FieldType.transform doesn't work on GPU tensors

[quva-lab]

FieldType.transform doesn't work on GPU tensors

[kristian-georgiev]

Not the author, but it seems like it might be easier/more efficient if np.einsum here and np.rot90 in the different ._basespace_actions are replaced with their torch equivalents. This way, there should be no need to move the tensor between the GPU and CPU(?)

[Gabri95]

hey there! Sorry for the late reply.

I agree that both .transform() methods should raise an exception in that case.

Currently, the backend of the library dealing with groups, representations and generation of the kernels only relies on numpy.
The e2cnn.nn subpackage, then, tries to provide an interface similar to PyTorch and wraps the np.ndarray in the other subpackages inside torch.Tensor. The rationale was that I wanted the backend of the library to be independent of PyTorch (such that one could also extract the filter bases for other applications).
I am now a bit more convinced that the backend should also support PyTorch, or at least the e2cnn.gspace should.

For the moment, I'd still recommend not directly using tensor.transform(el) when this implies some interpolation of the input. This is because one probably wants to have the freedom to customise more this operation (e.g. change the order of the interpolation).

Anyways, in a future release, I will try to fix the gspace as @kristian-georgiev suggested, using torch operations.

Thanks!
Gabriele