Unsupervised Learning of Anomaly Detection from Contaminated Image Data using Simultaneous Encoder Training

This repo contains the source code presented in the ArXiV paper with the same name:

A. Berg, J. Ahlberg, and M. Felsberg, Unsupervised Learning of Anomaly Detection from Contaminated Image Data using Simultaneous Encoder Training, https://arxiv.org/abs/1905.11034, (2019)

About the Code

The source code is an altered version of the Tensorflow version of the "Progressive Growing of GANs" code [1].

More information on system requirements, how to import and use pre-trained networks, how to prepare your own datasets etc. can be found here: https://github.com/tkarras/progressive_growing_of_gans

How to run the code

1. Open config.py
2. Alter the file depending on what you want to do (see below)
3. Run train.py

Training

Start by making sure that all rows below header "Utility scripts" are commented. Put your dataset in the dataset folder (need to be in .tfrecord format). I have included an mnist dataset as an example. Again, if you want to use another dataset, see https://github.com/tkarras/progressive_growing_of_gans for more information on how to pre-process training data. I have added a few functions in dataset_tool.py for MNIST, LSUN, CIFAR etc. In contrast to the functions in the original dataset_tool.py they also export validation data as .png's.

Rows 60 to 102 show a bunch of examples on how to configure which dataset to use. If you want to train with the example mnist dataset, just leave the following line uncommented:

desc += '-mnist_2_0.0';               dataset = EasyDict(tfrecord_dir='mnist_2_0.0')

Anomaly detection

Uncomment the anomaly detection example in config.py (row 153) and edit run_id (the id of the folder where the results are saved) and test_data_folder. The test_data_folder should contain .png files of the same dimensions as the images used for training.

MNIST example dataset

I have included an example dataset. Unpack the datasets.zip folder in order to use it. In this dataset, MNIST number 2 is considered to be normal samples and all other numbers anomalies. There are no anomalies in the training data.

Finally

Don't hesitate to contact me if you have any questions. Happy coding!

References

[1] T. Karras, T. Aila, S. Laine, and J. Lehtinen. Progressive Growing of GANs for Improved Quality, Stability, and Variation. In ICLR 2018, oct 2017.