README.md

Pretrained Model Weights and Configurations

Model	Config	Log
FastMaskVim-B.ckpt	FastMaskVim-B.yaml	FastMaskVim-B.csv
FastMaskVim-L.ckpt	FastMaskVim-L.yaml	FastMaskVim-L.csv
FastMaskVim-H.ckpt	FastMaskVim-H.yaml	FastMaskVim-H.csv
Vim-B.ckpt	Vim-B.yaml	Vim-B.csv
Vim-L.ckpt	Vim-L.yaml	Vim-L.csv

Notes:

• For reproducibility, make sure overall batch size remains 4096 across GPUs/Nodes. Flag accum_iter can be used.
• trainer/global_step in log files refers to gradient steps with batch size 4096.
• Modify imagenet_train_dir_path flag in datasets_mae.py.

Finetuned Model Weights and Configurations

Model	Top-1 Acc.	Config	Log
FastVim-B.ckpt	83.0	FastVim-B.yaml	FastVim-B.csv
FastVim-L.ckpt	84.9	FastVim-L.yaml	FastVim-L.csv
FastVim-H.ckpt	86.1	FastVim-H.yaml	FastVim-H.csv
FastVim-H_488.ckpt	86.7	FastVim-H_448.yaml	FastVim-H_448.csv
Vim-B.ckpt	83.3	Vim-B.yaml	Vim-B.csv
Vim-L.ckpt	85.1	Vim-L.yaml	Vim-L.csv

Notes:

• For reproducibility, make sure overall batch size remains 1024 across GPUs/Nodes. Flag accum_iter can be used.
• trainer/global_step in log files refers to gradient steps with batch size 1024.
• Modify imagenet_train_dir_path and imagenet_val_dir_path flags in datasets_finetune.py.