Possible invalid memory access?

[cpennington]

I think the im2col (and maybe col2im) can potentially read uninitialized memory: Running Grenade.Layers.Internal.Convolution.im2col 10 10 2 2 $ (5><6) [0..] gives a bunch of small float values, which seem likely to be because of reading invalid memory locations.

[HuwCampbell]

im2col isn't a total function, it only works where the kernel and stride "fit" for the input the matrix. So yes, it does look like that's what's happening.

This "fitting" constraint is captured in the proofs going into the convolution and deconvolution layers. I found that if I tried to push these through to the im2col function it lead to a considerable slowdown, so put them in an internal function set instead.

It could probably be better documented, and maybe even just stop with error inside im2col if these constraints are violated.

[cpennington]

Yeah, it seems like either erroring out in im2col when the arguments aren't valid would be a good safety net.