Problems with nn.ConcatTable

[davidstutz]

Hi,

I have difficulties when using net-toolkit with a network that contains a nn.ConcatTable layer. It usually fails in craftGradNBias at

module.gradInput = torch.Tensor():typeAs(module.output)

I tried several workarounds. I guess the problem is that gradInput as well as output might be tables. So I got some workarounds working, i.e. I am able to save models. But when I load them again, I get illegal memory accesses.

Any idea?

Thanks, David

[Atcold]

Can you provide me the smallest net and code snippet which will break the package? I can try to play around a bit.