Skip to content

Project which uses a GAN or an adversarial auto-encoder to censor a classifier from using a a sensitive variable for discrimination. Includes a Dockerized Web application and RESTful API with Continuous Integration.

License

Notifications You must be signed in to change notification settings

daniloefl/advdecorr

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

advdecorr

Build Status

This project is a test bench for classification studies, where a classifier is expected to separate signal against a background without discriminating on a variable S, referred to as systematic uncertainty throughout the code.

The training code is in the directory training, while a production web app and a server-side RESTful code are in the directories app and restful. docker-compose can be used to create a docker image for both the web app and the server-side RESTful code using the docker-compose.yml configuration. See more information about this in the next section.

Looking in the training directory, The training of the classifier can be done using the code in keras_gan.py, where a GAN topology is used to penalize the classifier (taken as the GAN's generator) if it is possible to learn the variable S. Another topology is implemented in keras_aae.py, where an adversarial auto-encoder is trained to censor a latent representation, before using such latent representation for classification.

The data used for tests can be a real data sample or a toy data can be produced with generateToys.py. The toy consists of two variables, A and B, which are produced according to two different generative models: nominal and systematic variation. The marginal distributions of the model can be seen below.

Variable A Variable B
A B

An example of the neural network output for the toy model can be seen in training/result_toy.

Without adversary With adversary
Network output without adversary training Network output with adversary training

Flask client-server model test for production

The code in production/ implements the production code, which receives data, loads the network and performs classification with the unseen input. It can be run as a Flask REST API server with:

python classify.py

By default it runs in port 5001 in the local machine.

This can be transformed in a docker image with:

docker build -t advdecorr-restful .
docker run -d -p 5001:5001 advdecorr-restful

Now connecting to port 5001 outside of the container would work, but if the argument -p 5001:5001 is omitted, this would not be possible in the host, but it would be possible from other containers, which could be beneficial for isolation between the RESTful server and the Web application.

The code in app/ implements the client-side web application, which communicates with the server-side to submit information and obtain new results. It can be run as a Flask application with:

python web.py

It runs in port 5000 by default. Open the browser and navigate to localhost:5000 and try to enter new data. After submitting, the client communicates with the server and collects the p-values back, which are then shown in the resulting form.

The web application can be transformed in a docker image with:

docker build -t advdecorr-app .
docker run -d -p 5000:5000 advdecorr-app

The two docker containers can be built and ran using docker-compose, so that they are able to communicate with each other. It can be done as follows, from the root directory of the project:

docker-compose up --build

About

Project which uses a GAN or an adversarial auto-encoder to censor a classifier from using a a sensitive variable for discrimination. Includes a Dockerized Web application and RESTful API with Continuous Integration.

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

 
 
 

Languages