NIR
NIR copied to clipboard
neuromorphs
Fix to `input_type` node of `CubaLIF` neurons
CubaLIF neurons would have an input_type node of [], instead of [1], when created from a node with scalar v_threshold. This pull request fixes this issue.
Also adds a squeeze() to the rockpool RNN notebook.
Looks good to me, thanks for the PR @DylanMuir!
Edit: actually, the v_threshold should not be a scalar, but an array of the same size as the neuron population. Can you explain where you get the NIR graph that contains a CubaLIF neuron with scalar v_threshold? We should fix it there. The current fix could also be okay but it assumes that CubaLIF neurons are 1D populations, but we may want to have multi-dim populations e.g. in SCNNs.
Hi @stevenabreu7 , it occurred precisely when the neuron population contains 1 neuron. In other cases v_threshold was a vector as expected. I did want to check with you about 1D or 2D populations.
The alternative fix would be to ensure v_threshold is a vector with a single element, when creating a single neuron.
Explicitly, it occurred when exporting a net to nir from snntorch, containing Synaptic modules.
in export_nir.py line 58, in _extract_snntorch_module():
vthr = module.threshold.detach().numpy()
The line looks fine, but becomes a scalar numpy array, instead of a single-element vector. So I guess it's up to you whether to fix this in NIR or ask Jason for a fix ;)
Thank you for following up on this! If I'm understanding this correctly, the scalar seems to be coming from snnTorch. But your addition, @DylanMuir, should work with both scalars and arrays alike, correct? In that case, I think it's an improvement compared to the previous version. Would you agree @stevenabreu7?
Regarding the snnTorch codebase, I opened an issue here: https://github.com/jeshraghian/snntorch/issues/334 I'm not so familiar with the codebase though, so I'm not sure I'm a good candidate to dive in for a fix.
Sorry about the delay.. As it is now, it breaks the types of 2D neuron populations. I pushed a simple change here to ensure that if v_threshold is scalar, it is cast into a numpy array, then we can also keep the v_treshold.shape at the end. Does this work for you @DylanMuir?
PR for snnTorch is also up here: https://github.com/jeshraghian/snntorch/pull/335