AMP-DA-Net

Supporting code for “Learned Digital Codes for Over-the-Air Computation in Federated Edge Learning” (2025).
This repo contains three end-to-end stages:

1. Dataset collection (simulate FL + wireless compression, save iFed bundles)
2. Pre-training the AMP-DA-Net decoder + URA codebook on the collected bundles
3. FEEL inference (federated training with the learned wireless compressor)

---

Quick start

# 1) Clone & enter
git clone https://github.com/tonytarizzo/AMP-DA-Net
cd AMP-DA-Net

# 2) Python env
python3.11 -m venv .venv && source .venv/bin/activate
python -m pip install --upgrade pip setuptools wheel
pip install -r requirements.txt

# 3) Run the recipes (from repo root)
bash recipes/run_collection.sh
bash recipes/run_pretrain.sh
bash recipes/run_inference.sh

# Or optionally run scripts directly (useful for testing/debugging)
python -m scripts.1_dataset_collection
python -m scripts.2_pretraining
python -m scripts.3_feel_inference

The paths for collected datasets and pretrained encoder-decoder pairs are automatically named. This ensures that the next script can locate them automatically, while ensuring the same setup is used for all 3 scripts. If you change one recipe, change the other two to match.

Citation

If you use this code in your work, please cite:

@article{tarizzo2025learneddigitalotacomputation,
      title={Learned Digital Codes for Over-the-Air Computation in Federated Edge Learning},
      author={Antonio Tarizzo and Mohammad Kazemi and Deniz Gündüz},
      year={2025},
      archivePrefix={arXiv},
      eprint={2512.19777},
      doi={10.48550/arXiv.2512.19777},
      url={https://arxiv.org/abs/2512.19777},
      note={Submitted to IEEE Journal on Selected Areas in Communications 2026}
}

Or for the earlier conference version:

@article{tarizzo2025learneddigitalcodesovertheair,
      title={Learned Digital Codes for Over-the-Air Federated Learning}, 
      author={Antonio Tarizzo and Mohammad Kazemi and Deniz Gündüz},
      year={2025},
      eprint={2509.16577},
      archivePrefix={arXiv},
      primaryClass={cs.LG},
      url={https://arxiv.org/abs/2509.16577},
      note={Initial version first submitted to ICASSP 2026; revised conference version submitted to SPAWC 2026}
}