Storing model checkpoints to continue training at a later point