Variational Autoencoder for Anomaly Detection: A Comparative Study

This repository contains the implementation of the paper titled "Variational Autoencoder for Anomaly Detection: A Comparative Study". This paper code base is developed based on the code from Hugo's original paper.

Getting Started

Before running the experiments, ensure you have downloaded the MVTec and MiAD datasets. Modify the dataset links in datasets.py accordingly to link these datasets to your experiment.

Prerequisites

The required libraries are listed in requirements.txt. Install them using the following command:

pip install -r requirements.txt

It is recommended to run the installation in a Conda environment, especially on Windows.

Training

To train the model, execute the following command:

sh vae_train.sh

For training Variational Autoencoder (VAE) and VAE-GRF as per our experiments, ensure the following parameters:

• batch_size: 8
• latent_image_size: 32
• latent_dim: 256
• image_size: 256

In vae_test.py, modify mad = mad.repeat(16, axis=0).repeat(16, axis=1) to mad = mad.repeat(8, axis=0).repeat(8, axis=1) to run VAE and VAE-GRF.

For training ViT-VAE as per our experiments, ensure the following parameters:

• batch_size: 8
• latent_image_size: 14
• latent_dim: 384
• image_size: 224

In vae_test.py, modify mad = mad.repeat(8, axis=0).repeat(8, axis=1) to mad = mad.repeat(16, axis=0).repeat(16, axis=1) to run ViT-VAE.

Testing

To test the model, use the following command:

sh vae_test.sh

Ensure to provide the appropriate parameters as mentioned above.

Built With

The code is built using PyTorch and other standard libraries.

More Information

For more details, please refer to the publication.

---