This repo is the official implementation of "Large-Scale Channel Prediction System for ICASSP 2023 Pathloss Radio Map Prediction Challenge".
- Our system achieves an RMSE of 0.02569 on the provided RadioMap3Dseer dataset, and 0.0383 on the challenge test set, placing it in the 1st-rank of the ML competition @ IEEE ICASSP 2023.
- RMSE: 0.02569
@inproceedings{lee2023large,
author={Lee, Ju-Hyung and Lee, Joohan and Lee, Seon-Ho and Molisch, Andreas F.},
booktitle={IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)},
title={PMNet: Large-Scale Channel Prediction System},
year={2023},
}
Install packages using the following instruction.
pip install -r requirements.txt
- Python 3.8.15
- torch cuda
sh eval.sh
python eval.py \
--input_dir './RadioMap3DSeer/png' \
--gt_dir './RadioMap3DSeer/gain' \
--first_index 631 \
--last_index 700 \
--pretrained_model './checkpoints/radiomapseer3d_pmnetV3_V2_model_0.00076.pt' \
--network_type 'pmnet_v3' \
--output_dir './outputs/RadioMap3DSeer_Test'
The RMSE and run-time will be printed in results.txt file and the terminal where you run.
-
input_dir : The directory where folders of input images such as antennasWHeight, buildingsWHeight exist. (It should be {your_directory}/png)
-
gt_dir : The directory where ground truth images exist. (It should be {your_directory}/gain)
-
first_index : The first index of map indices. For example, when we have 0_0.png ~ 83_79.png images, it should be 0 which is the first index of maps.
-
last_index : The last index of map indices. For example, in the above case, it should be 83 which is the last index of maps.
-
pretrained_model : The path of pretrained model.
-
network_type : The type of network. (pmnet_v3)
-
output_dir : The directory where predicted images are saved. (Default: {current_directory}/outputs)
-
Note: To use other datasets, one need to set arguments: input_dir, gt_dir, first_index, last_index