-
Notifications
You must be signed in to change notification settings - Fork 0
/
run_models.py
211 lines (159 loc) · 7.53 KB
/
run_models.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
import os
import time
import numpy as np
import tensorflow as tf
from beeprint import pp
from data_utils.corpus import DailyDialogCorpus
from data_utils.data_utils import DailyDataLoader
from data_utils.corpus import SWDADialogCorpus
from data_utils.data_utils import SWDADataLoader
tf.app.flags.DEFINE_string("word2vec_path", None, "The path to word2vec. Can be None.")
tf.app.flags.DEFINE_string("dataset", None, "dailydialog or switchboard")
tf.app.flags.DEFINE_string("model", None, "seq2seq or hred or vhred")
tf.app.flags.DEFINE_string("work_dir", "save", "Experiment results directory.")
tf.app.flags.DEFINE_bool("equal_batch", True, "Make each batch has similar length.")
tf.app.flags.DEFINE_bool("resume", False, "Resume from previous")
tf.app.flags.DEFINE_bool("forward_only", False, "Only do decoding")
tf.app.flags.DEFINE_bool("save_model", True, "Create checkpoints")
tf.app.flags.DEFINE_string("test_path", "", "the dir to load checkpoint for forward only")
FLAGS = tf.app.flags.FLAGS
def main():
if FLAGS.model == "seq2seq":
from config_utils import Seq2SeqConfig as Config
from models.vanilla import Seq2Seq as Model
elif FLAGS.model == "hred":
from config_utils import HREDConfig as Config
from models.hred import HRED as Model
elif FLAGS.model == "vhred":
from config_utils import VHREDConfig as Config
from models.vhred import VHRED as Model
elif FLAGS.model == "kgcvae":
from config_utils import KGCVAEConfig as Config
from models.kgcvae import KGCVAE as Model
# config for training
config = Config()
# config for validation
valid_config = Config()
valid_config.keep_prob = 1.0
valid_config.dec_keep_prob = 1.0
valid_config.batch_size = 60
# configuration for testing
test_config = Config()
test_config.keep_prob = 1.0
test_config.dec_keep_prob = 1.0
test_config.batch_size = 1
pp(config)
if FLAGS.dataset == "dailydialog":
api = DailyDialogCorpus(
"data/dailydialog/dailydialog_split.pkl",
word2vec=FLAGS.word2vec_path, word2vec_dim=config.embed_size)
elif FLAGS.dataset == "switchboard":
api = SWDADialogCorpus(
"data/switchboard/full_swda_clean_42da_sentiment_dialog_corpus.p",
word2vec=FLAGS.word2vec_path, word2vec_dim=config.embed_size)
else:
raise ValueError("dataset should be specified among [dailydialog or switchboard]")
dial_corpus = api.get_dialog_corpus()
meta_corpus = api.get_meta_corpus()
train_meta, valid_meta, test_meta = meta_corpus.get("train"), meta_corpus.get("valid"), meta_corpus.get("test")
train_dial, valid_dial, test_dial = dial_corpus.get("train"), dial_corpus.get("valid"), dial_corpus.get("test")
if FLAGS.dataset == "dailydialog":
train_feed = DailyDataLoader("Train", train_dial, train_meta, config)
valid_feed = DailyDataLoader("Valid", valid_dial, valid_meta, config)
test_feed = DailyDataLoader("Test", test_dial, test_meta, config)
elif FLAGS.dataset == "switchboard":
train_feed = SWDADataLoader("Train", train_dial, train_meta, config)
valid_feed = SWDADataLoader("Valid", valid_dial, valid_meta, config)
test_feed = SWDADataLoader("Test", test_dial, test_meta, config)
else:
raise ValueError("dataset should be specified among [dailydialog or switchboard]")
if FLAGS.forward_only or FLAGS.resume:
log_dir = os.path.join(FLAGS.work_dir, FLAGS.test_path)
else:
log_dir = os.path.join(FLAGS.work_dir, "run_"+FLAGS.model+"_"+FLAGS.dataset+"_"+str(int(time.time())))
with tf.Session() as sess:
initializer = tf.random_uniform_initializer(-1.0 * config.init_w, config.init_w)
scope = "model"
with tf.variable_scope(scope, reuse=None, initializer=initializer):
model = Model(sess, config, api, log_dir=None if FLAGS.forward_only else log_dir, forward=False, scope=scope)
with tf.variable_scope(scope, reuse=True, initializer=initializer):
valid_model = Model(sess, valid_config, api, log_dir=None, forward=False, scope=scope)
with tf.variable_scope(scope, reuse=True, initializer=initializer):
test_model = Model(sess, test_config, api, log_dir=None, forward=True, scope=scope)
print("Created computation graphs")
if api.word2vec is not None and not FLAGS.forward_only:
print("Loaded word2vec")
sess.run(model.embedding.assign(np.array(api.word2vec)))
# write config to a file for logging
if not FLAGS.forward_only:
with open(os.path.join(log_dir, "run.log"), "wb") as f:
f.write(pp(config, output=False))
# create a folder by force
ckp_dir = os.path.join(log_dir, "checkpoints")
if not os.path.exists(ckp_dir):
os.mkdir(ckp_dir)
ckpt = tf.train.get_checkpoint_state(ckp_dir)
print("Created models with fresh parameters.")
sess.run(tf.global_variables_initializer())
if ckpt:
print("Reading models parameters from %s" % ckpt.model_checkpoint_path)
model.saver.restore(sess, ckpt.model_checkpoint_path)
if not FLAGS.forward_only:
dm_checkpoint_path = os.path.join(ckp_dir, model.__class__.__name__+ ".ckpt")
global_t = 1
patience = 10 # wait for at least 10 epoch before stop
dev_loss_threshold = np.inf
best_dev_loss = np.inf
for epoch in range(config.max_epoch):
print(">> Epoch %d with lr %f" % (epoch, model.learning_rate.eval()))
# begin training
if train_feed.num_batch is None or train_feed.ptr >= train_feed.num_batch:
train_feed.epoch_init(config.batch_size, config.backward_size,
config.step_size, shuffle=True)
global_t, train_loss = model.train(global_t, sess, train_feed, update_limit=config.update_limit)
test_feed.epoch_init(test_config.batch_size, test_config.backward_size,
test_config.step_size, shuffle=True, intra_shuffle=False)
test_model.test(sess, test_feed, num_batch=5)
# begin validation
valid_feed.epoch_init(valid_config.batch_size, valid_config.backward_size,
valid_config.step_size, shuffle=False, intra_shuffle=False)
valid_loss = valid_model.valid("ELBO_VALID", sess, valid_feed)
done_epoch = epoch + 1
# only save a models if the dev loss is smaller
# Decrease learning rate if no improvement was seen over last 3 times.
if config.op == "sgd" and done_epoch > config.lr_hold:
sess.run(model.learning_rate_decay_op)
if valid_loss < best_dev_loss:
if valid_loss <= dev_loss_threshold * config.improve_threshold:
patience = max(patience, done_epoch * config.patient_increase)
dev_loss_threshold = valid_loss
# still save the best train model
if FLAGS.save_model:
print("Save model!!")
model.saver.save(sess, dm_checkpoint_path, global_step=epoch)
best_dev_loss = valid_loss
if config.early_stop and patience <= done_epoch:
print("!!Early stop due to run out of patience!!")
break
print("Best validation loss %f" % best_dev_loss)
print("Done training")
else:
# begin validation
# begin validation
valid_feed.epoch_init(valid_config.batch_size, valid_config.backward_size,
valid_config.step_size, shuffle=False, intra_shuffle=False)
valid_model.valid("ELBO_VALID", sess, valid_feed)
test_feed.epoch_init(valid_config.batch_size, valid_config.backward_size,
valid_config.step_size, shuffle=False, intra_shuffle=False)
valid_model.valid("ELBO_TEST", sess, test_feed)
dest_f = open(os.path.join(log_dir, "test.txt"), "wb")
test_feed.epoch_init(test_config.batch_size, test_config.backward_size,
test_config.step_size, shuffle=False, intra_shuffle=False)
test_model.test(sess, test_feed, num_batch=None, repeat=10, dest=dest_f)
dest_f.close()
if __name__ == "__main__":
if FLAGS.forward_only:
if FLAGS.test_path is None:
print("Set test_path before forward only")
exit(1)
main()