Skip to main content

DeepMIMO (LoS/NLoS & Beam Prediction)

Multi-scenario DeepMIMO channels packaged for downstream tasks in LoS/NLoS classification and beam prediction. WavesFM task ids: deepmimo-los, deepmimo-beam (dataset class DeepMIMO).

Tasks supported​

  • LoS/NLoS classification (binary).
  • Beam prediction (best beam index from a codebook; default 64 beams, configurable).

Class labels​

  • LoS/NLoS: 0 = NLoS, 1 = LoS.
  • Beam prediction: class indices 0..(n_beams-1) for the selected codebook.

Split​

  • No fixed split in the cache; use --val-split when training (default 0.2).

Preprocessing​

Install the DeepMIMOv3 generator before running the preprocessing script:

pip install DeepMIMOv3
  1. Generate DeepMIMOv3 channels for the selected scenarios and active base station.
  2. Filter invalid users (LoS = -1), scale channels by 1e6, and convert complex CSI to real/imag channels.
  3. Compute LoS labels and beam labels by sweeping a beamforming codebook (default 64 beams, 180-degree FOV).
  4. Optionally resize channel grids to 224x224 and normalize with dataset-wide mean/std.
  5. Write sample, label_los, label_beam_{n}, and scenario to HDF5, plus metadata (scenarios, beam options, class weights).

Script: preprocess_deepmimo.py

python preprocessing/preprocess_deepmimo.py \
  --output data/deepmimo.h5 \
  --dataset-folder <DEEP_MIMO_SCENARIOS_DIR> \
  --n-beams-list 16,32,64

Add --clone-scenarios to fetch the scenario repo into --dataset-folder (requires git-lfs). For beam prediction training, pass --deepmimo-n-beams <n> so the dataset selects label_beam_{n}.

Metric​

  • Accuracy and macro F1 for LoS/NLoS.
  • Accuracy and macro F1 for beam prediction.

Citations​

@article{alikhani2024largewirelessmodellwm,
      title={Large Wireless Model (LWM): A Foundation Model for Wireless Channels},
      author={Sadjad Alikhani and Gouranga Charan and Ahmed Alkhateeb},
      year={2024},
      journal={arXiv preprint arXiv:2411.08872},
      url={https://arxiv.org/abs/2411.08872},
}

@InProceedings{Alkhateeb2019,
      author = {Alkhateeb, A.},
      title = {{DeepMIMO}: A Generic Deep Learning Dataset for Millimeter Wave and Massive {MIMO} Applications},
      booktitle = {Proc. of Information Theory and Applications Workshop (ITA)},
      year = {2019},
      pages = {1-8},
      month = {Feb},
      Address = {San Diego, CA}, 
}