Zero grad on single parameter

[MariosOreo]

I think using with torch.no_grad() is a much better way than using .data which ptblck mentioned in this thread.