Regression with CGAN Code

This repository contains the code of the following paper

K Aggarwal, M Kirchmeyer, P Yadav, S Sathiya Keerthi, P Gallinari, "Regression with Conditional GAN"

Dependencies

In order to run, the code requires the following Python modules referenced in requirements.txt:

• numpy, jupyter, matplotlib, pandas
• sklearn
• tensorflow, keras
• GPy https://sheffieldml.github.io/GPy/

CGAN code is derived from https://github.com/eriklindernoren/Keras-GAN

Quickstart

• Install Miniconda
• Create conda environment: conda create --name ganRegression python=3.6 -y then source it source activate ganRegression
• Install the requirements in this environment pip install -r requirements.txt
• Install the package pip install -e . at the root
• Run the notebooks using jupyter-notebook

Notebooks

• Run notebook/synthetic_data.ipynb for synthetic data
• Run notebook/real_world_data.ipynb for real world data
• Notebooks will save figures in the figures folder for each data scenario

Datasets

• Synthetic datasets: linear, sinus, heteroscedastic, exp, multi-modal
• Real World datasets:
  • CA-housing taken from sklearn.datasets. CA-housing-single takes the most important feature from CA-housing (cf. study in the paper)
  • ailerons taken from http://www.dcc.fc.up.pt/~ltorgo/Regression/DataSets.html
  • comp-activ, pumadyn, bank, census-house, abalone taken from https://www.cs.toronto.edu/~delve/data/datasets.html

Config

• The Config class handles all parameters. These are set at the beginning of each notebook. Refer to config.py for more details
• Architectures are fixed in cgan_model.py or can be set in the Config object for custom experiments.

Results and uncertainty

• Results from the paper can be reproduced with an uncertainty smaller than 0.05 on NLPD + MAE for CGAN.