CV之IG：基于CNN网络架构+ResNet网络进行DIY图像生成网络

设计思路

实现代码

# 定义图像生成网络：image, training，两个参数

   # Less border effects when padding a little before passing through ..

   image = tf.pad(image, [[0, 0], [10, 10], [10, 10], [0, 0]], mode='REFLECT')

   with tf.variable_scope('conv1'):

       conv1 = relu(instance_norm(conv2d(image, 3, 32, 9, 1)))

   with tf.variable_scope('conv2'):

       conv2 = relu(instance_norm(conv2d(conv1, 32, 64, 3, 2)))

   with tf.variable_scope('conv3'):

       conv3 = relu(instance_norm(conv2d(conv2, 64, 128, 3, 2)))

   with tf.variable_scope('res1'):

       res1 = residual(conv3, 128, 3, 1)

   with tf.variable_scope('res2'):

       res2 = residual(res1, 128, 3, 1)

   with tf.variable_scope('res3'):

       res3 = residual(res2, 128, 3, 1)

   with tf.variable_scope('res4'):

       res4 = residual(res3, 128, 3, 1)

   with tf.variable_scope('res5'):

       res5 = residual(res4, 128, 3, 1)

   # print(res5.get_shape())

   with tf.variable_scope('deconv1'):

       # deconv1 = relu(instance_norm(conv2d_transpose(res5, 128, 64, 3, 2)))

       deconv1 = relu(instance_norm(resize_conv2d(res5, 128, 64, 3, 2, training)))

   with tf.variable_scope('deconv2'):

       # deconv2 = relu(instance_norm(conv2d_transpose(deconv1, 64, 32, 3, 2)))

       deconv2 = relu(instance_norm(resize_conv2d(deconv1, 64, 32, 3, 2, training)))

   with tf.variable_scope('deconv3'):

       # deconv_test = relu(instance_norm(conv2d(deconv2, 32, 32, 2, 1)))

       deconv3 = tf.nn.tanh(instance_norm(conv2d(deconv2, 32, 3, 9, 1)))

   y = (deconv3 + 1) * 127.5

   height = tf.shape(y)[1]

   width = tf.shape(y)[2]

   y = tf.slice(y, [0, 10, 10, 0], tf.stack([-1, height - 20, width - 20, -1]))

   return y