Model can't be moved to another device after inference

[mzemlyanikin]

If we create a model on cpu and run it, than copies of convolutions' weights (weight.original) are saved on cpu. If we run model.cuda(), they stay on cpu.