forked from Websail-NU/seqmodel
-
Notifications
You must be signed in to change notification settings - Fork 1
/
_main.py
147 lines (123 loc) · 6.36 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
import sys
import os
from collections import ChainMap
from functools import partial
import tensorflow as tf
sys.path.insert(0, '../')
import seqmodel as sq # noqa
def _main(opt, model_class, model_opt, data_fn, run_fn, logger, train_opt=None,
decode_opt=None, decode_batch_fn=None, eval_run_fn=None, pg=False):
is_training = opt['command'] == 'train'
is_testing = opt['command'] == 'eval'
is_init_only = opt['command'] == 'init'
is_decoding = opt['command'] == 'decode'
if eval_run_fn is None:
eval_run_fn = run_fn
logger.info('Loading data...')
data, batch_iter, vocabs = data_fn()
if opt['set_vocab_size']:
model_vocab_opt = model_class.get_vocab_opt(*(v.vocab_size for v in vocabs))
model_opt = ChainMap(model_vocab_opt, model_opt)
logger.info('Building graph...')
if is_training:
train_batch_iter = partial(batch_iter, *data[0])
valid_batch_iter = partial(batch_iter, *data[1])
train_model = model_class()
init_lr = train_opt['train:init_lr']
train_model.build_graph(model_opt)
if model_class == sq.SeqModel:
train_model.set_default_feed('train_loss_denom', opt['batch_size'])
else:
train_model.set_default_feed('dec.train_loss_denom', opt['batch_size'])
lr = tf.placeholder(tf.float32, shape=[], name='learning_rate')
if pg:
return_ph = tf.placeholder(tf.float32, shape=(None, None), name='return')
train_op = sq.create_pg_train_op(
train_model.nll, return_ph,
optim_class=train_opt['train:optim_class'],
learning_rate=lr, clip_gradients=train_opt['train:clip_gradients'])
else:
train_op = sq.create_train_op(
train_model.training_loss, optim_class=train_opt['train:optim_class'],
learning_rate=lr, clip_gradients=train_opt['train:clip_gradients'])
eval_batch_iter = partial(batch_iter, *data[-1])
eval_model = model_class()
eval_model.build_graph(model_opt, reuse=is_training, no_dropout=True)
logger.debug('Trainable Variables:')
for v in tf.trainable_variables():
logger.debug(f'{v.name}, {v.get_shape()}')
if is_init_only:
return
sess_config = tf.ConfigProto() if opt['gpu'] else tf.ConfigProto(device_count={'GPU': 0}) # noqa
with tf.Session(config=sess_config) as sess:
sess.run(tf.global_variables_initializer())
saver = tf.train.Saver()
if is_training:
logger.info('Training...')
success, train_state = sq.load_exp(sess, saver, opt['exp_dir'], latest=True,
checkpoint=opt['load_checkpoint'])
if success:
logger.info('Loaded model from checkpoint.')
if train_state is None:
logger.info('No experiment to resume.')
else:
logger.info('Resume experiment.')
if pg:
return_feed_fn = partial(train_model.set_default_feed, return_ph)
train_fn = partial(run_fn, sess, train_model, train_batch_iter, train_op,
return_feed_fn=return_feed_fn)
valid_fn = partial(run_fn, sess, eval_model, valid_batch_iter,
greedy=True)
eval_run_fn = partial(eval_run_fn, greedy=True)
else:
train_fn = partial(run_fn, sess, train_model, train_batch_iter, train_op)
valid_fn = partial(run_fn, sess, eval_model, valid_batch_iter)
begin_epoch_fn = partial(
sq.update_learning_rate, partial(train_model.set_default_feed, lr),
**sq.dict_with_key_startswith(train_opt, 'lr:'))
def end_epoch_fn(train_state):
sq.save_exp(sess, saver, opt['exp_dir'], train_state)
return sq.is_done_training_early(train_state, train_opt['lr:imp_wait'])
sq.train(train_fn, logger, max_epoch=train_opt['train:max_epoch'],
train_state=train_state, init_lr=init_lr,
valid_run_epoch_fn=valid_fn, begin_epoch_fn=begin_epoch_fn,
end_epoch_fn=end_epoch_fn)
checkpoint = None if is_training else opt['load_checkpoint']
if checkpoint is not None:
logger.info(f'Loading parameters from `{checkpoint}` ...')
else:
_m = 'latest' if opt['eval_latest'] else 'best'
logger.info(f'Loading parameters from {_m} checkpoint...')
success, __ = sq.load_exp(sess, saver, opt['exp_dir'], latest=opt['eval_latest'],
checkpoint=checkpoint)
if not success:
logger.warn('Loading model from checkpoint failed.')
if is_decoding:
logger.info('Decoding...')
for batch, samples in sq.decode_epoch(
sess, eval_model, eval_batch_iter,
greedy=decode_opt['decode:greedy'],
num_samples=decode_opt['decode:num_samples']):
decode_batch_fn(batch, samples, vocabs)
else:
logger.info('Evaluating...')
info = eval_run_fn(sess, eval_model, eval_batch_iter)
logger.info(info.summary('eval'))
def mle(opt, model_opt, train_opt, logger, data_fn, model_class, eval_run_fn=None):
_main(opt, model_class, model_opt, data_fn, sq.run_epoch, logger,
train_opt=train_opt, eval_run_fn=eval_run_fn)
def policy_gradient(opt, model_opt, train_opt, pg_opt, logger, data_fn, model_class,
reward_fn=None, pack_data_fn=None):
reward_fn = sq.reward_match_label if reward_fn is None else reward_fn
discount_factor = pg_opt['pg:discount']
run_fn = partial(sq.run_sampling_epoch, reward_fn=reward_fn,
with_score=pg_opt['pg:sample_logprob'], pack_data_fn=pack_data_fn,
discount_factor=discount_factor)
_main(opt, model_class, model_opt, data_fn, run_fn, logger,
train_opt=train_opt, pg=True)
def decode(opt, model_opt, decode_opt, decode_batch_fn, logger, data_fn, model_class):
_main(opt, model_class, model_opt, data_fn, sq.run_epoch, logger,
decode_opt=decode_opt, decode_batch_fn=decode_batch_fn)
if __name__ == '__main__':
import warnings
warnings.warn('This is not a main script to run. Please see other main files.')