Support scalar parameters in type inference

[fabio-innatera]

I am trying to import a NIR file exported by latest snnTorch from GitHub, and got an error in type inference.

Turns out that snnTorch uses numpy scalars like np.float32(4.0) to define parameters, which end up
as input_type={'input': ()} (empty tuple). This implies that the actual neuron count is not known on the neuron node.

Technically, snnTorch is wrong here, the type annotation for neurons call for a np.array, but because numpy scalar values like np.float32 have a .shape property, this actually worked just fine before we introduced type inference.

Thankfully we are already doing type inference, and in most cases we can infer the number of neurons and therefore input_type and output_type from preceding nodes, and this is what I am doing here.

[fabio-innatera]

Maybe we should also change the type annotations?

import typing
import numpy.typing as npt

typing.Union[np.floating, npt.NDArray[np.floating]]

[Jegp]

I would argue against this. One of the core motivations when building NIR was to disambiguate as much as possible. We had long discussios on broadcasting when we designed the specs and the general sentiment was to avoid it because it is, as you say, ambiguous. We wouldn't know how many neurons that appear in the graph which isn't kosher for evaluation purposes.

[fabio-innatera]

I agree that we should enforce usage of np.array instead of loosening the constraints.

I opened a new PR that does just that: #162

Digging around in snntorch, it looks like at least the tests already use array parameters: https://github.com/jeshraghian/snntorch/blob/master/tests/test_nir.py#L60