FilesExpand file tree

 master

/

tutorial_cifar10.py

Copy path

Blame

More file actions

Blame

More file actions

Latest commit

 

History

History

History

163 lines (134 loc) · 7.38 KB

 master

/

tutorial_cifar10.py

Copy path

Top

File metadata and controls

• Code
• Blame

163 lines (134 loc) · 7.38 KB

Raw

Copy raw file

Download raw file

Open symbols panel

Edit and raw actions

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

#! /usr/bin/python

# -*- coding: utf-8 -*-

import time

import numpy as np

import tensorflow as tf

import tensorlayer as tl

from tensorlayer.layers import *

sess = tf.InteractiveSession()

X_train, y_train, X_test, y_test = tl.files.load_cifar10_dataset(shape=(-1, 32, 32, 3), plotable=False)

def model(x, y_, reuse):

W_init = tf.truncated_normal_initializer(stddev=5e-2)

W_init2 = tf.truncated_normal_initializer(stddev=0.04)

b_init2 = tf.constant_initializer(value=0.1)

with tf.variable_scope("model", reuse=reuse):

net = InputLayer(x, name='input')

net = Conv2d(net, 64, (5, 5), (1, 1), act=tf.nn.relu, padding='SAME', W_init=W_init, name='cnn1')

net = MaxPool2d(net, (3, 3), (2, 2), padding='SAME', name='pool1')

net = LocalResponseNormLayer(net, depth_radius=4, bias=1.0, alpha=0.001 / 9.0, beta=0.75, name='norm1')

net = Conv2d(net, 64, (5, 5), (1, 1), act=tf.nn.relu, padding='SAME', W_init=W_init, name='cnn2')

net = LocalResponseNormLayer(net, depth_radius=4, bias=1.0, alpha=0.001 / 9.0, beta=0.75, name='norm2')

net = MaxPool2d(net, (3, 3), (2, 2), padding='SAME', name='pool2')

net = FlattenLayer(net, name='flatten')

net = DenseLayer(net, 384, act=tf.nn.relu, W_init=W_init2, b_init=b_init2, name='d1relu')

net = DenseLayer(net, 192, act=tf.nn.relu, W_init=W_init2, b_init=b_init2, name='d2relu')

net = DenseLayer(net, 10, act=tf.identity, W_init=W_init2, name='output')

y = net.outputs

ce = tl.cost.cross_entropy(y, y_, name='cost')

# L2 for the MLP, without this, the accuracy will be reduced by 15%.

L2 = 0

for p in tl.layers.get_variables_with_name('relu/W', True, True):

L2 += tf.contrib.layers.l2_regularizer(0.004)(p)

cost = ce + L2

correct_prediction = tf.equal(tf.argmax(y, 1), y_)

acc = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))

return net, cost, acc

def model_batch_norm(x, y_, reuse, is_train):

""" Batch normalization should be placed before rectifier. """

W_init = tf.truncated_normal_initializer(stddev=5e-2)

W_init2 = tf.truncated_normal_initializer(stddev=0.04)

b_init2 = tf.constant_initializer(value=0.1)

with tf.variable_scope("model", reuse=reuse):

net = InputLayer(x, name='input')

net = Conv2d(net, 64, (5, 5), (1, 1), padding='SAME', W_init=W_init, b_init=None, name='cnn1')

net = BatchNormLayer(net, is_train, act=tf.nn.relu, name='batch1')

net = MaxPool2d(net, (3, 3), (2, 2), padding='SAME', name='pool1')

net = Conv2d(net, 64, (5, 5), (1, 1), padding='SAME', W_init=W_init, b_init=None, name='cnn2')

net = BatchNormLayer(net, is_train, act=tf.nn.relu, name='batch2')

net = MaxPool2d(net, (3, 3), (2, 2), padding='SAME', name='pool2')

net = FlattenLayer(net, name='flatten') # output: (batch_size, 2304)

net = DenseLayer(net, 384, act=tf.nn.relu, W_init=W_init2, b_init=b_init2, name='d1relu')

net = DenseLayer(net, 192, act=tf.nn.relu, W_init=W_init2, b_init=b_init2, name='d2relu')

net = DenseLayer(net, 10, act=tf.identity, W_init=W_init2, name='output')

y = net.outputs

ce = tl.cost.cross_entropy(y, y_, name='cost')

# L2 for the MLP, without this, the accuracy will be reduced by 15%.

L2 = 0

for p in tl.layers.get_variables_with_name('relu/W', True, True):

L2 += tf.contrib.layers.l2_regularizer(0.004)(p)

cost = ce + L2

correct_prediction = tf.equal(tf.argmax(y, 1), y_)

acc = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))

return net, cost, acc

def distort_fn(x, is_train=False):

"""

The images are processed as follows:

.. They are cropped to 24 x 24 pixels, centrally for evaluation or randomly for training.

.. They are approximately whitened to make the model insensitive to dynamic range.

For training, we additionally apply a series of random distortions to

artificially increase the data set size:

.. Randomly flip the image from left to right.

.. Randomly distort the image brightness.

"""

# print('begin',x.shape, np.min(x), np.max(x))

x = tl.prepro.crop(x, 24, 24, is_random=is_train)

# print('after crop',x.shape, np.min(x), np.max(x))

if is_train:

# x = tl.prepro.zoom(x, zoom_range=(0.9, 1.0), is_random=True)

# print('after zoom', x.shape, np.min(x), np.max(x))

x = tl.prepro.flip_axis(x, axis=1, is_random=True)

# print('after flip',x.shape, np.min(x), np.max(x))

x = tl.prepro.brightness(x, gamma=0.1, gain=1, is_random=True)

# print('after brightness',x.shape, np.min(x), np.max(x))

# tmp = np.max(x)

# x += np.random.uniform(-20, 20)

# x /= tmp

# normalize the image

x = (x - np.mean(x)) / max(np.std(x), 1e-5) # avoid values divided by 0

# print('after norm', x.shape, np.min(x), np.max(x), np.mean(x))

return x

x = tf.placeholder(dtype=tf.float32, shape=[None, 24, 24, 3], name='x')

y_ = tf.placeholder(dtype=tf.int64, shape=[None], name='y_')

## using local response normalization

# network, cost, _ = model(x, y_, False)

# _, cost_test, acc = model(x, y_, True)

## you may want to try batch normalization

network, cost, _ = model_batch_norm(x, y_, False, is_train=True)

_, cost_test, acc = model_batch_norm(x, y_, True, is_train=False)

# train

n_epoch = 50000

learning_rate = 0.0001

print_freq = 1

batch_size = 128

train_params = network.all_params

train_op = tf.train.AdamOptimizer(learning_rate, beta1=0.9, beta2=0.999, epsilon=1e-08,

use_locking=False).minimize(cost, var_list=train_params)

tl.layers.initialize_global_variables(sess)

network.print_params(False)

network.print_layers()

print(' learning_rate: %f' % learning_rate)

print(' batch_size: %d' % batch_size)

for epoch in range(n_epoch):

start_time = time.time()

for X_train_a, y_train_a in tl.iterate.minibatches(X_train, y_train, batch_size, shuffle=True):

X_train_a = tl.prepro.threading_data(X_train_a, fn=distort_fn, is_train=True) # data augmentation for training

sess.run(train_op, feed_dict={x: X_train_a, y_: y_train_a})

if epoch + 1 == 1 or (epoch + 1) % print_freq == 0:

print("Epoch %d of %d took %fs" % (epoch + 1, n_epoch, time.time() - start_time))

# train_loss, train_acc, n_batch = 0, 0, 0

# for X_train_a, y_train_a in tl.iterate.minibatches(

# X_train, y_train, batch_size, shuffle=True):

# X_train_a = tl.prepro.threading_data(X_train_a, fn=distort_fn, is_train=False) # central crop

# err, ac = sess.run([cost_test, acc], feed_dict={x: X_train_a, y_: y_train_a})

# train_loss += err; train_acc += ac; n_batch += 1

# print(" train loss: %f" % (train_loss/ n_batch))

# print(" train acc: %f" % (train_acc/ n_batch))

test_loss, test_acc, n_batch = 0, 0, 0

for X_test_a, y_test_a in tl.iterate.minibatches(X_test, y_test, batch_size, shuffle=False):

X_test_a = tl.prepro.threading_data(X_test_a, fn=distort_fn, is_train=False) # central crop

err, ac = sess.run([cost_test, acc], feed_dict={x: X_test_a, y_: y_test_a})

test_loss += err

test_acc += ac

n_batch += 1

print(" test loss: %f" % (test_loss / n_batch))

print(" test acc: %f" % (test_acc / n_batch))