GroupAD.jl

Benchmarking of Generative Anomaly Detection for Multiple Instance Learning problems. Inspired by GenerativeAD.jl.

Installation

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

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

3. Install all packages and download datasets.

]instantiate
using GroupAD
data = GroupAD.load_data("Fox")
# the last line should ask for permission to download datasets

4. Running a single experiment of VAE with 5-fold crossvalidation on the Tiger dataset.

cd scripts/experiments_mill
julia vae_basic.jl 5 Tiger

5. You can quickly evaluate the results using this recursive script.

julia GroupAD.jl/scripts/evaluate_performance_single.jl path/to/results

Project structure

Source files can be found in src. There are multiple modules used for utilities, and the model implementations and be found here.

Since every experiments is a little bit different, each group has its own experimental folder in the scripts folder:

• MIL datasets
• LHCO dataset
• point cloud datasets (MNIST, ModelNet10)

Each model has its own run script and bash script to submit to the cluster. Scripts for submitting experiments to run in parallel are also present. Always submit the run script from its script folder.

Running experiments on the RCI cluster

Note: Since LHCO dataset, Python is needed for data loading. Use Python/3.8 to install pandas.

0. First, load Julia and Python modules.

ml Julia
ml Python/3.8

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.

cd GroupAD.jl/scripts/experiments_mill
./run_parallel.sh vae_basic 20 5 10 datasets_mill.txt