Awesome
A-GCRNN: Attention Graph Convolution Recurrent Neural Network for Multi-band Spectrum Prediction
This is a PyTorch implementation of A-GCRNN in the following paper: A-GCRNN: Attention Graph Convolution Recurrent Neural Network for Multi-band Spectrum Prediction(https://ieeexplore.ieee.org/document/10251662).
Structure
<div align=center><img src="https://github.com/TLTLHILL/A-GCRNN-for-spectrum-prediction/blob/main/photo/A-GCRNN.png" width="500" height="470" /></div>Dataset
The dataset for this project comes from the open source platform: https://electrosense.org
| Dataset parameters | Value |
|---|---|
| Dataset source | https://electrosense.org |
| Sensor location | Madrid, Spain |
| Frequency band | 500 MHz–800 MHz |
| Monitoring time | 2021.5.28–2021.6.28 |
| Frequency resolution | 2 MHz |
| Time resolution | 15 minutes |
| The dimensionality of samples | 151 × 2880 |
！！！The opening time of sensors on this platform is uncertain, and there may be some sensors shutdown.
Usage
File description
- Data: Datasets storage file
- models: Models storage file
- photo: Models and some experimental results image storage file
- tasks: Tasks storage file
- utils: Key function code storage file
- adj_create.py: Code for constructing adjacency matrix
- main.py: Training Code
- tesy_main.py: Test Code
Requirements
- Numpy
- torch
- pytorch-lightning
- pandas
- matplotlib
Model training
python main.py --model_name AGCRNN --max_epochs 3000 --learning_rate 0.0001 --batch_size 64 --hidden_dim 100 --settings supervised --gpus 1
Model test
python test_main.py --model_name AGCRNN --max_epochs 3000 --learning_rate 0.0001 --batch_size 64 --hidden_dim 100 --settings supervised --gpus 1
Parameters description
！！！These parameters can be adjusted independently.
| Parameters | Description |
|---|---|
| --data | The name of the dataset |
| --seq_len | Required historical data length |
| --pre_len | Predicted data length |
| --split_ratio | Dataset spliting ratio |
| --hidden_dim | Number of GRU hidden layers |
Run tensorboard --logdir lightning_logs/version_0 --samples_per_plugin scalars=999999999 in terminal to view the prediction results and experimental indicators.
Citation
Please cite the following paper if you use the code in your work:
@ARTICLE{ZhangTVT2023a,
author={Zhang, Xile and Guo, Lantu and Ben, Cui and Peng, Yang and Wang, Yu and Shi, Shengnan and Lin, Yun and Gui, Guan},
journal={IEEE Transactions on Vehicular Technology},
title={A-GCRNN: Attention Graph Convolution Recurrent Neural Network for Multi-Band Spectrum Prediction},
year={2023},
doi={10.1109/TVT.2023.3315450}}