RadioGUNet: Group Equivariant Convolutional Networks for Pathloss Estimation

This repository contains code for training and evaluating the RadioGUNet model in pathloss estimation on RadioMapSeer dataset

Getting Started

Prerequisites

• requirements.txt

Dataset

Download the RadioMapSeer dataset as described in the paper.

---

Reproducing Results

The experiments can be reproduced by running the training and testing scripts with the appropriate parameters:

• model_path: Where to save model checkpoints and logs
• dataset_path: Path to the RadioMapSeer dataset
• experiment_type:
  • DPM_no_car
  • DPM_cars
  • IRT_no_car
  • IRT_cars
• symmetry_group:
  • C2, D2, C4, D4, C8, D8 (as in the paper, can be extended to C16 D16 so long as it is supported in e2cnn)

---

Training

Replace <exp_type> and <group> with your choices:

python3 train.py \
  --model_path ./results/<exp_type>_<group> \
  --dataset_path ./RadioMapSeer \
  --experiment_type <exp_type> \
  --symmetry_group <group>

Example: DPM with cars, D8 group

python3 train.py \
  --model_path ./results/DPM_cars_D8 \
  --dataset_path ./RadioMapSeer \
  --experiment_type DPM_cars \
  --symmetry_group D8

---

Evaluation

After training, run evaluation on the test split:

python3 test.py \
  --model_path ./results/<exp_type>_<group> \
  --dataset_path ./RadioMapSeer \
  --experiment_type <exp_type> \
  --symmetry_group <group>

Example: IRT without cars, D4 group

python3 test.py \
  --model_path ./results/IRT_no_car_D4 \
  --dataset_path ./RadioMapSeer \
  --experiment_type IRT_no_car \
  --symmetry_group D4

---

Notes

• The scripts automatically configure the dataset loader and model based on your selected experiment and symmetry group.
• Results, logs, and model checkpoints will be saved in the specified model_path.
• For additional settings (e.g., batch size, learning rate, model size), edit train.py directly.

---

Citation