trustorno/alphagan

Alphagan

This an adapted implementation of AlphaGan for Zero Shot outlier detection in https://www.kaggle.com/mlg-ulb/creditcardfraud dataset using only negative samples for training

Create conda environment with all dependencies

conda env create -f environment.yaml

Activate conda environment

source activate alpha

Add your comet-ml api key and project in config.py

run main.py

python main.py

Example experiment https://www.comet.ml/dragosnicolae5555/c-alpha/279172963e4041e79764b6a25d18c71b

General Framework :

1. Train GAN to generate only normal data points (negative samples).
2. When predicting anomaly, use GAN to reconstruct the input images of both normal and abnormal images (negative and positive samples).
3. Compute reconstruction and discrimination losses.

Discriminate between normal and abnormal cases using these statistics: 
Reconstruction loss are the differences between original and reconstructed images.
Discrimination loss is simply the output of the Discriminator.