C++ Inference Accuracy is row..

[seungtaek94]

I trained my classification model(5 classes) in python(This model).
And i tried to load that model in c++ environment.

I got 91% Accuracy in pyhton,
But i got less then 40% Accuracy in C++.

I followed this inference example exactly.

c++ inference example log:

[11:40:50] D:\WORK\AI\mxnet-v1.7.x\src\io\iter_image_recordio_2.cc:178: ImageRecordIOParser2: C:\Users\admin\Desktop\test.rec, use 6 threads for decoding..
[11:40:51] C:\Users\admin\Desktop\mxnet_cpp_2\mxnet_cpp\mxnet_cpp.cpp:307: Loading the model from C:\Users\admin\Desktop\mxnet_cpp_2\x64\Release\cspdarknet53-symbol.json

[11:40:51] D:\WORK\AI\mxnet-v1.7.x\src\nnvm\legacy_json_util.cc:209: Loading symbol saved by previous version v1.6.0. Attempting to upgrade...
[11:40:51] D:\WORK\AI\mxnet-v1.7.x\src\nnvm\legacy_json_util.cc:217: Symbol successfully upgraded!
[11:40:51] C:\Users\admin\Desktop\mxnet_cpp_2\mxnet_cpp\mxnet_cpp.cpp:322: Loading the model parameters from C:\Users\admin\Desktop\mxnet_cpp_2\x64\Release\cspdarknet53smooth_133(best).params

[11:41:14] D:\WORK\AI\mxnet-v1.7.x\src\operator\nn\cudnn\./cudnn_algoreg-inl.h:97: Running performance tests to find the best convolution algorithm, this can take a while... (set the environment variable MXNET_CUDNN_AUTOTUNE_DEFAULT to 0 to disable)
[11:42:19] C:\Users\admin\Desktop\mxnet_cpp_2\mxnet_cpp\mxnet_cpp.cpp:502: INFO:Dataset for inference: C:\Users\admin\Desktop\test.rec
[11:42:19] C:\Users\admin\Desktop\mxnet_cpp_2\mxnet_cpp\mxnet_cpp.cpp:503: INFO:label_name = dense0_bias
[11:42:20] C:\Users\admin\Desktop\mxnet_cpp_2\mxnet_cpp\mxnet_cpp.cpp:504: INFO:rgb_mean: (123.68, 116.779, 103.939)
[11:42:22] C:\Users\admin\Desktop\mxnet_cpp_2\mxnet_cpp\mxnet_cpp.cpp:506: INFO:rgb_std: (58.393, 57.12, 57.375)
[11:42:23] C:\Users\admin\Desktop\mxnet_cpp_2\mxnet_cpp\mxnet_cpp.cpp:508: INFO:Image shape: (3, 224, 224)
[11:42:24] C:\Users\admin\Desktop\mxnet_cpp_2\mxnet_cpp\mxnet_cpp.cpp:510: INFO:Finished inference with: 16384 images
[11:42:25] C:\Users\admin\Desktop\mxnet_cpp_2\mxnet_cpp\mxnet_cpp.cpp:512: INFO:Batch size = 256 for inference
[11:42:25] C:\Users\admin\Desktop\mxnet_cpp_2\mxnet_cpp\mxnet_cpp.cpp:513: INFO:Accuracy: 0.390137
[11:42:26] C:\Users\admin\Desktop\mxnet_cpp_2\mxnet_cpp\mxnet_cpp.cpp:514: INFO:Throughput: 249.635 images per second

And below exception occurred when i run that example.

Exception thrown at 0x00007FF9D3D7D759 in mxnet_cpp.exe: Microsoft C++ exception: `anonymous namespace'::ExifParsingError at memory location 0x0000003C6D7FE6E0.
Exception thrown at 0x00007FF9D3D7D759 in mxnet_cpp.exe: Microsoft C++ exception: `anonymous namespace'::ExifParsingError at memory location 0x0000003C6DBFEBF0.
Exception thrown at 0x00007FF9D3D7D759 in mxnet_cpp.exe: Microsoft C++ exception: `anonymous namespace'::ExifParsingError at memory location 0x0000003C6D8FEC90.

Env:

training

• ubuntu 18.04
• python
• mxnet 1.6.0
• titanxp
• cuda10.2

inference

• windows10
• c++
• mxnet 1.7.0
• rtx 2080 super
• cuda10.2

[szha]

cc @access2rohit

[cao-nv]

Hi @seungtaek94 , have you solved this problem yet?
I'm facing various problems when running a Python trained model on C++.

[seungtaek94]

Hi, Sorry for late response :(

Actually, i dont have time to focusing on this problem.. so i didnt solve this problem yet.

[cao-nv]

Hi, Sorry for late response :(

Actually, i dont have time to focusing on this problem.. so i didnt solve this problem yet.

Fortunately, even results from Python and C++ are not identical, I finally got the same performance.
The most annoying issue for me is the order of element when converting output tensors for post-processing.

[LIU-liu-66]

Hi @seungtaek94 , do you know what the problem is and how to solve it?