-
Notifications
You must be signed in to change notification settings - Fork 17
/
dfb_train.py
190 lines (168 loc) · 7.62 KB
/
dfb_train.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
# -*- coding:utf-8 -*-
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
import sys
import tensorflow as tf
from tensorflow.contrib import slim
from models.nets.resnet import resnet_v2
from models.nets import resnet_utils
import os
import time
import numpy as np
from time import time
from dataset import reader
from models.dfb import dfb
from dfb_opt import train
import dfb_utils
from PIL import Image
from config import *
M = FLAGS.num_class
k = 10
def main(argv=None):
is_training = True
input_images = tf.placeholder(dtype=tf.float32,
shape=[FLAGS.batch_size, 448, 448, 3],
name="input_images")
y_true = tf.placeholder(dtype=tf.float32,
shape=[FLAGS.batch_size, FLAGS.num_class],
name="y_true")
keep_prob = tf.placeholder(dtype=tf.float32,
name="dropout")
learning_rate = tf.placeholder(dtype=tf.float64,
name="learning_rate")
clip_value = tf.placeholder(dtype=tf.float32,
name="clip_value")
"""
""inference
"""
fc_obj, fc_part, fc_ccp, base_var_list, t_var_list = dfb(input_images, keep_prob, is_training=is_training)
# Predictions
t_predictions = (tf.nn.softmax(fc_part) + 0.1 * tf.nn.softmax(fc_ccp) + tf.nn.softmax(fc_obj)) / 3.
"""
""Loss
"""
obj_loss = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(labels=y_true, logits=fc_obj))
part_loss = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(labels=y_true, logits=fc_part))
ccp_loss = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(labels=y_true, logits=fc_ccp))
loss = 0.1 * ccp_loss + part_loss + obj_loss
acc_top1 = dfb_utils.accuracy_top1(y_true, t_predictions)
acc_top5 = dfb_utils.accuracy_top5(y_true, t_predictions)
"""
""Summary
"""
tf.summary.scalar("t_obj_loss", obj_loss)
tf.summary.scalar("t_part_loss", part_loss)
tf.summary.scalar("t_ccp_loss", ccp_loss)
tf.summary.scalar("t_loss", loss)
tf.summary.scalar("acc_top1", acc_top1)
tf.summary.scalar("acc_top5", acc_top5)
train_op = train(loss, base_var_list, t_var_list, learning_rate, clip_value)
print("Setting up summary op...")
summary_op = tf.summary.merge_all()
"""
" Loading Data
"""
print("Loading Data......")
if FLAGS.mode == 'train':
generator_train = reader.generator(FLAGS.train_dir, FLAGS.train_json)
print("\tLoaded Train Data......")
generator_valid = reader.generator(FLAGS.valid_dir, FLAGS.valid_json)
print("\tLoaded Validation Data......")
"""
" Setting up Saver
"""
print("Setting up Saver...")
sess = tf.Session(config=tf.ConfigProto(gpu_options=tf.GPUOptions(allow_growth=True)))
saver = tf.train.Saver(t_var_list, max_to_keep=3)
train_writer = tf.summary.FileWriter(os.path.join(FLAGS.logs_dir, 'train'),
sess.graph)
valid_writer = tf.summary.FileWriter(os.path.join(FLAGS.logs_dir, 'valid'),
sess.graph)
print("Initialize global variables")
sess.run(tf.global_variables_initializer())
"""
" Resume
"""
ckpt = tf.train.get_checkpoint_state(FLAGS.logs_dir)
if ckpt and ckpt.model_checkpoint_path:
saver.restore(sess, ckpt_t.model_checkpoint_path)
print("Model restored...")
"""
" Training...
"""
if FLAGS.mode == 'train':
train_batch = int(FLAGS.num_train / FLAGS.batch_size)
valid_batch = int(FLAGS.num_valid / FLAGS.batch_size)
last_loss = 10000.
patience = 0
best_acc = 0.0
clipvalue = 1e-4
current = 1e-4
for epoch in range(1, FLAGS.epoches if FLAGS.debug else 1):
print("Epoch %i ----> Starting......" % epoch)
start_time = time()
"""
" Build learning rate
"""
if epoch <= 10:
lr = 1e-3 / 10.0 * epoch
current = lr
elif epoch > 10:
lr = current
for step in range(train_batch):
batch_x, batch_y = next(generator_train)
summary, _ , loss_t, loss_ct, loss_ot, loss_pt, acc_1t, acc_5t = \
sess.run([summary_op, train_op, loss, ccp_loss, obj_loss, part_loss, acc_top1, acc_top5],
feed_dict={input_images: batch_x,
y_true: batch_y,
keep_prob: 0.3,
learning_rate: lr,
clip_value: clipvalue})
print("Step %i, Loss %0.4f, ccp_loss %0.4f,obj_loss %0.4f,part_loss %0.4f, acc_1 %0.4f,acc_5 %0.4f, lr %.7f,clip:%.7f" %
((epoch-1) * train_batch + step, loss_t, loss_ct, loss_ot, loss_pt, acc_1t, acc_5t, lr, clipvalue))
train_writer.add_summary(summary, step + train_batch * (epoch-1))
acc1_reg = []
acc5_reg = []
loss_reg = []
for step in range(valid_batch):
batch_x, batch_y = next(generator_valid)
loss_v, acc_1v, acc_5v, summary = sess.run([loss, acc_top1, acc_top5, summary_op],
feed_dict={input_images: batch_x,
y_true: batch_y,
keep_prob: 1.,
learning_rate: lr,
clip_value: clipvalue})
valid_writer.add_summary(summary, step + valid_batch * (epoch-1))
acc1_reg.append(acc_1v)
acc5_reg.append(acc_5v)
loss_reg.append(loss_v)
avg_acc1 = np.mean(np.array(acc1_reg))
avg_acc5 = np.mean(np.array(acc5_reg))
avg_loss = np.mean(np.array(loss_reg))
print("Valid_loss ----> %0.4f Valid_acc ----> %0.4f, %0.4f" % (avg_loss, avg_acc1, avg_acc5))
"""
" Save the best model
"""
if avg_acc1 > best_acc:
best_acc = avg_acc1
saver.save(sess=sess,
save_path=os.path.join(FLAGS.teacher_logs_dir, 'teacher'),
global_step=epoch)
print("Save the best model with val_acc %0.4f" % best_acc)
else:
print("Val_acc stay with val_acc %0.4f" % best_acc)
if patience >= FLAGS.patience:
patience = 0
last_loss = 10000
current = current * 0.8
clipvalue = clipvalue * 0.8
print("Lr decay, update the learning rate when lr = %0.4f" % lr)
end_time = time()
print("Epoch %i ----> Ended in %0.4f" % (epoch, end_time - start_time))
train_writer.close()
valid_writer.close()
print("......Ended")
print("Ending......")
if __name__ == "__main__":
tf.app.run()