RadioTrace: Transmitter-Aware Diffusion for Radio Map Estimation without Deployment-Time Fine-Tuning

This repository provides code for RadioTrace, a noval radio map estimation framework that reconstructs RSS radio maps from sparse measurements without deployment-time fine-tuning by integrating a pre-trained diffusion prior with transmitter location estimation inside the denoising loop. An optional propagation-guided K-means initialization improves robustness under restricted-area sampling.

Related Publications: 📄 Conference Paper |   •   📄 Journal Extended Version |

---

Results

Quick Start

0) Prepare the environment

conda create -y -n radiotrace python=3.13 -c conda-forge
conda activate radiotrace
pip install -r requirements.txt

1) Download and place the dataset (RadioMapSeer)

1. Download RadioMapSeer.
2. Unzip it into the repository root directory.

---

2) Download the pretrained diffusion model (Google Drive)

Download the pretrained checkpoint from Google Drive and place it anywhere you prefer (recommended: ./).

Note: The pretrained model is trained as a generic diffusion prior and does not need to see diverse sampling patterns during training. Sampling patterns are handled at inference via RadioTrace.

---

3) Run RadioTrace inference / testing

python sample_radiotrace.py \
  --data_dir path/to/dataset/ \
  --model_path path/to/model/file \
  --type restrict_wo_BS \
  --rate 0.01 \
  --cluster_init

---

Optional: Train the diffusion model yourself

If you do not use a provided pretrained checkpoint, you can train the conditional diffusion model with: python train_cond_ddpm.py

---

Command-Line Arguments (inference)

• --data_dir
  Path to the dataset root directory (RadioMapSeer).

• --model_path
  Path to the pretrained diffusion checkpoint file (.pt / .pth).

• --type
  Sampling / evaluation mode. Example used in the paper: restrict_wo_BS (restricted-area sampling without base-station region coverage).

• --rate
  Sampling rate (e.g., 0.01 means 1% measurements are observed).

• --cluster_init
  Enables propagation-guided K-means initialization for Tx coordinates to reduce poor local minima and improve robustness.

To list all options: python sample_radiotrace.py -h

---

Citation

@ARTICLE{yang2026radiotrace,
  author={Yang, Liu and Li, Qiang and Cao, Zhuo and Xiong, Weijie and Sun, Guomin and Lin, Jingran},
  journal={IEEE Transactions on Wireless Communications}, 
  title={RadioTrace: Transmitter-Aware Diffusion for Radio Map Estimation Without Deployment-Time Fine-Tuning}, 
  year={2026},
  volume={25},
  number={},
  pages={21105-21118},
  doi={10.1109/TWC.2026.3716965}
}

@inproceedings{yang2025radiotrace,
  title={Radiotrace: Bridging Diffusion Priors and RSS Measurements for Accurate Radio Map Estimation},
  author={Yang, Liu and Li, Qiang and Cao, Zhuo and Lin, Jingran},
  booktitle={2025 IEEE 35th International Workshop on Machine Learning for Signal Processing (MLSP)},
  pages={1--6},
  year={2025},
  organization={IEEE}
}

---

Acknowledgements

This implementation builds upon and is inspired by the RadioDiff repository: