Model soups: averaging weights of multiple fine-tuned models improves accuracy without increasing inference time
