If I have a dataset with hundreds of pics, can I get away with training less steps/epochs?