Toybrick

标题: 转换caffe模型出错 [打印本页]

---

作者: had_in    时间: 2020-6-29 16:14
标题: 转换caffe模型出错
求助。。
转换车牌识别库hyperlpr的caffe模型出错，原模型prototxt文件如下所示

1. input: "data"
   
2. input_dim: 1
   
3. input_dim: 3
   
4. input_dim: 16
   
5. input_dim: 66
   
6. layer {
   
7.   name: "conv1"
   
8.   type: "Convolution"
   
9.   bottom: "data"
   
10.   top: "conv1"
    
11.   convolution_param {
    
12.     num_output: 10
    
13.     bias_term: true
    
14.     pad: 0
    
15.     kernel_size: 3
    
16.     stride: 1
    
17.   }
    
18. }
    
19. layer {
    
20.   name: "relu1"
    
21.   type: "ReLU"
    
22.   bottom: "conv1"
    
23.   top: "conv1"
    
24. }
    
25. layer {
    
26.   name: "max_pooling2d_3"
    
27.   type: "Pooling"
    
28.   bottom: "conv1"
    
29.   top: "max_pooling2d_3"
    
30.   pooling_param {
    
31.     pool: MAX
    
32.     kernel_size: 2
    
33.     stride: 2
    
34.     pad: 0
    
35.   }
    
36. }
    
37. layer {
    
38.   name: "conv2"
    
39.   type: "Convolution"
    
40.   bottom: "max_pooling2d_3"
    
41.   top: "conv2"
    
42.   convolution_param {
    
43.     num_output: 16
    
44.     bias_term: true
    
45.     pad: 0
    
46.     kernel_size: 3
    
47.     stride: 1
    
48.   }
    
49. }
    
50. layer {
    
51.   name: "relu2"
    
52.   type: "ReLU"
    
53.   bottom: "conv2"
    
54.   top: "conv2"
    
55. }
    
56. layer {
    
57.   name: "conv3"
    
58.   type: "Convolution"
    
59.   bottom: "conv2"
    
60.   top: "conv3"
    
61.   convolution_param {
    
62.     num_output: 32
    
63.     bias_term: true
    
64.     pad: 0
    
65.     kernel_size: 3
    
66.     stride: 1
    
67.   }
    
68. }
    
69. layer {
    
70.   name: "relu3"
    
71.   type: "ReLU"
    
72.   bottom: "conv3"
    
73.   top: "conv3"
    
74. }
    
75. layer {
    
76.   name: "flatten_2"
    
77.   type: "Flatten"
    
78.   bottom: "conv3"
    
79.   top: "flatten_2"
    
80. }
    
81. layer {
    
82.   name: "dense"
    
83.   type: "InnerProduct"
    
84.   bottom: "flatten_2"
    
85.   top: "dense"
    
86.   inner_product_param {
    
87.     num_output: 2
    
88.   }
    
89. }
    
90. layer {
    
91.   name: "relu4"
    
92.   type: "ReLU"
    
93.   bottom: "dense"
    
94.   top: "dense"
    
95. }
    

复制代码我改成

1. # input: "data"
   
2. # input_dim: 1
   
3. # input_dim: 3
   
4. # input_dim: 16
   
5. # input_dim: 66
   
6. layer {
   
7.   name: "data"
   
8.   type: "Input"
   
9.   top: "data"
   
10.   input_param { shape: { dim: 1 dim: 3 dim: 16 dim: 66 } }
    
11. }
    
12. layer {
    
13.   name: "conv1"
    
14.   type: "Convolution"
    
15.   bottom: "data"
    
16.   top: "conv1"
    
17.   convolution_param {
    
18.     num_output: 10
    
19.     bias_term: true
    
20.     pad: 0
    
21.     kernel_size: 3
    
22.     stride: 1
    
23.   }
    
24. }
    
25. layer {
    
26.   name: "relu1"
    
27.   type: "ReLU"
    
28.   bottom: "conv1"
    
29.   top: "conv1"
    
30. }
    
31. layer {
    
32.   name: "max_pooling2d_3"
    
33.   type: "Pooling"
    
34.   bottom: "conv1"
    
35.   top: "max_pooling2d_3"
    
36.   pooling_param {
    
37.     pool: MAX
    
38.     kernel_size: 2
    
39.     stride: 2
    
40.     pad: 0
    
41.   }
    
42. }
    
43. layer {
    
44.   name: "conv2"
    
45.   type: "Convolution"
    
46.   bottom: "max_pooling2d_3"
    
47.   top: "conv2"
    
48.   convolution_param {
    
49.     num_output: 16
    
50.     bias_term: true
    
51.     pad: 0
    
52.     kernel_size: 3
    
53.     stride: 1
    
54.   }
    
55. }
    
56. layer {
    
57.   name: "relu2"
    
58.   type: "ReLU"
    
59.   bottom: "conv2"
    
60.   top: "conv2"
    
61. }
    
62. layer {
    
63.   name: "conv3"
    
64.   type: "Convolution"
    
65.   bottom: "conv2"
    
66.   top: "conv3"
    
67.   convolution_param {
    
68.     num_output: 32
    
69.     bias_term: true
    
70.     pad: 0
    
71.     kernel_size: 3
    
72.     stride: 1
    
73.   }
    
74. }
    
75. layer {
    
76.   name: "relu3"
    
77.   type: "ReLU"
    
78.   bottom: "conv3"
    
79.   top: "conv3"
    
80. }
    
81. layer {
    
82.   name: "flatten_2"
    
83.   type: "Flatten"
    
84.   bottom: "conv3"
    
85.   top: "flatten_2"
    
86. }
    
87. layer {
    
88.   name: "dense"
    
89.   type: "InnerProduct"
    
90.   bottom: "flatten_2"
    
91.   top: "dense"
    
92.   inner_product_param {
    
93.     num_output: 2
    
94.   }
    
95. }
    
96. layer {
    
97.   name: "relu4"
    
98.   type: "ReLU"
    
99.   bottom: "dense"
    
100.   top: "dense"
     
101. }

复制代码模型在加载阶段rknn.load_caffe就出错了，报错信息如下所示

1. --> config model
   
2. done
   
3. --> Loading model
   
4. E Catch exception when loading caffe model: ./HorizonalFinemapping.prototxt!
   
5. E Traceback (most recent call last):
   
6. E   File "rknn\api\rknn_base.py", line 441, in rknn.api.rknn_base.RKNNBase.load_caffe
   
7. E   File "rknn\base\RKNNlib\converter\caffeloader.py", line 1042, in rknn.base.RKNNlib.converter.caffeloader.CaffeLoader.load_blobs
   
8. E   File "rknn\base\RKNNlib\converter\caffeloader.py", line 935, in rknn.base.RKNNlib.converter.caffeloader.CaffeLoader.parse_blobs
   
9. E   File "rknn\base\RKNNlib\converter\caffeloader.py", line 358, in rknn.base.RKNNlib.converter.caffeloader.proc_blobs_convolution
   
10. E   File "G:\ProgramData\Anaconda3\envs\rknn\lib\site-packages\numpy\core\fromnumeric.py", line 292, in reshape
    
11. E     return _wrapfunc(a, 'reshape', newshape, order=order)
    
12. E   File "G:\ProgramData\Anaconda3\envs\rknn\lib\site-packages\numpy\core\fromnumeric.py", line 56, in _wrapfunc
    
13. E     return getattr(obj, method)(*args, **kwds)
    
14. E ValueError: cannot reshape array of size 0 into shape (10,3,3,3)
    
15. Load model failed! Ret = -1

复制代码模型文件地址是https://github.com/zeusees/Hyper ... nception.caffemodel
请问这可能是什么地方出问题，看输出信息，问题在第一个卷积层就出现了，但是我实在看不出来我输入层有什么问题，求指导。
我的环境是windows10，rknn版本是1.3.2

---

作者: had_in    时间: 2020-6-29 16:21
https://github.com/zeusees/Hyper ... emapping.caffemodel

不好意思，模型地址发错了

---

作者: had_in    时间: 2020-6-29 16:54
日志信息

1. D Save log info to: ./caffe.log
   
2. --> config model
   
3. done
   
4. --> Loading model
   
5. I Set caffe proto to caffe
   
6. I Load caffe model ./HorizonalFinemapping.prototxt
   
7. I Parsing net parameters ...
   
8. D import clients finished
   
9. I Parsing layer parameters ...
   
10. D Convert layer image
    
11. D Convert layer conv1
    
12. D Convert layer relu1
    
13. D Convert layer max_pooling2d_3
    
14. D Convert layer conv2
    
15. D Convert layer relu2
    
16. D Convert layer conv3
    
17. D Convert layer relu3
    
18. D Convert layer flatten_2
    
19. D Convert layer dense
    
20. D Convert layer relu4
    
21. I Parsing connections ...
    
22. D Connect: image_0,0 to conv1_1,0
    
23. D Connect: conv1_1,0 to relu1_2,0
    
24. D Connect: relu1_2,0 to max_pooling2d_3_3,0
    
25. D Connect: max_pooling2d_3_3,0 to conv2_4,0
    
26. D Connect: conv2_4,0 to relu2_5,0
    
27. D Connect: relu2_5,0 to conv3_6,0
    
28. D Connect: conv3_6,0 to relu3_7,0
    
29. D Connect: relu3_7,0 to flatten_2_8,0
    
30. D Connect: flatten_2_8,0 to dense_9,0
    
31. D Connect: dense_9,0 to relu4_10,0
    
32. D Connect: relu4_10,0 to output_11,0,
    
33. I Load net complete.
    
34. 2020-06-29 16:50:26.078880: I tensorflow/core/platform/cpu_feature_guard.cc:141] Your CPU supports instructions that this TensorFlow binary was not compiled to use: AVX2
    
35. D Process image_0 ...
    
36. D RKNN output shape(input): (0 16 66 3)
    
37. D Process conv1_1 ...
    
38. D RKNN output shape(convolution): (0 14 64 10)
    
39. D Process relu1_2 ...
    
40. D RKNN output shape(relu): (0 14 64 10)
    
41. D Process max_pooling2d_3_3 ...
    
42. D RKNN output shape(pooling): (0 7 32 10)
    
43. D Process conv2_4 ...
    
44. D RKNN output shape(convolution): (0 5 30 16)
    
45. D Process relu2_5 ...
    
46. D RKNN output shape(relu): (0 5 30 16)
    
47. D Process conv3_6 ...
    
48. D RKNN output shape(convolution): (0 3 28 32)
    
49. D Process relu3_7 ...
    
50. D RKNN output shape(relu): (0 3 28 32)
    
51. D Process flatten_2_8 ...
    
52. D RKNN output shape(flatten): (0 2688)
    
53. D Process dense_9 ...
    
54. D RKNN output shape(fullconnect): (0 2)
    
55. D Process relu4_10 ...
    
56. D RKNN output shape(relu): (0 2)
    
57. D Process output_11 ...
    
58. D RKNN output shape(output): (0 2)
    
59. I Build  complete.
    
60. I Load blobs from caffe model ./HorizonalFinemapping.caffemodel
    
61. I Check if ./HorizonalFinemapping.caffemodel contains 'type' value
    
62. I Parsing net blobs ...
    
63. D Load blobs of conv1
    
64. E Catch exception when loading caffe model: ./HorizonalFinemapping.prototxt!
    
65. E Traceback (most recent call last):
    
66. E   File "rknn\api\rknn_base.py", line 441, in rknn.api.rknn_base.RKNNBase.load_caffe
    
67. E   File "rknn\base\RKNNlib\converter\caffeloader.py", line 1042, in rknn.base.RKNNlib.converter.caffeloader.CaffeLoader.load_blobs
    
68. E   File "rknn\base\RKNNlib\converter\caffeloader.py", line 935, in rknn.base.RKNNlib.converter.caffeloader.CaffeLoader.parse_blobs
    
69. E   File "rknn\base\RKNNlib\converter\caffeloader.py", line 358, in rknn.base.RKNNlib.converter.caffeloader.proc_blobs_convolution
    
70. E   File "G:\ProgramData\Anaconda3\envs\rknn\lib\site-packages\numpy\core\fromnumeric.py", line 292, in reshape
    
71. E     return _wrapfunc(a, 'reshape', newshape, order=order)
    
72. E   File "G:\ProgramData\Anaconda3\envs\rknn\lib\site-packages\numpy\core\fromnumeric.py", line 56, in _wrapfunc
    
73. E     return getattr(obj, method)(*args, **kwds)
    
74. E ValueError: cannot reshape array of size 0 into shape (10,3,3,3)
    
75. Load model failed! Ret = -1

复制代码

---

欢迎光临 Toybrick (https://t.rock-chips.com/)

Powered by Discuz! X3.3