How do you choose between transformers, CNNs, or hybrid models for a computer vision task under hardware limits?
