- images are now loaded for each batch instead of all at the start in lambda_coco.py - checkpoint location is now checked for existing weights and loaded if available