inference runs on CPU if no GPU available

[kemaleren]

I wanted to test this model on a device with no GPU. Even though it is much slower and would not be a good idea in production, for testing small images it still runs in a reasonable amount of time.

[Parskatt]

That's cool! I know other people have also expressed interest on running on CPU/other devices so this seems like a good addition. I'll just verify that your changes don't break anything first :)

[Parskatt]

I think it's probably better practice to have device as inputs to functions/properties of nn.Modules rather than defined outside the local scope, but that's a minor issue.

[kemaleren]

I thought so too, but the PyTorch blog has a section on writing device-agnostic code, that that's what they recommended: https://pytorch.org/blog/pytorch-0_4_0-migration-guide/#writing-device-agnostic-code

[kemaleren]

Also I'm sure there's things that still won't work. If they need to be fixed I can do that. I made these changes to get the demo_match.py script working.

[Parskatt]

I'll merge it and fix any issues later :p