Below you will find the steps to follow to obtain the results from the paper "Anomaly Detection With Partitioning Overfitting Autoencoder Ensembles". This has only been tested on ubuntu 18.04, but should run on any architecture supporting anaconda3 and tensorflow 2 installation.

After cloning the repository, you need to make sure that you have the necessary libraries installed. We use anaconda3 with tensorflow 2.

Then, on a command line, change to the project directory, and then type:

python -c 'import run; run.run()'

This will first create random datasets from Mnist and FMnist as described in the paper: there will be 50 datasets with the digit 0 the inliers and 50 datasets with the digit 1 the inliers. The same will be done for the FMnist dataset taking first class 1 (images of T-shirts/tops), then class 2 (images of trousers) as inliers. In total, this will create four different dataset collections of 50 datasets each. Unless you change the configurations, those dataset collections will be placed in the subdirectory generatedData/datasets.

Next, the above call will apply both a regularized deep convolutional autoencoder and a POTATOES model to those datasets, as described in the paper. By default, the evaluation results will be saved inside the data collection subdirectories next to the data. I.e. if the data for the data collection ova_mnist_bc0_rm0.005_s50 is saved under

generatedData/datasets/ova_mnist_bc0_rm0.005_s50/data

then the belonging evaluation data is saved under

generatedData/datasets/ova_mnist_bc0_rm0.005_s50/eval

as csv files.

Finally, the above call will take the four evaluation files from those four dataset collections and plot them as boxplots, so this should look similar to the evaluation plots in the paper, although, of course, not exactly like those, having used other randomly generated datasets. The combined evaluation data is saved in a csv file under generatedData/datasets together with the belonging plot, which is a pdf file with the same file name stem.

The above call then exits. You will need a GPU equipped machine to run this. If you just want to check whether everything is properly installed and don't want to weight several days for the program to finish, you should call:

python -c 'import run; run.run_small()'

which does the same as the first call, but with much smaller datasets and only very short training. So the results will be useless, this is only for checking your setup. It still took about 20 minutes to finish on my desktop machine.

Everything is configurable. Just have a look at the module run.py, it mainly contains the high level configuration code, so it should be easy to follow and adapt.