NLN: Nearest-Latent-Neighbours

A repository containing the implementation of the paper entitled Improving Novelty Detection using the Reconstructions of Nearest Neighbours

Installation

Install conda environment by:

    conda create --name nln python=3.7

Run conda environment by:

    conda activate nln

Install dependancies by running:

    pip install -r dependancies

Additionally for training on a GPU run:

    conda install -c anaconda tensorflow-gpu=2.2.0

Replication of results in paper

Run the following to replicate the results for MNIST, CIFAR-10, Fashion-MNIST and MVTec-AD respectively

    sh experiments/run_mnist.sh
    sh experiments/run_cifar.sh
    sh experiments/run_fmnist.sh
    sh experiments/run_mvtec.sh

Or to execute all experiments sequentially the following script can be run:

    sh experiments/run_all.sh

MVTec-AD usage

You will need to download the MVTec anomaly detection dataset and specify the its path using -mvtec_path command line option.

Training

Run the following:

    python main.py -anomaly_class <0,1,2,3,4,5,6,7,8,9,bottle,cable,...> \
                   -percentage_anomaly <float> \
                   -limit <int> \
                   -epochs <int> \
                   -latent_dim <int> \
                   -data <MNIST,FASHION_MNIST,CIFAR10,MVTEC> \
                   -mvtec_path <str>\
                   -neighbors <int(s)> \
                   -algorithm <knn> \
		   -patches <True, False> \
		   -crop <True, False> \
		   -rotate <True, False> \
		   -patch_x <int> \    
		   -patch_y <int> \    
		   -patch_x_stride <int> \    
		   -patch_y_stride <int> \    
		   -crop_x <int> \    
		   -crop_y <int> \    

Reporting Results

Run the following given the correctly generated results files:

    python report.py -data <MNIST,CIFAR10,FASHION_MNIST,MVTEC> -seed <filepath-seed>

Licensing

Source code of NLN is licensed under the MIT License.