Error with SpatialFullConvolution when running on GPU on Ubuntu 16.04 with Cuda 7.5 about dcgan.torch

[soumith]

should be fixed if you install the latest cudnn.torch bindings: luarocks install torch
fixed via commit: soumith/cudnn.torch@556c272

[ili3p]

Actually, it's not fixed. Or maybe is re-introduced. Anyway, I get the same error with the latest versions.

The problem is self.adjW and self.adjH are still not initialized when calling SpatialFullConvolution:createIODescriptors(...). SpatialFullConvolution:read(...) is never called when the network is created via torch.load(...), so the fix doesn't solve the problem.

One way to solve this is to add the two lines:

self.adjW = self.adjW or 0
self.adjH = self.adjH or 0

to cudnn.SpatialFullConvolution.lua line 111.

Although I'm not sure if this is the correct way and does not introduce new problems.