cudnn.convert() isn't converting ReLU() modules back to nn about cudnn.torch

[szagoruyko]

self.mode in ReLU should not be a problem, it is not used in nn and does not depend on cudnn. If you have cudnn modules inside rnn's that's probably the issue, clearState is not propagating into rnn modules.

[willfrey]

Thanks for the response!

Do you have any suggestions for manually fixing this? I'm digging around myself but I'd still consider myself a Torch novice.

Edit: I've saved my cudnn-tainted CPU model as an ascii file and I'm going to use sed to replace all instances of cudnn.ReLU with nn.ReLU. We'll see if it works.

[willfrey]

Closing this as it probably doesn't matter much with the release of R5 and the native cudnn RNN modules.