I am using Keras under Tensorflow for CNN training. Some computationally-expensive preprocessing is to be done on the training batches. The size of batches is dynamic, nonetheless, this can be handled to be static for the time being. Can PyTorch add anything in terms of development convenience and/or performance?