Model | #param. | Top-1 Acc. | Config | Log |
---|---|---|---|---|
FastVim-T.ckpt | 7M | 75.4 | FastVimT.yaml | FastVimT.csv |
FastVim-S.ckpt | 26M | 81.1 | FastVimS.yaml | FastVimS.csv |
FastVim-B.ckpt | 98M | 82.6 | FastVimB.yaml | FastVimB.csv |
Vim-B w/ LN.ckpt | 98M | 82.6 | VimB.yaml | VimB.csv |
Notes:
- For reproducibility, make sure overall batch size remains 1024 across GPUs/Nodes.
- trainer/global_step in log files refers to gradient steps with batch size 1024.
- Modify
imagenet_train_dir_path
andimagenet_val_dir_path
flags in datasets_supervised.py code.