π Welcome to the repository for the paper "Reverse Ordering Techniques for Attention-Based Channel Prediction"! This repository contains the code to reproduce the main results of our research work (see our paper and poster for details).
In our study, we introduce two models for predicting time-varying channels: Transformer-RPE and Seq2Seq-attn-R. Both outperform existing methods in channel prediction accuracy across different noise levels and generalize to unseen sequence lengths.
The channel dataset can be downloaded here as a .zip file.
After you have downloaded and extracted the dataset, you have put it in the same folder of the source code.
You can reproduce the results with these commands:
python main.py transformer-rpe
or
python main.py seq2seq-attn-r
You will see this on your terminal for the Transformer-RPE:
Click to expand test results
Testing l=16 and delta=4 (same as training): SNR=-5dB NMSE=0.4747 Testing l=8 and delta=2: SNR=-5dB NMSE=0.5138 Testing l=14 and delta=6: SNR=-5dB NMSE=0.574 --------------------------------------------------------------------------- Testing l=16 and delta=4 (same as training): SNR=0dB NMSE=0.2915 Testing l=8 and delta=2: SNR=0dB NMSE=0.3225 Testing l=14 and delta=6: SNR=0dB NMSE=0.4014 --------------------------------------------------------------------------- Testing l=16 and delta=4 (same as training): SNR=5dB NMSE=0.1808 Testing l=8 and delta=2: SNR=5dB NMSE=0.1742 Testing l=14 and delta=6: SNR=5dB NMSE=0.289 --------------------------------------------------------------------------- Testing l=16 and delta=4 (same as training): SNR=10dB NMSE=0.1127 Testing l=8 and delta=2: SNR=10dB NMSE=0.0945 Testing l=14 and delta=6: SNR=10dB NMSE=0.203 --------------------------------------------------------------------------- Testing l=16 and delta=4 (same as training): SNR=15dB NMSE=0.0711 Testing l=8 and delta=2: SNR=15dB NMSE=0.0583 Testing l=14 and delta=6: SNR=15dB NMSE=0.1373 --------------------------------------------------------------------------- Testing l=16 and delta=4 (same as training): SNR=20dB NMSE=0.0448 Testing l=8 and delta=2: SNR=20dB NMSE=0.0388 Testing l=14 and delta=6: SNR=20dB NMSE=0.1074 ---------------------------------------------------------------------------
And this for the Seq2Seq-attn-R:
Click to expand test results
Testing l=16 and delta=4 (same as training): SNR=-5dB NMSE=0.5075 Testing l=8 and delta=2: SNR=-5dB NMSE=0.519 Testing l=14 and delta=6: SNR=-5dB NMSE=0.6161 --------------------------------------------------------------------------- Testing l=16 and delta=4 (same as training): SNR=0dB NMSE=0.3164 Testing l=8 and delta=2: SNR=0dB NMSE=0.3041 Testing l=14 and delta=6: SNR=0dB NMSE=0.4243 --------------------------------------------------------------------------- Testing l=16 and delta=4 (same as training): SNR=5dB NMSE=0.1993 Testing l=8 and delta=2: SNR=5dB NMSE=0.1876 Testing l=14 and delta=6: SNR=5dB NMSE=0.2988 --------------------------------------------------------------------------- Testing l=16 and delta=4 (same as training): SNR=10dB NMSE=0.1217 Testing l=8 and delta=2: SNR=10dB NMSE=0.118 Testing l=14 and delta=6: SNR=10dB NMSE=0.2083 --------------------------------------------------------------------------- Testing l=16 and delta=4 (same as training): SNR=15dB NMSE=0.0739 Testing l=8 and delta=2: SNR=15dB NMSE=0.0747 Testing l=14 and delta=6: SNR=15dB NMSE=0.148 --------------------------------------------------------------------------- Testing l=16 and delta=4 (same as training): SNR=20dB NMSE=0.0466 Testing l=8 and delta=2: SNR=20dB NMSE=0.0523 Testing l=14 and delta=6: SNR=20dB NMSE=0.119 ---------------------------------------------------------------------------
π If you are using this code and/or the provided dataset for your research, please cite
@article{rizzello2024reverse, author={Rizzello, Valentina and B{\"o}ck, Benedikt and Joham, Michael and Utschick, Wolfgang}, journal={IEEE Open Journal of Signal Processing}, title={Reverse Ordering Techniques for Attention-Based Channel Prediction}, year={2024}, volume={5}, number={}, pages={248-256}, doi={10.1109/OJSP.2023.3344024} }