Identity Curvature Laplace Approximation

This repo contains official implementation of Identity Curvature Laplace Approximation for Improved Out-of-Distribution Detection [WACV 2025] by Maksim Zhdanov, Stanislav Dereka and Sergey Kolesnikov. The implementation is based on OpenOOD.

Preparing the environment

1. Build Miniconda environment from env_laplace.yml by running conda env create -f env_laplace.yml.
2. Alternatively, you can build Docker image from Dockerfile.
3. Additionally, you need to install asdfghjkl backend by running pip install git+https://[email protected]/wiseodd/asdl@asdfghjkl.
4. Download the datasets:

bash /scripts/download/download.sh

Note. If some links in download.sh are broken, download the datasets by yourself from the official repositories and refer to ./scripts/imglist_generator.py to prepare image lists in OpenOOD format.

Reproducing the results

As we use wandb as a monitoring tool, you need to set wandb environment variables in .env file:

WANDB_API_KEY=<YOUR_KEY>

Note: ICLA and LLLA variant are all post-hoc methods which do not require any training.

Note: all the experiments on cifar10, cifar100, imagenet-200 are guaranteed to run on a single Nvidia A100 GPU.

Evaluation script parameters of interest

In evaluation .sh scripts, you can tweak the following parameters:

1. EXP_NAME='name' - the name of your experiment in Wandb.
2. export CUDA_VISIBLE_DEVICES=0 - ID of your CUDA device.
3. --postprocessor.postprocessor_args.llla_type 'type' - type of Laplace Approximation to use. Supported variants are ef, ggn, k-fac and icla.
4. --postprocessor.postprocessor_args.optimize_precision true - optimize prior precision or not. This corresponds to ICLA and ICLA (zero) algorithms in the paper.
5. --recorder.project icla - the name of your Wandb project.

Run one of the following commands to reproduce the results from the paper. You can use LLLA with EF, GGN, K-FAC or ICLA (ours).

CIFAR-10

Empirical Fisher (EF)

bash scripts/ood/icla/cifar10_llla_ef_eval.sh 

Generalized Gauss-Newton (GGN).=

bash scripts/ood/icla/cifar10_llla_ggn_eval.sh

Kronecker-Factored Approximate Curvature (K-FAC)

bash scripts/ood/icla/cifar10_llla_kfac_eval.sh

Identity Curvature Laplace Approximation (ICLA)

bash scripts/ood/icla/cifar10_llla_icla_eval.sh

Identity Curvature Laplace Approximation (ICLA) with Precision Zero Hessian Optimization

bash scripts/ood/icla/cifar10_llla_icla_zero_eval.sh

CIFAR-100

Empirical Fisher (EF)

bash scripts/ood/icla/cifar100_llla_ef_eval.sh 

Generalized Gauss-Newton (GGN)

bash scripts/ood/icla/cifar100_llla_ggn_eval.sh

Kronecker-Factored Approximate Curvature (K-FAC)

bash scripts/ood/icla/cifar100_llla_kfac_eval.sh

Identity Curvature Laplace Approximation (ICLA)

bash scripts/ood/icla/cifar100_llla_icla_eval.sh

Identity Curvature Laplace Approximation (ICLA) with Precision Zero Hessian Optimization

bash scripts/ood/icla/cifar100_llla_icla_zero_eval.sh

ImageNet-200

Empirical Fisher (EF).

bash scripts/ood/icla/im200_llla_ef_eval.sh 

Generalized Gauss-Newton (GGN).

bash scripts/ood/icla/im200_llla_ggn_eval.sh

Kronecker-Factored Approximate Curvature (K-FAC).

bash scripts/ood/icla/im200_llla_kfac_eval.sh

Identity Curvature Laplace Approximation (ICLA)

bash scripts/ood/icla/im200_llla_icla_eval.sh

Identity Curvature Laplace Approximation (ICLA) with Precision Zero Hessian Optimization

bash scripts/ood/icla/im200_llla_icla_zero_eval.sh

Citing

If you used this code for your research or a project, please cite us as:

@misc{zhdanov2024identitycurvaturelaplaceapproximation,
      title={Identity Curvature Laplace Approximation for Improved Out-of-Distribution Detection}, 
      author={Maksim Zhdanov and Stanislav Dereka and Sergey Kolesnikov},
      year={2024},
      eprint={2312.10464},
      archivePrefix={arXiv},
      primaryClass={cs.LG},
      url={https://arxiv.org/abs/2312.10464}, 
}