Thiago Crepaldi

> @thiagocrepaldi looks like ORT uses Col2imND for ConvTranspose: https://cs.github.com/microsoft/onnxruntime/blob/2dfd81b9bb097c90388010e5b7d298498274f8d9/onnxruntime/core/providers/cpu/nn/conv_transpose.cc?q=repo%3Amicrosoft%2Fonnxruntime+Col2Im#L192 > > So if they already have an N-dimensional Col2im implementation, the question is can we generalize the ONNX op...

> @thiagocrepaldi I think my original question is still the important one: can we generalize the ONNX op spec to support Col2imND such that it's easy to re-use the existing...

> I agree with @garymm 's points above. One clarification though: it doesn't matter if the existing implementation uses Im2col internally ... but it would be useful to verify that...

> @thiagocrepaldi please write a test case for N-dimensional col2im in ORT to make sure their current implementation is correct. If so then we have confidence we can implement this...

https://github.com/microsoft/onnxruntime/pull/12311 is now ready and implements Col2Im as contrib op on ORT

@askhade Could you dismiss your review? All comments were addressed back in the day, but I have no permission to dismiss nor request another review. Feel free to give it...

> @askhade Could you dismiss your review? All comments were addressed back in the day, but I have no permission to dismiss nor request another review. Feel free to give...

Would https://github.com/pytorch/pytorch/issues/69674 leverage this? We need Pytorch' s `::grid_sampler` to export a detectron2 model

@gramalingam @jackwish Have you seen this?

@gramalingam @zhenhuaw-me gentle ping