pytorch转caffe成功之后的验证

东方耀

pytorch转caffe成功之后的验证

思路：用同一张图片作为输入，经过相同的预处理
各自模型的前向计算，比较输出的结果是否相同？

1. 
   
2. import torch
   
3. import numpy as np
   
4. from config import get_config_lite
   
5. from Learner import face_learner
   
6. import torch.nn.functional as F
   
7. from torch.nn import CrossEntropyLoss
   
8. from torchvision import transforms as trans
   
9. import os
   
10. import cv2
    
11. 
    
12. os.environ["CUDA_DEVICE_ORDER"] = "PCI_BUS_ID"
    
13. # 选择哪一块gpu,如果是-1，就是调用cpu
    
14. os.environ['CUDA_VISIBLE_DEVICES'] = "0"
    
15. 
    
16. 
    
17. # 看同一张图片 输入后  torch与caffe的输出的512维度向量
    
18. img_path = "/home/jingyun/py2_caffe_face_works/onnx2caffe-master/model/10.jpg"
    
19. 
    
20. # conf = get_config_dfy(training=False)
    
21. conf = get_config_lite(training=False)
    
22. # conf = get_config_dfy(training=False)
    
23. print("easydict的配置参数：", conf)
    
24. learner = face_learner(conf, inference=True)
    
25. 
    
26. fixed_str = "2020-08-09-11-00_accuracy:0.9094341157427415_step:1324780_final.pth"
    
27. learner.load_state(conf, fixed_str, model_only=True, from_save_folder=True)
    
28. learner.model.eval()
    
29. 
    
30. test_transform = trans.Compose([
    
31.     trans.ToTensor(),
    
32.     trans.Normalize(mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5])
    
33. ])
    
34. 
    
35. 
    
36. img = cv2.imread(img_path)
    
37. print("cv2.imread之后：", img.shape)
    
38. # 图片的预处理
    
39. img = test_transform(img)
    
40. 
    
41. 
    
42. print("torch预处理后：", img.size())
    
43. img = torch.unsqueeze(img, dim=0)
    
44. print("torch 扩展维度后：", img.size(), type(img))
    
45. 
    
46. 
    
47. with torch.no_grad():
    
48.     result = learner.model(img.to(conf.device))
    
49.     print("result_512:", result.size())
    
50.     result = result.cpu().numpy()[0]
    
51.     print(type(result), result.shape, result)
    
52.     # result是否归一化了： 25.143478
    
53.     print("result是否归一化了：", np.linalg.norm(result))
    
54.     # numpy l2_norm()的实现
    
55.     result = np.divide(result, np.linalg.norm(result))
    
56.     print(result)
    
57. 
    
58. # torch的结果：
    
59. #  4.18923460e-02 -4.78874184e-02 -2.89320014e-02 -5.47427386e-02
    
60. #  -2.47906484e-02  2.65899673e-03 -5.64723015e-02 -1.81384571e-02
    
61. #  -5.86921759e-02 -1.66410357e-02  5.65116517e-02 -5.67275733e-02
    
62. #   2.84908544e-02 -3.57359536e-02 -2.56096441e-02 -1.55471719e-03
    
63. 
    
64. 
    
65. 
    
66. # caffe的结果：
    
67. # 4.18923870e-02 -4.78875041e-02 -2.89320182e-02 -5.47428243e-02
    
68. #  -2.47906558e-02  2.65889615e-03 -5.64723164e-02 -1.81384161e-02
    
69. #  -5.86921349e-02 -1.66409668e-02  5.65117151e-02 -5.67275919e-02
    
70. #   2.84908172e-02 -3.57360132e-02 -2.56096572e-02 -1.55465060e-03]
    
71. 
    

复制代码

1. # -*- coding: utf-8 -*-
   
2. from __future__ import print_function, division
   
3. __author__ = u'东方耀 微信：dfy_88888'
   
4. __date__ = '2020/8/10 上午10:34'
   
5. __product__ = 'PyCharm'
   
6. __filename__ = 'caffe_inference_extract_512_feature'
   
7. 
   
8. 
   
9. import caffe
   
10. import os
    
11. import cv2
    
12. import numpy as np
    
13. 
    
14. os.environ["CUDA_DEVICE_ORDER"] = "PCI_BUS_ID"
    
15. # 选择哪一块gpu,如果是-1，就是调用cpu
    
16. os.environ['CUDA_VISIBLE_DEVICES'] = "0"
    
17. 
    
18. 
    
19. deploy = "./caffe_model/FaceRecog_dfy06-sim.prototxt"
    
20. trained_model = "./caffe_model/FaceRecog_dfy06-sim.caffemodel"
    
21. caffe.set_mode_gpu()
    
22. caffe.set_device(0)
    
23. 
    
24. net_caffe = caffe.Net(deploy, 1, weights=trained_model)
    
25. 
    
26. 
    
27. # 看同一张图片 输入后  torch与caffe的输出的512维度向量
    
28. img_path = "/home/jingyun/py2_caffe_face_works/onnx2caffe-master/model/10.jpg"
    
29. 
    
30. img = cv2.imread(img_path)
    
31. print("cv2.imread after", img.shape)
    
32. # brg (0-255)
    
33. 
    
34. # test_transform = trans.Compose([
    
35. #     trans.ToTensor(),
    
36. #     trans.Normalize(mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5])
    
37. # ])
    
38. 
    
39. 
    
40. # 图片的预处理
    
41. img2 = img.copy()
    
42. # img = test_transform(img)
    
43. 
    
44. img2 = img2 / 255.0
    
45. img2 = (img2 - 0.5) / 0.5
    
46. img2 = np.transpose(img2, axes=[2, 0, 1])
    
47. 
    
48. # cv2.imread after (112, 112, 3)
    
49. # torch预处理后： (3, 112, 112)
    
50. # print("torch预处理后：", img.size(), img)
    
51. # [-0.7333, -0.8588, -0.8824,  ..., -0.2863,  0.4196,  0.8588],
    
52. #          [-0.7176, -0.8431, -0.8745,  ..., -0.3569,  0.3569,  0.8588],
    
53. #          [-0.7176, -0.8431, -0.8745,  ..., -0.4667,  0.2471,  0.8196],
    
54. #          ...,
    
55. #          [ 0.5765,  0.5765,  0.5529,  ...,  0.6941,  0.7020,  0.7176],
    
56. #          [ 0.5765,  0.5686,  0.5529,  ...,  0.7020,  0.7020,  0.7255],
    
57. #          [ 0.5686,  0.5686,  0.5529,  ...,  0.7020,  0.7098,  0.7255]
    
58. img2 = np.expand_dims(img2, axis=0)
    
59. print("numpy预处理后：", img2.shape)
    
60. # img = torch.unsqueeze(img, dim=0)
    
61. # print("torch 扩展维度后：", img.size(), type(img))
    
62. 
    
63. net_caffe.blobs["input0"].data[...] = img2
    
64. 
    
65. result = net_caffe.forward()['output0'][0]
    
66. 
    
67. print("result_caffe:", result.shape, type(result))
    
68. # result是否归一化了： 25.14348
    
69. print("result是否归一化了：", np.linalg.norm(result))
    
70. result = np.divide(result, np.linalg.norm(result))
    
71. print(result)
    
72. 
    
73. # 4.18923870e-02 -4.78875041e-02 -2.89320182e-02 -5.47428243e-02
    
74. #  -2.47906558e-02  2.65889615e-03 -5.64723164e-02 -1.81384161e-02
    
75. #  -5.86921349e-02 -1.66409668e-02  5.65117151e-02 -5.67275919e-02
    
76. #   2.84908172e-02 -3.57360132e-02 -2.56096572e-02 -1.55465060e-03]
    
77. 
    
78. 
    
79. 
    
80. 
    

复制代码

东方耀

1. #coding:utf-8
   
2. 
   
3. import caffe
   
4. import cv2
   
5. import numpy as np
   
6. from scipy.spatial.distance import pdist
   
7. import numpy as np
   
8. from PIL import Image
   
9. from collections import namedtuple
   
10. import time
    
11. net = caffe.Net('FaceRecog_dfy06-sim.prototxt', 'FaceRecog_dfy06-sim.caffemodel',caffe.TEST)
    
12. def caffeGetFeature(imgPath):
    
13.         bgr = cv2.imread(imgPath)
    
14.         #cv2.imshow("BGR",img)
    
15.         #cv2.waitKey(0)
    
16. 
    
17.         # BGR 0 1 2
    
18.         # RGB 2
    
19.         rgb = bgr[...,::-1]
    
20.         # scale = 1 / 128.0 = 0.0078125
    
21.         rgb = (rgb - 128.0) / 128.0
    
22.         rgb = np.transpose(rgb, axes=[2, 0, 1])
    
23.         # rgb = rgb.transpose((2,1,0))
    
24.         # rgb = np.swapaxes(rgb, 0, 2)
    
25.         # rgb = np.swapaxes(rgb, 1, 2)
    
26.         # rgb = rgb[None,:] # add singleton dimension
    
27.         rgb = np.expand_dims(rgb, axis=0)
    
28. 
    
29.         #cv2.imshow("RGB",rgb)
    
30.         #cv2.waitKey(0)
    
31.         #print (rgb)
    
32.         # out = net.forward_all( data = rgb ) # out is probability
    
33.         net.blobs["input0"].data[...] = rgb
    
34. 
    
35.         result = net.forward()['output0'][0]
    
36.         #print(out['fc1'][0])
    
37.         # a = out['output0'][0]
    
38.         return result
    
39. 
    
40. 
    
41. print("caffe forword result:", caffeGetFeature("./images/10.jpg"))

复制代码

东方耀

东方耀 发表于 2020-8-10 14:57

-2.36872983e+00  -1.75416946e+00   2.15635824e+00  -4.48371619e-02
  -5.19178569e-01  -7.63961196e-01   4.42535505e-02   4.14647192e-01
   4.16591078e-01   4.52785224e-01   6.10753894e-01  -1.32915735e+00
   2.55063891e+00   4.76892054e-01  -3.57504308e-01  -1.64712965e+00
  -3.39873940e-01   1.36864388e+00  -1.14875889e+00  -2.29086727e-01
   1.20023048e+00   3.08750719e-01   5.54423034e-01   1.23881590e+00
   9.36651051e-01   7.87511617e-02   1.13178682e+00   7.52464712e-01
   9.71228063e-01  -9.95164633e-01  -3.82107556e-01   7.82528222e-01
   9.88112748e-01  -8.99936497e-01  -1.86793208e-01  -2.27842927e+00
  -8.61048102e-01   4.04678315e-01  -1.56575155e+00  -1.47956216e+00
  -1.26751435e+00  -7.26402700e-02   1.32661295e+00  -2.19057918e+00
   7.20586538e-01  -5.92324138e-01  -7.70517468e-01  -1.42683581e-01