Deep Combinatorial Aggregation

This repository contains the source code for the NeurIPS 2022 paper "Deep Combinatorial Aggregation" by Yuesong Shen and Daniel Cremers.

[paper] [Arxiv link]

Usage

Dependencies

This implementation has been tested with

• python 3.9.7
• pytorch 1.10.0 (cuda 11.3)
• torchvision 0.11.1

It should work with Python >= 3.6 and latest pytorch / torchvision.

Setup

1. Prepare a python environment with the above dependencies.

2. Install the dca package locally by running

pip install -e .

in this folder.

3. Checkout the Python scripts in folder "experiments" to run the respective experiments:

• cifar10: In-domain CIFAR-10 experiments;
• svhn: In-domain SVHN experiments;
• cifar10c: Distributional shift experiments on CIFAR-10-C;
• ood: Out of distribution experiments;

Both the CIFAR-10-C and the OOD experiments require pre-trained models on CIFAR-10.

Citation

@InProceedings{shen2022dca,
title={Deep Combinatorial Aggregation},
author={Yuesong Shen and Daniel Cremers},
booktitle = {NeurIPS},
year = {2022}
}