-
Notifications
You must be signed in to change notification settings - Fork 7
/
05_convolutional_net2.py
127 lines (106 loc) · 5.85 KB
/
05_convolutional_net2.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
#-*- coding:utf-8 -*-
#!/usr/bin/env python
# 卷积神经网络
# 权重(卷积核) + 偏置 AdamOptimizer
import tensorflow as tf
import numpy as np
from tensorflow.examples.tutorials.mnist import input_data# 读取mnist手写字体数据工具
batch_size = 128#每一次训练 batch大小 即每次读入128张图片进行训练
test_size = 256 #每一次测试 batch大小
## 初始化权重 #####
def init_weights(shape):
return tf.Variable(tf.random_normal(shape, stddev=0.01))
## 初始化偏置 #####
def init_bias(shape):
return tf.Variable(tf.constant(0.1, shape=shape))
## 卷积 ##
def conv2d(x, W):
return tf.nn.conv2d(x, W, strides=[1, 1, 1, 1], padding='SAME') #stride size=buchang=1 padding size=bianju=0 bu 0
## 池化 ##
def max_pool_2x2(x): # 2x2 max pool img size halve
return tf.nn.max_pool(x, ksize=[1, 2, 2, 1],
strides=[1, 2, 2, 1], padding='SAME')
## 模型 ##
def model(X, w, w2, w3, w4, w_o, p_keep_conv, p_keep_hidden):
## 第一 ##
l1a = tf.nn.relu(conv2d(X, w)) # 卷积层 l1a shape=(?, 28, 28, 32) 图像大小 28*28 32个卷积核 outputs
l1 = max_pool_2x2(l1a) # 池化层 l1 shape=(?, 14, 14, 32) 图像大小 14*14 32 outputs
l1 = tf.nn.dropout(l1, p_keep_conv)# dropout 层 部分神经元激活
## 第二 ##
l2a = tf.nn.relu(conv2d(l1, w2)) # 卷积层 l2a shape=(?, 14, 14, 64) img size 14*14 64 outputs
l2 = max_pool_2x2(l2a) # 池化层 l2 shape=(?, 7, 7, 64) img size 7*7 64 outputs
l2 = tf.nn.dropout(l2, p_keep_conv)# dropout 层 部分神经元激活
## 第三 ##
l3a = tf.nn.relu(conv2d(l2, w3,)) # 卷积层 l3a shape=(?, 7, 7, 128) img size 7*7 128 outputs
l3 = max_pool_2x2(l3a) # 池化层 l3 shape=(?, 4, 4, 128) img size 4*4 128 outputs
l3 = tf.reshape(l3, [-1, w4.get_shape().as_list()[0]])#全连接 reshape to (?, 2048) 4*4*128 -> 1*2048
l3 = tf.nn.dropout(l3, p_keep_conv)# dropout 层 部分神经元激活
## 第四 ##
l4 = tf.nn.relu(tf.matmul(l3, w4)) # 1*2048 .* 2048*625 -> 1*625
l4 = tf.nn.dropout(l4, p_keep_hidden) # dropout 层 部分神经元激活
## 输出层 ##
pred_yx = tf.matmul(l4, w_o) # 输出层 1*625 .* 625*10 -> 1*10
return pred_yx
## 数据处理 data preprocessing
mnist = input_data.read_data_sets("../minist/MNIST_data/", one_hot=True)
# 训练数据 训练数据标签 测试数据 测试数据标签
trX, trY, teX, teY = mnist.train.images, mnist.train.labels, mnist.test.images, mnist.test.labels
# reshape成 卷积网络需要的 图片格式
trX = trX.reshape(-1, 28, 28, 1) # 训练数据 28x28x1 input img
teX = teX.reshape(-1, 28, 28, 1) # 测试数据 28x28x1 input img
## 训练数据的 占位符
X = tf.placeholder("float", [None, 28, 28, 1])
Y = tf.placeholder("float", [None, 10])
### 模型参数初始化 ####
w = init_weights([3, 3, 1, 32]) # 3x3x1 卷积 conv, 32 个卷积核outputs
w2 = init_weights([3, 3, 32, 64]) # 3x3x32 conv, 64 个卷积核outputs
w3 = init_weights([3, 3, 64, 128]) # 3x3x32 conv, 128 个卷积核outputs
w4 = init_weights([128 * 4 * 4, 625]) # 全连接层 FC 128 * 4 * 4 inputs, 625 outputs
w_o = init_weights([625, 10]) # FC 625 inputs, 10 outputs (labels)
## dropout参数配置 占位符 优化时再 调入
p_keep_conv = tf.placeholder("float")
p_keep_hidden = tf.placeholder("float")
## 模型
pred_yx = model(X, w, w2, w3, w4, w_o, p_keep_conv, p_keep_hidden)
## 计算代价函数 softmax回归 后 对数 信息增益 均值
cost = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(pred_yx, Y))
# 优化函数 全局学习速率 0.001
# http://blog.csdn.net/u014595019/article/details/52989301
#train_op = tf.train.RMSPropOptimizer(0.001, 0.9).minimize(cost)
train_op = tf.train.AdamOptimizer(1e-4).minimize(cost)
## 预测值
predict_op = tf.argmax(pred_yx, 1)
## 变量初始化
# init = tf.initialize_all_variables()#旧的
init =tf.global_variables_initializer()
## 启动 图 回话 Launch the graph in a session
with tf.Session() as sess:
# 初始化变量
#tf.initialize_all_variables().run() 旧版
#tf.global_variables_initializer().run()
sess.run(init)
for i in range(5):#训练5次
training_batch = zip(range(0, len(trX), batch_size), #range(start,end,scan):
range(batch_size, len(trX)+1, batch_size))
# start=(0 batch_size batch_size+batch_size...)
# end =(batch_size batch_size+batch_size batch_size+batch_size+batch_size ...)
# 开始训练
for start, end in training_batch:
sess.run(train_op, feed_dict={X: trX[start:end], Y: trY[start:end], #0:banch_size
p_keep_conv: 0.8, p_keep_hidden: 0.5})#加入 dropout参数
test_indices = np.arange(len(teX)) # Get A Test Batch -> array[1:len(teX)]
np.random.shuffle(test_indices) # disorganize the Test Batch test_indices
test_indices = test_indices[0:test_size] # 测试数据大小
print(i, np.mean(np.argmax(teY[test_indices], axis=1) ==
sess.run(predict_op, feed_dict={X: teX[test_indices],
Y: teY[test_indices],
p_keep_conv: 1.0,## 测试时关闭 dropout
p_keep_hidden: 1.0})))#计算每次训练的正确率
# np.random.random() -> 0~1.0
# np.random.uniform(a,b) -> a.0~b.0 / b.0 ~ a.0
# np.random.randint(a,b) -> a~b / b ~ a
# np.random.randrange(start, stop, step) == random.choice(range(start, stop, step)
# np.random.choice(sequence) -> random select one element
# np.random.shuffle(sequence) -> disorganize the sequence
# np.random.sample(sequence, k) -> random select k element
#