VIbCReg

This repository accompanies the VIbCReg paper. It contains code to reproduce the experiments made in the paper, and to use the VIbCReg methodology for other experiments.

Usage

Before running the experiments, set up a conda environment running

conda env create -f environment.yaml 

Then activate the environment by running

conda activate vibcreg 

Then download the dataset used in the experiment by running

python -m vibcreg.data.download_data 

Now, you are ready to run experiments. Note that we use weights and biases (https://wandb.ai/) for logging the experiments, and you may need to configure wandb before actually running the experiments.

The default experiment can be run as follows

python -m vibcreg.examples.learn_representations 

And to run a specific configuration, run

python -m vibcreg.examples.learn_representations --config_dataset config_dataset_filepath --config_framework config_framework_filepath --device_ids "0"

After training the model, it is possible to evaluate the model by running

python -m vibcreg.examples.evaluations

with appropriate configuration files, if not using the default configs.