Study of "AE-FLOW: Autoencoders with Normalizing Flows for Anomaly Detection"

The main branch conforms to the required structure, and was thoroughly tested to work before submission. However, if any difficult errors occur, please refer to the 'andre' branch if any errors are encountered while running scripts.

This repository contains our work for a mini-project for the DL2 course at the University of Amsterdam (https://uvadl2c.github.io/, 2023), aiming to firstly replicate the work "AE-FLOW: Autoencoders with Normalizing Flows for Anomaly Detection" by Yuzhong Zhao, Qiaoqiao Ding and Xiaoqun Zhang (2023), and secondly to extend this work by determining it's generalizability, by applying it to other data-sets. We furthermore extend the AE-FLOW model by implementing uncertainty quantification using Deep Ensembles.

‼️ WandB Authentication ‼️

Our repository uses WandB to log runs. Please generate a WandB key and put it into a file in the ./src folder called 'wandbkey.txt'. For further instructions, go to the WandB guide.

Getting started

The directory 'src/script/job_files' contains all files needed to install the environment using our supplied 'src/script/job_files/environment.yml' file, and run the experiments. Using these will be enough to run the repository on a SLURM-based cluster such as LISA. For setting/defining parameters both with regard to model-setup and running experiments we currently refer to the train.py file.

Please install any missing requirements on your own.

Data

Please download the datasets into src/data/ using the following structure:

repo
|
|-- src
    |--data
        |--chest_xray
            |--test
            |--train
            |--val
        |--OCT2017
            |--test
            |--train
            |--val
        |--miic
            |--test
            |--train

Chest X-ray: https://www.kaggle.com/datasets/paultimothymooney/chest-xray-pneumonia

OCT2017: https://www.kaggle.com/datasets/paultimothymooney/kermany2018

MIIC: https://researchdata.ntu.edu.sg/dataset.xhtml?persistentId=doi:10.21979/N9/WBLTFI

The BTAD and MVTec dataset are downloaded and loaded using the Anomalib library.

Running locally

Multiple files can be used to run the model (located in src):

• train.py: used for running the autoencoder, Fastflow, AE-FLOW models. Please see the main function for details on arguments.
• generate_ue_results.py: can be ran out of the box. It will evaluate the deep ensemble and provide results. No models are required to run this file, as the model results and thresholds are pickled and available in the repository.
• generate_experiment_files.py: generates the job files for the final experiments. No practical use outside of our experiment.

Deliverables

Please refer to the blogpost.md for our further research efforts for the course

Some trained models

Chest X-Ray dataset, AE-FLOW, Seed 1

Chest X-Ray dataset, AE-FLOW, Seed 59

Chest X-Ray dataset, AE-FLOW, Seed 85