Support for torch.Tensor.unfold or torch.nn.Unfold

[Russell-Izadi-Bose]

❓Question

Is there any plan for adding unfold ops or is there an alternative to perform such operation?

[IlyaKrotov]

Hi! Any updates?

[Karllas]

Hello. Will this be implemented?

[memset-c]

Is there any plan for implementing this operator?

[TobyRoseman]

Supporting torch.Tensor.unfold with our current set of MIL ops might be a bit tricky. One idea is to use a convolution followed by an interleave concat. The stride of the convolution would be equal to step. The size of the kernels should be size.

Supporting torch.nn.Unfold seem a little different. However using a convolution here too is probably also the key.

If anyone can help me fill in the details here, please let me know.

[fwcd]

Unless I am missing something, coremltools currently only supports torch.nn.functional.unfold. It would be pretty great to have the more general Tensor.unfold too though.

[TobyRoseman]

Unless I am missing something, coremltools currently only supports torch.nn.functional.unfold. It would be pretty great to have the more general Tensor.unfold too though.

Yes, I think that's correct; Tensor.unfold is still not supported. Please create a new issue with a minimal PyTorch model that fails to convert because Tensor.unfold is not supported.