LBC

This is the code for the IJCAI 2024 Paper: Learning Robust Classifiers with Self-Guided Spurious Correlation Mitigation.

Preparation

Download datasets

• Waterbirds
• CelebA (metadata)
• NICO
• ImageNet (train ,val)
• ImageNet-A

Unzip the dataset files into individual folders.

In the config.py file, set {dataset}_DATA_ROOT to your corresponding dataset folder.

Prepare metadata.csv for each dataset

• Waterbirds and CelebA provide metadata.csv files.
• For the ImageNet-9 and ImageNet-A datasets, run the following code

  from datasets.in9_data import prepare_imagenet9_metadata, prepare_imageneta_metadata
  base_dir = "path/to/imagenet/folder"
  prepare_imagenet9_metadata(base_dir)
  data_root = "path/to/imagenet-a/folder"
  prepare_imageneta_metadata(data_root)

• For the NICO dataset, run the following code

  from datasets.nico_data import prepare_metadata
  prepare_metadata()

Extracting attributes

For each dataset, run the following code:

python extract_attributes.py --dataset waterbirds
python extract_attributes.py --dataset celeba
python extract_attributes.py --dataset nico
python extract_attributes.py --dataset imagenet-9

Pretrain ERM models

For example, run the following code for the ImageNet-9 dataset

python pretrain.py --dataset imagenet-9 --lr 0.001 --num_epochs 100 --pretrained_model

LBC training

python lbc_train.py --dataset imagenet-9 --lr 0.0001 --num_batches 100 --epoch 50 --K 4 --backbone resnet18
al

At the end of training, look for the corresponding results:

• Waterbirds, CelebA, and NICO: Find results starting with [pseudo_val_unbiased].
• ImageNet-9: Find results starting with [val_avg].

Citation

Please consider citing this paper if you find the code helpful.

@inproceedings{zheng2024learning,
 title={Learning Robust Classifiers with Self-Guided Spurious Correlation Mitigation},
 author={Zheng, Guangtao and Ye, Wenqian and Zhang, Aidong},
 booktitle={The 33rd International Joint Conference on Artificial Intelligence},
 year={2024}
}