CSI4Free: GAN-Augmented mmWave CSI for Improved Pose Classification

This repository contains the code for the paper:

CSI4Free: GAN-Augmented mmWave CSI for Improved Pose Classification
Nabeel Nisar Bhat; Rafael Berkvens; Jeroen Famaey
Link to Paper

In this work, we demonstrate stable GAN training on mmWave CSI data. Using a Wasserstein GAN (WGAN), we can generate synthetic CSI samples to augment limited real-world datasets, improving performance in pose classification tasks.

A subset of the GAN-generated dataset is available here:
Zenodo Dataset

---

Dataset

Place your CSI dataset in the data/ folder:

Features

• WGAN-GP for stable training
• Conditional GAN (cWGAN) for class-specific sample generation
• Code for:
  • Loading datasets
  • Training GANs
  • Saving generated CSI data
  • Tracking generator/discriminator losses

---

Configuration

All dataset and hyperparameter options can be set in the config.yaml file, including:

Usage: python train.py

Citation

If you use this repository or dataset, please cite:

@inproceedings{bhat2024csi4free, title={CSI4Free: GAN-Augmented mmWave CSI for Improved Pose Classification}, author={Bhat, Nabeel Nisar and Berkvens, Rafael and Famaey, Jeroen}, booktitle={2024 IEEE 4th International Symposium on Joint Communications & Sensing (JC&S)}, pages={1--6}, year={2024}, organization={IEEE} }