Preprocessing on GPU

[zclawr]

Try converting np arrays to PyTorch tensors and running preprocessing on these, then convert back to np array for passing into the model. Ideally get some number comparisons in terms of performance between the CPU preprocessing and GPU preprocessing.