The code of the paper "FairViT: Fair Vision Transformer via Adaptive Masking", this paper will appear at ECCV 2024.
Step 1: Create the conda environment through vit2024.yaml. The command is "conda env create -f vit2024.yaml".
Step 2: Download the CelebA dataset and place the decompressed Anno, Eval and img_align_celeba in "../nfs3/datasets/celeba". The file structure should be:
- nfs3/
datasets/
celeba/
Anno/
Eval/
img_align_data/ - Fair-Vision-Transformer/
README.md
...
Step 3: You can run the code by "python finetune_transformer.py" for debugging. To implement the default tasks in the paper, here are some quick commands:
python finetune_transformer.py --feature Smiling_Male -e 30 -r 1 -ta Smiling -sa Male -d celeba --gpu_en 0 -g 0.5 -s 19 -alr 3
python finetune_transformer.py --feature Attractive_Male -e 30 -r 1 -ta Attractive -sa Male -d celeba --gpu_en 0 -g 0.5 -s 19 -alr 3
python finetune_transformer.py --feature Attractive_Brown_Hair -e 30 -r 1 -ta Attractive -sa Brown_Hair -d celeba --gpu_en 0 -g 0.5 -s 19 -alr 3
To get further helps, please refer to "help" options in finetune_transformer.py.