This repository contains the code for our work Dissecting Supervised Contrastive Learning which was accepted at ICML'21 (see also the updated arxiv version).

Installation

This code requires a pytorch installation (tested on version 1.13.1).

Additionally, the following packages are required: scikit-learn, fastprogress, pandas,

    pip install scikit-learn fastprogress pandas

Application

To reproduce the experiments from Section 5.2 (Theory vs. Practivce), run

    python run_exp_performance.py 

To reproduce the random label experiments from Section 5.3, run

    python run_exp_noisy_labels.py 

Experiments can be evaluated with the notebooks results_performance.ipynb, or results_noisy_labels.ipynb, respectively.

Notebooks to reproduce Figures 1, 5 and 6 can be found in the notebooks/ directory.
These notebooks further include animations to visualize the convergence towards the simplex configuration when optimizing the representations directly.

Reference

@inproceedings{Graf21a,
  author          = {Graf, Florian and Hofer, Christoph and Niethammer, Marc and Kwitt, Roland},
  title           = {Dissecting Supervised Contrastive Learning},
  booktitle       = {ICML},
  year            = {2021}
}