ood-mode-ensemble

PyTorch implementation of the paper Revisiting Deep Ensemble for Out-of-Distribution Detection: A Loss Landscape Perspective. journal, arxiv.

If our work is helpful for your research, please consider citing:

@article{Fang2024,
author={Fang, Kun and Tao, Qinghua and Huang, Xiaolin and Yang, Jie},
title={Revisiting Deep Ensemble for Out-of-Distribution Detection: A Loss Landscape Perspective},
journal={International Journal of Computer Vision},
year={2024},
month={Jul},
day={15},
issn={1573-1405},
doi={10.1007/s11263-024-02156-x},
url={https://doi.org/10.1007/s11263-024-02156-x}
}

Introduction

Our work is summarized as follows:

• Models trained independently w.r.t different random seeds converge to isolated modes.
• These independent modes, which all reach low-loss regions with in-distribution data, yet yield significantly different loss landscapes with out-distribution data, which further suggests significantly fluctuating OoD detection performance across independent modes and has long been ignored by the research community.
• Motivated by such diversities on OoD loss landscape across modes, we revisit the deep ensemble method for OoD detection through mode ensemble, and design corresponding ensemble strategies for different types of OoD detectors, leading to improved performance and benefiting the OoD detector with reduced variances.

The following table appears as an example to show the high FPR variances across modes and the improved performance of mode ensemble, where the RankFeat detector is executed on 5 independent modes (DN121-ImageNet).

modes	iNaturalist	SUN	Places	Texture
mode-1	66.01	$\underline{75.53}$	$\underline{79.95}$	43.60
mode-2	58.49	$\underline{34.70}$	$\underline{50.70}$	32.73
mode-3	59.53	50.07	63.27	40.64
mode-4	$\underline{84.70}$	69.57	76.45	$\underline{49.89}$
mode-5	$\underline{46.58}$	44.46	58.95	$\underline{22.48}$
ensemble	39.32	39.48	55.61	15.98

Overview of this repository

A description on the files contained in this repository.

Training

1. train_c10.py: training isolated modes w.r.t different random seeds on CIFAR10
2. train_imgnet.py: training isolated modes w.r.t different random seeds on ImageNet

Evaluation

1. eval_clean.py and eval_clean_ensemble.py: evaluation the clean accracy of single modes and ensembling modes, respectively
2. eval_ood.py and eval_ood_ensemble.py: evaluation the OoD detection performance of single modes and ensembling modes, respectively

Others

1. utils_ood.py: A collection on the utility functions of OoD detectors
2. utils.py: Utility functions
3. utils_knn/: Utility functions on the kNN method
4. utils_mahalanobis/: Utility functions on the Mahalanobis method

Getting started

Install dependencies

conda create -n ood python=3.8
conda activate ood
conda install pytorch torchvision cudatoolkit=11.3 -c pytorch # for Linux
pip install pandas, scipy, scikit-learn, tensorboard
pip install statsmodels

Install faiss package following its docs.

dataset preparation

• InD datasets are the CIFAR10 and ImageNet-1K, respectively.
• OoD datasets are based on the KNN method. Follow the intructions in KNN to prepare the OoD datasets.

A full collection of all the training and evaluation commands can be found in EXPERIMENTS.md.

Released trained-models

Our models trained w.r.t different random seeds, including R18-C10, WRN28X10-C10, R50-ImgNet, DN121-ImgNet and T2T-ViT-14-ImgNet, are released here.

Download these models and put them in ./save/ as follows

ood-mode-ensemble
├── model
├── utils_knn
├── utils_mahalanobis
├── save
|   ├── CIFAR10
|   |   ├── R18
|   |   └── WRN28X10
|   |       ├── seed-1000
|   |       |   └── epoch150.pth
|   |       ├── ...
|   |       └── seed-2400
|   └── ImageNet
|       ├── DN121
|       |   ├── seed-1000
|       |   |   └──checkpoint.pth.tar
|       |   ├── seed-2000
|       |   ├── ...
|       |   └── seed-5000
|       ├── R50
|       └── t2tvit 
├── ...

Additional references

The loss landscape visualization techniques follow mode-connectivity and loss-surface.

If u have problems about the code or paper, u could contact me (fanghenshao@sjtu.edu.cn) or raise issues here.

If the code benefits ur researches, welcome to fork and star ⭐ this repo and cite our paper! :)