This repo is the official PyTorch implementation for the paper Diffusion Denoising Process for Perceptron Bias in Out-of-distribution Detection

by Luping Liu, Yi Ren, Xize Cheng, Zhou Zhao (Zhejiang University).

In this work, we provide a new perceptron bias assumption that the discriminator models are more sensitive to some subareas to explain the overconfidence problem. Our detection methods combine the discriminator and generation models, which uses a ResNet to extract features and the diffusion denoising process of a diffusion model (with classifier-free guidance) to reduce overconfidence areas. Our methods get competitive OOD detection results with SOTA methods.

• Model: DDIM, iDDPM
• Dataset: Cifar10, Cifar100
• Runner:
  • runner: basic training and sampling
  • ood_detection: diffusion-based OOD detection

Run the following to install necessary python packages for our code.

pip install -r requirements.txt

Train the diffusion models through main.py.

python main.py --runner training --config config/ddim_cifar10_cond.yml --train_path temp/train/base_multi
torchrun --nproc_per_node 2 main.py --runner training --config config/ddim_cifar10_cond.yml --train_path temp/train/base_multi

• runner: choose the mode of runner
• device: choose the device to use
• config: choose the config file
• train_path: choose the path to save training status

Evaluate the diffusion models through main.py.

python main.py --runner fid --method PNDM4 --sample_step 50 --device cuda --config config/ddim_cifar10_cond.yml \
  --image_path temp/sample --model_path temp/models/ddim/ema_cifar10.ckpt

• method: choose the numerical methods
• sample_step: control the total generation step
• image_path: choose the path to save images
• model_path: choose the path of diffusion model

Generate samples for OOD detection through main.py.

python main.py --runner detection --config config/32_cifar10_cond.yml --method DDIM --sample_step 50 \
  --model_path temp/model/ddim_cifar10_cond.ckpt --disc_path temp/model/res18_cifar10_disc.ckpt --repeat_size 4

• disc_path: choose the path of discriminator model
• repeat_size: choose the number of repeat samples

Compute OOD detection results through detect.py.

python detect/ood_detect.py --id_name cifar10 --space logit  --repeat_size 4

• id_name: choose the name for in-distribution datasets
• space: choose the detection space generated by the discriminator model

All datasets, precalculated statistics for FID and checkpoints of models are provided in this Google Drive.

If you find the code useful for your research, please consider citing:

@misc{liu2022diffusion,
      title={Diffusion Denoising Process for Perceptron Bias in Out-of-distribution Detection}, 
      author={Luping Liu and Yi Ren and Xize Cheng and Zhou Zhao},
      year={2022},
      eprint={2211.11255},
      archivePrefix={arXiv},
      primaryClass={cs.CV}
}

This work is built upon some previous papers which might also interest you:

• Song, Jiaming, Chenlin Meng, and Stefano Ermon. "Denoising Diffusion Implicit Models." International Conference on Learning Representations. 2020.
• Liu, Luping, et al. "Pseudo Numerical Methods for Diffusion Models on Manifolds." International Conference on Learning Representations. 2021.
• Yang, Jingkang, et al. "OpenOOD: Benchmarking Generalized Out-of-Distribution Detection." Thirty-sixth Conference on Neural Information Processing Systems Datasets and Benchmarks Track.