Experiments for "Class-wise and Reduced Calibration Methods"

This repository contains the experiments for the paper by Panchenko, Benmerzoug and de Benito Delgado, Class-wise and reduced calibration methods, submitted to the 21st IEEE International Conference on Machine Learning and Applications (ICMLA 2022).

Pre-requisites

This project uses Poetry for dependency management. More specifically version 1.2.0 of Poetry.

Start by installing it and then proceed to installing the requirements:

poetry install --no-root

And then activating the created virtual environment:

poetry shell

As an alternative you could build a docker image from the Dockerfile:

docker build . -t classwise-calibration:local

And then simply start a container:

docker container run -it --rm classwise-calibration:local

To start the notebooks from within the container use:

docker run -it --rm -p 8888:8888 classwise-calibration:local jupyter notebook --NotebookApp.default_url=/lab/ --ip=0.0.0.0 --port=8888

Experiments

Experiment	Binder
Random Forest on Synthetic Data
LightGBM on Sorel20M
ResNet56 on CIFAR10
DistilBERT on IMDB
DeiT on RVL-CDIP

Running the Experiments

To run the experiments use:

python -m src.experiments.<Experiment Module>

Where you would replace with the name of one of experiments' module.

For example, to run the Random Forest experiment with Synthetic Data use:

python -m src.experiments.random_forest_synthetic

Notebooks

You could also use the notebooks if we to interactively run the experiments.

They are generated from the experiment scripts as follows:

bash scripts/generate_notebooks.sh

License

This repository is distributed under LGPL-3.0. A complete version can be found in two files: here and here.

All contributions will be distributed under this license.