Skip to content

Latest commit

 

History

History

IrisProtocolValidation

Folders and files

NameName
Last commit message
Last commit date

parent directory

..
Model
Model
 
 
images
images
 
 
README.md
README.md
 
 
WGAN.ipynb
WGAN.ipynb
 
 

README.md

Iris Tabular Data Generation using PyTorch

This project explores synthetic data generation for the Iris dataset using different architectures for data generation including Variational Auto Encoder (VAE), Generative Adversarial Network (GAN) and Wasserstein GAN WGAN. The goal is to generate high-quality synthetic tabular data that preserves the statistical properties and class-specific relationships of the original dataset. This work is part of a larger initiative to generate synthetic medical data (tabular and imaging) and is a validation of a protocol to be followed in data generation tasks.

Generated vs Real Data Comparison Example of synthetic vs real data distributions.

Table of Contents

  1. Libraries
  2. Data Preparation
  3. Exploring Architectures
  4. Optimizing Architecture
  5. Wasserstein Conditional GAN
  6. Data Augmentation
  7. Final Evaluation
  8. Installation
  9. Usage
  10. Results
  11. Conclusion
  12. References

Libraries

The project uses the following Python libraries:

  • PyTorch for model implementation.
  • Pandas/NumPy for data manipulation.
  • Scikit-learn for metrics and preprocessing.
  • Matplotlib/Seaborn for visualization.
  • SciPy for statistical tests (Kolmogorov-Smirnov, Wasserstein distance).
import pandas as pd
import numpy as np
from scipy import stats
import matplotlib.pyplot as plt
import seaborn as sns
import torch

Data Preparation

The Iris dataset is loaded and preprocessed:

  • Normalization: Features scaled to [0, 1].
  • Class Conditioning: Labels encoded and provided as input to generators.

Exploring Architectures

GAN vs VAE

Two architectures were compared:

  1. Variational Autoencoder (VAE): Learned latent representations but struggled with class-specific feature relationships.
  2. Generative Adversarial Network (GAN): Outperformed VAE in capturing feature characteristics as shown by images.
AVG Metric GAN VAE
Wasserstein ↓ 0.32274 0.25224
KS Test p-value ↑ 0.00053 2.5e-6

Simple VAE Q-Q plots Simple VAE Q-Q plots for each features. Simple GAN Q-Q plots Simple GAN Q-Q plots for each features.

Optimizing Architecture

Conditional GAN (cGAN)

  • Grid Search:
    • Noise Vector Size: Tested [2, 4, 8, 16, 32, 64]. Optimal: 16 (lowest Wasserstein distance).
    • Network Size: Small, Medium, Large. Optimal: Medium.
Small Medium Big
2 0.797k 3.069k 11.709k
4 0.861k 3.197k 11.965k
8 0.989k 3.453k 12.477k
16 1.245k 3.965k 13.501k
32 1.757k 4.989k 15.549k
64 2.781k 7.037k 19.645k

Number of parameters of each explored model

Mean Absolute Correlations Differences Medium GAN with 16 Noise Vector Big GAN with 16 Noise Vector
Overall 0.0788 0.0682
Setosa 0.1294 0.4591
Versicolor 0.1593 0.4322
Virginica 0.2694 0.2571

Introducing the WGAN

The Wasserstein GAN (WGAN) with gradient penalty improved stability and feature fidelity:

  • Critic Updates: 5 critic steps per generator step.
  • Gradient Penalty: λ = 10.

Data Augmentation

To improve the class-specific performance of the WGAN, two new features were created using those in the original dataset.

$$ \text{Sepal Area} \rightarrow \text{Sepal Length} \times \text{Sepal Width} $$

$$ \text{Petal Area} \rightarrow \text{Petal Length} \times \text{Petal Width} $$

These features led to the achievement of excellent overall and class-specific performance by bringing all mean absolute differences between correlations to be < 0.1.

Feature Wasserstein Distance Kolmonorov-Smirnov's D KS p-value
sepal length (cm) 0.064483 0.060000 0.951111
sepal width (cm) 0.042666 0.126667 0.180395
petal length (cm) 0.062556 0.060000 0.951111
petal width (cm) 0.061056 0.120000 0.230782

Final Evaluation

Train on Real Test on Synthetic and Train on Synthetic Test on Real protocol was used to assess the quality of the generated data understood as their indistinguishability in the eyes of a well-trained classifier.

  • Training of two classifier models, one on the true dataset one on the synthetic dataset.
  • Evaluation using test sets.
  • Cross evaluation from the two models.

Confusion Matrices Confusion Matrices

Installation

To use this script, follow the steps below:

  1. Clone the repository
    Open the terminal and clone the GitHub repository with the command:

    git clone https://github.com/Sim98B/TabularDataGeneration.git
cd TabularDataGeneration/IrisProtocolValidation/Model

  2. Install requirements

    pip install -r requirements.txt

Usage

To generate a synthetic dataset based on the Iris dataset, use the following command:

python GenerateIrisData.py --weights WGAN_weights.pth --scaler scaler.pkl

Required Arguments:

  • --weights: Path to the generator weights file (.pth).
  • --scaler: Path to the saved scaler file (.pkl).

Options:

  • --setosa: Number of samples to generate for the setosa class (default: 50).
  • --versicolor: Number of samples to generate for the versicolor class (default: 50).
  • --virginica: Number of samples to generate for the virginica class (default: 50).
  • --seed: Set a seed for reproducibility.
  • --output: Name of the output CSV file (default: SyntheticIris.csv).

Results

  • Pre-trained model to generate data is saved in the Model/ directory.
  • Full notebook analysis is stored the WGAN.ipynb.
  • All functions used in the implementation of the project are in the utils.py file.

Conclusion

The Wasserstein cGAN combined with engineered features (sepal_area and petal_area) achieved the best performance in generating synthetic Iris data. The evaluation—both statistical (KS test, Wasserstein distance) and practical (cross-evaluation using classifiers)—confirms that the synthetic data closely mimics the real data distribution. This approach can be extended to more complex datasets, including those in the medical domain, with appropriate privacy safeguards.

References