Device placement decision

[raffeyC]

Unlike tensorflow, pytorch has a device system where models and data need to be placed explicitly on device before training can happen.
So, either the model or the application needs to accommodate this.
Pytorch models don't inherently have device parameters, so we would need to add our own in this case. The issue here is that if we load a saved model and move to a different device, we would also need to explicitly change the self.device parameter. This might get annoying and lead to unexpected results when training with multiple gpus (if this occurs)
On the other hand, we could add device parameter to the app, this is much cleaner and probably the best way forward. However, in this case, old users will have to add device parameter when declaring the app, so again not perfect.