STORM: Spatial TransfOrmer for Radio Map Estimation

This repository contains the implementation of STORM and the code to reproduce the numerical experiments of the paper "Spatial Transformer for Radio Map Estimation" by Pham Q. Viet and Daniel Romero. The paper was presented in the International Conference on Communications (ICC), 2025.

Overview

• Python version: you may need Python 3.12 or later.

• The experiments (code) to train and evaluate STORM are in experiments/transformer_experiments.py.

• Training and testing data are available inside folder data. We will work with three datasets, referred to as USRP, Gradiant, and ray-tracing; for more information, see "Radio Map Estimation: Empirical Validation and Analysis".

Set up

After cloning the repository, do the following steps.

1. Install gsim: git submodule init git submodule update bash gsim/install.sh For more information, see gsim.

2. In gsim_conf.py, change module_name = "experiments.example_experiments" to module_name = "experiments.transformer_experiments".

3. An experiment, e.g. experiment_1000 can be run as follows: python run_experiment 1000

4. This step is optional as it is to download the trained weights for the benchmark estimators. The folders inside this link should be downloaded to output/trained_estimators.

Radio map estimation

1. Create training datasets:
   • You can use Experiments 1000, 1005, or 1010 to create a training dataset for STORM.
   • Experiment 1015 creates a training set with ray-tracing data for other DNN benchmarks.

The obtained training datasets will be stored inside folder output/datasets.

2. To train STORM on USRP, Gradiant, and ray-tracing data, run Experiments 2000, 2005, and 2010, respectively.

3. Numerical results:

   • Run Experiment 3112 to obtain Figure 3.
   • Run Experiment 3105 to obtain Figure 4.
   • Run Experiment 3107 to obtain Figure 5.

Active sensing

• Run Experiments 4000 and 4010 to obtain Figure 6.

Contact: - Viet Pham: viet.q.pham@uia.no - Daniel Romero: daniel.romero@uia.no