tgt_vocab.END_TOK = tgt_vocab[END_TOKEN]
tgt_vocab.add_unk(args.unk_thresh)
print("Source vocabulary of size", src_vocab.size, "and target vocab of size", tgt_vocab.size)
print("Creating model...")
s2s = S2SModel(model, src_vocab, tgt_vocab, args)
print("...done!")
if args.extract_lvs:
print("Writing language vectors...")
embeddings.write_embeddings(s2s.src_lookup, src_vocab, args.extract_lvs)
sys.exit("...done.")
# create log file for training
if args.log_output:
outfile = open(args.log_output, 'w')
outfile.write("")
outfile.close()
if args.eval_all:
print("Evaluating for all languages..")
s2s.all_langs_translate()
# only evaluate existing model on test data
if args.eval:
print("Evaluating model...")
test_data_src = list(util.get_reader(args.reader_mode)(args.test_src, mode=args.reader_mode, begin=BEGIN_TOKEN, end=END_TOKEN))
test_data_tgt = list(util.get_reader(args.reader_mode)(args.test_tgt, mode=args.reader_mode, begin=BEGIN_TOKEN, end=END_TOKEN))
with open(args.test_src) as f:
lines = f.readlines()
#combined = list(zip(test_data_src, test_data_tgt))
#random.shuffle(combined)
#test_data_src[:], test_data_tgt[:] = zip(*combined)
test_data = zip(test_data_src, test_data_tgt)
s2s.translate(lines, test_data, args.results_filename, "final")
if args.test:
s2s.evaluate(test_data)
sys.exit("...done.")
else: raise Exception("Test file path argument missing")
sys.exit("...done.")
if args.plot_embeddings:
s2s.tsne_embeddings(test_data)
sys.exit("...done.")
# Shuffle training data for bible model
#combined = list(zip(train_data_src, train_data_tgt))
#random.shuffle(combined)
#train_data_src[:], train_data_tgt[:] = zip(*combined)
# Store starting index of each minibatch
if args.minibatch_size != 0:
train_order = [x*args.minibatch_size for x in range(int(len(train_data_src)/args.minibatch_size + (1 if len(train_data_src)%args.minibatch_size != 0 else 0)))]
valid_order = [x*args.minibatch_size for x in range(int(len(valid_data_src)/args.minibatch_size + (1 if len(valid_data_src)%args.minibatch_size != 0 else 0)))]
else:
train_order = range(len(train_data))
valid_order = range(len(valid_data))
# run training loop
word_count = sent_count = cum_loss = 0.0
try:
for ITER in range(args.epochs):
s2s.epoch = ITER
random.shuffle(train_order)
sample_num = 0
_start = time.time()
log_start = time.time()
for i, sid in enumerate(train_order):
# Retrieving batch from training data
batched_src = train_data_src[sid : sid + args.minibatch_size]
batched_tgt = train_data_tgt[sid : sid + args.minibatch_size]
sample_num = 1 + i
if sample_num % (int(args.log_train_every_n/args.minibatch_size)) == 0:
print("[training_set] Epoch:", ITER, "Batch:", sample_num)
trainer.status()
print("Loss:", cum_loss / word_count, "Time elapsed:", (time.time() - _start) ,"WPS:", word_count/(time.time() - log_start))
# sample = lm.beam_search_generate(src, beam_n=args.beam_size)
# sample = s2s.generate(src, sampled=False)
word_count = sent_count = cum_loss = 0.0
log_start = time.time()
print
# end of test logging
if sample_num % (int(args.log_valid_every_n/args.minibatch_size)) == 0:
v_word_count = v_sent_count = v_cum_loss = v_cum_bleu = v_cum_em = 0.0
v_start = time.time()
for vid in valid_order:
batched_v_src = valid_data_src[vid : vid + args.minibatch_size]
batched_v_tgt = valid_data_tgt[vid : vid + args.minibatch_size]
v_loss = s2s.get_batch_loss(batched_v_src, batched_v_tgt)
v_cum_loss += v_loss.scalar_value()
# v_cum_em += s2s.get_em(batched_v_src, batched_v_tgt)
# v_cum_bleu += s2s.get_bleu(v_src, v_tgt, args.beam_size)
v_word_count += sum([(len(tgt_sent) - 1) for tgt_sent in batched_v_tgt])
v_sent_count += args.minibatch_size
print("[Validation Set", str(sample_num) + "]\t" + \
"Loss:", str(v_cum_loss / v_word_count) + "\t" + \
"Perplexity:", str(np.exp(v_cum_loss / v_word_count)) + "\t" + \
# "BLEU: "+str(v_cum_bleu / v_sent_count) + "\t" + \
# "EM: " +str(v_cum_em / v_sent_count) + "\t" + \
"Time elapsed:", str(time.time() - v_start))
if args.log_output:
print("(logging to", args.log_output + ")")
with open(args.log_output, "a") as outfile:
outfile.write(str(ITER) + "\t" + \
str(sample_num) + "\t" + \
str(v_cum_loss / v_word_count) + "\t" + \
str(np.exp(v_cum_loss / v_word_count)) + "\n")
# str(v_cum_em / v_sent_count) + "\t" + \
# str(v_cum_bleu / v_sent_count) + "\n")
s2s.translate(zip(valid_data_src, valid_data_tgt), args.results_filename, sample_num, ITER)
if args.save:
print("saving checkpoint...")
s2s.save(args.save + ".checkpoint")
# end of validation logging
loss = s2s.get_batch_loss(batched_src, batched_tgt)
loss_value = loss.value()
cum_loss += loss_value * args.minibatch_size
word_count += sum([(len(tgt_sent) - 1) for tgt_sent in batched_tgt])
sent_count += args.minibatch_size
ppl = np.exp((loss_value * args.minibatch_size) / word_count)
loss.backward()
trainer.update()
### end of batch train loop
trainer.update_epoch()
### end of epoch
### end of training loop
except KeyboardInterrupt:
if args.save:
print("saving...")
model.save(args.save)
print("Unexpected error:", sys.exc_info()[0])
raise
if args.save:
print("saving...")
s2s.save(args.save)