zeligism / padpaf

PaDPaF

Official repo for the paper "PaDPaF: Partial Disentanglement with Partially-Federated GANs" (Published in TMLR 2024).

For federated learning problems, we introduce a personalized generative model that shows the benefits of having local adaptors, such as a local style generator. In the image below, the local adaptor for this regression task is simply chosen to be an additive bias term.

Running Experiments

Install the requirements:

pip install -r requirements. txt

The main code is in run_gan.py:

cd fl_sim
python run_gan.py <args>

If on SLURM cluster, you can submit a job:

sbatch job-<task>.sh

Otherwise, just run the command in the job script directly.

Notebooks

Notebooks are mainly for smaller experiments and visualization. You have to have a pre-trained model to run them.