Reproduce Results

Use the selector below to jump straight to a single recipe.

Seeds

Commands are shown with --seed 0 for readability. To reproduce the reported averages, run seeds 0, 1, 2 and average the metrics (individual runs should land in the same ballpark).

Environment

• Python: 3.10+
• PyTorch: 2.x with CUDA 11/12 (match your GPU driver)
• Hardware: 1x A100 or equivalent; adjust batch size for smaller GPUs.

Install

requirements.txt pins the core dependencies for reproducibility, including default (CPU) PyTorch versions. If you need CUDA, reinstall torch/torchvision afterward using the wheels from pytorch.org to match your GPU driver.

git clone https://github.com/AhmedTarek62/wavesfm.git
cd wavesfm
python -m venv .venv
source .venv/bin/activate
pip install -U pip
pip install -r requirements.txt
# If you need CUDA, reinstall torch/torchvision from pytorch.org after this step.
# DeepMIMO preprocessing dependency (only needed for DeepMIMO recipes):
pip install DeepMIMOv3

Recipe browser

Clicking a finetuning label on the Benchmarks page will open the matching recipe here.

Releasev1.0RecipePOWDER (RFP) - LPPOWDER (RFP) - FT2POWDER (RFP) - LoRA (R=32, alpha=64)ICARUS (INTD) - LPICARUS (INTD) - FT2ICARUS (INTD) - LoRA (R=32, alpha=64)ICARUS (INTC) - LPICARUS (INTC) - FT2ICARUS (INTC) - LoRA (R=32, alpha=64)EfficientFi (HAS) - LPEfficientFi (HAS) - FT2EfficientFi (HAS) - LoRA (R=32, alpha=64)CommRad RF (RFS) - LPCommRad RF (RFS) - FT2CommRad RF (RFS) - LoRA (R=32, alpha=64)5G NR Positioning (POS) - LP5G NR Positioning (POS) - FT25G NR Positioning (POS) - LoRA (R=32, alpha=64)RML2022 - LPRML2022 - FT2RML2022 - LoRA (R=32, alpha=64)UWB Indoor Positioning and Tracking (UWB-Indoor) - LPUWB Indoor Positioning and Tracking (UWB-Indoor) - FT2UWB Indoor Positioning and Tracking (UWB-Indoor) - LoRA (R=32, alpha=64)UWB Industrial Localization (UWB-Industrial) - LPUWB Industrial Localization (UWB-Industrial) - FT2UWB Industrial Localization (UWB-Industrial) - LoRA (R=32, alpha=64)DeepMIMO (LoS/NLoS) - LPDeepMIMO (LoS/NLoS) - FT2DeepMIMO (LoS/NLoS) - LoRA (R=32, alpha=64)DeepMIMO (Beam Prediction) - LPDeepMIMO (Beam Prediction) - FT2DeepMIMO (Beam Prediction) - LoRA (R=32, alpha=64)RADCOM - LPRADCOM - FT2RADCOM - LoRA (R=32, alpha=64)

POWDER (RFP) - LP

Dataset: POWDER (RFP) | Task: rfp | Finetuning: LP

Preprocess

python preprocessing/preprocess_rfp.py --data-path <POWDER_DIR> --output data/rfp.h5

Train

python main_finetune.py \
  --task rfp \
  --train-data data/rfp.h5 \
  --val-split 0.2 \
  --output-dir runs/v1.0/rfp/lp \
  --finetune <WAVESFM_BASE_CKPT> \
  --model vit_multi_small \
  --use-conditional-ln \
  --warmup-epochs 5 \
  --num-workers 2 \
  --seed 0 \
  --smoothing 0.1 \
  --epochs 10 --batch-size 256

Checkpoint: replace --finetune <WAVESFM_BASE_CKPT> with --download-pretrained --hf-repo ahmedaboulfo/wavesfm --hf-file wavesfm-v1p0.pth if you don't have it locally.

Evaluate

python main_finetune.py \
  --task rfp \
  --train-data data/rfp.h5 \
  --val-split 0.2 \
  --finetune <FINETUNED_CKPT> \
  --model vit_multi_small \
  --use-conditional-ln \
  --seed 0 \
  --eval-only

Need dataset details? See the dataset pages.