Fix problems with loading ONNX networks with parametrized inputs#18145
Fix problems with loading ONNX networks with parametrized inputs#18145opencv-pushbot merged 1 commit intoopencv:3.4from
Conversation
|
@alalek Can you please check test Layer_Test_Slice.slice_channels_and_batch_17762 ? My changes should not affect this test case at all, but anyway it fails on MYRIAD device |
|
@dkurt Can you please review the changes? Also, I think I need your help with some tests on Custom machines. For some reason, there are failed cases on MYRIAD/CPU with DLIE, even though some of them are not related to my code (for example, Layer_Test_Slice.slice_channels_and_batch_17762). I can't reproduce it locally on my machine, maybe you have any ideas about failed tests. |
2edc42d to
00ab83f
Compare
|
BTW, avoid using |
thank you! |
|
Ignore this ABI error. It belongs to
|
Thanks! Is there some article/wiki page about such details of OpenCV's CI configuration? I am asking just for the future work, so there will be less problems like this |
|
ABI checks are not intended to be configured externally. It is a part of Buildbot configuration (updates require restart, GitHub code may be a bit outdated). |
|
@sl-sergei Could you check if these added onnx do pass with this PR? |
|
@sl-sergei Please squash commits and rebase on latest 3.4 branch |
d245084 to
ad0fbf4
Compare
|
@dkurt friendly reminder |
resolves #17914 and #18072
merge with extra: opencv/opencv_extra#798
Current limitations for fix: Expand, Shape and ConstantFill layers are not supported, since their parameters use constBlobs, which are not accessible during allocation of layers memory
There is a problem with loading ONNX networks with input shapes denoted by some string parameter (e.g. "width", "height"). It seems that OpenCV parser replaces these parameters to zeroes by default.

One of the examples is MaskRCNN in #17914 . I believe that converting from PyTorch to ONNX with "dynamic_axes" will produce the same issues.