Multi-Class Hypersphere Anomaly Detection

This Repository contains the source code for the paper Multi-Class Hypersphere Anomaly Detection as presented at ICPR 2022.

You can find a minimal implementation here. A blogpost is here.

This repository is a fork of the lightning-hydra-template, so you might want to read their excellent instructions on how to use this software stack. Most of the implemented methods and datasets are taken from pytorch-ood.

# setup environment
conda env create --name mchad -f environment.yaml
conda activate mchad

# these would lead to conflicts or have been installed later
pip install aiohttp==3.7 async-timeout==3.0.1 tensorboardX==2.5.1

Experiments are defined in config/experiments. To run MCHAD on CIFAR10 run:

python run.py experiment=cifar10-mchad

Each experiment will create a results.csv file that contains metrics for all datasets, as well as a CSV log of the metrics during training, and a TensorBoard log.

You can override configuration parameters via the command line, such as:

python run.py experiment=cifar10-mchad trainer.gpus=1

to train on the GPU.

You can run experiments for multiple random seeds in parallel with hydra sweeps:

python run.py -m experiment=cifar10-mchad trainer.gpus=1 seed="range(1,22)"

We configured the Ray Launcher for parallelization. Per default, we run experiments in parallel on 21 GPUs. You might have to adjust config/hydra/launcher/ray.yaml.

To visualize the embeddings of MCHAD, you can use the following callback:

python run.py experiment=cifar10-gmchad callbacks=mchad_embeds.yaml

This callback will save the embeddings to the tensorboard in TSV format.

Download Pre-Trained Weights used for models:

wget -P data "https://github.com/hendrycks/pre-training/raw/master/uncertainty/CIFAR/snapshots/imagenet/cifar10_excluded/imagenet_wrn_baseline_epoch_99.pt"

Experiments can be replicated by running bash/run-rexperiments.sh, which also accepts command line overrides, such as:

bash/run-rexperiments.sh dataset_dir=/path/to/your/dataset/directory/

All datasets will be downloaded automatically to the given dataset_dir.

Results for each run will be written to csv files which have to be aggregated. You can find the scripts in notebooks/eval.ipynb.

To replicate the ablation experiments, run:

bash/run-ablation.sh dataset_dir=/path/to/your/dataset/directory/

We average all results over 21 seed replicates and several benchmark outlier datasets.

experiment=svhn-mchad trainer.gpus=1 model.weight_center=10.0 trainer.min_epochs=100 model.n_embedding=2

experiment=svhn-gmchad trainer.gpus=1 model.weight_center=10.0 trainer.min_epochs=100  model.n_embedding=2

If you use this code, please consider citing us:

@article{kirchheim2022multi,
	author = {Kirchheim, Konstantin and Filax, Marco and Ortmeier, Frank},
	journal = {International Conference on Pattern Recognition},
	number = {},
	pages = {},
	publisher = {IEEE},
	title = {Multi-Class Hypersphere Anomaly Detection},
	year = {2022}
}