Disentangled Latent Spaces Facilitate Data-Driven Auxiliary Learning

Official implementation of the paper Disentangled Latent Spaces Facilitate Data-Driven Auxiliary Learning accepted at the 23rd International Conference on Image Analysis and Processing (ICIAP 2025).

Installation

1. Repository setup:

• $ git clone https://github.com/intelligolabs/Detaux
• $ cd Detaux
• $ git clone https://github.com/google-research/disentanglement_lib.git
• $ mv disentanglement_lib disentanglement_library
• $ cp -r disentanglement_lib_patch/* disentanglement_library/disentanglement_lib/
• Download 3dshapes.h5 from https://console.cloud.google.com/storage/browser/3d-shapes;tab=objects?prefix=&forceOnObjectsSortingFiltering=false

2. Conda enviroment setup:

• $ conda create -n detaux python=3.7
• $ conda activate detaux
• $ python -m pip install pytorch-lightning==1.9.4
• $ cd disentanglement_library/
• $ python -m pip install -v -e .
• $ python -m pip install tensorflow-gpu==1.14
• $ python -m pip install --upgrade tensorboard
• $ cd ../
• $ python -m pip install wandb
• $ pip install torchvision

Run Detaux

1. To run the disentanglement part, use the file detaux.py. In particular, launch_dis.sh it contains one example of a launch script that you can use to modify the default configuration directly.
2. To run the clustering part, use the file clustering.py.
3. Finally, with the file aux_learning.py, you will be able to perform the auxiliary learning phase with the new labels discovered in step 2.

Credits

We want to thank Marco Fumero for the repository PMPdisentanglement, which provides us with the scripts used to manage the disentanglement part.

Authors

Geri Skenderi¹, Luigi Capogrosso², Andrea Toaiari², Matteo Denitto³, Franco Fummi², Simone Melzi⁴

¹ Bocconi University, Bocconi Institute for Data Science and Analytics, Milan, Italy

² University of Verona, Dept. of Engineering for Innovation Medicine, Verona, Italy

³ HUMATICS - SYS-DAT Group, Verona, Italy

⁴ University of Milano-Bicocca, Dept. of Informatics, Systems and Communication, Milan, Italy

Citation

If you use Detaux, please, cite the following paper:

@Article{skenderi2023disentangled,
  title   = {{Disentangled Latent Spaces Facilitate Data-Driven Auxiliary Learning}},
  author  = {Skenderi, Geri and Capogrosso, Luigi and Toaiari, Andrea and Denitto, Matteo and Fummi, Franco and Melzi, Simone and Cristani, Marco},
  journal = {arXiv preprint arXiv:2310.09278},
  year    = {2023}
}