What If the Input is Expanded in OOD Detection?

This repository contains the source codes for reproducing the results of NeurIPS'24 paper: What If the Input is Expanded in OOD Detection?.

Author List: *Boxuan Zhang, *Jianing Zhu, Zengmao Wang, Tongliang Liu, Bo Du, Bo Han.

(* Equal Contribution)

Required Packages

Our experiments are conducted on Ubuntu Linux 20.04 with Python 3.10 and Pytorch 2.2. Besides, the following packages are required to be installed:

• transformers
• scipy
• scikit-learn
• matplotlib
• seaborn
• pandas
• tqdm

Checkpoints

For DNN models, our reported results are based on checkpoints of ResNet-50. The official checkpoints can be downloaded from pytorch. In addition, ASH provide ResNet-50 checkpoints for its implementation, please refer to its code repository. Please download the corresponding checkpoints in DNNs/checkpoints/.

For CLIP models, our reported results are based on checkpoints provided by Open-CLIP. Similar results can be obtained with checkpoints in the codebase by OpenAI.

Data Preparation

Please download or create the datasets in folder:

./datasets/

Original In-distribution Datasets

We consider the following (in-distribution) datasets:

• ImageNet-1k, ImageNet-10, ImageNet-20, ImageNet-100

The ImageNet-1k dataset (ILSVRC-2012) can be downloaded here.

For ImageNet-10, 20, and 100 in hard OOD detection, please follow the instructions in MCM.

Original Out-of-Distribution Datasets

We consider the following (out-of-distribution) datasets:

• iNaturalist, SUN, Places, Texture

Following MOS, we use the large-scale OOD datasets iNaturalist, SUN, Places, and Texture. To download these four test OOD datasets, one could follow the instructions in the code repository of MOS.

Corrupted ID and OOD datasets

We use the corruptions defined in Hendrycks et al. 2019 to expand the original single input dimension into a multi-dimensional one. The official ImageNet-C dataset can be downloaded here, which has all 1000 classes where each image is the standard size. We also provide a copy of the official implementation code for 18 types of corruptions at 5 severity levels in utils/imagenet_c/make_imagenet_c.py.

This implementation can also be used to implement the four corrupted OOD datasets, e.g. iNaturalist-C, SUN-C, Places-C, Texture-C and other ID datasets, e.g. ImageNet-10-C, ImageNet-20-C, ImageNet-100-C.

To create a Corruption dataset, the following script can be used:

python utils/imagenet_c/make_imagenet_c.py

We use the SVHN dataset as the validation set to determine the most effective corruption types for each method in all experiments. The SVHN-C can also be created by the above script.

Overall Structure

After introducing the corrupted datasets for input expansion, the overall file structure is as follows:

CoVer
|-- datasets
    |-- ImageNet
    |-- DistortedImageNet
    |-- ImageNet_OOD_dataset
        |-- Distorted
          |-- iNaturalist
          |-- dtd
          |-- SUN
          |-- Places 
        |-- iNaturalist
        |-- dtd
        |-- SUN
        |-- Places
    ...

OOD Detection Evaluation

The main script for evaluating OOD detection performance is eval_ood_detection.py. Here are the list of arguments:

• --name: A unique ID for the experiment, can be any string
• --score: The OOD detection score, which accepts any of the following:
  • CoVer: our CoVer mode
  • energy: The Energy score
• --seed: A random seed for the experiments
• --gpu: The index of the GPU to use. For example --gpu=0
• --in_dataset: The in-distribution dataset
• -b, --batch_size: Mini-batch size
• --CLIP_ckpt: Specifies the pre-trained CLIP encoder to use
  • Accepts: ViT-B/32, ViT-B/16, ViT-L/14.

The OOD detection results will be generated and stored in results/in_dataset/score/CLIP_ckpt/name/.

Furthermore, the corruptions are selected in the main script eval_ood_detection.py, a variable named imagenet_c. We also recommend several corruption types through the validation set verification, please feel free to try in eval_ood_detection.py.

We provide bash scripts to help reproduce the numerical results of our paper.

To evaluate the performance of CoVer on ImageNet-1k based on CLIP-B/16:

sh scripts/eval_mcm.sh CLIP eval_ood ImageNet CoVer

To evaluate the performance of CoVer on ImageNet-1k based on ResNet-50:

sh scripts/eval_mcm.sh ResNet50 eval_ood ImageNet CoVer

Citation

If you find our paper and repo useful, please cite our paper:

@inproceedings{zhang2024what,
  title={What If the Input is Expanded in OOD Detection?},
  author={Zhang, Boxuan and Zhu, Jianing and Wang, Zengmao and Liu, Tongliang and Du, Bo and Han, Bo},
  booktitle={The Thirty-Eighth Conference on Neural Information Processing Systems},
  year={2024},
}