Multiresolution Knowledge Distillation for Anomaly Detection

This repository contains code for training and evaluating the proposed method in our paper Multiresolution Knowledge Distillation for Anomaly Detection.

Citation

If you find this useful for your research, please cite the following paper:

@article{salehi2020distillation,
  title={Multiresolution Knowledge Distillation for Anomaly Detection},
  author={Salehi, Mohammadreza and Sadjadi, Niousha and Baselizadeh, Soroosh and Rohban, Mohammad Hossein and Rabiee, Hamid R},
  year={2020},
  eprint={2011.11108},
  archivePrefix={arXiv},
  primaryClass={cs.CV}
}

1- Clone this repo:

git clone https://github.com/Niousha12/Knowledge_Distillation_AD.git
cd Knowledge_Distillation_AD

2- Datsets:

This repository performs Novelty/Anomaly Detection in the following datasets: MNIST, Fashion-MNIST, CIFAR-10, MVTecAD, and 2 medical datasets (Head CT hemorrhage and Brain MRI Images for Brain Tumor Detection).

Furthermore, Anomaly Localization have been performed on MVTecAD dataset.

MNIST, Fashion-MNIST and CIFAR-10 datasets will be downloaded by Torchvision. You have to download MVTecAD, Retina, Head CT Hemorrhage, and Brain MRI Images for Brain Tumor Detection, and unpack them into the Dataset folder.

For Localization test you should remove the good folder in {mvtec_class_name}/test/ folder.

3- Train the Model:

Start the training using the following command. The checkpoints will be saved in the folder outputs/{experiment_name}/{dataset_name}/checkpoints.

Train parameters such as experiment_name, dataset_name, normal_class, batch_size and etc. can be specified in configs/config.yaml.

python train.py --config configs/config.yaml

4- Test the Trained Model:

Test parameters can also be specified in configs/config.yaml.

python test.py --config configs/config.yaml

✨ For the transformer-based implementation of KDAD, please check out GeneralAD/kdad_vit.