-
Notifications
You must be signed in to change notification settings - Fork 9
/
model_seq2seq.py
491 lines (416 loc) · 23.8 KB
/
model_seq2seq.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
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
import os
import numpy as np
import tensorflow as tf
import random
import argparse
import math
from colors import *
from tqdm import *
from handler import Batch, DatasetBase, DatasetTrain, DatasetEval, DatasetTest
import util
from gensim.models import Word2Vec
FLAGS = None
random.seed(0)
np.random.seed(0)
tf.set_random_seed(0)
#filename = '/tmp'
#total_line_num = 100
#train_line_num = 80
#eval_line_num = 20
#filename = '/xaa'
#total_line_num = 50000
#train_line_num = 45000
#eval_line_num = 5000
#filename = '/clr_conversation.txt'
#total_line_num = 2842478
#train_line_num = 2840000
#eval_line_num = 2478
filename = '/conversation.txt'
total_line_num = 3599478
train_line_num = 3587000
eval_line_num = 12478
emb_size = 300
PKL_EXIST = True
MAX_SENTENCE_LENGTH = 15 # longest
special_tokens = {'<PAD>': 0, '<BOS>': 1, '<EOS>': 2, '<UNK>': 3}
special_tokens_to_word = ['<PAD>', '<BOS>', '<EOS>', '<UNK>']
modes = {'train': 0, 'eval': 1, 'test': 2}
class Seq2Seq:
def __init__(self, voc, idx2word, mode, att, lr=None):
self.num_layers = 2
self.rnn_size = 2048
self.keep_prob = 1.0
self.vocab_num = voc
self.with_attention = att
self.mode = mode
self.lr = lr
self.idx2word = idx2word
self.embedding_size = emb_size
self.hist_summary = []
def _create_rnn_cell(self):
def single_rnn_cell():
cell = tf.contrib.rnn.GRUCell(self.rnn_size)
#cell = tf.contrib.rnn.LSTMCell(self.rnn_size, initializer=tf.orthogonal_initializer())
if self.mode == modes['train']:
cell = tf.contrib.rnn.DropoutWrapper(cell, input_keep_prob=self.keep_prob,
output_keep_prob=self.keep_prob)
return cell
cell = tf.contrib.rnn.MultiRNNCell([single_rnn_cell() for _ in range(self.num_layers)])
return cell
def build_model(self, emb=None):
self.encoder_inputs = tf.placeholder(tf.int32, [None, None], name='encoder_inputs')
self.encoder_inputs_length = tf.placeholder(tf.int32, [None], name='encoder_inputs_length')
with tf.device("/cpu:0"):
embed = tf.constant_initializer(emb, dtype=tf.float32)
embedding = tf.get_variable(
initializer=embed, shape=emb.shape, dtype=tf.float32, trainable=True, name='embedding')
#embedding = tf.get_variable('embedding', [self.vocab_num, self.embedding_size])
self.hist_summary.append(tf.summary.histogram(embedding.name + '/embed', embedding))
self.batch_size = tf.placeholder(tf.int32, [], name='batch_size')
self.sampling_prob = tf.placeholder(tf.float32, [], name='sampling_prob')
if self.mode != modes['test']:
self.decoder_targets = tf.placeholder(tf.int32, [None, None], name='decoder_targets')
self.decoder_targets_length = tf.placeholder(tf.int32, [None], name='decoder_targets_length')
self.max_target_sequence_length = tf.reduce_max(self.decoder_targets_length, name='max_target_len')
self.mask = tf.sequence_mask(self.decoder_targets_length, self.max_target_sequence_length,
dtype=tf.float32, name='masks')
with tf.variable_scope('encoder'):
encoder_cell = self._create_rnn_cell()
with tf.device("/cpu:0"):
encoder_inputs_embedded = tf.nn.embedding_lookup(embedding, self.encoder_inputs)
encoder_outputs, encoder_state = tf.nn.dynamic_rnn(encoder_cell, encoder_inputs_embedded,
sequence_length=self.encoder_inputs_length,
dtype=tf.float32)
with tf.variable_scope('decoder'):
encoder_inputs_length = self.encoder_inputs_length
decoder_cell = self._create_rnn_cell()
batch_size = self.batch_size
if self.with_attention:
print('wrapped with bahdanau attention...')
attention_mechanism = tf.contrib.seq2seq.BahdanauAttention(
num_units=self.rnn_size, memory=encoder_outputs,
memory_sequence_length=encoder_inputs_length)
decoder_cell = tf.contrib.seq2seq.AttentionWrapper(
cell=decoder_cell, attention_mechanism=attention_mechanism,
attention_layer_size=self.rnn_size, name='Attention_Wrapper')
decoder_initial_state = decoder_cell.zero_state(batch_size=batch_size,
dtype=tf.float32).clone(cell_state=encoder_state)
else:
decoder_initial_state = encoder_state
projection_layer = tf.layers.Dense(
self.vocab_num, kernel_initializer=tf.truncated_normal_initializer(mean=0.0, stddev=0.1))
#self.vocab_num, kernel_initializer=tf.contrib.layers.xavier_initializer())
if self.mode == modes['train']:
ending = tf.strided_slice(self.decoder_targets, [0, 0], [self.batch_size, -1], [1, 1])
decoder_input = tf.concat([tf.fill([self.batch_size, 1], special_tokens['<BOS>']), ending], 1)
with tf.device("/cpu:0"):
decoder_inputs_embedded = tf.nn.embedding_lookup(embedding, decoder_input)
training_helper = tf.contrib.seq2seq.ScheduledEmbeddingTrainingHelper(
inputs=decoder_inputs_embedded,
sequence_length=self.decoder_targets_length,
embedding=embedding,
time_major=False,
sampling_probability=self.sampling_prob,
name='training_helper')
training_decoder = tf.contrib.seq2seq.BasicDecoder(cell=decoder_cell, helper=training_helper,
initial_state=decoder_initial_state, output_layer=projection_layer)
decoder_outputs, _, final_seq_len = tf.contrib.seq2seq.dynamic_decode(decoder=training_decoder,
impute_finished=True,
maximum_iterations=self.max_target_sequence_length)
#pad_size = self.max_target_sequence_length - tf.reduce_max(final_seq_len)
#pad_rnn_output = tf.pad(decoder_outputs.rnn_output, [[0, 0], [0, pad_size], [0, 0]])
#self.decoder_logits_train = tf.identity(pad_rnn_output)
self.decoder_logits_train = tf.identity(decoder_outputs.rnn_output)
self.decoder_predict_train = tf.argmax(self.decoder_logits_train, axis=-1, name='decoder_pred_train')
print(self.decoder_predict_train)
self.train_loss = tf.contrib.seq2seq.sequence_loss(logits=self.decoder_logits_train,
targets=self.decoder_targets, weights=self.mask)
self.train_summary = tf.summary.scalar('training loss', self.train_loss)
elif self.mode == modes['eval']:
start_tokens = tf.ones([self.batch_size, ], tf.int32) # * special_tokens['<BOS>']
end_token = special_tokens['<EOS>']
decoding_helper = tf.contrib.seq2seq.GreedyEmbeddingHelper(embedding=embedding,
start_tokens=start_tokens, end_token=end_token)
inference_decoder = tf.contrib.seq2seq.BasicDecoder(cell=decoder_cell, helper=decoding_helper,
initial_state=decoder_initial_state,
output_layer=projection_layer)
decoder_outputs, _, final_seq_len = tf.contrib.seq2seq.dynamic_decode(decoder=inference_decoder,
maximum_iterations=self.max_target_sequence_length)
# pad to same shape in order to calculate loss
pad_size = self.max_target_sequence_length - tf.reduce_max(final_seq_len)
pad_rnn_output = tf.pad(decoder_outputs.rnn_output, [[0, 0], [0, pad_size], [0, 0]])
# self.decoder_logits_eval = tf.identity(decoder_outputs.rnn_output)
self.decoder_logits_eval = tf.identity(pad_rnn_output)
self.decoder_predict_eval = tf.argmax(self.decoder_logits_eval, axis=-1, name='decoder_pred_eval')
print(self.decoder_predict_eval)
self.eval_loss = tf.contrib.seq2seq.sequence_loss(logits=pad_rnn_output,
targets=self.decoder_targets, weights=self.mask)
self.eval_summary = tf.summary.scalar('validation loss', self.eval_loss)
elif self.mode == modes['test']:
start_tokens = tf.ones([self.batch_size, ], tf.int32) # * special_tokens['<BOS>']
end_token = special_tokens['<EOS>']
decoding_helper = tf.contrib.seq2seq.GreedyEmbeddingHelper(embedding=embedding,
start_tokens=start_tokens, end_token=end_token)
inference_decoder = tf.contrib.seq2seq.BasicDecoder(cell=decoder_cell, helper=decoding_helper,
initial_state=decoder_initial_state,
output_layer=projection_layer)
decoder_outputs, _, final_seq_len = tf.contrib.seq2seq.dynamic_decode(decoder=inference_decoder,
maximum_iterations=MAX_SENTENCE_LENGTH)
self.decoder_logits_eval = tf.identity(decoder_outputs.rnn_output)
self.decoder_predict_eval = tf.argmax(self.decoder_logits_eval, axis=-1, name='decoder_pred_eval')
print(self.decoder_predict_eval)
def build_optimizer(self, lr):
optimizer = tf.train.GradientDescentOptimizer(lr)
#optimizer = tf.train.AdamOptimizer(lr)#.minimize(self.train_loss)
#optimizer = tf.train.AdamOptimizer(lr, epsilon=0.1)#.minimize(self.train_loss)
trainable_params = tf.trainable_variables()
gradients = tf.gradients(self.train_loss, trainable_params)
gradients, _ = tf.clip_by_global_norm(gradients, 5.0)
gradients = list(zip(gradients, trainable_params))
self.train_op = optimizer.apply_gradients(gradients)
for grad, var in gradients:
self.hist_summary.append(tf.summary.histogram(var.name + '/gradient', grad))
self.hist_summary.append(self.train_summary)
self.train_summary = tf.summary.merge(self.hist_summary)
def train(self, sess, batch, print_pred, summary_writer, add_global, prob):
feed_dict = {self.encoder_inputs: batch.encoder_inputs,
self.encoder_inputs_length: batch.encoder_inputs_length,
self.decoder_targets: batch.decoder_targets,
self.decoder_targets_length: batch.decoder_targets_length,
self.batch_size: len(batch.encoder_inputs),
self.sampling_prob: prob}
if print_pred:
run_options = tf.RunOptions(trace_level=tf.RunOptions.FULL_TRACE)
_, loss, pred, summary, current_step, print_lr = sess.run([self.train_op, self.train_loss,
self.decoder_predict_train, self.train_summary, add_global, self.lr],
feed_dict=feed_dict, options=run_options)
i = np.random.randint(0, len(batch.encoder_inputs))
util.decoder_print(self.idx2word, batch.encoder_inputs[i], batch.encoder_inputs_length[i],
batch.decoder_targets[i], batch.decoder_targets_length[i], pred[i], 'yellow')
summary_writer.add_summary(summary, global_step=current_step)
else:
_, loss, current_step, print_lr = sess.run([self.train_op, self.train_loss,
add_global, self.lr], feed_dict=feed_dict)
return loss, calc_perplexity(loss), current_step, print_lr
def eval(self, sess, batch, summary_writer, current_step):
feed_dict = {self.encoder_inputs: batch.encoder_inputs,
self.encoder_inputs_length: batch.encoder_inputs_length,
self.decoder_targets: batch.decoder_targets,
self.decoder_targets_length: batch.decoder_targets_length,
self.batch_size: len(batch.encoder_inputs)}
loss, pred, summary = sess.run([self.eval_loss,
self.decoder_predict_eval, self.eval_summary], feed_dict=feed_dict)
print_num = 3
print_more = random.sample(range(len(batch.encoder_inputs)), print_num)
#repeat bug: np.random.randint(len(batch.encoder_inputs), size=(print_num))
for i in print_more:
util.decoder_print(self.idx2word, batch.encoder_inputs[i], batch.encoder_inputs_length[i],
batch.decoder_targets[i], batch.decoder_targets_length[i], pred[i], 'green')
summary_writer.add_summary(summary, global_step=current_step)
return loss, calc_perplexity(loss)
def inference(self, sess, batch, txt):
feed_dict = {self.encoder_inputs: batch.encoder_inputs,
self.encoder_inputs_length: batch.encoder_inputs_length,
self.batch_size: len(batch.encoder_inputs)}
pred = sess.run([self.decoder_predict_eval], feed_dict=feed_dict)
pred = pred[0]
for i in range(len(pred)):
sent = util.decoder_inference(self.idx2word, batch.encoder_inputs[i],
batch.encoder_inputs_length[i], pred[i])
txt.write(sent + "\n")
def calc_perplexity(loss):
return math.exp(float(loss)) if loss < 300 else float('inf')
def train():
datasetTrain = DatasetTrain()
print('start build dict...')
train_data, eval_data = datasetTrain.build_dict(FLAGS.data_dir, filename,
FLAGS.min_counts, train_line_num, eval_line_num, emb_size, PKL_EXIST)
print('build dict done!')
datasetTrain.prep(train_data)
datasetEval = DatasetEval()
datasetEval.load_dict()
datasetEval.prep(eval_data)
word2vec_model = datasetTrain.model
embeddings = np.zeros([datasetTrain.vocab_num, emb_size],dtype=np.float32)
for word in word2vec_model.wv.vocab:
index = word2vec_model.wv.vocab[word].index
vec = word2vec_model.wv[word]
embeddings[index] = vec
if word in special_tokens_to_word:
print('special word: ', word, ', index: ', index)
pad = np.random.normal(size=[emb_size])
unk = np.random.normal(size=[emb_size])
embeddings[0] = pad
embeddings[3] = unk
#print(embeddings)
# np.save('embeddings.npy', embeddings)
train_graph = tf.Graph()
eval_graph = tf.Graph()
gpu_config = tf.ConfigProto()
gpu_config.gpu_options.allow_growth = True
print('start building train graph...')
with train_graph.as_default():
global_step = tf.Variable(0, trainable=False)
lr = tf.train.exponential_decay(FLAGS.learning_rate,
global_step=global_step,
decay_steps=3250, decay_rate=0.95, staircase=True)
add_global = global_step.assign_add(1)
model = Seq2Seq(voc=datasetTrain.vocab_num, idx2word=datasetTrain.idx2word,
mode=modes['train'], att=FLAGS.with_attention, lr=lr)
model.build_model(embeddings)
model.build_optimizer(FLAGS.learning_rate)
model.saver = tf.train.Saver(max_to_keep = 3)
init = tf.global_variables_initializer()
train_sess = tf.Session(graph=train_graph, config=gpu_config)
print('start building eval graph...')
with eval_graph.as_default():
model_eval = Seq2Seq(voc=datasetEval.vocab_num, idx2word=datasetEval.idx2word,
mode=modes['eval'], att=FLAGS.with_attention)
model_eval.build_model(embeddings)
model_eval.saver = tf.train.Saver(max_to_keep = 3)
#init_eval = tf.global_variables_initializer()
eval_sess = tf.Session(graph=eval_graph, config=gpu_config)
ckpt = tf.train.get_checkpoint_state(FLAGS.save_dir)
if FLAGS.load_saver and ckpt and tf.train.checkpoint_exists(ckpt.model_checkpoint_path):
print('Reloading model parameters..')
model.saver.restore(train_sess, ckpt.model_checkpoint_path)
#model_eval.saver.restore(eval_sess, ckpt.model_checkpoint_path)
print(ckpt.model_checkpoint_path)
else:
print('Created new model parameters..')
train_sess.run(init)
#eval_sess.run(init_eval)
ckpts_path = FLAGS.save_dir + "chatbot.ckpt"
summary_writer = tf.summary.FileWriter(FLAGS.log_dir + '/train')
summary_writer.add_graph(train_graph)
summary_writer.add_graph(eval_graph)
num_steps = int( len(datasetTrain.data) / FLAGS.batch_size )
pbar = tqdm(range(FLAGS.num_epochs))
pt = 0
total_samp = FLAGS.num_epochs * 3
samp_prob = util.inv_sigmoid(total_samp)
current_step = 0 # after first round, assign global_step to current_step
for epo in pbar:
print(color('start: epoch ' + str(epo), fg='blue', bg='white'))
for i in range(num_steps):
batch = datasetTrain.next_batch(FLAGS.batch_size, shuffle=True)
print_pred = False
if current_step % FLAGS.num_display_steps == 0 and current_step != 0:
print_pred = True
loss, perp, current_step, print_lr = model.train(train_sess, batch, print_pred,
summary_writer, add_global, samp_prob[pt])
if current_step % FLAGS.num_saver_steps == 0 and current_step != 0:
ckpt_path = model.saver.save(train_sess, ckpts_path, global_step=current_step)
print(color("\nSaver saved: " + ckpt_path, fg='white', bg='green', style='bold'))
model_eval.saver.restore(eval_sess, ckpt_path)
print(color("\n[Eval. Prediction] Epoch " + str(epo) + ", step " + str(i) + "/" \
+ str(num_steps) + "......", fg='white', bg='green', style='underline'))
batch_eval = datasetEval.next_batch(FLAGS.batch_size, shuffle=True)
loss_eval, perp_eval = model_eval.eval(eval_sess, batch_eval, summary_writer, current_step)
print(color("Epoch " + str(epo) + ", step " + str(i) + "/" + str(num_steps) + \
", (Evaluation Loss: " + "{:.4f}".format(loss_eval) + \
", Perplexity: " + "{:.4f}".format(perp_eval) + ")", fg='white', bg='green'))
pbar.set_description("S " + str(i) + "/" + \
str(num_steps) + "(" + str(current_step) + ")" + \
", (L: " + "{:.4f}".format(loss) + ", P: " + "{:.1f}".format(perp) + ", Samp: "+ \
"{:.4f}".format(samp_prob[pt]) + ", lr: "+ "{:.6f}".format(print_lr) + ")" )
if i % int(num_steps / 3) == 0 and i != 0:
pt += 1
print(color('sampling pt: ' + str( pt ) + '/' + str(total_samp), fg='white', bg='red'))
print(color('sampling pt: ' + str( pt ) + '/' + str(total_samp), fg='white', bg='red'))
def test():
datasetTest = DatasetTest()
print('start build dict...')
datasetTest.load_dict()
print('build dict done!')
test_data = datasetTest.load_test_line(FLAGS.test_dir, FLAGS.test_file)
datasetTest.prep(test_data)
# word2vec_model = datasetTest.model
embeddings = np.zeros([datasetTest.vocab_num, emb_size],dtype=np.float32)
# for word in word2vec_model.wv.vocab:
# index = word2vec_model.wv.vocab[word].index
# vec = word2vec_model.wv[word]
# embeddings[index] = vec
# if word in special_tokens_to_word:
# print('special word: ', word, ', index: ', index)
# pad = np.random.normal(size=[emb_size])
# unk = np.random.normal(size=[emb_size])
# embeddings[0] = pad
# embeddings[3] = unk
#print(embeddings)
test_graph = tf.Graph()
gpu_config = tf.ConfigProto()
gpu_config.gpu_options.allow_growth = True
print('start building test graph...')
with test_graph.as_default():
model_test = Seq2Seq(voc=datasetTest.vocab_num, idx2word=datasetTest.idx2word,
mode=modes['test'], att=FLAGS.with_attention)
model_test.build_model(embeddings)
model_test.saver = tf.train.Saver(max_to_keep = 3)
test_sess = tf.Session(graph=test_graph, config=gpu_config)
ckpt = tf.train.get_checkpoint_state(FLAGS.save_dir)
if FLAGS.load_saver and ckpt and tf.train.checkpoint_exists(ckpt.model_checkpoint_path):
print('Reloading model parameters..')
model_test.saver.restore(test_sess, ckpt.model_checkpoint_path)
print(ckpt.model_checkpoint_path)
else:
print('ERROR: you should load model for testing!')
exit(0)
ckpts_path = FLAGS.save_dir + "chatbot.ckpt"
# You have to run this write_graph to get your testing model!
# (since that train/test model structure are different)
tf.train.write_graph(test_sess.graph_def, "./load", "test.pb", False) #proto
#exit(0)
num_steps = int( len(datasetTest.test_data) / FLAGS.batch_size )
txt = open(FLAGS.output_filename, 'w')
for i in range(num_steps):
batch = datasetTest.next_batch(FLAGS.batch_size)
model_test.inference(test_sess, batch, txt)
print(str(i) + '/' + str(num_steps) + ' steps...done')
print('\n\nTesting finished.')
print('\nSave file: ' + FLAGS.output_filename)
txt.close()
def main(_):
if FLAGS.test_mode == False:
#print(color('remove directory: ' + FLAGS.log_dir, fg='red'))
#if tf.gfile.Exists(FLAGS.log_dir):
# tf.gfile.DeleteRecursively(FLAGS.log_dir)
#tf.gfile.MakeDirs(FLAGS.log_dir)
print('train mode: start')
train()
else:
if FLAGS.load_saver == True:
print('load saver!!')
else:
print('ERROR: you cannot run test without saver...')
exit(0)
print('test mode: start')
test()
if __name__ == '__main__':
parser = argparse.ArgumentParser()
parser.add_argument('-lr', '--learning_rate', type=float, default=0.005)
parser.add_argument('-mi', '--min_counts', type=int, default=100) #50 -> 15000 words
parser.add_argument('-e', '--num_epochs', type=int, default=30)
parser.add_argument('-b', '--batch_size', type=int, default=250)
parser.add_argument('-t', '--test_mode', type=int, default=0)
parser.add_argument('-d', '--num_display_steps', type=int, default=120)
parser.add_argument('-ns', '--num_saver_steps', type=int, default=200)
parser.add_argument('-s', '--save_dir', type=str, default='save/')
parser.add_argument('-l', '--log_dir', type=str, default='logs/')
parser.add_argument('-o', '--output_filename', type=str, default='output.txt')
parser.add_argument('-lo', '--load_saver', type=int, default=0)
parser.add_argument('-at', '--with_attention', type=int, default=1)
parser.add_argument('--data_dir', type=str,
default=('/home/data/mlds_hw2_2_data')
)
parser.add_argument('--test_dir', type=str,
default=('/home/data/mlds_hw2_2_data')
)
parser.add_argument('--test_file', type=str,
default=('/test_input.txt')
)
FLAGS, unparsed = parser.parse_known_args()
tf.app.run(main=main, argv=[sys.argv[0]] + unparsed)