Awesome

Out-of-Distribution Detection via Conditional Kernel Independence Model

This repository is the official PyTorch implementation of Conditional-i method.

0 Requirements

• Python 3.8
• PyTorch install = 1.8.0
• torchvision install = 0.9.0
• CUDA 10.2
• Other dependencies: numpy, sklearn, six, pickle, lmdb

1 Experiments on IN1K (inliers) and IN22K (outliers)

1.1 Training

We release a demo for the proposed Conditional-i method. The model is built based on ResNet-18 architecture.

To train Conditional-i for 100 epochs on ImageNet1K and ImageNet21K, run:

DATASET='in1k'
MODEL='r18_bank'
DIRNAME=${DATASET}_${MODEL}_conditional_i

python train.py \
    ${DATASET} \
    --model ${MODEL} \
    --hsic-sigma 4 \
    --cond-i-weight 0.06 \
    --shuffle-ood 1 \
    --sample-cls 1 \
    --save ./outputs/${DIRNAME}

1.2 Evaluation

We present a demo for our novel evaluation metric.

DIRNAME=dirname_demo

python test.py \
    --method_name ${DIRNAME} \
    --save dirname_demo \
    --load dirname_demo/checkpoints/ckp-99.pth \
    --num_to_avg 10

2 Experiments on CIFAR-100 (inliers) and 300K Random Images (outliers)

The 80 Million Tiny Images dataset seems to be suspended recently. We therefore will supplement the results of Table 1 by training Conditional-i on CIFAR-100 and 300K Random Images (A cleaned subset of the original 80 Million Tiny Images) instead. The results will come soon.

3 Citation

@inproceedings{wangout,
  title={Out-of-Distribution Detection via Conditional Kernel Independence Model},
  author={Wang, Yu and Zou, Jingjing and Lin, Jingyang and Ling, Qing and Pan, Yingwei and Yao, Ting and Mei, Tao},
  booktitle={Advances in Neural Information Processing Systems},
  year={2022}
}