VGG16提取图像特征 （torch7）

VGG16提取图像特征 （torch7）

VGG16

loadcaffe

torch7

1. 下载pretrained model,保存到当前目录下

1. th> caffemodel_url = 'http://www.robots.ox.ac.uk/~vgg/software/very_deep/caffe/VGG_ILSVRC_16_layers.caffemodel' 

   ---

2. th> proto_url='https://gist.github.com/ksimonyan/211839e770f7b538e2d8#file-vgg_ilsvrc_16_layers_deploy-prototxt' 

   ---

3. th> os.execute('wget VGG_ILSVRC_16_layers.caffemodel' .. caffemodel_url) 

   ---

4. th> os.execute('wget VGG_ILSVRC_16_layers_deploy.prototxt' .. proto_url) 

   ---

2. 使用loadcaffe提取图像特征

1.  

   ---

2. require 'torch' -- 使用th命令，可忽略 

   ---

3. require 'nn' -- 修改model用到nn包 

   ---

4. require 'loadcaffe' -- 加在caffe训练的包 

   ---

5. require 'image' -- 加载图像，处理图像，可以使用cv中函数替代 

   ---

6.  

   ---

7. local loadSize = {3,256,256} -- 加载图像scale尺寸 

   ---

8. local sampleSize = {3,224,224} -- 样本尺寸，其实就是选取scale后图像的中间一块 

   ---

9.  

   ---

10. local function loadImage(input) 

    ---

11. -- 将图像缩放到loadSize尺寸，为了保证aspect ratio不变，将短边缩放后，长边按比例缩放 

    ---

12. if input:size(3) < input:size(2) then 

    ---

13. input = image.scale(input,loadSize[2],loadSize[3]*input:size(2)/input:size(3)) 

    ---

14. -- 注意image.scale(src,width,height),width对应的是input:size[3],height对应的是input:size[2] 

    ---

15. else 

    ---

16. input = image.scale(input,loadSize[2]*input:size(3)/input:size(2),loadSize[3]) 

    ---

17. end 

    ---

18. return input 

    ---

19. end 

    ---

20.  

    ---

21. local bgr_means = {103.939,116.779,123.68} --VGG预训练中的均值 

    ---

22. local function vggPreProcessing(img) 

    ---

23. local img2=img:clone() 

    ---

24. img2[{{1}}] =img2[{{3}}] -- image.load 加载图像是rgb格式，需转化维bgr 

    ---

25. img2[{{3}}] = img[{{1}}] 

    ---

26. img2:mul(255) -- image.load()加载的图像 pixel value \in [0,1] 

    ---

27. for i=1,3 do 

    ---

28. img2[i]:add(-bgr_means[i]) -- 中心化 

    ---

29. end 

    ---

30. return img2 

    ---

31. end 

    ---

32.  

    ---

33. local function centerCrop(input) 

    ---

34. local oH = sampleSize[2] 

    ---

35. local oW = sampleSize[3] 

    ---

36. local iW = input:size(3) 

    ---

37. local iH = input:size(2) 

    ---

38. local w1 = math.ceil((iW-oW)/2) 

    ---

39. local h1 = math.ceil((iH-oH)/2) 

    ---

40. local out = image.crop(input,w1,h1,w1+oW,h1+oH) 

    ---

41. return out 

    ---

42. end 

    ---

43.  

    ---

44. local function getPretrainedModel() 

    ---

45. local proto = 'VGG_ILSVRC_16_layers_deploy.prototxt' 

    ---

46. local caffemodel = '/home/zwzhou/modelZoo/VGG_ILSVRC_16_layers.caffemodel' 

    ---

47.  

    ---

48. local model = loadcaffe.load(proto,caffemodel,'nn') -- 加载pretrained model 

    ---

49. for i=1,3 do -- 将最后3层舍掉 

    ---

50. model.modules[#model.modules]=nil 

    ---

51. end 

    ---

52. -- 删除pretrained model的一些层官方方法 

    ---

53. -- ========================== 

    ---

54. -- for i= 40,38,-1 do 

    ---

55. -- model:remove(i) 

    ---

56. -- end 

    ---

57. -- ========================== 

    ---

58. model:add(nn.Normalize(2)) -- 添加一层正则化层，将输出向量归一化 

    ---

59.  

    ---

60. model:evaluate() -- self.training=false ，非训练，让网络参数不变 

    ---

61. return model 

    ---

62. end 

    ---

63.  

    ---

64. torch.setdefaulttensortype('torch.FloatTensor') 

    ---

65. model = getPretrainedModel() 

    ---

66.  

    ---

67. filepath = '/home/zwzhou/MOT16/train/MOT16-02/img1/000001.jpg' 

    ---

68. local img1=image.load(filepath) -- rgb图像 

    ---

69. local input = image.crop(img1,910,480,910+97,480+110) -- 里面参数时选择原图像的一个区域，boundingbox 

    ---

70.  

    ---

71. input = loadImage(input) 

    ---

72. local vggPreProcessed = vggPreProcessing(input) 

    ---

73. local out = centerCrop(vggPreProcessed) 

    ---

74.  

    ---

75. local outputs = model:forward(out) 

    ---

76.  

    ---

77. print(outputs) 

    ---

78. print(#outputs) 

    ---

3. 参考项目
   VGG-19-feature-extractor
   Torch 7 利用已有VGG模型提取图片特征