Hi @xadupre , can I ask whether there's a way to call the python implementation of an existing op when writing it for a new op in onnx/reference/ops? For example in this case, I would like to make use of the existing implementation for GridSample and Conv. Thanks in advance!

Page https://onnx.ai/onnx/api/reference.html shows an example which uses an implementation of new custom operator. The same mechanism can be used to replace an existing op. Is that what you are looking for?

Hi @gramalingam @p-wysocki , do you have any more comments or questions about this PR?

Thanks for this PR @yuanyao-nv ! feel free merging it anytime. I have 2 follow up questions on implementation of this op in ORT:
(1) I see you added 2 references to DeformConv models. I wonder is there any new models use the op after these 2 publications. I searched for any uses of pytorch DeformConv2d (https://pytorch.org/vision/main/generated/torchvision.ops.DeformConv2d.html) but did not find any use of it on internet.
(2) Now we have many ops extended to 3D and ND, any examples of using DeformDonv for 3D data is very helpful.
There is a plan to add DeformConv CPU kernels to ORT. It will be very helpful if you have any pointers to above 2 questions. thanks!

Hi @liqunfu , to answer your questions:

1. Besides the original papers, there are some other models mentioned in the issues I linked as well, such as Kernel Transformer Network https://arxiv.org/abs/1812.03115, and https://arxiv.org/abs/1911.08947.
2. Examples of 3D case: https://arxiv.org/abs/2004.02803 https://arxiv.org/abs/1911.01898

Hi @liqunfu , to answer your questions:

1. Besides the original papers, there are some other models mentioned in the issues I linked as well, such as Kernel Transformer Network https://arxiv.org/abs/1812.03115.
2. Examples of 3D case: https://arxiv.org/abs/2004.02803 https://arxiv.org/abs/1911.01898

Thank you very much @yuanyao-nv !