TODOs Fix load_checkpoint and save_checkpoint to work with greedy pre-training Refactor code into two repos: fastor fastor_experiments Add color augmentation Make color augmentation faster