【深度学习】VGGNet的tensorflow实现code

在上一篇博客中对VGGNet论文做了学习笔记以及网络结构记录，链接见此http://blog.csdn.net/qq_29340857/article/details/71440674，本篇博客的代码都是基于上一篇的理解写的，话不多说，code如下。

---

输入数据集定义（因为使用随机梯度下降，所以使用placeholder定义训练数据）：

  tf_train_data=tf.placeholder(tf.float32,shape=(batch_size,img_size,img_size,3))
  tf_train_label=tf.placeholder(tf.float32,shape=(batch_size,img_num))
  tf_valid_data=tf.constant(valid_data)
  tf_test_data=tf.constant(test_data)

参数初始化（不算pool层，fc层+conv层一共是16层神经网络）：

  w={
  'w1':tf.get_variable('w1',[3,3,3,64],initializer=tf.contrib.layers.xavier_initializer_conv2d()),
  'w2':tf.get_variable('w2',[3,3,64,64],initializer=tf.contrib.layers.xavier_initializer_conv2d()),
  'w3':tf.get_variable('w3',[3,3,64,128],initializer=tf.contrib.layers.xavier_initializer_conv2d()),
  'w4':tf.get_variable('w4',[3,3,128,128],initializer=tf.contrib.layers.xavier_initializer_conv2d()),
  'w5':tf.get_variable('w5',[3,3,128,256],initializer=tf.contrib.layers.xavier_initializer_conv2d()),
  'w6':tf.get_variable('w6',[3,3,256,256],initializer=tf.contrib.layers.xavier_initializer_conv2d()),
  'w7':tf.get_variable('w7',[3,3,256,256],initializer=tf.contrib.layers.xavier_initializer_conv2d()),
  'w8':tf.get_variable('w8',[3,3,256,512],initializer=tf.contrib.layers.xavier_initializer_conv2d()),
  'w9':tf.get_variable('w9',[3,3,512,512],initializer=tf.contrib.layers.xavier_initializer_conv2d()),
  'w10':tf.get_variable('w10',[3,3,512,512],initializer=tf.contrib.layers.xavier_initializer_conv2d()),
  'w11':tf.get_variable('w11',[3,3,512,512],initializer=tf.contrib.layers.xavier_initializer_conv2d()),
  'w12':tf.get_variable('w12',[3,3,512,512],initializer=tf.contrib.layers.xavier_initializer_conv2d()),
  'w13':tf.get_variable('w13',[3,3,512,512],initializer=tf.contrib.layers.xavier_initializer_conv2d()),
  'w14':tf.Variable(tf.random_normal([img_size/32*img_size/32*512,4096])),
  'w15':tf.Variable(tf.random_normal([4096,4096])),
  'w16':tf.Variable(tf.random_normal([4096,1000])),
  }

  b={
  'b1':tf.Variable(tf.zeros([64])),
  'b2':tf.Variable(tf.zeros([64])),
  'b3':tf.Variable(tf.zeros([128])),
  'b4':tf.Variable(tf.zeros([128])),
  'b5':tf.Variable(tf.zeros([256])),
  'b6':tf.Variable(tf.zeros([256])),
  'b7':tf.Variable(tf.zeros([256])),
  'b8':tf.Variable(tf.zeros([512])),
  'b9':tf.Variable(tf.zeros([512])),
  'b10':tf.Variable(tf.zeros([512])),
  'b11':tf.Variable(tf.zeros([512])),
  'b12':tf.Variable(tf.zeros([512])),
  'b13':tf.Variable(tf.zeros([512])),
  'b14':tf.Variable(tf.zeros([4096])),
  'b15':tf.Variable(tf.zeros([4096])),
  'b16':tf.Variable(tf.zeros([1000])),
  }

论文原文中提到两种参数初始化方法：
1. pre-training：训练A结构网络，直接把其前四层conv的参数和fc的前两层参数直接赋给D结构网络，就是本文代码实现的
2. no pre-training:利用Xaiver Initialization初始化参数
在此采用Xavier Initialization。

网络结构搭建：

  #13层卷积层+5层max_pool+3层FC层
  def model(input_data):
    conv1=tf.nn.conv2d(input_data,w['w1'],[1,1,1,3],padding="SAME")#img:224*224*64
    h1=tf.nn.relu(conv1+b['b1'])
    conv2=tf.nn.conv2d(h1,w['w2'],[1,1,1,64],padding="SAME")
    h2=tf.nn.relu(conv2+b['b2'])  

    max1=tf.nn.max_pool(h2,[2,2,64,64],[1,2,2,64],padding="VALID")

    conv3=tf.nn.conv2d(max1,w['w3'],[1,1,1,64],padding="SAME")#img:112*112*128
    h3=tf.nn.relu(conv3+b['b3'])
    conv4=tf.nn.conv2d(h3,w['w4'],[1,1,1,128],padding="SAME")
    h4=tf.nn.relu(conv4+b['b4'])

    max2=tf.nn.max_pool(h4,[2,2,128,128],[1,2,2,128],padding="VALID")

    conv5=tf.nn.conv2d(max2,w['w5'],[1,1,1,128],padding="SAME")#img:56*56*256
    h5=tf.nn.relu(conv5+b['b5'])
    conv6=tf.nn.conv2d(h5,w['w6'],[1,1,1,256],padding="SAME")
    h6=tf.nn.relu(conv6+b['b6'])
    conv7=tf.nn.conv2d(h6,w['w7'],[1,1,1,256],padding="SAME")
    h7=tf.nn.relu(conv7)+b['b7']

    max3=tf.nn.max_pool(h7,[2,2,256,256],[1,2,2,256],padding="VALID")

    conv8=tf.nn.conv2d(max3,w['w8'],[1,1,1,256],padding="SAME")#img:28*28*512
    h8=tf.nn.relu(conv8+b['b8'])
    conv9=tf.nn.conv2d(h8,w['w9'],[1,1,1,512],padding="SAME")
    h9=tf.nn.relu(conv9+b['b9'])
    conv10=tf.nn.conv2d(h9,w['w10'],[1,1,1,512],padding="SAME")
    h10=tf.nn.relu(conv10+b['b10'])

    max4=tf.nn.max_pool(h10,[2,2,512,512],[1,2,2,512],padding="VALID")

    conv11=tf.nn.conv2d(max4,w['w11'],[1,1,1,512],padding="SAME")#img:14*14*512
    h11=tf.nn.relu(conv11+b['b11'])
    conv12=tf.nn.conv2d(h11,w['w12'],[1,1,1,512],padding="SAME")
    h12=tf.nn.relu(conv12+b['b12'])
    conv13=tf.nn.conv2d(h12,w['w13'],[1,1,1,512],padding="SAME")
    h13=tf.nn.relu(conv13+b['b13'])

    max5=tf.nn.max_pool(h13,[2,2,512,512],[1,2,2,512],padding="VALID")

    shapes=max5.get_shape().as_list()#img:7*7*512
    reshape=tf.reshape(max5,[shapes[0],shapes[1]*shapes[2]*shapes[3]])
    fc1=tf.matmul(reshape,w['w14'])+b['b14']
    h14=tf.nn.dropout(tf.nn.relu(fc1),drop_param)
    fc2=tf.matmul(h14,w['w15'])+b['b15']
    h15=tf.nn.dropout(tf.nn.relu(fc2),drop_param)
    fc3=tf.matmul(h15,w['w16'])+b['b16']
    return fc3

L2正则化：

  l2_loss=None
  for i in np.arange(1,17):
    k="w"+str(i)
    l2_loss+=l2_param*tf.nn.l2_loss(w[k])
  loss=tf.reduce_mean(tf.softmax_cross_entropy_with_logits(logits,tf_train_label))+l2_loss

Momentum Optimizer更新参数：

  optimizer=tf.train.MomentumOptimizer(learning_rate,momentum_param)

对Momentum Optimizer不了解的小伙伴和想了解更多参数更新方法的小伙伴，可以参见我之前写的博客http://blog.csdn.net/qq_29340857/article/details/71353751

train,validation,test数据集预测：

  train_predictions=tf.nn.softmax(logits)
  valid_predictions=tf.nn.softmax(model(tf_valid_data))
  test_predictions=tf.nn.softmax(model(tf_test_data))

至此，VGGNet的搭建工作差不多完成，在下一篇博客中会记录实际训练时遇到的一些问题以及解决方案，欢迎继续关注！