-
Notifications
You must be signed in to change notification settings - Fork 0
/
vgg_flower_17.py
141 lines (126 loc) · 5.88 KB
/
vgg_flower_17.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
#!/usr/bin/env python
# -*- coding: UTF-8 -*-
from tflearn.datasets import oxflower17
import tensorflow as tf
# mnist oxflower17 cifar10 cifar100
X,Y = oxflower17.load_data(one_hot = True)
print(X.shape) # 224,224,3
print(Y.shape) # 17
# 图像分类
# 0.读取数据
# 1.设计超参数
# （1）训练有关的参数
train_epoch = 1000 # 训练迭代次数（训练数据遍历完一次，表示一次迭代）
batch_size = 10 # 每个批次训练多少数据
learning_rate = tf.placeholder(tf.float32) # 准备设计一个动态的学习率
display_epoch = 100 # 每迭代100次，进行一次评估
n_class = Y.shape[1]
# （2）网络有关的参数
# 权重
weight = {
'wc1_1': tf.Variable(tf.random_normal([3,3,3,64])),
'wc2_1': tf.Variable(tf.random_normal([3,3,64,128])),
'wc3_1': tf.Variable(tf.random_normal([3,3,128,256])),
'wc3_2': tf.Variable(tf.random_normal([3,3,256,256])),
'wc4_1': tf.Variable(tf.random_normal([3,3,256,512])),
'wc4_2': tf.Variable(tf.random_normal([3,3,512,512])),
'wc5_1': tf.Variable(tf.random_normal([3,3,512,512])),
'wc5_2': tf.Variable(tf.random_normal([3,3,512,512])),
'wfc_1': tf.Variable(tf.random_normal([7*7*512,4096])),
'wfc_2': tf.Variable(tf.random_normal([4096,4096])),
'wfc_3': tf.Variable(tf.random_normal([4096,n_class])),
}
# 偏置量
biase = {
'bc1_1': tf.Variable(tf.random_normal([64])),
'bc2_1': tf.Variable(tf.random_normal([128])),
'bc3_1': tf.Variable(tf.random_normal([256])),
'bc3_2': tf.Variable(tf.random_normal([256])),
'bc4_1': tf.Variable(tf.random_normal([512])),
'bc4_2': tf.Variable(tf.random_normal([512])),
'bc5_1': tf.Variable(tf.random_normal([512])),
'bc5_2': tf.Variable(tf.random_normal([512])),
'bfc_1': tf.Variable(tf.random_normal([4096])),
'bfc_2': tf.Variable(tf.random_normal([4096])),
'bfc_3': tf.Variable(tf.random_normal([n_class])),
}
x = tf.placeholder(tf.float32, [None,224,224,3]) # 输入的数据
y = tf.placeholder(tf.float32, [None,n_class]) # 标签数据
# 2.设计网络
def vgg_network():
# input reshape 输入层的预处理
# conv_1
# filter 卷积核
# strides 步长
net = tf.nn.conv2d(input=x, filter=weight['wc1_1'], strides=[1,1,1,1], padding='VALID') #卷积
print("Conv", net)
net = tf.nn.leaky_relu(tf.nn.bias_add(net, biase['bc1_1'])) #激励
print("Relu", net)
net = tf.nn.lrn(net) # 局部归一化
print("LRN", net)
# ksize 窗口大小
# strides 步长
net= tf.nn.max_pool(value=net, ksize=[1,2,2,1],strides=[1,2,2,1], padding='VALID') # 池化
print("Max pooling", net)
# conv_2
net = tf.nn.conv2d(net, weight['wc2_1'], [1, 1, 1, 1], padding='VALID')
net = tf.nn.leaky_relu(tf.nn.bias_add(net, biase['bc2_1']))
net = tf.nn.max_pool(net, [1, 2, 2, 1], [1, 2, 2, 1])
# conv_3
net = tf.nn.conv2d(net, weight['wc3_1'], [1, 1, 1, 1], padding='VALID')
net = tf.nn.leaky_relu(tf.nn.bias_add(net, biase['bc3_1']))
net = tf.nn.conv2d(net, weight['wc3_2'], [1, 1, 1, 1], padding='VALID')
net = tf.nn.leaky_relu(tf.nn.bias_add(net, biase['bc3_2']))
net = tf.nn.max_pool(net, [1, 2, 2, 1], [1, 2, 2, 1], padding='VALID')
# conv_4
net = tf.nn.conv2d(net, weight['wc4_1'], [1, 1, 1, 1], padding='VALID')
net = tf.nn.leaky_relu(tf.nn.bias_add(net, biase['bc4_1']))
net = tf.nn.conv2d(net, weight['wc4_2'], [1, 1, 1, 1], padding='VALID')
net = tf.nn.leaky_relu(tf.nn.bias_add(net, biase['bc4_2']))
net = tf.nn.max_pool(net, [1, 2, 2, 1], [1, 2, 2, 1], padding='VALID')
# conv_5
net = tf.nn.conv2d(net, weight['wc5_1'], [1, 1, 1, 1], padding='VALID')
net = tf.nn.leaky_relu(tf.nn.bias_add(net, biase['bc5_1']))
net = tf.nn.conv2d(net, weight['wc5_2'], [1, 1, 1, 1], padding='VALID')
net = tf.nn.leaky_relu(tf.nn.bias_add(net, biase['bc5_2']))
net = tf.nn.max_pool(net, [1, 2, 2, 1], [1, 2, 2, 1], padding='VALID')
# 拉伸
net = tf.reshape(net,shape=[-1, weight['wfc_1'].get_shape().to_list()[0]])
# 全连接（BP网络）
# fc1
net = tf.nn.relu(tf.matmul(net, weight['wfc_1']) + biase['bfc_1'])
# fc2
net = tf.nn.relu(tf.matmul(net, weight['wfc_2']) + biase['bfc_2'])
# out
# return tf.nn.softmax(tf.matmul(net, weight['wfc_3']) + biase['bfc_3'])
return tf.matmul(net, weight['wfc_3']) + biase['bfc_3']
# 3.设计损失函数和优化器，建立评估函数
pred = vgg_network()
loss = tf.nn.softmax_cross_entropy_with_logits_v2(logits=pred, labels=y) #损失函数
# 用于sigmoid
# tf.nn.sigmoid_cross_entropy_with_logits()
# 优化器由分类器和损失函数决定
opt = tf.train.GradientDescentOptimizer(learning_rate=learning_rate).minimize(loss) #优化器
# one_hot 17位
# [True,False,True,...] X
# True
acc_tf = tf.equal(tf.argmax(pred,1),tf.argmax(y,1))
acc = tf.reduce_mean(tf.cast(acc_tf,tf.float32))
# 4.训练
with tf.Session() as sess:
sess.run(tf.global_variables_initializer()) # 初始化全局变量
base_lr = 0.01
for epoch in range(train_epoch):
# batch_size=10,total_data=1360 => total_batch
# 取数据
total_batch = X.shape[0]//batch_size
for i in range(total_batch):
X_train,Y_train = X[i*batch_size:i*batch_size+batch_size],Y[i*batch_size:i*batch_size+batch_size]
# learning_rate = batch_size*((1-epoch)/train_epoch)**2
sess.run(opt,feed_dict={x:X_train,y:Y_train,learning_rate:base_lr})
if (i+1)*(epoch+1)%display_epoch ==0:
# 进行评估
cost, accuaray = sess.run(opt,feed_dict={x:X_train,y:Y_train,learning_rate:base_lr})
print("Step: %d, loss: %d, acc: %d" % (str(epoch)+"-"+str(i),cost,accuaray))
# 动态修改学习率
learning_rate = base_lr * ((1 - epoch) / train_epoch) ** 2