Flows4Ad

Repository for Anomaly Detection with Normalizing Flows.

Data preparation

Clone ADBench repository (somewhere):

git clone https://github.com/Minqi824/ADBench

Data is contained in datasets directory. Make a symbolic link to the root of the current project:

ln -s <ADBench_dir/datasets> datasets

Datasets are expected to be in .npz format with X and y keys.

Environment configuration

In order to run experiments verify the existence of working installation of torch, joblib, matplotlib, seaborn, scikit-learn.

The versions used in experiments:

torch 1.12.1 joblib 1.1.1 matplotlib 3.6.2 seaborn 0.12.1 scikit-learn 1.1.3

For logging with W&B install wandb:

pip install wandb

For running hyperparameter search with Optuna install optuna

pip install optuna

Experiment running

In order to train VAE launch run_train_encoder.py script with the specific .yaml config (look at the example in configs/encoder/_reference_config.yaml).

In order to train flow for AD detection launch run_train_detector.py script with the specific .yaml config (look at the example in configs/detector/_reference_config.yaml). In order to change parameters of experiment edit the config file.

In order to run training with hyperparam search launch run_train_detector_optuna.py.

For RealNVP it is necessary to choose either channel_wise or checkerboard split or both.

Results and report

materials contains results discussed in report.