A Scalable and Generalizable Pathloss Map Prediction

This repo is the implementation of our paper "A Scalable and Generalizable Pathloss Map Prediction", as well as the follow-ups.

Highlights

PMNet (NN tailored for Pathloss Map Prediction (PMP)) is described in arxiv, which serves as a backbone for the PMP task.
PMNet achieves strong performance on the PMP task ($10^{-2}$ level RMSE on val), surpassing other models by a large margin.
Proposed a method to predict pathloss in unseen network scenarios using transfer learning (TL) with three pre-trained models: VGG16 and two PMNet models. Our PMNet pre-trained model generalizes well, adapting to new scenarios 5.6× faster and using 4.5× less data, while maintaining high accuracy (RMSE $10^{-2}$ level).
This repository includes the training/test dataset and pre-trained model/checkpoints.

Citation

@article{lee2024scalable,
    title={A Scalable and Generalizable Pathloss Map Prediction}, 
    author={Ju-Hyung Lee and Andreas F. Molisch},
    year={2024},
    journal={arXiv preprint arXiv:2312.03950}
}

@inproceedings{lee2023pmnet,
    title={PMNet: Robust Pathloss Map Prediction via Supervised Learning},
    author={Ju-Hyung Lee and Omer Gokalp Serbetci and Dheeraj Panneer Selvam and Andreas F. Molisch},
    year={2023},
    month={December},
    booktitle={Proc. of IEEE Global Communicaions Conference (GLOBECOM)},
}

Dataset: Ray-Tracing (RT)-based Channel Measurement

#	Dataset (Map)	# of samples	Download Link
1	USC	4754	Download
2	UCLA	3776	Download
3	Boston	3143	Download

Available checkpoints for PMNet

#	Feature Size	Data-Augmentation	Fine-Tuning	RMSE	Download Link
1	16/H X 16/W	4-way rotation	-	0.012599	Download
2	8/H X 8/W	4-way rotation	-	0.010570	Download
3	16/H X 16/W	-	UCLA Dataset	0.031449	Download
4	16/H X 16/W	-	Boston Dataset	0.009875	Download

#3,4 checkpoints were fine-tuned using (1), a pre-trained model with USC Dataset.

Train

To train PMNet, please refer to train.sh

python train.py -d [dataset-root] -n [network-type] -c [config-class-name]
# e.g., python train.py -d '/USC/' -n 'pmnet_v3' -c 'config_USC_pmnetV3_V2'

Evaluation

To evaluate above models, refer to the following commands. Or, you can run eval.sh

python eval.py \
    --data_root [dataset-directory] \
    --network [network-type] \ # pmnet_v1 or pmnet_v3
    --model_to_eval [model-to-eval] \
    --config [config-class-name]
# e.g.,
# python eval.py \
#    --data_root '/USC/' \
#    --network 'pmnet_v3' \
#    --model_to_eval 'config_USC_pmnetV3_V2_epoch30/16_0.0001_0.5_10/model_0.00012.pt' \
#    --config 'config_USC_pmnetV3_V2'

abman23 / pmnet