机器学习进阶笔记之六 | 深入理解Fast Neural Style

Transform Network网络结构

1. import tensorflow as tf
2. 
   
3. def conv2d(x, input_filters, output_filters, kernel, strides, padding=’SAME’):
4. with tf.variable_scope('conv') as scope:
5. 
   
6.     shape = [kernel, kernel, input_filters, output_filters]
7.     weight = tf.Variable(tf.truncated_normal(shape, stddev=0.1), name='weight')
8.     convolved = tf.nn.conv2d(x, weight, strides=[1, strides, strides, 1], padding=padding, name='conv')
9. 
   
10.     normalized = batch_norm(convolved, output_filters)
11. 
    
12.     return normalized
13. 
    
14. def conv2d_transpose(x, input_filters, output_filters, kernel, strides, padding=’SAME’):
15. with tf.variable_scope('conv_transpose') as scope:
16. 
    
17.     shape = [kernel, kernel, output_filters, input_filters]
18.     weight = tf.Variable(tf.truncated_norm
19. al(shape, stddev=0.1), name='weight')
20. 
    
21.     batch_size = tf.shape(x)[0]
22.     height = tf.shape(x)[1] * strides
23.     width = tf.shape(x)[2] * strides
24.     output_shape = tf.pack([batch_size, height, width, output_filters])
25.     convolved = tf.nn.conv2d_transpose(x, weight, output_shape, strides=[1, strides, strides, 1], padding=padding, name='conv_transpose')
26. 
    
27.     normalized = batch_norm(convolved, output_filters)
28.     return normalized
29. 
    
30. def batch_norm(x, size):
31. 
    
32. batch_mean, batch_var = tf.nn.moments(x, [0, 1, 2], keep_dims=True)
33. beta = tf.Variable(tf.zeros([size]), name='beta')
34. scale = tf.Variable(tf.ones([size]), name='scale')
35. epsilon = 1e-3
36. return tf.nn.batch_normalization(x, batch_mean, batch_var, beta, scale, epsilon, name='batch')
37. 
    
38. def residual(x, filters, kernel, strides, padding=’SAME’):
39. with tf.variable_scope('residual') as scope:
40.   
41.     conv1 = conv2d(x, filters, filters, kernel, strides, padding=padding)
42.     conv2 = conv2d(tf.nn.relu(conv1), filters, filters, kernel, strides, padding=padding)
43. 
    
44.     residual = x + conv2
45. 
    
46.     return residual
47. 
    
48. def net(image):
49. 
    
50. with tf.variable_scope('conv1'):
51.     conv1 = tf.nn.relu(conv2d(image, 3, 32, 9, 1))
52. with tf.variable_scope('conv2'):
53.     conv2 = tf.nn.relu(conv2d(conv1, 32, 64, 3, 2))
54. with tf.variable_scope('conv3'):
55.     conv3 = tf.nn.relu(conv2d(conv2, 64, 128, 3, 2))
56. with tf.variable_scope('res1'):
57.     res1 = residual(conv3, 128, 3, 1)
58. with tf.variable_scope('res2'):
59.     res2 = residual(res1, 128, 3, 1)
60. with tf.variable_scope('res3'):
61.     res3 = residual(res2, 128, 3, 1)
62. with tf.variable_scope('res4'):
63.     res4 = residual(res3, 128, 3, 1)
64. with tf.variable_scope('res5'):
65.     res5 = residual(res4, 128, 3, 1)
66. with tf.variable_scope('deconv1'):
67.     deconv1 = tf.nn.relu(conv2d_transpose(res5, 128, 64, 3, 2))
68. with tf.variable_scope('deconv2'):
69.     deconv2 = tf.nn.relu(conv2d_transpose(deconv1, 64, 32, 3, 2))
70. with tf.variable_scope('deconv3'):
71.     deconv3 = tf.nn.tanh(conv2d_transpose(deconv2, 32, 3, 9, 1))
72. 
    
73. y = deconv3 * 127.5
74. 
    
75. return y

       使用deep residual network来训练COCO数据集，能够在保证性能的前提下，训练更深的模型。 而Loss Network是有pretrained的VGG网络来计算，网络结构：

1.  import tensorflow as tf
2.  import numpy as np
3.  import scipy.io
4.  from scipy import misc
5. 
   
6. 
   
7.  def net(data_path, input_image):
8.      layers = (
9.          'conv1_1', 'relu1_1', 'conv1_2', 'relu1_2', 'pool1',
10. 
    
11.          'conv2_1', 'relu2_1', 'conv2_2', 'relu2_2', 'pool2',
12. 
    
13.          'conv3_1', 'relu3_1', 'conv3_2', 'relu3_2', 'conv3_3',
14.          'relu3_3', 'conv3_4', 'relu3_4', 'pool3',
15. 
    
16.          'conv4_1', 'relu4_1', 'conv4_2', 'relu4_2', 'conv4_3',
17.          'relu4_3', 'conv4_4', 'relu4_4', 'pool4',
18. 
    
19.          'conv5_1', 'relu5_1', 'conv5_2', 'relu5_2', 'conv5_3',
20.          'relu5_3', 'conv5_4', 'relu5_4'
21.      )
22. 
    
23.      data = scipy.io.loadmat(data_path)
24.      mean = data['normalization'][0][0][0]
25.      mean_pixel = np.mean(mean, axis=(0, 1))
26.      weights = data['layers'][0]
27. 
    
28.      net = {}
29.      current = input_image
30.      for i, name in enumerate(layers):
31.          kind = name[:4]
32.          if kind == 'conv':
33.              kernels, bias = weights[i][0][0][0][0]
34.              # matconvnet: weights are [width, height, in_channels, out_channels]
35.              # tensorflow: weights are [height, width, in_channels, out_channels]
36.              kernels = np.transpose(kernels, (1, 0, 2, 3))
37.              bias = bias.reshape(-1)
38.              current = _conv_layer(current, kernels, bias, name=name)
39.          elif kind == 'relu':
40.              current = tf.nn.relu(current, name=name)
41.          elif kind == 'pool':
42.              current = _pool_layer(current, name=name)
43.          net[name] = current
44. 
    
45.      assert len(net) == len(layers)
46.      return net, mean_pixel
47. 
    
48. 
    
49.  def _conv_layer(input, weights, bias, name=None):
50.      conv = tf.nn.conv2d(input, tf.constant(weights), strides=(1, 1, 1, 1),
51.                          padding='SAME', name=name)
52.      return tf.nn.bias_add(conv, bias)
53. 
    
54. 
    
55.  def _pool_layer(input, name=None):
56.      return tf.nn.max_pool(input, ksize=(1, 2, 2, 1), strides=(1, 2, 2, 1),
57.                            padding='SAME', name=name)
58. 
    
59. 
    
60.  def preprocess(image, mean_pixel):
61.      return image - mean_pixel
62. 
    
63. 
    
64.  def unprocess(image, mean_pixel):
65.      return image + mean_pixel