Contrastive Corpus Attribution (COCOA)

Code repository for Contrastive Corpus Attribution for Explaining Representations.

Environment setup

1. Git clone or download this repository.
2. cd cl-explainability.
3. Create and activate the specified conda environment by running

   conda env create -f environment.yml
   conda activate cl-explain-env
   

4. Install the cl_explain package and the necessary dependencies for development by running pip install -e ".[dev]".

Set up project paths

Modify global constants in scripts/constants.py for paths where the image data, encoder models, and results are stored.

Run experiments

• To train a ResNet18 model for MURA, execute python scripts/train_classifier.py. Run python scripts/train_classifier.py --help to see how to use each command line argument. Please see our paper for how to obtain a trained SimCLR model for ImageNet and a trained SimSiam model for CIFAR-10.
• To run feature attributions, execute python scripts/attribute.py. Run python scripts/attribute.py --help to see how to use each command line argument.
• To evaluate feature attributions, execute python scripts/eval.py. Run python scripts/eval.py --help to see how to use each command line argument.