triani/PhotonicDeepBeam-dataset

PhotonicDeepBeam Dataset

This repository hosts the datasets used in the QuantBeam project — an adaptation of the DeepBeam architecture towards Photonic-Aware Neural Networks (PANN) for millimeter-wave (mmWave) beam classification.

⚠️ The datasets are NOT ours. They are the original experimental datasets collected and published by the DeepBeam / WiNES Lab team (Polese, Restuccia, Melodia — Northeastern University). We re-host them here solely to facilitate reproducibility of our project. All credit and intellectual property belongs to the original authors.

---

Original Dataset Sources (DeepBeam)

Dataset	Description	Original Link
Tx_5-beams.h5	5 TX beams, 3 gain values	D20410052
Tx_12-beams_*.h5	12 TX beams, 3 gain values	D20410075

For the original DeepBeam documentation and full dataset suite, refer to:
📄 http://hdl.handle.net/2047/D20410105
💻 https://github.com/wineslab/deepbeam

---

Repository Contents

PhotonicDeepBeam-dataset/
├── Dataset/
│   ├── Tx_5-beams.h5               # 5-beam dataset (Pi-Radio TXB)
│   ├── Tx_12-beams_<config>.h5     # 12-beam dataset (single config used in our work)
│   └── Indexes/
│       ├── Tx_5-beams_5percent/    # Pre-generated split indexes (seed 42–46, 5% undersampling)
│       ├── Tx_5-beams_20percent/   # Pre-generated split indexes (seed 42–46, 20% undersampling)
│       ├── Tx_5-beams_100percent/  # Pre-generated split indexes (seed 42–46, full dataset)
│       └── Tx_12-beams_20percent/  # Pre-generated split indexes (seed 42–46, 20% undersampling)
└── Modelli addestrati/
    ├── DNN/
    │   ├── 5 beams/                # Trained DNN weights for 5-beam task
    │   └── 12 beams/               # Trained DNN weights for 12-beam task
    └── PANN/
        ├── 5 beams/                # Trained PANN weights (various configs)
        └── 12 beams/               # Trained PANN weights (various configs)

Index files (.pkl)

Pre-computed index files ensure that the exact same train/validation/test split is used across devices and runs. Each .pkl file contains a list [train_indexes, valid_indexes, test_indexes] where each element is a list of frame indices into the corresponding .h5 file.

Filename convention:
i_<undersample%>_<seed>_<train%>_<valid%>_<test%>.pkl
Example: i_20_42_70_15_15.pkl → 20% undersampling, seed 42, 70/15/15 split.

---

Dataset Description

The data consists of real-world mmWave waveform measurements collected using the DeepBeam experimental testbeds. Signals are represented as In-Phase and Quadrature (I/Q) samples of the raw baseband waveform from the receiver RF chain.

5-beam dataset

• Testbed: Pi-Radio TXB
• TX beams: 5
• Transmitter gain values: 3
• Blocks per frame: 100 (5 used per sample)
• Samples per block: 512
• Frames per (beam, gain) pair: 10,000
• Input tensor shape: (batch, 5, 512, 2) — 5 blocks × 512 samples × [I, Q]

12-beam dataset

• Testbed: Multi-RF-chain
• TX beams: 12
• Receiver gain values: 3 (40 dB, 50 dB, 60 dB → SNR range ≈ −15 dB to +20 dB)
• Blocks per frame: 15 (5 used per sample)
• Samples per block: 2048
• Frames per (beam, gain) pair: 10,000
• Input tensor shape: (batch, 5, 2048, 2) — 5 blocks × 2048 samples × [I, Q]

Note: In our project we used only a single .h5 file for the 12-beam task (the simplest available configuration) and trained on 20% of the data due to hardware constraints.

---

How to Use in the QuantBeam Notebook

The notebook (PhotonicDeepBeam_Notebook.ipynb) handles dataset loading automatically. To use this HuggingFace repository, initialize the storage using the Persistent_Storage class like in the notebook.

Citation

Please cite the original DeepBeam paper when using this data:

@inproceedings{polese2021deepbeam,
  author    = {Polese, Michele and Restuccia, Francesco and Melodia, Tomaso},
  title     = {DeepBeam: Deep Waveform Learning for Coordination-Free Beam Management in mmWave Networks},
  booktitle = {Proceedings of ACM MobiHoc},
  year      = {2021}
}