-
Notifications
You must be signed in to change notification settings - Fork 4
/
train_mobilenet_lstm.py
188 lines (173 loc) · 7.64 KB
/
train_mobilenet_lstm.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
import sys,time,os
sys.path.append('Net')
import CNNLSTM
import mobilenet
from mobilenet import *
from CNNLSTM import *
import input_data
import tensorflow as tf
import numpy as np,time,os
# predefine
model_save_dir="./modelsmobile2"
use_pretrained_model=True
batchsize=24
time_steps=6
hidden_size=150
classes=19
max_steps=30000
def train(train_root,train_txt,valid_root,valid_txt):
if not os.path.exists(model_save_dir):
os.makedirs(model_save_dir)
model_filename = "./models/mobilenet_lstm_model_0.51-4199"
model_filename = ""
if len(model_filename)!=0:
start_steps=int(model_filename.strip().split('-')[-1])
else:
start_steps=0
graph = tf.Graph()
with graph.as_default():
X,Y,endpoints,features,predict,loss,accuracy = inference_attention_mobilenet_lstm(batchsize=batchsize,
time_steps=time_steps,
hidden_size=hidden_size,
classes=classes,
loss='softmax_loss')
learning_rate_value = 0.0002
learning_rate = tf.Variable(learning_rate_value, trainable=False)
optimizer = tf.train.AdamOptimizer(learning_rate=learning_rate)
weight_decay_loss = tf.add_n(tf.get_collection('weight_decay_loss'))
total_loss = loss+weight_decay_loss
train_op = optimizer.minimize(total_loss)
saver = tf.train.Saver(max_to_keep=15,keep_checkpoint_every_n_hours=1)
init = tf.global_variables_initializer()
load_op = load_pretrained_mobilenet_model_ops()
#tf.summary.image('input_images', X, 4)
merged = tf.summary.merge_all()
with tf.Session(config=tf.ConfigProto(
allow_soft_placement=True
#gpu_options = tf.GPUOptions(per_process_gpu_memory_fraction=0.9)
),graph=graph) as sess:
sess.run(init)
if model_filename == "":
sess.run(load_op)
print('load_op')
pass
else:
saver.restore(sess, model_filename)
train_writer = tf.summary.FileWriter('./visual_logs/train/attention', sess.graph)
# train process
next_batch_start = -1
last_acc = 0
lines=None
epoch = int(start_steps/(1900/(batchsize)))
summ_loss=0
epoch_steps=0
best_acc = 0
for step in range(start_steps,max_steps):
start_time = time.time()
epoch_steps+=1
if epoch>=10:
# open data augmentation
status = 'TRAIN'
else:
# close data augmentation
status = 'TEST'
startprocess_time = time.time()
train_images, train_labels, next_batch_start, _, _,lines = input_data.read_clip_and_label(
rootdir = train_root,
filename= train_txt,
batch_size=batchsize,
lines=lines,
start_pos=next_batch_start,
num_frames_per_clip=time_steps,
crop_size=(CNNLSTM.HEIGHT,CNNLSTM.WIDTH),
shuffle=False,
phase=status
)
train_images = train_images.reshape([-1,CNNLSTM.HEIGHT,CNNLSTM.WIDTH,CNNLSTM.CHANNELS])
endprocess_time = time.time()
#preprocess_time = ((endprocess_time-startprocess_time)/(batchsize*1.0))
#print("preprocess per time :%f"%preprocess_time)
#endpoints_res = sess.run(endpoints, feed_dict={
# X: train_images,
# Y: train_labels
# })
#for e in endpoints_res:
# print(e.shape)
_,losses,wd_losses = sess.run([train_op,loss,weight_decay_loss], feed_dict={
X: train_images,
Y: train_labels
})
summ_loss+=(losses+wd_losses)
##train_writer.add_summary(summary, step)
duration = time.time() - start_time
print('Epoch: %d Step %d: %.3f sec' % (epoch, step, duration))
print ("lr:%f loss: %.4f wd_loss: %.4f avg_loss: %.8f"%(sess.run(learning_rate),losses,wd_losses,summ_loss*1.0/epoch_steps))
# Save a checkpoint and evaluate the model periodically.
#if step%2000==0 and step!=0 and step!=start_steps or step+1 == max_steps:
# saver.save(sess, os.path.join(model_save_dir, 'mobilenet_lstm_model'), global_step=step)
# print('Model Saved.')
# print('Training Data Eval:')
pChangeLR = False
if next_batch_start == -1:
epoch+=1
summ_loss=0
epoch_steps=0
if epoch==450:
print("Learning Done.")
break
if epoch%5==0 and epoch!=0:
# test
test_batch_start = -1
sum_acc=0
total_num=0
patients=0
while True:
test_lines = None
val_images, val_labels, test_batch_start, _, _,test_lines = input_data.read_clip_and_label(
rootdir= valid_root,
filename= valid_txt,
batch_size=1,
lines=test_lines,
start_pos=test_batch_start,
num_frames_per_clip=time_steps,
crop_size=(CNNLSTM.HEIGHT, CNNLSTM.WIDTH),
shuffle=False,
phase='TEST'
)
val_images=np.array([val_images[0,:]]*batchsize,dtype=np.float32)
val_labels=np.array([val_labels]*batchsize,dtype=np.int64).ravel()
val_images = val_images.reshape([-1,CNNLSTM.HEIGHT,CNNLSTM.WIDTH,CNNLSTM.CHANNELS])
[acc] = sess.run(
[ accuracy],
feed_dict={
X: val_images,
Y: val_labels
})
sum_acc+=acc
total_num+=1
if test_batch_start == -1:
acc = sum_acc*1.0/total_num
print('Epoch: %d test accuracy: %f'%(epoch,acc))
break
if acc<last_acc*1.05:
patients+=1
if patients>=2:
pChangeLR=True
patients=0
if epoch>=100 and acc>best_acc:
best_acc=acc
saver.save(sess, os.path.join(model_save_dir, 'mobilenet_lstm_model_%.2f'%best_acc), global_step=step)
if pChangeLR or (step%8000==0 and step!=0):
learning_rate_value = sess.run(learning_rate)
learning_rate_value = learning_rate_value*0.5
assign_op = tf.assign(learning_rate, learning_rate_value)
sess.run(assign_op)
print("acc:%f < last_acc*1.1:%f"%(acc,last_acc*1.05))
print("learning_rate changed to %f"%sess.run(learning_rate))
last_acc = acc
print("Done")
train_writer.flush()
train_writer.close()
if __name__ == '__main__':
train(train_root='E:/dataset/VIVA_avi_group/VIVA_avi_part2/train',train_txt='E:/dataset/VIVA_avi_group/VIVA_avi_part2/gen_train_shuffle.txt',
valid_root='E:/dataset/VIVA_avi_group/VIVA_avi_part2/val',valid_txt='E:/dataset/VIVA_avi_group/VIVA_avi_part2/val.txt')