增加基于tensorflow的VGG16网络搭建

ch12_网络搭建及训练/第十二章_网络搭建及训练.md

@@ -293,6 +293,77 @@ input_label=tf.placeholder(tf.float32,[None,2],"input_label")
 
   shape为[None,2]，说明数据第一个维度是不确定的，然后TensorFlow会根据我们传递的数据动态推断第一个维度，这样我们就可以在运行时改变batch的大小。比如一个数据是2维，一次传递10个数据对应的tensor的shape就是[10,2]。可不可以把多个维度指定为None？理论上不可以！
 
+## 12.1.6 如何基于tensorflow搭建VGG16
+
+​	介绍完关于tensorflow的基础知识，是时候来一波网络搭建实战了。虽然网上有很多相关教程，但我想从最标准的tensorflow代码和语法出发（而不是调用更高级的API，失去了原来的味道），向大家展示如何搭建其标准的VGG16网络架构。话不多说，上代码：
+
+```python
+import numpy as np
+import tensorflow as tf
+
+def get_weight_variable(shape):
+    return tf.get_variable('weight', shape=shape, initializer=tf.truncated_normal_initializer(stddev=0.1))
+
+def get_bias_variable(shape):
+    return tf.get_variable('bias', shape=shape, initializer=tf.constant_initializer(0))
+
+def conv2d(x, w, padding = 'SAME', s=1):
+    x = tf.nn.conv2d(x, w, strides=[1, s, s, 1], padding = padding)
+    return x
+
+def maxPoolLayer(x):
+    return tf.nn.max_pool(x, ksize = [1, 2, 2, 1],
+                          strides = [1, 2, 2, 1], padding = 'SAME')
+
+def conv2d_layer(x,in_chs, out_chs, ksize, layer_name):
+    with tf.variable_scope(layer_name):
+        w = get_weight_variable([ksize, ksize, in_chs, out_chs])
+        b = get_bias_variable([out_chs])
+        y = tf.nn.relu(tf.bias_add(conv2d(x,w,padding = 'SAME', s=1), b))
+    return y
+
+def fc_layer(x,in_kernels, out_kernels, layer_name):
+    with tf.variable_scope(layer_name):
+        w = get_weight_variable([in_kernels,out_kernels])
+        b = get_bias_variable([out_kernels])
+        y = tf.nn.relu(tf.bias_add(tf.matmul(x,w),b))
+    return y
+        
+def VGG16(x):
+    conv1_1 = conv2d_layer(x,tf.get_shape(x).as_list()[-1], 64, 3, 'conv1_1')
+    conv1_2 = conv2d_layer(conv1_1,64, 64, 3, 'conv1_2')
+    pool_1 = maxPoolLayer(conv1_2)
+    
+    conv2_1 = conv2d_layer(pool1,64, 128, 3, 'conv2_1')
+    conv2_2 = conv2d_layer(conv2_1,128, 128, 3, 'conv2_2')
+    pool2 = maxPoolLayer(conv2_2)
+    
+	conv3_1 = conv2d_layer(pool2,128, 256, 3, 'conv3_1')
+    conv3_2 = conv2d_layer(conv3_1,256, 256, 3, 'conv3_2')
+	conv3_3 = conv2d_layer(conv3_2,256, 256, 3, 'conv3_3')
+    pool3 = maxPoolLayer(conv3_3)
+    
+	conv4_1 = conv2d_layer(pool3,256, 512, 3, 'conv4_1')
+    conv4_2 = conv2d_layer(conv4_1,512, 512, 3, 'conv4_2')
+	conv4_3 = conv2d_layer(conv4_2,512, 512, 3, 'conv4_3')
+    pool4 = maxPoolLayer(conv4_3)
+    
+	conv5_1 = conv2d_layer(pool4,512, 512, 3, 'conv5_1')
+    conv5_2 = conv2d_layer(conv5_1,512, 512, 3, 'conv5_2')
+	conv5_3 = conv2d_layer(conv5_1,512, 512, 3, 'conv5_3')
+    pool5 = maxPoolLayer(conv5_3)
+    
+	pool5_flatten_dims = int(np.prod(pool5.get_shape().as_list()[1:]))
+    pool5_flatten = tf.reshape(pool5,[-1,pool5_flatten_dims])
+    
+    fc_6 = fc_layer(pool5_flatten, pool5_flatten_dims, 4096, 'fc6')
+	fc_7 = fc_layer(fc_6, 4096, 4096, 'fc7')
+	fc_8 = fc_layer(fc_7, 4096, 10, 'fc8')
+    
+    return fc_8
+    
+```
+
 
 
 # 12.2 Pytorch