Does it matter what is the difference in size of needed inputs for inference vs. training?