GenerativeAD.jl

Generative models for anomaly detection. This Julia package contains code for the paper "Comparison of Anomaly Detectors: Context Matters" arXiv preprint.

Installation

1. Clone this repo somewhere.
2. Run Julia in the cloned dir.

cd path/to/repo/GenerativeAD.jl
julia --project

3. Install all the packages using ]instantiate and compile the package.

(@julia) pkg> instantiate
(@julia) using GenerativeAD

Some of the bash scripts are calling julia without --project flag and uses @quickactivate macro to activate the environment, however this fails, unless DrWatson is installed in the base julia environment. In order to avoid these problems install DrWatson in your base environment.

cd ~
julia -e 'using Pkg; Pkg.add("DrWatson");'

Python instalation

Some models (PIDforest, scikit-learn, PyOD) are available only through PyCall with appropriate environment active. With upcoming bayesian optimisation from scikit-optimize every model will require an active environment, which can be setup in following way using python's venv module. (Most of the scripts have hardcoded path to this environment, though this can be easily changed).

cd ~
python -m venv sklearn-env

source ${HOME}/sklearn-env/bin/activate
export PYTHON="${HOME}/sklearn-env/bin/python"

Then install requirements inside this repository

cd path/to/repo/GenerativeAD.jl

pip install -r requirements.txt
pip install git+https://github.com/janfrancu/pidforest.git # not registerd anywhere

julia --project -e 'using Pkg; Pkg.build("PyCall");' # rebuilds PyCall.jl to point to the current environment

Running experiments on the RCI cluster

0. First, load Julia and Python modules.

ml Julia
ml Python

1. Install the package somewhere on the RCI cluster.
2. Then the experiments can be run via slurm. This will run 20 experiments with the basic VAE model, each with 5 crossvalidation repetitions on all datasets in the text file with 10 parallel processes for each dataset. All data will be saved in GenerativeAD.jl/data/experiments/tabular

cd GenerativeAD.jl/scripts/experiments_tabular
./run_parallel.sh vae 20 5 10 datasets_tabular.txt

Data:

Only UCI datasets are available upon installation via the UCI package. Remaining tabular and image datasets are downloaded upon first request (e.g. via the GenerativeAD.Datasets.load_data(dataset) function). First download requires user input to accept download terms for individual datasets. If you want to avoid this, do

export DATADEPS_ALWAYS_ACCEPT=true

before running Julia.