-
Notifications
You must be signed in to change notification settings - Fork 2
/
Copy pathtrain_sdtm.py
169 lines (147 loc) · 8.28 KB
/
train_sdtm.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
#!/usr/bin/python
#coding:utf-8
from __future__ import division
from attention_model import Topical_attention_model
import tensorflow as tf
import time
import numpy as np
import random
import os
from utilities import *
from sklearn.utils import shuffle
import sys
if sys.argv[1] != 'yelp' and sys.argv[1]!= 'yelp_helpful' and sys.argv[1] != '20news' and sys.argv[1] != 'stackexchange':
print('dataset name wrong')
sys.exit()
dataset_name = sys.argv[1]
train_x_rnn, train_y, train_x_bow, num_train_docs, valid_x_rnn, valid_y, valid_x_bow, num_valid_docs, test_x_rnn, test_y, test_x_bow, num_dev_docs, new_vocab, wv_matrix, _ = load_data(dataset_name)
num_classes = test_y.shape[1]
assert(test_y.shape[1]==train_y.shape[1])
# hyperparameters
tf.flags.DEFINE_integer("vocab_size", 2000, "vocabulary size for nueral topic model")
tf.flags.DEFINE_integer("num_classes", num_classes, "number of classes")
tf.flags.DEFINE_integer("embedding_size", EMBEDDING_SIZE, "Dimensionality of word embedding (default: 100)")
tf.flags.DEFINE_integer("RNN_hidden_size", 64, "Dimensionality of GRU hidden layer (default: 50)")
tf.flags.DEFINE_integer("topic_hidden_size", 64, "Dimensionality of GRU hidden layer (default: 64)")
tf.flags.DEFINE_integer("batch_size", 64, "Batch Size (default: 64)")
tf.flags.DEFINE_integer("num_epochs", 200, "Number of training epochs (default: 50)")
tf.flags.DEFINE_float("learning_rate", 1e-3, "learning rate")
tf.flags.DEFINE_integer("word_num", MAX_NUM_WORD, "the max number of words in a document")
tf.flags.DEFINE_integer("num_topics", 50, "number of topics for nueral topic model")
tf.flags.DEFINE_integer("pretrain_epoch", 0, "for loading pretrain weights and continue training")
tf.flags.DEFINE_string("ckpt_name", '', "checkpoint name")
tf.flags.DEFINE_string("timestamp", '', "only useful for continue training multiple runs")
tf.flags.DEFINE_boolean("train_embed", True, "whether make word embeddings trainable or not. Default not trainable.")
tf.flags.DEFINE_float("threshold", 0.1, "threshold value")
FLAGS = tf.flags.FLAGS
def train_step(sess, model, x_rnn_batch, x_bow_batch, y_batch, mode="train_all"):
feed_dict = {
model.input_x: x_rnn_batch,
model.vtm.x: x_bow_batch,
model.input_y: y_batch,
model.is_training: True,
model.vtm.is_training: True
}
if mode=="train_clf":
train_op = model.train_clf_op
elif mode=="train_vtm":
train_op = model.train_vtm_op
else:
train_op = model.train_op
if FLAGS.threshold > 0:
_, step, summaries, cost, acc, likelihood, kl, perp, thr = sess.run(
[train_op, model.global_step, model.train_summary_op, model.clf_loss, model.acc, model.generative_loss, model.inference_loss, model.vtm.perp, tf.reduce_mean(model.threshold)], feed_dict)
print("step {}, cross-entro {:.2f}, acc {:.4f}; likelihood loss {:.2f}, kl_div loss {:.2f}, perplexity {:.2f}, threshold {:.4f}".format(step, cost, acc, likelihood, kl, perp, thr))
else:
_, step, summaries, cost, acc, likelihood, kl, perp = sess.run(
[train_op, model.global_step, model.train_summary_op, model.clf_loss, model.acc, model.generative_loss, model.inference_loss, model.vtm.perp], feed_dict)
print("step {}, cross-entro {:.2f}, acc {:.4f}; likelihood loss {:.2f}, kl_div loss {:.2f}, perplexity {:.2f}".format(step, cost, acc, likelihood, kl, perp))
model.train_summary_writer.add_summary(summaries, step)
return step
def dev_step(sess, model, x_rnn, x_bow, y, acc_record, perp_record, epoch, mode="batch"):
if mode == "whole":
feed_dict = {
model.input_x: x_rnn,
model.vtm.x: x_bow,
model.input_y: y,
model.is_training:False,
model.vtm.is_training: False
}
step, summaries, perp, acc = sess.run([model.global_step, model.dev_summary_op, model.vtm.perp, model.acc], feed_dict)
elif mode=="batch":
perp_list= []
acc_list = []
num_batch = len(y)//FLAGS.batch_size
for i in range(num_batch):
feed_dict = {
model.input_x: x_rnn[i*FLAGS.batch_size: (i+1)*FLAGS.batch_size],
model.vtm.x: x_bow[i*FLAGS.batch_size: (i+1)*FLAGS.batch_size],
model.input_y: y[i*FLAGS.batch_size: (i+1)*FLAGS.batch_size],
model.is_training:False,
model.vtm.is_training: False
}
perp, acc = sess.run([model.vtm.perp, model.acc], feed_dict)
perp_list.append(perp)
acc_list.append(acc)
perp = np.nanmean(perp_list)
acc = np.nanmean(acc_list)
summaries = sess.run(model.dev_summary_op, feed_dict={model.dev_perp: perp, model.dev_acc: acc})
print("+++++++++++++dev+++++++++++++: epoch {}, acc {:.4f}; perplexity {:.2f} ".format(epoch, acc, perp))
model.dev_summary_writer.add_summary(summaries, epoch)
if acc > acc_record:
acc_record = acc
print("new best acc: ", acc_record)
model.best_acc_saver.save(sess, model.checkpoint_dir + '/best-acc-model-acc={:.4f}-perp={:.2f}-epoch{}.ckpt'.format(acc_record, perp, epoch))
else:
print("the best dev acc: ", acc_record)
if perp < perp_record:
perp_record = perp
print("new best perplexity: ", perp_record)
model.best_perp_saver.save(sess, model.checkpoint_dir + '/best-perp-model-acc={:.4f}-perp={:.2f}-epoch{}.ckpt'.format(acc, perp_record, epoch))
else:
print("the best dev perplexity: ", perp_record)
model.current_saver.save(sess, model.checkpoint_dir + '/cur-model-acc={:.4f}-prep={:.2f}-epoch{}.ckpt'.format(acc, perp, epoch))
#if epoch % 100 == 0:
# model.recent_saver.save(sess, model.checkpoint_dir + '/model-acc={:.4f}-prep={:.2f}-epoch{}.ckpt'.format(acc, perp, epoch))
return acc_record, perp_record
def train_topical(acc_record, perp_record):
config = tf.ConfigProto()
config.gpu_options.allow_growth=True # allocate only as much GPU memory based on runtime allocations
with tf.Graph().as_default(), tf.Session(config=config) as sess:
model = Topical_attention_model(
reduced_vocab_size=FLAGS.vocab_size,
num_topic=FLAGS.num_topics,
num_classes=FLAGS.num_classes,
pretrained_embed=wv_matrix,
embedding_size=FLAGS.embedding_size,
RNN_hidden_size=FLAGS.RNN_hidden_size,
topic_hidden_size=FLAGS.topic_hidden_size,
dropout_keep_proba=0.8,
max_word_num=FLAGS.word_num,
threshold=FLAGS.threshold,
train_embed = FLAGS.train_embed
)
if FLAGS.train_embed:
out_dir = 'sDTM_new/'+dataset_name+'/'
else:
out_dir = 'sDAM_no_embed'
out_dir += "/topics={}_threshold={}_batch={}_".format(FLAGS.num_topics, FLAGS.threshold, FLAGS.batch_size)
model.train_settings(out_dir, FLAGS.learning_rate, sess, FLAGS.pretrain_epoch, FLAGS.ckpt_name)
for epoch in range(FLAGS.pretrain_epoch+1, FLAGS.num_epochs+1):
X_rnn_train, X_bow_train, Y_train = shuffle(train_x_rnn, train_x_bow, train_y)
#X_rnn_test, X_bow_test, Y_test = shuffle(test_x_rnn, test_x_bow, test_y)
X_rnn_valid, X_bow_valid, Y_valid = shuffle(valid_x_rnn, valid_x_bow, valid_y)
print('current epoch %s' % (epoch))
mode = "train_all"
print("train mode: ", mode)
for i in range(train_y.shape[0]//FLAGS.batch_size):
x_rnn_batch = X_rnn_train[i*FLAGS.batch_size: (i+1)*FLAGS.batch_size]
x_bow_batch = X_bow_train[i*FLAGS.batch_size: (i+1)*FLAGS.batch_size]
y_batch = Y_train[i*FLAGS.batch_size: (i+1)*FLAGS.batch_size]
step = train_step(sess, model, x_rnn_batch, x_bow_batch, y_batch, mode)
#acc_record, perp_record = dev_step(sess, model, X_rnn_test, X_bow_test, Y_test, acc_record, perp_record, epoch, "batch")
acc_record, perp_record = dev_step(sess, model, X_rnn_valid, X_bow_valid, Y_valid, acc_record, perp_record, epoch, "batch")
if __name__ == '__main__':
acc_record = 0.0
perp_record = np.Infinity
train_topical(acc_record, perp_record)