/
main.py
280 lines (235 loc) · 10.6 KB
/
main.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
#!/usr/bin/env python3
# coding: utf-8
import sys
import random
import math
from argparse import ArgumentParser
from collections import defaultdict
import numpy as np
from primitiv import Device, Graph, Optimizer
from primitiv import devices as D
from primitiv import optimizers as O
from utils import (
make_vocab, load_corpus, load_corpus_ref, count_labels, make_batch,
save_ppl, make_inv_vocab, line_to_sent, load_ppl
)
from bleu import get_bleu_stats, calculate_bleu
from bahdanau_encdec import EncoderDecoder
SRC_TRAIN_FILE = "data/train.en"
TRG_TRAIN_FILE = "data/train.ja"
SRC_VALID_FILE = "data/dev.en"
TRG_VALID_FILE = "data/dev.ja"
SRC_TEST_FILE = "data/test.en"
REF_TEST_FILE = "data/test.ja"
# Training encode decode model.
def train(encdec, optimizer, args, best_valid_ppl):
prefix = args.model
max_epoch = args.epoch
batch_size = args.minibatch
# Registers all parameters to the optimizer.
optimizer.add(encdec)
# Loads vocab.
src_vocab = make_vocab(SRC_TRAIN_FILE, args.src_vocab)
trg_vocab = make_vocab(TRG_TRAIN_FILE, args.trg_vocab)
inv_trg_vocab = make_inv_vocab(trg_vocab)
print("#src_vocab:", len(src_vocab))
print("#trg_vocab:", len(trg_vocab))
# Loads all corpus
train_src_corpus = load_corpus(SRC_TRAIN_FILE, src_vocab)
train_trg_corpus = load_corpus(TRG_TRAIN_FILE, trg_vocab)
valid_src_corpus = load_corpus(SRC_VALID_FILE, src_vocab)
valid_trg_corpus = load_corpus(TRG_VALID_FILE, trg_vocab)
test_src_corpus = load_corpus(SRC_TEST_FILE, src_vocab)
test_ref_corpus = load_corpus_ref(REF_TEST_FILE, trg_vocab)
num_train_sents = len(train_trg_corpus)
num_valid_sents = len(valid_trg_corpus)
num_test_sents = len(test_ref_corpus)
num_train_labels = count_labels(train_trg_corpus)
num_valid_labels = count_labels(valid_trg_corpus)
print("train:", num_train_sents, "sentences,", num_train_labels, "labels")
print("valid:", num_valid_sents, "sentences,", num_valid_labels, "labels")
# Sentence IDs
train_ids = list(range(num_train_sents))
valid_ids = list(range(num_valid_sents))
# Train/valid loop.
for epoch in range(max_epoch):
# Computation graph.
g = Graph()
Graph.set_default(g)
print("epoch %d/%d:" % (epoch + 1, max_epoch))
print(" learning rate scale = %.4e" % optimizer.get_learning_rate_scaling())
# Shuffles train sentence IDs.
random.shuffle(train_ids)
# Training.
train_loss = 0.
for ofs in range(0, num_train_sents, batch_size):
print("%d" % ofs, end="\r", flush=True)
batch_ids = train_ids[ofs:min(ofs+args.minibatch, num_train_sents)]
src_batch = make_batch(train_src_corpus, batch_ids, src_vocab)
trg_batch = make_batch(train_trg_corpus, batch_ids, trg_vocab)
g.clear()
encdec.encode(src_batch, True)
loss = encdec.loss(trg_batch, True)
train_loss += loss.to_float() * len(batch_ids)
optimizer.reset_gradients()
loss.backward()
optimizer.update()
train_ppl = math.exp(train_loss / num_train_labels)
print(" train PPL = %.4f" % train_ppl)
# Validation.
valid_loss = 0.
for ofs in range(0, num_valid_sents, batch_size):
print("%d" % ofs, end="\r", flush=True)
batch_ids = valid_ids[ofs:min(ofs+batch_size, num_valid_sents)]
src_batch = make_batch(valid_src_corpus, batch_ids, src_vocab)
trg_batch = make_batch(valid_trg_corpus, batch_ids, trg_vocab)
g.clear()
encdec.encode(src_batch, False)
loss = encdec.loss(trg_batch, False)
valid_loss += loss.to_float() * len(batch_ids)
valid_ppl = math.exp(valid_loss/num_valid_labels)
print(" valid PPL = %.4f" % valid_ppl)
# Calculates test BLEU.
stats = defaultdict(int)
for ofs in range(0, num_test_sents, batch_size):
print("%d" % ofs, end="\r", flush=True)
src_batch = test_src_corpus[ofs:min(ofs + batch_size, num_test_sents)]
ref_batch = test_ref_corpus[ofs:min(ofs + batch_size, num_test_sents)]
hyp_ids = test_batch(encdec, src_vocab, trg_vocab,
src_batch, args.generation_limit)
for hyp_line, ref_line in zip(hyp_ids, ref_batch):
for k, v in get_bleu_stats(ref_line[1:-1], hyp_line).items():
stats[k] += v
bleu = calculate_bleu(stats)
print(" test BLEU = %.2f" % (100 * bleu))
# Saves best model/optimizer.
if valid_ppl < best_valid_ppl:
best_valid_ppl = valid_ppl
print(" saving model/optimizer ... ", end="", flush=True)
encdec.save(prefix+".model")
optimizer.save(prefix+".optimizer")
save_ppl(prefix+".valid_ppl", best_valid_ppl)
print("done.")
else:
# Learning rate decay by 1/sqrt(2)
new_scale = .7071 * optimizer.get_learning_rate_scaling()
optimizer.set_learning_rate_scaling(new_scale)
def test_batch(encdec, src_vocab, trg_vocab, lines, generation_limit):
g = Graph()
Graph.set_default(g)
src_batch = make_batch(lines, list(range(len(lines))), src_vocab)
encdec.encode(src_batch, False)
trg_ids = [np.array([trg_vocab["<bos>"]] * len(lines))]
eos_id = trg_vocab["<eos>"]
eos_ids = np.array([eos_id] * len(lines))
while (trg_ids[-1] != eos_ids).any():
if len(trg_ids) > generation_limit + 1:
print("Warning: Sentence generation did not finish in", generation_limit,
"iterations.", file=sys.stderr)
trg_ids.append(eos_ids)
break
y = encdec.decode_step(trg_ids[-1], False)
trg_ids.append(np.array(y.argmax(0)).T)
return [hyp[:np.where(hyp == eos_id)[0][0]] for hyp in np.array(trg_ids[1:]).T]
# Generates translation by consuming stdin.
def test(encdec, args):
# Loads vocab.
src_vocab = make_vocab(SRC_TRAIN_FILE, args.src_vocab)
trg_vocab = make_vocab(TRG_TRAIN_FILE, args.trg_vocab)
inv_trg_vocab = make_inv_vocab(trg_vocab)
for line in sys.stdin:
sent = [line_to_sent(line.strip(), src_vocab)]
trg_ids = test_batch(encdec, src_vocab, trg_vocab,
sent, args.generation_limit)[0]
# Prints the result.
print(" ".join(inv_trg_vocab[wid] for wid in trg_ids))
def get_arguments():
src_vocab = 4000
trg_vocab = 5000
embed_size = 512
hidden_size = 512
epoch = 30
minibatch_size = 64
generation_limit = 32
dropout = 0.5
parser = ArgumentParser()
parser.add_argument("mode", help="'train', 'resume', or 'test'")
parser.add_argument("model", help='model file prefix')
parser.add_argument("--gpu", default=-1, metavar='INT', type=int,
help='GPU device ID to be used (default: %(default)d [use CPU])')
parser.add_argument("--src-vocab", default=src_vocab, metavar='INT', type=int,
help="source vocabulary size (default: %(default)d)")
parser.add_argument("--trg-vocab", default=trg_vocab, metavar='INT', type=int,
help="target vocabulary size (default: %(default)d)")
parser.add_argument("--embed", default=embed_size, metavar='INT', type=int,
help="embedding layer size (default: %(default)d)")
parser.add_argument("--hidden", default=hidden_size, metavar='INT', type=int,
help="hidden layer size (default: %(default)d)")
parser.add_argument("--epoch", default=epoch, metavar='INT', type=int,
help="number of training epoch (default: %(default)d)")
parser.add_argument("--minibatch", default=minibatch_size, metavar="INT", type=int,
help="minibatch size (default: %(default)d)")
parser.add_argument("--generation-limit", default=generation_limit, metavar="INT", type=int,
help="maximum number of words to be generated for test input (default: %(default)d)")
parser.add_argument("--dropout", default=dropout_rate, metavar="FLOAT", type=flaot, help="dropout rate")
args = parser.parse_args()
try:
if args.mode not in ("train", "resume", "test"):
raise ValueError("you must set mode = 'train', 'resume', or 'test'")
if args.gpu < 0:
raise ValueError("you must set --gpu >= 0")
if args.src_vocab < 1:
raise ValueError("you must set --src-vocab >= 1")
if args.trg_vocab < 1:
raise ValueError("you must set --trg-vocab >= 1")
if args.embed < 1:
raise ValueError("you must set --embed >= 1")
if args.hidden < 1:
raise ValueError("you must set --hidden >= 1")
if args.epoch < 1:
raise ValueError("you must set --epoch >= 1")
if args.minibatch < 1:
raise ValueError("you must set --minibatch >= 1")
if args.generation_limit < 1:
raise ValueError("you must set --generation-limit >= 1")
if args.dropout < 0 or args.dropout > 1:
raise ValueError("you must set --dropout in [0, 1]")
except Exception as ex:
parser.print_usage(file=sys.stderr)
print(ex, file=sys.stderr)
sys.exit()
for (key, val) in vars(args).items():
print("%s: %s" % (key, val))
return args
def main():
args = get_arguments()
print("initializing device ... ", end="", file=sys.stderr, flush=True)
dev = D.Naive() if args.gpu < 0 else D.CUDA(args.gpu)
Device.set_default(dev)
print("done.", file=sys.stderr)
mode = args.mode
prefix = args.model
if mode == "train":
encdec = EncoderDecoder(args.dropout)
encdec.init(args.src_vocab, args.trg_vocab, args.embed, args.hidden)
optimizer = O.Adam()
optimizer.set_weight_decay(1e-6)
optimizer.set_gradient_clipping(5)
train(encdec, optimizer, args, 1e10)
elif mode == "resume":
print("loading model/optimizer ... ", end="", file=sys.stderr, flush=True)
encdec = EncoderDecoder(args.dropout)
encdec.load(prefix+".model")
optimizer = O.Adam()
optimizer.load(prefix + ".optimizer")
valid_ppl = load_ppl(prefix + ".valid_ppl")
print("done.", file=sys.stderr)
train(encdec, optimizer, args, valid_ppl)
else:
print("loading model ... ", end="", file=sys.stderr, flush=True)
encdec = EncoderDecoder(args.dropout)
encdec.load(prefix+".model")
print("done.", file=sys.stderr)
test(encdec, args)
if __name__ == "__main__":
main()