Out of data error on Windows with 2 CUDA capable cards

[TheDailySpank]

Long story short, Torch doesn't like working with multiple GPUs on Windows (shared storage related issue).

Fix is to set n_gpu = 1 in train()

Any chance we could set n_gpu=1 as a default when running on Windows?

import os
if os.name == 'nt':
    n_gpu = 1

[minimaxir]

Sure, can do something like that. (w/ an appropriate log message)

[TheDailySpank]

Awesome. Thank you!

I had almost stopped trying to get it to work after about 3 days of tracking down the cause.