Comment by code-is-code

Training in parallel on more then one GPU is pretty hard. Distributing the training over many random computers and adding at least some kind of validation, seems to kill the performance. So it would always be easier to rend cloud GPUs.