Skip to content
/ gcd Public

Official implementation of Global Counterfactual Directions accepted to ECCV 2024.

Notifications You must be signed in to change notification settings

sobieskibj/gcd

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

2 Commits
 
 
 
 
 
 
 
 
 
 

Repository files navigation

[ECCV 2024] Global Counterfactual Directions

Official implementation of Global Counterfactual Directions accepted to ECCV 2024.

teaser

Virtual environment

We provide requirements.txt file generated with pip to create a virtual environment with all required packages.

Required files

To run the experiments, it is required to download:

  • DiffAE weights
  • Classifier weights
  • Dataset

We provide weights through the attached link. We do not include the FFHQ DiffAE checkpoint as it can be obtained through the original paper's repository. For datasets, we refer to the following sources:

Our codebase assumes that images are placed as .png files, with each file in a separate folder.

We also include example gradient-based directions used in computing the results included in the paper.

To run proxy training and/or direction transfer, move to gcd. To create attribution maps, move to bblig.

Proxy training

We include Hydra configuration files that allow for running the proxy training loop (see configs directory). Each field with ??? is required and must be filled by the user for the code to work correctly. To run the experiment, use

python src/main.py --config-path ../configs/single_image_gmc_mlp_proxy_training/DATASET/MODEL/CLASS

where DATASET, MODEL and CLASS are specified by the subdirectories in the configs directory.

Direction transfer

Proxy training phase automatically creates a log subdirectory in the outputs directory. To perform direction transfer on a different image, move to gcd directory (aka cd gcd) and use

python src/direction_transfer.py --log-dir-path LOG_DIR_PATH --direction-path DIRECTION_PATH --img-path IMG_PATH --subdir-name SUBDIR_NAME

where LOG_DIR_PATH is a path to a log directory created as a subdirectory of outputs. Directions found throughout proxy training phase can be found in the proxy subdirectory of the log directory, and DIRECTION_PATH should point to a .pt file with such direction. The IMG_PATH argument is a path to .png file with the image on which you want to perform direction transfer. The SUBDIR_NAME argument gives a name to a subdirectory where the results of the transfer are stored (inside the direction_transfer subdirectory of the log directory).

BB-LIG

To produce attribution maps with BB-LIG, use

python src/main.py --config-name sig

You may also customize the configs by following Hydra documentation.

Citation

@InProceedings{sobieski2024global,
author="Sobieski, Bartlomiej and Biecek, Przemyslaw",
title="Global Counterfactual Directions",
booktitle="European Conference on Computer Vision",
year="2024",
publisher="Springer Nature Switzerland",
}

About

Official implementation of Global Counterfactual Directions accepted to ECCV 2024.

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages