tensorflow卷积层可视化的实现。基于VGG-19.

1. #导入包

2. import scipy.io

3. import numpy as np

4. import os

5. import scipy.misc

6. import matplotlib.pyplot as plt

7. import tensorflow as tf

8. #import seaborn as sns #用来画热度图

9.  

10. #定义网络结构，本文用的是VGG-19

11. def _conv_layer(input, weights, bias):

12. conv = tf.nn.conv2d(input, tf.constant(weights), strides=(1, 1, 1, 1),

13. padding='SAME')

14. return tf.nn.bias_add(conv, bias)

15. def _pool_layer(input):

16. return tf.nn.max_pool(input, ksize=(1, 2, 2, 1), strides=(1, 2, 2, 1),

17. padding='SAME')

18. def preprocess(image, mean_pixel):

19. return image - mean_pixel

20. def unprocess(image, mean_pixel):

21. return image + mean_pixel

22. def imread(path):

23. return scipy.misc.imread(path).astype(np.float)

24. def imsave(path, img):

25. img = np.clip(img, 0, 255).astype(np.uint8)

26. scipy.misc.imsave(path, img)

27. print ("Functions for VGG ready")

28.  

29. def net(data_path, input_image):

30. layers = (

31. 'conv1_1', 'relu1_1', 'conv1_2', 'relu1_2', 'pool1',

32. 'conv2_1', 'relu2_1', 'conv2_2', 'relu2_2', 'pool2',

33. 'conv3_1', 'relu3_1', 'conv3_2', 'relu3_2', 'conv3_3',

34. 'relu3_3', 'conv3_4', 'relu3_4', 'pool3',

35. 'conv4_1', 'relu4_1', 'conv4_2', 'relu4_2', 'conv4_3',

36. 'relu4_3', 'conv4_4', 'relu4_4', 'pool4',

37. 'conv5_1', 'relu5_1', 'conv5_2', 'relu5_2', 'conv5_3',

38. 'relu5_3', 'conv5_4', 'relu5_4'

39. )

40. data = scipy.io.loadmat(data_path)

41. mean = data['normalization'][0][0][0]

42. mean_pixel = np.mean(mean, axis=(0, 1))

43. weights = data['layers'][0]

44. net = {}

45. current = input_image

46. for i, name in enumerate(layers):

47. kind = name[:4]

48. if kind == 'conv':

49. kernels, bias = weights[i][0][0][0][0]

50. # matconvnet: weights are [width, height, in_channels, out_channels]

51. # tensorflow: weights are [height, width, in_channels, out_channels]

52. kernels = np.transpose(kernels, (1, 0, 2, 3))

53. bias = bias.reshape(-1)

54. current = _conv_layer(current, kernels, bias)

55. elif kind == 'relu':

56. current = tf.nn.relu(current)

57. elif kind == 'pool':

58. current = _pool_layer(current)

59. net[name] = current

60. assert len(net) == len(layers)

61. return net, mean_pixel, layers

62. print ("Network for VGG ready")

63.  

64. #展示VGG-19每一层的特征图。一共有36层，就有36张特征图。

65. cwd = os.getcwd()

66. VGG_PATH = cwd+"\\imagenet-vgg-verydeep-19.mat" #导入数据，初始化网络。

67. IMG_PATH = cwd+"\\123.jpg" #导入图片

68. input_image = imread(IMG_PATH)

69. shape = (1,input_image.shape[0],input_image.shape[1],input_image.shape[2])

70. with tf.Session() as sess:

71. image = tf.placeholder('float', shape=shape)

72. nets, mean_pixel, all_layers = net(VGG_PATH, image)

73. input_image_pre = np.array([preprocess(input_image, mean_pixel)])

74. layers = all_layers # For all layers

75. # layers = ('conv1_1', 'conv2_2','conv3_2','conv5_2')

76. for i, layer in enumerate(layers):

77. print ("[%d/%d] %s" % (i+1, len(layers), layer))

78. features = nets[layer].eval(feed_dict={image: input_image_pre})

79.  

80. print (" Type of 'features' is ", type(features))

81. print (" Shape of 'features' is %s" % (features.shape,))

82. # Plot response

83. if 1:

84. plt.figure(i+1, figsize=(10, 5))

85. #heatmap = sns.heatmap(features[0, :, :, 0]) #此处显示的是热度图。

86. plt.matshow(features[0, :, :, 0], fignum=i+1,cmap=plt.cm.gray) #显示灰度图

87. plt.title("" + layer)

88. plt.colorbar()

89. plt.show()