-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathtrain.py
166 lines (127 loc) · 6.5 KB
/
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
#!/usr/bin/env python
# -*- coding_utf-8 -*-
# ===========================
# Author : LZH
# Time : 2019-05-31
# Language : Python3
# ===========================
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
import cv2
import os
import sys
import argparse
import tensorflow as tf
import numpy as np
from dataset import CifarData
from mobilenet import Mobilenet
os.chdir(os.getcwd())
tf.reset_default_graph()
slim = tf.contrib.slim
def main(args):
print_summary(args)
x = tf.placeholder(dtype=tf.float32, shape=(None,3072),name='input_data')
y_true = tf.placeholder(dtype=tf.int64, shape=(None),name='input_label' )
is_train = tf.placeholder(dtype=tf.bool,name='is_train')
global_step = tf.Variable(0, trainable=False)
reshape_x = tf.reshape(x, [-1, 3, 32, 32])
reshape_x = tf.transpose(reshape_x, perm=[0, 2, 3, 1])
with tf.variable_scope("MobileNet"):
# the class Mobilenet will ouputs 3 variable, you should ignore the first two.
# you are expect to get the ouput from Mobilenet backbone
# and i use a avg_pooling and two fully connected layer, because i think the number of neural from 1024 to 10
# has a big gap, that will lead to information loss, so i use a transition FC layer to
# trainsit the information, and use leaky_relu to activate ouput from the first FC layer.
_,_,output = Mobilenet(reshape_x, is_train).outputs
avg_pooling = tf.nn.avg_pool(output,ksize=[1,7,7,1],strides=[1,1,1,1],padding="SAME",name="Avg_pooling")
dense1 = tf.layers.dense(inputs=avg_pooling, units=512, activation=None,
kernel_initializer=tf.random_normal_initializer(stddev=0.01), trainable=True,name="dense1")
bn1 = tf.layers.batch_normalization(dense1, beta_initializer=tf.zeros_initializer(),
gamma_initializer=tf.ones_initializer(),
moving_mean_initializer=tf.zeros_initializer(),
moving_variance_initializer=tf.ones_initializer(), training=is_train,
name='bn1')
relu1 = tf.nn.leaky_relu(bn1,0.1)
dense2 = tf.layers.dense(inputs=relu1, units=10,
kernel_initializer=tf.random_normal_initializer(stddev=0.01), trainable=True,name="dense2")
sqz = tf.squeeze(dense2,[1,2],name='sqz')
prediction = tf.nn.softmax(sqz,name='prediction')
loss = tf.reduce_mean(tf.nn.sparse_softmax_cross_entropy_with_logits(labels=y_true, logits=sqz))
predict = tf.argmax(prediction, 1)
correct_prediction = tf.equal(predict, y_true)
accuracy = tf.reduce_mean(tf.cast(correct_prediction, tf.float64))
moving_ave = tf.train.ExponentialMovingAverage(0.99).apply(tf.trainable_variables())
train_filename = [os.path.join('./cifar', 'data_batch_%d' % i) for i in range(1, 6)]
test_filename = [os.path.join('./cifar', 'test_batch')]
saver = tf.train.Saver(max_to_keep=2)
with tf.control_dependencies(tf.get_collection(tf.GraphKeys.UPDATE_OPS)):
with tf.control_dependencies([moving_ave]):
train_op = tf.train.AdamOptimizer( args.lr ).minimize(loss)
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
if args.load_pretrain == '1':
saver.restore(sess,args.pretrain_path)
print('Load pre_train model from: %s ' % args.pretrain_path)
print('Start Training...')
for epoch in range(1,args.epochs[0]+1):
train_data = CifarData(train_filename, True)
for i in range(10000):
batch_data, batch_labels, = train_data.next_batch(args.batch_size[0])
loss_val, acc_val, _ ,predict1,yt= sess.run([loss, accuracy, train_op,predict,y_true],
feed_dict={x: batch_data, y_true: batch_labels, is_train: True})
if (i + 1) % 1000 == 0:
print('[Train] Epoch: %d Step: %d, loss: %4.5f, acc: %.3f' % (epoch, i + 1, loss_val, acc_val))
if (i + 1) % 2000 == 0:
test_data = CifarData(test_filename, False)
test_acc_sum = []
for j in range(100):
test_batch_data, test_batch_labels = test_data.next_batch(args.batch_size[0])
test_acc_val = sess.run([accuracy],
feed_dict={x: test_batch_data, y_true: test_batch_labels, is_train: False})
test_acc_sum.append(test_acc_val)
test_acc = np.mean(test_acc_sum)
print('[Test ] acc: %4.5f' % (test_acc))
ckpt_file = "./ckpt/mobileNet_test_acc=%.4f.ckpt" % test_acc
print('Save model to: %s \n' % ckpt_file)
saver.save(sess, ckpt_file)
def print_summary(args):
print('*'*30)
print('learning rate : {}'.format(args.lr))
print('Batch size : {}'.format(args.batch_size[0]))
print('Epoch : {}'.format(args.epochs[0]))
if args.load_pretrain == '1':
print('load_pretrain : YES')
print('pretrain_path : {}'.format(args.pretrain_path))
else:
print('load_pretrain : No')
print('*' * 30)
def parse(argv):
parser = argparse.ArgumentParser()
parser.add_argument('--lr',
type=float,
help='set lr',
default=1e-2)
parser.add_argument('--batch_size',
type=int,
nargs='+',
help='Batch Size to train',
default=16)
parser.add_argument('--epochs',
type=int,
nargs='+',
default=10,
help='Train Epochs')
parser.add_argument('--load_pretrain',
type=str,
help='load_pretrain',
nargs='+',
default='1')
parser.add_argument('--pretrain_path',
type=str,
nargs='+',
help='the path of pretrain model',
default='./ckpt/MobileNet.ckpt')
return parser.parse_args(argv)
if __name__ == '__main__':
main(parse(sys.argv[1:]))