The whole logic for learning, loading/saving pretrained weights requires proper interpretation of the argparser arguments,
Try to parametrize everything, so that basic learning is rather light without irrelevant logs (includes calculation of intermediate avg losses, losses per minibatch, ...).
argparser
arguments,