The discriminative step is now limited to the capacity of one GPU We must be able to perform the training of all single models on several GPUs.