-
Notifications
You must be signed in to change notification settings - Fork 0
/
mt_s2s_encdec.py
316 lines (265 loc) · 10.4 KB
/
mt_s2s_encdec.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
310
311
312
313
314
315
import sys
import numpy
from argparse import ArgumentParser
from chainer import Chain, Variable, cuda, functions, links, optimizer, optimizers, serializers
import util.generators as gens
from util.functions import trace, fill_batch
from util.vocabulary import Vocabulary
def parse_args():
def_gpu_device = 0
def_vocab = 1000
def_embed = 100
def_hidden = 200
def_epoch = 10
def_minibatch = 64
def_generation_limit = 128
p = ArgumentParser(
description='Encoder-decoder neural machine trainslation',
usage=
'\n %(prog)s train [options] source model'
'\n %(prog)s test source model'
'\n %(prog)s -h',
)
p.add_argument('mode', help='\'train\' or \'test\'')
p.add_argument('source', help='[in] source corpus')
p.add_argument('target', help='[in/out] target corpus')
p.add_argument('model', help='[in/out] model file')
p.add_argument('--use-gpu', action='store_true', default=False,
help='use GPU calculation')
p.add_argument('--gpu-device', default=def_gpu_device, metavar='INT', type=int,
help='GPU device ID to be used (default: %(default)d)')
p.add_argument('--vocab', default=def_vocab, metavar='INT', type=int,
help='vocabulary size (default: %(default)d)')
p.add_argument('--embed', default=def_embed, metavar='INT', type=int,
help='embedding layer size (default: %(default)d)')
p.add_argument('--hidden', default=def_hidden, metavar='INT', type=int,
help='hidden layer size (default: %(default)d)')
p.add_argument('--epoch', default=def_epoch, metavar='INT', type=int,
help='number of training epoch (default: %(default)d)')
p.add_argument('--minibatch', default=def_minibatch, metavar='INT', type=int,
help='minibatch size (default: %(default)d)')
p.add_argument('--generation-limit', default=def_generation_limit, metavar='INT', type=int,
help='maximum number of words to be generated for test input (default: %(default)d)')
args = p.parse_args()
# check args
try:
if args.mode not in ['train', 'test', 'server']: raise ValueError('you must set mode = \'train\' or \'test\'')
if args.vocab < 1: raise ValueError('you must set --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')
except Exception as ex:
p.print_usage(file=sys.stderr)
print(ex, file=sys.stderr)
sys.exit()
return args
class XP:
__lib = None
@staticmethod
def set_library(args):
if args.use_gpu:
XP.__lib = cuda.cupy
cuda.get_device(args.gpu_device).use()
else:
XP.__lib = numpy
@staticmethod
def __zeros(shape, dtype):
return Variable(XP.__lib.zeros(shape, dtype=dtype))
@staticmethod
def fzeros(shape):
return XP.__zeros(shape, XP.__lib.float32)
@staticmethod
def __array(array, dtype):
return Variable(XP.__lib.array(array, dtype=dtype))
@staticmethod
def iarray(array):
return XP.__array(array, XP.__lib.int32)
@staticmethod
def farray(array):
return XP.__array(array, XP.__lib.float32)
class Encoder(Chain):
def __init__(self, vocab_size, embed_size, hidden_size):
super(Encoder, self).__init__(
xe = links.EmbedID(vocab_size, embed_size),
eh = links.Linear(embed_size, 4 * hidden_size),
hh = links.Linear(hidden_size, 4 * hidden_size),
)
def __call__(self, x, c, h):
e = functions.tanh(self.xe(x))
return functions.lstm(c, self.eh(e) + self.hh(h))
class Decoder(Chain):
def __init__(self, vocab_size, embed_size, hidden_size):
super(Decoder, self).__init__(
ye = links.EmbedID(vocab_size, embed_size),
eh = links.Linear(embed_size, 4 * hidden_size),
hh = links.Linear(hidden_size, 4 * hidden_size),
hf = links.Linear(hidden_size, embed_size),
fy = links.Linear(embed_size, vocab_size),
)
def __call__(self, y, c, h):
e = functions.tanh(self.ye(y))
c, h = functions.lstm(c, self.eh(e) + self.hh(h))
f = functions.tanh(self.hf(h))
return self.fy(f), c, h
class EncoderDecoder(Chain):
def __init__(self, vocab_size, embed_size, hidden_size):
super(EncoderDecoder, self).__init__(
enc = Encoder(vocab_size, embed_size, hidden_size),
dec = Decoder(vocab_size, embed_size, hidden_size),
)
self.vocab_size = vocab_size
self.embed_size = embed_size
self.hidden_size = hidden_size
def reset(self, batch_size):
self.zerograds()
self.c = XP.fzeros((batch_size, self.hidden_size))
self.h = XP.fzeros((batch_size, self.hidden_size))
def encode(self, x):
self.c, self.h = self.enc(x, self.c, self.h)
def decode(self, y):
y, self.c, self.h = self.dec(y, self.c, self.h)
return y
def save_spec(self, filename):
with open(filename, 'w') as fp:
print(self.vocab_size, file=fp)
print(self.embed_size, file=fp)
print(self.hidden_size, file=fp)
@staticmethod
def load_spec(filename):
with open(filename) as fp:
vocab_size = int(next(fp))
embed_size = int(next(fp))
hidden_size = int(next(fp))
return EncoderDecoder(vocab_size, embed_size, hidden_size)
def forward(src_batch, trg_batch, src_vocab, trg_vocab, encdec, is_training, generation_limit):
batch_size = len(src_batch)
src_len = len(src_batch[0])
trg_len = len(trg_batch[0]) if trg_batch else 0
src_stoi = src_vocab.stoi
trg_stoi = trg_vocab.stoi
trg_itos = trg_vocab.itos
encdec.reset(batch_size)
x = XP.iarray([src_stoi('</s>') for _ in range(batch_size)])
encdec.encode(x)
for l in reversed(range(src_len)):
x = XP.iarray([src_stoi(src_batch[k][l]) for k in range(batch_size)])
encdec.encode(x)
t = XP.iarray([trg_stoi('<s>') for _ in range(batch_size)])
hyp_batch = [[] for _ in range(batch_size)]
if is_training:
loss = XP.fzeros(())
for l in range(trg_len):
y = encdec.decode(t)
t = XP.iarray([trg_stoi(trg_batch[k][l]) for k in range(batch_size)])
loss += functions.softmax_cross_entropy(y, t)
output = cuda.to_cpu(y.data.argmax(1))
for k in range(batch_size):
hyp_batch[k].append(trg_itos(output[k]))
return hyp_batch, loss
else:
while len(hyp_batch[0]) < generation_limit:
y = encdec.decode(t)
output = cuda.to_cpu(y.data.argmax(1))
t = XP.iarray(output)
for k in range(batch_size):
hyp_batch[k].append(trg_itos(output[k]))
if all(hyp_batch[k][-1] == '</s>' for k in range(batch_size)):
break
return hyp_batch
def train(args):
trace('making vocabularies ...')
src_vocab = Vocabulary.new(gens.word_list(args.source), args.vocab)
trg_vocab = Vocabulary.new(gens.word_list(args.target), args.vocab)
trace('making model ...')
encdec = EncoderDecoder(args.vocab, args.embed, args.hidden)
if args.use_gpu:
encdec.to_gpu()
for epoch in range(args.epoch):
trace('epoch %d/%d: ' % (epoch + 1, args.epoch))
trained = 0
gen1 = gens.word_list(args.source)
gen2 = gens.word_list(args.target)
gen3 = gens.batch(gens.sorted_parallel(gen1, gen2, 100 * args.minibatch), args.minibatch)
opt = optimizers.AdaGrad(lr = 0.01)
opt.setup(encdec)
opt.add_hook(optimizer.GradientClipping(5))
for src_batch, trg_batch in gen3:
src_batch = fill_batch(src_batch)
trg_batch = fill_batch(trg_batch)
K = len(src_batch)
hyp_batch, loss = forward(src_batch, trg_batch, src_vocab, trg_vocab, encdec, True, 0)
loss.backward()
opt.update()
for k in range(K):
trace('epoch %3d/%3d, sample %8d' % (epoch + 1, args.epoch, trained + k + 1))
trace(' src = ' + ' '.join([x if x != '</s>' else '*' for x in src_batch[k]]))
trace(' trg = ' + ' '.join([x if x != '</s>' else '*' for x in trg_batch[k]]))
trace(' hyp = ' + ' '.join([x if x != '</s>' else '*' for x in hyp_batch[k]]))
trained += K
trace('saving model ...')
prefix = args.model + '.%03.d' % (epoch + 1)
src_vocab.save(prefix + '.srcvocab')
trg_vocab.save(prefix + '.trgvocab')
encdec.save_spec(prefix + '.spec')
serializers.save_hdf5(prefix + '.weights', encdec)
trace('finished.')
def test(args):
trace('loading model ...')
src_vocab = Vocabulary.load(args.model + '.srcvocab')
trg_vocab = Vocabulary.load(args.model + '.trgvocab')
encdec = EncoderDecoder.load_spec(args.model + '.spec')
if args.use_gpu:
encdec.to_gpu()
serializers.load_hdf5(args.model + '.weights', encdec)
trace('generating translation ...')
generated = 0
with open(args.target, 'w') as fp:
for src_batch in gens.batch(gens.word_list(args.source), args.minibatch):
src_batch = fill_batch(src_batch)
K = len(src_batch)
trace('sample %8d - %8d ...' % (generated + 1, generated + K))
hyp_batch = forward(src_batch, None, src_vocab, trg_vocab, encdec, False, args.generation_limit)
for hyp in hyp_batch:
hyp.append('</s>')
hyp = hyp[:hyp.index('</s>')]
print(' '.join(hyp), file=fp)
generated += K
trace('finished.')
import msgpackrpc
class Serve(object):
def __init__(self, args):
trace('loading model ...')
self.args = args
self.src_vocab = Vocabulary.load(args.model + '.srcvocab')
self.trg_vocab = Vocabulary.load(args.model + '.trgvocab')
self.encdec = EncoderDecoder.load_spec(args.model + '.spec')
if args.use_gpu:
self.encdec.to_gpu()
serializers.load_hdf5(args.model + '.weights', self.encdec)
trace('generating translation ...')
def dqn(self, s):
words = s.strip().split()
v = ([b for b in gens.batch([words], 1)])
src_batch = fill_batch(v[0])
hyp_batch = forward(src_batch, None, self.src_vocab, self.trg_vocab, self.encdec, False, self.args.generation_limit)
result = ""
for hyp in hyp_batch:
hyp.append('</s>')
hyp = hyp[:hyp.index('</s>')]
print(' '.join(hyp))
result = ' '.join(hyp)
return result
def serve(args):
server = msgpackrpc.Server(Serve(args), pack_encoding='utf-8', unpack_encoding='utf-8')
server.listen(msgpackrpc.Address('localhost', 5000))
server.start()
def main():
args = parse_args()
XP.set_library(args)
if args.mode == 'train': train(args)
elif args.mode == 'test': test(args)
elif args.mode == 'server': serve(args)
if __name__ == '__main__':
main()