Transformers don't seem to optimal as an architecture for vision tasks, but I don't think pure convnets are either. Either way, having the right data is 10x as

Transformers don't seem to optimal as an architecture for vision tasks, but I don't think pure convnets are either. Either way, having the right data is 10x as important.