-
Notifications
You must be signed in to change notification settings - Fork 1
/
run_faa.py
283 lines (234 loc) · 10.4 KB
/
run_faa.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
import os
import argparse
import numpy as np
import tensorflow as tf
import pandas as pd
from pandas import Series, DataFrame
from model.gra import GRA
import config as cfg
import util.dataset as dt
def train(epoch, dataset, config, log_dir):
"""Train model for one epoch."""
model_config = config['model']
train_config = config['train']
sess_config = config['session']
with tf.Graph().as_default():
model = GRA(model_config)
model.build_inference()
model.build_loss(train_config['reg_coeff'], train_config['shu_coeff'])
model.build_train(train_config['learning_rate'])
with tf.Session(config=sess_config) as sess:
sum_dir = os.path.join(log_dir, 'summary')
# create event file for graph
if not os.path.exists(sum_dir):
summary_writer = tf.summary.FileWriter(sum_dir, sess.graph)
summary_writer.close()
summary_writer = tf.summary.FileWriter(sum_dir)
ckpt_dir = os.path.join(log_dir, 'checkpoint')
ckpt_path = tf.train.latest_checkpoint(ckpt_dir)
saver = tf.train.Saver()
if ckpt_path:
print('load checkpoint {}.'.format(ckpt_path))
saver.restore(sess, ckpt_path)
else:
print('no checkpoint.')
if not os.path.exists(ckpt_dir):
os.makedirs(ckpt_dir)
sess.run(tf.global_variables_initializer())
stats_dir = os.path.join(log_dir, 'stats')
stats_path = os.path.join(stats_dir, 'train.json')
if os.path.exists(stats_path):
print('load stats file {}.'.format(stats_path))
stats = pd.read_json(stats_path, 'records')
else:
print('no stats file.')
if not os.path.exists(stats_dir):
os.makedirs(stats_dir)
stats = pd.DataFrame(columns=['epoch', 'loss', 'acc'])
# train iterate over batch
batch_idx = 0
total_loss = 0
total_acc = 0
batch_total = np.sum(dataset.train_batch_total)
while dataset.has_train_batch:
vgg, c3d, question, answer = dataset.get_train_batch()
feed_dict = {
model.appear: vgg,
model.motion: c3d,
model.question_encode: question,
model.answer_encode: answer
}
_, loss, acc = sess.run(
[model.train, model.loss, model.acc], feed_dict)
total_loss += loss
total_acc += acc
if batch_idx % 100 == 0:
print('[TRAIN] epoch {}, batch {}/{}, loss {:.5f}, acc {:.5f}.'.format(
epoch, batch_idx, batch_total, loss, acc))
batch_idx += 1
loss = total_loss / batch_total
acc = total_acc / batch_total
print('\n[TRAIN] epoch {}, loss {:.5f}, acc {:.5f}.\n'.format(
epoch, loss, acc))
summary = tf.Summary()
summary.value.add(tag='train/loss', simple_value=float(loss))
summary.value.add(tag='train/acc', simple_value=float(acc))
summary_writer.add_summary(summary, epoch)
record = Series([epoch, loss, acc], ['epoch', 'loss', 'acc'])
stats = stats.append(record, ignore_index=True)
saver.save(sess, os.path.join(ckpt_dir, 'model.ckpt'), epoch)
stats.to_json(stats_path, 'records')
dataset.reset_train()
return loss, acc
def val(epoch, dataset, config, log_dir):
"""Validate model."""
model_config = config['model']
sess_config = config['session']
answerset = pd.read_csv(
os.path.join(config['preprocess_dir'], 'answer_set.txt'), header=None)[0]
with tf.Graph().as_default():
model = GRA(model_config)
model.build_inference()
with tf.Session(config=sess_config) as sess:
sum_dir = os.path.join(log_dir, 'summary')
summary_writer = tf.summary.FileWriter(sum_dir)
ckpt_dir = os.path.join(log_dir, 'checkpoint')
save_path = tf.train.latest_checkpoint(ckpt_dir)
saver = tf.train.Saver()
if save_path:
print('load checkpoint {}.'.format(save_path))
saver.restore(sess, save_path)
else:
print('no checkpoint.')
exit()
stats_dir = os.path.join(log_dir, 'stats')
stats_path = os.path.join(stats_dir, 'val.json')
if os.path.exists(stats_path):
print('load stats file {}.'.format(stats_path))
stats = pd.read_json(stats_path, 'records')
else:
print('no stats file.')
if not os.path.exists(stats_dir):
os.makedirs(stats_dir)
stats = pd.DataFrame(columns=['epoch', 'acc'])
# val iterate over examples
correct = 0
while dataset.has_val_example:
vgg, c3d, question, answer = dataset.get_val_example()
feed_dict = {
model.appear: [vgg],
model.motion: [c3d],
model.question_encode: [question],
}
prediction = sess.run(model.prediction, feed_dict=feed_dict)
prediction = prediction[0]
if answerset[prediction] == answer:
correct += 1
acc = correct / float(dataset.val_example_total)
print('\n[VAL] epoch {}, acc {:.5f}.\n'.format(epoch, acc))
summary = tf.Summary()
summary.value.add(tag='val/acc', simple_value=float(acc))
summary_writer.add_summary(summary, epoch)
record = Series([epoch, acc], ['epoch', 'acc'])
stats = stats.append(record, ignore_index=True)
stats.to_json(stats_path, 'records')
dataset.reset_val()
return acc
def test(dataset, config, log_dir):
"""Test model, output prediction as json file."""
model_config = config['model']
sess_config = config['session']
answerset = pd.read_csv(
os.path.join(config['preprocess_dir'], 'answer_set.txt'), header=None)[0]
with tf.Graph().as_default():
model = GRA(model_config)
model.build_inference()
with tf.Session(config=sess_config) as sess:
ckpt_dir = os.path.join(log_dir, 'checkpoint')
checkpoints=['model.ckpt-' +str(i+10) for i in range(19)]
print checkpoints
for i in range(len(checkpoints)):
ckpt_name = os.path.join(ckpt_dir, checkpoints[i])
#save_path = tf.train.get_checkpoint_state(ckpt_dir)
saver = tf.train.Saver()
saver.restore(sess, ckpt_name)
#saver = tf.train.Saver()
#if save_path:
#print('load checkpoint {}.'.format(save_path))
#saver.restore(sess, save_path)
#else:
# print('no checkpoint.')
# exit()
# test iterate over examples
result = DataFrame(columns=['id', 'answer'])
correct = 0
while dataset.has_test_example:
vgg, c3d, question, answer, example_id = dataset.get_test_example()
feed_dict = {
model.appear: [vgg],
model.motion: [c3d],
model.question_encode: [question],
}
prediction, channel_weight, appear_weight, motion_weight = sess.run(
[model.prediction, model.channel_weight, model.appear_weight, model.motion_weight], feed_dict=feed_dict)
prediction = prediction[0]
channel_weight = channel_weight[0]
appear_weight = appear_weight[0]
motion_weight = motion_weight[0]
result = result.append(
{'id': example_id, 'answer': answerset[prediction]}, ignore_index=True)
if answerset[prediction] == answer:
correct += 1
#print(answer, example_id, channel_weight)
# print(appear_weight)
# print(motion_weight)
print correct
print dataset.test_example_total
result.to_json(os.path.join(
log_dir, 'prediction.json'), 'records')
acc = correct / float(dataset.test_example_total)
print('\n[TEST] acc {:.5f}.\n'.format(acc))
dataset.reset_test()
print str(i+10)+' '+str(acc)
return acc
def main():
"""Main script."""
parser = argparse.ArgumentParser()
parser.add_argument('--mode', required=True,
help='train/test')
parser.add_argument('--gpu', required=True,
help='gpu id')
parser.add_argument('--log', required=True,
help='log directory')
parser.add_argument('--dataset', required=True,
help='dataset name, msvd_qa/msrvtt_qa')
parser.add_argument('--config', required=True,
help='config id')
args = parser.parse_args()
config = cfg.get('gra', args.dataset, args.config, args.gpu)
if args.dataset == 'msvd_qa':
dataset = dt.MSVDQA(
config['train']['batch_size'], config['preprocess_dir'])
elif args.dataset == 'msrvtt_qa':
dataset = dt.MSRVTTQA(
config['train']['batch_size'], config['preprocess_dir'])
if args.mode == 'train':
best_val_acc = -1
val_acc = 0
not_improved = -1
for epoch in range(0, 30):
if val_acc > best_val_acc:
best_val_acc = val_acc
not_improved = 0
else:
not_improved += 1
#if not_improved == 10:
#print('early stopping.')
#break
train(epoch, dataset, config, args.log)
val_acc = val(epoch, dataset, config, args.log)
elif args.mode == 'test':
print('start test.')
test(dataset, config, args.log)
if __name__ == '__main__':
main()