-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathmain.py
170 lines (131 loc) · 7.09 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
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
import tensorflow as tf
import logging
from tqdm import tqdm
import numpy as np
import time
import os
from RAM import RAMNetwork
from input_fn import get_data
from utility import Utility, auto_adjust_flags
from Visualization import Visualization
def eval(model, sess, FLAGS, handle, num_batches, writer, prefix, is_training=False):
'''
:param dataset: tuple of (init_op, data, placeholder)
'''
fetch = [model.summary, model.global_step, model.accuracy, model.accuracy_MC, model.reward, model.loss, model.xent,
model.RL_loss, model.baselines_mse, model.learning_rate, model.unknown_accuracy]
vars = ['acc', 'acc_MC', 'reward', 'loss', 'xent', 'RL_loss', 'b_mse', 'lr', 'acc_unknown']
averages = np.zeros(len(vars))
assert(len(fetch) - 2 == len(vars))
for _ in range(num_batches):
out = sess.run(fetch, feed_dict={model.is_training: is_training,
model.handle: handle})
averages += np.array(out[2:]) / num_batches
summs = [tf.Summary.Value(tag="batch/" + var, simple_value=avg) for var, avg in zip(vars, averages)]
batch_values = tf.Summary(value=summs)
step = out[1]
writer.add_summary(batch_values, step)
writer.add_summary(out[0], step)
strs = [item for pair in zip(vars, averages) for item in pair]
s = 'step ' + str(step) + ' - ' + prefix + len(vars) * ' {}: {:.3f}'
logging.info(s.format(*strs))
report = dict(zip(vars, averages))
report['step'] = step
report['MC_sampling'] = FLAGS.MC_samples
report['loc_std'] = FLAGS.loc_std
return report
def train_model(model, FLAGS):
with tf.Session() as sess:
train_writer, valid_writer, test_writer, train_handle, valid_handle, test_handle = model.setup(sess, train_data, valid_data, test_data)
Visual = Visualization(model, FLAGS)
epochs_completed = 0
best_acc_validation = 0.
print("t - before train: {:.2f}s".format(time.time() - start_time))
while epochs_completed < FLAGS.num_epochs:
# Evaluate and visualize
if epochs_completed % FLAGS.eval_step_interval == 0:
Visual(sess, 'epoch_{}'.format(epochs_completed), valid_handle)
_ = eval(model, sess, FLAGS, train_handle, FLAGS.batches_per_eval_valid, train_writer,
prefix='TRAIN: ', is_training=True)
report = eval(model, sess, FLAGS, valid_handle, FLAGS.batches_per_eval_valid, valid_writer,
prefix='VALIDATION: ')
if report['acc'] > best_acc_validation:
best_acc_validation = report['acc']
model.saver_best.save(sess, FLAGS.path + "/cp_best.ckpt", write_meta_graph=False)
# Train
fetch = [model.train_op, model.global_step]
feed = {model.is_training: True,
model.handle: train_handle}
for i in tqdm(range(FLAGS.train_batches_per_epoch), desc='Train'):
if i % 100 == 0:
_, step, summary = sess.run(fetch + [model.summary], feed_dict=feed)
train_writer.add_summary(summary, step)
else:
_, step = sess.run(fetch, feed_dict=feed)
# Checkpoint
if (epochs_completed % 1 == 0) or (epochs_completed - 1 == FLAGS.num_epochs):
model.saver.save(sess, FLAGS.path + "/cp.ckpt") # global_step=model.global_step)
epochs_completed += 1
# Test set
logging.info('FINISHED TRAINING, {} EPOCHS COMPLETED\n'.format(epochs_completed))
eval(model, sess, FLAGS, test_handle, FLAGS.batches_per_eval_test, test_writer, prefix='TEST - LAST MODEL: ')
Visual(sess, 'epoch_final', test_handle)
model.saver.restore(sess, FLAGS.path + "/cp_best.ckpt")
logging.info('Best validation accuracy: {:.3f}'.format(best_acc_validation))
eval(model, sess, FLAGS, test_handle, FLAGS.batches_per_eval_test, test_writer, prefix='TEST - BEST MODEL: ')
train_writer.close()
valid_writer.close()
test_writer.close()
def select_hyper_para_random(FLAGS):
FLAGS.learning_rate_decay_factor = np.round(np.random.uniform(0.93, 0.99), 3)
FLAGS.learning_rate_decay_steps = np.round(np.random.uniform(350, 650), 0)
FLAGS.loc_std = np.round(np.random.uniform(0.075, 0.125), 3)
FLAGS.learning_rate_RL = np.round(np.random.uniform(1., 1.4), 3)
logging.info('CHOSEN PARAMS:\n'
'batch size:\t{}\n'
'MC samples:\t{}\n'
'loc std:\t{:.3f}\n'
'lrate decay\t{:.3f}\n'
'lrate decay step\t{}\n'
'lrate RL\t{:.3f}\n'.format(
FLAGS.batch_size, FLAGS.MC_samples, FLAGS.loc_std,
FLAGS.learning_rate_decay_factor, FLAGS.learning_rate_decay_steps,
FLAGS.learning_rate_RL))
return FLAGS
if __name__ == "__main__":
logging.getLogger().setLevel(logging.INFO)
# Parsing experimental set up
FLAGS, _ = Utility.parse_arg()
# set img_shape, padding, num_classes according to dataset (ignoring cl inputs!)
auto_adjust_flags(FLAGS)
max_runs = 1
if FLAGS.random_search:
max_runs = 20
for r in range(max_runs):
if FLAGS.random_search:
FLAGS = select_hyper_para_random(FLAGS)
if not FLAGS.start_checkpoint:
unknown_suffix = '_uk{}'.format(FLAGS.n_unknown_train) if FLAGS.open_set else ''
experiment_name = '{}gl_bs{}_MC{}_std{}_dcay{}_step{}_lr{}_lrRL{}_{}sc{}_{}{}_Tanh{}_{}A'.format(
FLAGS.num_glimpses, FLAGS.batch_size, FLAGS.MC_samples, FLAGS.loc_std,
FLAGS.learning_rate_decay_factor, FLAGS.learning_rate_decay_steps,
FLAGS.learning_rate, FLAGS.learning_rate_RL, len(FLAGS.scale_sizes),
FLAGS.scale_sizes[0], FLAGS.cell, FLAGS.size_rnn_state, unknown_suffix,
FLAGS.exp_name_suffix)
else:
logging.info('CONTINUE TRAINING\n')
experiment_name = FLAGS.start_checkpoint
t_sz = (str(FLAGS.translated_size) if FLAGS.translated_size else "")
FLAGS.path = FLAGS.summaries_dir + '/' + FLAGS.dataset + t_sz + '/' + experiment_name
logging.info('\nPATH: ' + FLAGS.path + '\nCURRENT MODEL: ' + experiment_name + '\n')
start_time = time.time()
# load datasets
train_data, valid_data, test_data = get_data(FLAGS)
print("t - data loaded: {:.2f}s".format(time.time() - start_time))
with tf.device('/device:GPU:*'):
RAM = RAMNetwork(FLAGS=FLAGS,
full_summary=False)
print("t - built graph: {:.2f}s".format(time.time() - start_time))
train_model(RAM, FLAGS)
# tensorboard --logdir=logs/
# tensorboard --logdir=\\?\path for long paths due to windows max_len restriction