ahmedaboulfo / wavesfm

WavesFM v1.0

Summary

WavesFM is a multimodal wireless foundation model for IQ streams and image-like modalities (spectrograms, CSI), plus CIR. It uses a ViT backbone with modality-specific input embeddings and a masked wireless modeling objective.

Model Details

• Architecture: ViT encoder + modality-specific input embeddings + lightweight task heads.
• Modalities: IQ, spectrograms, CSI, CIR.
• Objective: masked wireless modeling (MAE-style reconstruction).

Links

• Benchmarks: wavesfm.waveslab.ai/benchmarks
• Reproduce: wavesfm.waveslab.ai/docs/reproduce
• GitHub: github.com/AhmedTarek62/wavesfm
• Quickstart notebook: wavesfm_demo.ipynb

Intended Use

• Research on multimodal wireless representation learning.
• Fine-tuning on downstream wireless tasks.

Downstream Tasks

• RF fingerprinting (IQ)
• Interference detection/classification (IQ)
• Human activity sensing (CSI)
• RF signal classification (spectrograms)
• 5G NR positioning (CSI)
• DeepMIMO LoS/NLoS and beam prediction (CSI)
• RADCOM signal/modulation classification (IQ)
• UWB indoor and industrial localization (CIR)

HF Usage

Clone the training repo:

git clone https://github.com/AhmedTarek62/wavesfm.git
cd wavesfm

Download the checkpoint:

pip install -U huggingface_hub
huggingface-cli download ahmedaboulfo/wavesfm wavesfm-v1p0.pth --local-dir ./checkpoints

Fine-tune:

python main_finetune.py \
  --task <task_name> \
  --train-data <path/to/train.h5> \
  --val-split 0.2 \
  --finetune ./checkpoints/wavesfm-v1p0.pth \
  --output-dir <run_dir>

Evaluate:

python main_finetune.py \
  --task <task_name> \
  --train-data <path/to/train.h5> \
  --val-split 0.2 \
  --eval-only \
  --download-pretrained \
  --hf-repo ahmedaboulfo/wavesfm \
  --hf-file wavesfm-v1p0.pth

Refer to the WavesFM website for more detailed instructions.

Citation

@article{aboulfotouh2025multimodal,
  title = {Multimodal Wireless Foundation Models},
  author = {Aboulfotouh, Ahmed and Abou-Zeid, Hatem},
  journal = {arXiv preprint arXiv:2511.15162},
  year = {2025},
  url = {https://arxiv.org/abs/2511.15162}
}