ethan-keller / fct-gan

The first GAN-based tabular data synthesizer integrating the Fourier Neural Operator for global dependency imitation

FCT-GAN

FCT-GAN is the first GAN-based tabular data synthesizer integrating the Fourier Neural Operator to improve global dependency imitation. This repository contains the source code for FCT-GAN, and an example dataset called Adult. Additionally, two Jupyter notebooks have been added to this repository. The first notebook will run FCT-GAN on the Adult dataset. The second one plots the individual losses to evaluate the training process.

Prerequisites

The required python packages:

numpy
torch
pandas
sklearn
dython==0.6.4.post1
scipy

Example Jupyter notebooks

Adult notebook

Experiment_Script_Adult.ipynb is an example Jupyter notebook to train FCT-GAN on the Adult dataset. The dataset is included in the Real_Datasets directory. The evaluation code is also provided.

Loss notebook

plot_losses.ipynb is an example Jupyter notebook to plot losses. If you want to plot losses during training, uncomment lines 586 to 597 in model/synthesizer/fctgan_synthesizer.py. This will plot individual losses after every epoch.

Large datasets

If your dataset has a large number of columns, you may encounter the problem that the current code cannot encode all of your data. What you can do is change lines 404 and 411 in model/synthesizer/fctgan_synthesizer.py:

sides = [4, 8, 16, 24, 32]

sides is the side size of an embedded image in FCT-GAN. You can enlarge the list as shown below to accept larger datasets:

# sides = [4, 8, 16, 24, 32]

# enlarged sides
sides = [4, 8, 16, 24, 32, 64]

# or even larger
sides = [4, 8, 16, 24, 32, 64, 128]

Affiliation

This research is part of the Research Project coordinated by TU Delft