There are many neural networks to choose from, and in one category, there are many mutants of networks. It is pointless to evaluate all mutants and parameters, so a few are seleted randomly.

In the market, "Select" and "Wagyu" are hard to find. The imbalance of training dataset may influence models.

It is hard to grade steak after cooking. It will be interesting if the grade can be determined from the side surface of cooked steak.

Involving or augmenting samples in various lighting conditions

Add a new category "Unknown" for other meats and bad inputs

Change to patch-based networks to focus the marbling texture more