Test Cases with Input Shape (-1, 7)

[DarrenZhang01]

Hi, could you confirm that the input is (-1, 7) in https://github.com/google/neural-tangents/blob/master/tests/stax_test.py#L1017? Since this is not reshaping, what does -1 represent here as the first dimension? Thanks!

[romanngg]

I think this just means it's not being used, or used in places where the actualized size at axis 0 can be inferred from the context. E.g. by default the leading dimension is the batch dimension, that does not influence how parameters of the Dense layer are initialized. I imagine it could error if any other dimension is -1, or if batch_axis != 0 (but it might also work, if again, the size where -1 is used can be inferred from the context).

It may be cleaner to just pass X0_1.shape there though.

[DarrenZhang01]

I see, thanks!