Possible problem with moving_window in Spikemax?

[Michaeljurado42]

Hello and good day to all.

I was looking at the Spikemax implementation in the case that a moving window is specified. It makes sense to me that with a moving window implementation you would have to have a label for every timestep for every class. The comments seem to support this by specifying a (batch, num_class, time) format. However, I notice that in the MovingWindow confidence function (here) there is a reshape usage that goes like this:
sliding_rate = self.rate(spike).reshape( spike.shape[0], spike.shape[-1] )

This causes an error since the self.rate function produces a 3d tensor which makes the reshape invalid. I made this discussion post and not an issue because of the possibility that I may be using the loss function in the wrong way. If anybody can provide guidance on this I would appreciate it.

[bamsumit]

@Michaeljurado42 it looks like an error. It should have been
sliding_rate = self.rate(spike).reshape( spike.shape[0], -1, spike.shape[-1] )
Can you test if that works?

[Michaeljurado42]

I think that the reshape may be superflous because self.rate(spike).reshape(spike.shape[0], -1, spike.shape[-1] ) seems to be equivalent to self.rate(spike). Either way, I will make an issue and pull request for this fix and one other problem I recently noticed in the moving_window case soon.

[bamsumit]

Thanks @Michaeljurado42 Reshape is needed when you don't have dense layers at the end. In this case, you will have a 5 dimensional tensor. So it's better to enforce it.