Fixing RNG seed

[shashankg7]

Hi,
I am using this code as benchmark in one of my papers. Now I need to fix the random seed in order to do so. I am trying the seed flag, - '-seed 42'. But even with this flag I am getting different results each time.

Can you guide me on this?

Thanks

[jeffreyling]

Are you using cudnn? This seems to be a known issue with convolution: soumith/cudnn.torch#92
I haven't found a great way to get around the problem myself.

[shashankg7]

I am using cudnn. So the only way to get around this problem is to use cpu instead of gpu?
This is sad.