Skip to main content

Anomaly Toolbox Powered by GANs

Project description

Anomaly Toolbox


Anomaly Toolbox Powered by GANs.

This is the accompanying toolbox for the paper "A Survey on GANs for Anomaly Detection" (

The toolbox is meant to be used by the user to explore the performance of different GAN based architectures (in our work aka "experiments"). It also already provides some datasets to perform experiments on:

We provided the MNIST dataset because the original works extensively use it. On the other hand, we have also added the previously listed datasets both because used by a particular architecture and because they contribute a good benchmark for the models we have implemented.

All the architectures were tested on commonly used datasets such as MNIST, FashionMNIST, CIFAR-10, and KDD99. Some of them were even tested on more specific datasets, such as an X-Ray dataset that, however, we could not provide because of the impossibility of getting the data (privacy reasons).

The user can create their own dataset and use it to test the models.

Quick Start

  • First thing first, install the toolbox
pip install anomaly-toolbox

Then you can choose what experiment to run. For example:

  • Run the GANomaly experiment (i.e., the GANomaly architecture) with hyperparameters tuning enabled, the pre-defined hyperparameters file hparams.json and the MNIST dataset: --experiment GANomalyExperiment --hps-path path/to/config/hparams.json --dataset 
  • Otherwise, you can run all the experiments using the pre-defined hyperparameters file hparams. json and the MNIST dataset: --run-all --hps-path path/to/config/hparams.json --dataset MNIST 

For any other information, feel free to check the help: --help


This work is completely open source, and we would appreciate any contribution to the code. Any merge request to enhance, correct or expand the work is welcome.


The structures of the models inside the toolbox come from their respective papers. We have tried to respect them as much as possible. However, sometimes, due to implementation issues, we had to make some minor-ish changes. For this reason, you could find out that, in some cases, some features such as the number of layers, the size of kernels, or other such things may differ from the originals.

However, you don't have to worry. The heart and purpose of the architectures have remained intact.


pip install anomaly-toolbox


  --experiment [AnoGANExperiment|DeScarGANExperiment|EGBADExperiment|GANomalyExperiment]
                                  Experiment to run.
  --hps-path PATH                 When running an experiment, the path of the
                                  JSON file where all the hyperparameters are
                                  located.  [required]
  --tuning BOOLEAN                If you want to use hyperparameters tuning,
                                  use 'True' here. Default is False.
  --dataset TEXT                  The dataset to use. Can be a ready to use
                                  dataset, or a .py file that implements the
                                  AnomalyDetectionDataset interface
  --run-all BOOLEAN               Run all the available experiments
  --help                          Show this message and exit.

Datasets and Custom Datasets

The provided datasets are:

and are automatically downloaded when the user makes a specific choice: ["MNIST", "CorruptedMNIST", "SurfaceCracks","MVTecAD"].

The user can also add its own specific dataset. To do this, the new dataset should inherit from the AnomalyDetectionDataset abstract class implementing its own configure method. For a more detailed guide, the user can refer to the file inside the src/anomaly_toolbox/datasets folder. Moreover, in the examples folder, the user can find a module with the basic skeleton code to implement a dataset.


Project details

Download files

Download the file for your platform. If you're not sure which to choose, learn more about installing packages.

Source Distribution

anomaly_toolbox-0.1.2.tar.gz (35.2 kB view hashes)

Uploaded Source

Built Distribution

anomaly_toolbox-0.1.2-py3-none-any.whl (61.1 kB view hashes)

Uploaded Python 3

Supported by

AWS AWS Cloud computing and Security Sponsor Datadog Datadog Monitoring Fastly Fastly CDN Google Google Download Analytics Microsoft Microsoft PSF Sponsor Pingdom Pingdom Monitoring Sentry Sentry Error logging StatusPage StatusPage Status page