onnx export model differs too much from pytorch model

Thanks for your error report and we appreciate it a lot.

Describe the bug I trained faster-rcnn on a custom coco dataset and exported it to onnx using pytorch2onnx.py. The onnx model results differs too much from pytorch model.

I retried training and exporting using retinanet and this problem did not occur?

Reproduction

1. What command or script did you run?

python /home/jovyan/work/mmdetection/tools/train.py\
/home/jovyan/work/vdetection/work_dirs/chess_faster_mmdet2/faster.py

python /home/jovyan/work/mmdetection/tools/deployment/pytorch2onnx.py\
/home/jovyan/work/vdetection/work_dirs/chess_faster_mmdet2/faster.py\
/home/jovyan/work/vdetection/work_dirs/chess_faster_mmdet2/epoch_12.pth\
--output-file /home/jovyan/work/vdetection/work_dirs/chess_faster_mmdet2/tmp.onnx --verify\
--input-img /home/jovyan/work/vdetection/work_dirs/chess_faster_mmdet2/472613.jpg

I provide the links to the config, ckpt , log and onnx file in this link.

Environment

2021-09-20 10:51:43,820 - mmdet - INFO - Environment info:

sys.platform: linux Python: 3.8.8 (default, Feb 24 2021, 21:46:12) [GCC 7.3.0] CUDA available: True GPU 0,1: NVIDIA GeForce GTX TITAN X CUDA_HOME: /usr/local/cuda NVCC: Build cuda_11.3.r11.3/compiler.29745058_0 GCC: gcc (Ubuntu 9.3.0-17ubuntu1~20.04) 9.3.0 PyTorch: 1.9.0a0+2ecb2c7 PyTorch compiling details: PyTorch built with:

• GCC 9.3
• C++ Version: 201402
• Intel® Math Kernel Library Version 2019.0.4 Product Build 20190411 for Intel® 64 architecture applications
• Intel® MKL-DNN v1.8.0 (Git Hash N/A)
• OpenMP 201511 (a.k.a. OpenMP 4.5)
• NNPACK is enabled
• CPU capability usage: AVX2
• CUDA Runtime 11.3
• NVCC architecture flags: -gencode;arch=compute_52,code=sm_52;-gencode;arch=compute_60,code=sm_60;-gencode;arch=compute_61,code=sm_61;-gencode;arch=compute_70,code=sm_70;-gencode;arch=compute_75,code=sm_75;-gencode;arch=compute_80,code=sm_80;-gencode;arch=compute_86,code=sm_86;-gencode;arch=compute_86,code=compute_86
• CuDNN 8.2
• Magma 2.5.2
• Build settings: BLAS_INFO=mkl, BUILD_TYPE=Release, CUDA_VERSION=11.3, CUDNN_VERSION=8.2.0, CXX_COMPILER=/usr/bin/c++, CXX_FLAGS= -fvisibility-inlines-hidden -DUSE_PTHREADPOOL -fopenmp -DNDEBUG -DUSE_FBGEMM -DUSE_QNNPACK -DUSE_PYTORCH_QNNPACK -DUSE_XNNPACK -O2 -fPIC -Wno-narrowing -Wall -Wextra -Werror=return-type -Wno-missing-field-initializers -Wno-type-limits -Wno-array-bounds -Wno-unknown-pragmas -Wno-sign-compare -Wno-unused-parameter -Wno-unused-variable -Wno-unused-function -Wno-unused-result -Wno-unused-local-typedefs -Wno-strict-overflow -Wno-strict-aliasing -Wno-error=deprecated-declarations -Wno-stringop-overflow -Wno-psabi -Wno-error=pedantic -Wno-error=redundant-decls -Wno-error=old-style-cast -fdiagnostics-color=always -faligned-new -Wno-unused-but-set-variable -Wno-maybe-uninitialized -fno-math-errno -fno-trapping-math -Werror=format -Werror=cast-function-type -Wno-stringop-overflow, FORCE_FALLBACK_CUDA_MPI=1, LAPACK_INFO=mkl, PERF_WITH_AVX=1, PERF_WITH_AVX2=1, PERF_WITH_AVX512=1, TORCH_VERSION=1.9.0, USE_CUDA=ON, USE_CUDNN=ON, USE_EXCEPTION_PTR=1, USE_GFLAGS=OFF, USE_GLOG=OFF, USE_MKL=ON, USE_MKLDNN=ON, USE_MPI=ON, USE_NCCL=ON, USE_NNPACK=ON, USE_OPENMP=ON,

TorchVision: 0.9.0a0 OpenCV: 4.5.3 MMCV: 1.3.13 MMCV Compiler: GCC 9.3 MMCV CUDA Compiler: 11.3 MMDetection: 2.16.0+0d66ba3

Error traceback If applicable, paste the error trackback here.

Traceback (most recent call last):
  File "/home/jovyan/work/mmdetection/tools/deployment/pytorch2onnx.py", line 325, in <module>
    pytorch2onnx(
  File "/home/jovyan/work/mmdetection/tools/deployment/pytorch2onnx.py", line 197, in pytorch2onnx
    np.testing.assert_allclose(
  File "/opt/conda/lib/python3.8/site-packages/numpy/testing/_private/utils.py", line 1528, in assert_allclose
    assert_array_compare(compare, actual, desired, err_msg=str(err_msg),
  File "/opt/conda/lib/python3.8/site-packages/numpy/testing/_private/utils.py", line 761, in assert_array_compare
    raise AssertionError(msg)
AssertionError: 
Not equal to tolerance rtol=0.001, atol=1e-05
The numerical values are different between Pytorch and ONNX, but it does not necessarily mean the exported ONNX model is problematic.
(shapes (4, 5), (0, 5) mismatch)
 x: array([[-2., -2., -2., -2.,  0.],
       [-2., -2., -2., -2.,  0.],
       [-2., -2., -2., -2.,  0.],
       [-2., -2., -2., -2.,  0.]], dtype=float32)
 y: array([], shape=(0, 5), dtype=float32)

onnx results pytorch model results

Bug fix

[RunningLeon]

@Hussni-V Could try with PyTorch==1.8.0 and ONNXRuntime==1.8.0. Tested OK on my machine

sys.platform: linux
Python: 3.7.5 (default, Oct 25 2019, 15:51:11) [GCC 7.3.0]
CUDA available: True
GPU 0: NVIDIA GeForce RTX 2080
CUDA_HOME: /usr/local/cuda
NVCC: Build cuda_11.1.TC455_06.29190527_0
GCC: gcc (Ubuntu 7.5.0-3ubuntu1~16.04) 7.5.0
PyTorch: 1.8.0
PyTorch compiling details: PyTorch built with:
  - GCC 7.3
  - C++ Version: 201402
  - Intel(R) oneAPI Math Kernel Library Version 2021.2-Product Build 20210312 for Intel(R) 64 architecture applications
  - Intel(R) MKL-DNN v1.7.0 (Git Hash 7aed236906b1f7a05c0917e5257a1af05e9ff683)
  - OpenMP 201511 (a.k.a. OpenMP 4.5)
  - NNPACK is enabled
  - CPU capability usage: AVX2
  - CUDA Runtime 11.1
  - NVCC architecture flags: -gencode;arch=compute_37,code=sm_37;-gencode;arch=compute_50,code=sm_50;-gencode;arch=compute_60,code=sm_60;-gencode;arch=compute_61,code=sm_61;-gencode;arch=compute_70,code=sm_70;-gencode;arch=compute_75,code=sm_75;-gencode;arch=compute_80,code=sm_80;-gencode;arch=compute_86,code=sm_86;-gencode;arch=compute_37,code=compute_37
  - CuDNN 8.0.5
  - Magma 2.5.2
  - Build settings: BLAS_INFO=mkl, BUILD_TYPE=Release, CUDA_VERSION=11.1, CUDNN_VERSION=8.0.5, CXX_COMPILER=/opt/rh/devtoolset-7/root/usr/bin/c++, CXX_FLAGS= -Wno-deprecated -fvisibility-inlines-hidden -DUSE_PTHREADPOOL -fopenmp -DNDEBUG -DUSE_KINETO -DUSE_FBGEMM -DUSE_QNNPACK -DUSE_PYTORCH_QNNPACK -DUSE_XNNPACK -O2 -fPIC -Wno-narrowing -Wall -Wextra -Werror=return-type -Wno-missing-field-initializers -Wno-type-limits -Wno-array-bounds -Wno-unknown-pragmas -Wno-sign-compare -Wno-unused-parameter -Wno-unused-variable -Wno-unused-function -Wno-unused-result -Wno-unused-local-typedefs -Wno-strict-overflow -Wno-strict-aliasing -Wno-error=deprecated-declarations -Wno-stringop-overflow -Wno-psabi -Wno-error=pedantic -Wno-error=redundant-decls -Wno-error=old-style-cast -fdiagnostics-color=always -faligned-new -Wno-unused-but-set-variable -Wno-maybe-uninitialized -fno-math-errno -fno-trapping-math -Werror=format -Wno-stringop-overflow, LAPACK_INFO=mkl, PERF_WITH_AVX=1, PERF_WITH_AVX2=1, PERF_WITH_AVX512=1, TORCH_VERSION=1.8.0, USE_CUDA=ON, USE_CUDNN=ON, USE_EXCEPTION_PTR=1, USE_GFLAGS=OFF, USE_GLOG=OFF, USE_MKL=ON, USE_MKLDNN=ON, USE_MPI=OFF, USE_NCCL=ON, USE_NNPACK=ON, USE_OPENMP=ON, 

TorchVision: 0.9.0
OpenCV: 4.5.2
MMCV: 1.3.13
MMCV Compiler: GCC 7.5
MMCV CUDA Compiler: 11.1
MMDetection: 2.16.0+5ef56c1

[RunningLeon]

@jshilong We have in the doc: https://github.com/open-mmlab/mmdetection/blob/master/docs/tutorials/pytorch2onnx.md#list-of-supported-models-exportable-to-onnx

Notes:

Minimum required version of MMCV is 1.3.5

All models above are tested with Pytorch==1.6.0 and onnxruntime==1.5.1, except for CornerNet. For more details about the torch version when exporting CornerNet to ONNX, which involves mmcv::cummax, please refer to the Known Issues in mmcv.