Error in population_encode when encoding higher-dimensional Tensor about norse

[Jegp]

Thank you for the report @StrasserFlorian! This is definitely not expected behavior, and I can see there's a bug in the way we concatenate tensors if we get unexpected inputs with different dimensionality.

I was wondering, is the expected output really shaped 100, 10, 3 or is it 100, 3, 10? That is, should the additional dimension be inserted between the outer and inner dimensions, or is it desirable to append the additional dimension in case we wish to process higher-dimensional (>3) tensors.

[StrasserFlorian]

Hi @Jegp,
i think your suggestion of appending the additional dimension at the end makes more sense. I only thought it might be on purpose to insert it in the middle, since that is how the size of the new Tensor is defined right now in the code:
size = (input_values.size(0), out_features) + input_values.size()[1:]

[Jegp]

I've written a fix in this PR: #351

Note that this changes the location where the new dimension is inserted. I do think that the new version makes more sense. Always inserting at dimension 1 may be somewhat arbitrary. For example, if the input tensor shape is batch-width-height, it'll be inserted between batch-width, but if the input tensor is time-batch-width, it'll be between time-batch.

Would you mind checking out the PR code to see if that correctly addresses the problem? :-)

[StrasserFlorian]

Yes, i think implementing it this way makes the most sense. Looks good! 👍

[Jegp]

Apologies for the delay, but this should now be fixed with #351 . Please don't hesitate to reopen this if you're still experiencing issues.