TLDR

Official Implementation of TLDR: Text Based Last-layer Retraining for Debiasing Image Classifiers (WACV 2025) (Proceedings: https://openaccess.thecvf.com/content/WACV2025/papers/Park_TLDR_Text_Based_Last-Layer_Retraining_for_Debiasing_Image_Classifiers_WACV_2025_paper.pdf)

This code is based on the SpuCo repository (https://github.com/BigML-CS-UCLA/SpuCo)

Environment Setup

conda env create --file tldr.yaml

Training

• The code will automatically download Waterbirds, CelebA, and SpucoAnimals dataset.
• The code assume that COCO-2014 dataset is located at /data1/fiftyone/coco-2014. You can modify this path by changing https://github.com/beotborry/TLDR/blob/63bdf5ca1e5b4e6fa192f5096a5819fff6f61a6f/datasets/coco_dataset.py#L21C9-L21C48.

bash script.sh