Variational Autoencoders and Normalizing Flows for Anomaly Detection

CS 6362 | Final Project
Alexander Lin & Allan Zhang

Set Up

1. Create a Python virtual environment.
2. Run pip install -r requirements.txt to load to required modules.
3. Run pip install -e . to allow imports across directories.

Data Acquisition

CIFAR-10

1. Run wget https://www.cs.toronto.edu/~kriz/cifar-10-python.tar.gz in the data directory and untar the downloaded file.
2. Execute data/join_data.py,data/assemble_sets.py and then data/assemble_dataloaders.py to create the PyTorch DataLoaders necessary for model training and testing.

MNIST

1. Download MNIST data from https://drive.google.com/file/d/11ZiNnV3YtpZ7d9afHZg0rtDRrmhha-1E/view into the data directory.
2. Execute the data/assemble_sets_mnist.py script to create the MNIST data used for anomaly testing.

Model Training

Run any of the training Python scripts to train a given model. The hyperparameters of the model can be edited by parameterizing the model constructor.

Results

Experiment 1

anomaly.py and latent_graph.py were used to generate Figure 1 in the paper. Substitute any of models labeled 1 in final_models into the script to review their specific results. These may vary from what is presented in the paper since it is dependent on randomized data.

Experiment 2

anomaly_cifar.py and latent_graph_cifar.py were used to generate Figure 1 in the paper. Substitute any of models labeled 2 in final_models into the script to review their specific results. These may vary from what is presented in the paper since it is dependent on randomized data.