The repository has implemented the Anomaly Detection with GAN and has been applied to the Flickr Face HQ Dataset.

Requirements

• python 3.6
• tensorflow-gpu==1.14
• pillow
• matplotlib

Architecture

I made the kernel size different from the picture above when i implemented it.

• config.py : A file that stores various parameters and path settings.
• model.py : ANOGAN's network implementation file
• train.py : This file load the data and learning with GAN.
• train_anogan.py : This file load the GAN model and then learns the detector.
• utils.py : Various functions such as loading data*

1. Download Flickr Face HQ Dataset

2. The read_images function on utils.py has a subfolder that has an image corresponding to each class in the root folder.

   ROOT_FOLDER
      |   
      |--------SUBFOLDER (Class 0)   
      |          |------image1.jpg   
      |          |------image2.jpg   
      |          |------etc..   
      |--------SUBFOLDER (Class 1)   
      |          |------image1.jpg   
      |          |------image2.jpg   
      |          |------etc..
   

   Please create a folder to store learning images and insert learning images to this standard. I used 10,000 images in thumbnails128x128 folder. (from 00000 folder to 09000 folder) The path i used is as follows.

   Example: Copy thumbnails128x128 to D:/data and Write D:/data/thumbnails128x128 on config.py
   

3. Run train.py

4. Run train_anogan.py

5. The result images are stored per epoch in the temp

• Upload result images

• Schlegl, Thomas, et al. "Unsupervised anomaly detection with generative adversarial networks to guide marker discovery." International conference on information processing in medical imaging. Springer, Cham, 2017.
• https://github.com/tkwoo/anogan-keras