AS-ViT Implement of Adaptive Sparse ViT (IJCAI23) requirement torch==1.9.0 torchvision==0.10.1 timm==0.4.12 training sh finetune.sh inference python3 infer.py --model-path path_to_model.pth --batch-size 256 --arch deit_small --data-path path_to_imagenet