In pytorch I have to tensors of dimensions [K,L,M] and [M,L,N]. I want to perform a standard tensor convolution product of those tensors along the middle two dimensions to obtain a [K,N] tensor. I couldn't find official documentation on how to perform those operations, perhaps it should be better done in some other library and then reconverted to pytorch tensor?
If by convolution you actually mean something like contraction, you're probably looking for torch.tensordot. You can specify the indices that should be contracted.