This is the Tensorflow implementation of the paper: Jianhua Guo, Lei Wang, Feng Li, and Jiang Xue, "CSI Feedback with Model-Driven Deep Learning of Massive MIMO Systems" .
- Python 3.6
- Tensorflow 1.15.0
- tflearn 0.3.2
- Numpy
In OFDM system, we use the channel state information (CSI) matirx is generated by the COST2100 channel model. Specifically, we use the pre-processed CSI dataset provided by Chao-Kai Wen, Wan-Ting Shih, and Shi Jin in the repository and you can download the dataset from Google Drive and put it in DATA/ folder.
Train the FISTA-Net from the scratch in different CRs and scenarios with
python FISTA_Net.py
The CSI reconstruction results by FISTA-Net are presented as follows. We also provides the pre-trained model to reproduce this results with
python FISTA_Net_test.py
| CR | Methods | Indoor | Outdoor | Trainable Params | MACC | |
|---|---|---|---|---|---|---|
| NMSE | NMSE | Encoder | Decoder | |||
| 1/4 | CsiNet | -17.36 | -8.75 | 2.10M | 1.09M | 4.39M |
| CsiNet+ | -27.37 | -12.4 | 2.12M | 1.45M | 23.26M | |
| FISTA | -10.46 | -6.35 | - | 1.05M | 41.94M | |
| FISTA-Net | -36.76 | -22.4 | 1.09M | 1.05M | 74.71M | |
| 1/8 | CsiNet | -12.7 | -7.61 | 1.05M | 0.56M | 3.86M |
| CsiNet+ | -18.29 | -8.72 | 1.07M | 0.93M | 22.73M | |
| FISTA | -6.39 | -2.91 | - | 0.52M | 20.97M | |
| FISTA-Net | -26.5 | -13.65 | 0.56M | 0.52M | 53.74M | |
| 1/16 | CsiNet | -8.65 | -4.51 | 0.53M | 0.30M | 3.60M |
| CsiNet+ | -14.14 | -5.73 | 0.55M | 0.67M | 22.47M | |
| FISTA | -3.18 | -1.15 | - | 0.26M | 10.49M | |
| FISTA-Net | -17.51 | -7.57 | 0.30M | 0.26M | 43.26M | |
| 1/32 | CsiNet | -6.24 | -2.81 | 0.27M | 0.17M | 3.47M |
| CsiNet+ | -10.43 | -3.4 | 0.29M | 0.54M | 22.34M | |
| FISTA | -1.11 | -0.35 | - | 0.13M | 5.24 M | |
| FISTA-Net | -12.01 | -4.41 | 0.17M | 0.13M | 38.01M | |
| 1/64 | CsiNet | -5.84 | -1.93 | 0.14M | 0.11M | 3.40M |
| CsiNet+ | -5.99 | -2.22 | 0.16M | 0.47M | 22.27M | |
| FISTA | -0.29 | -0.05 | - | 0.07M | 2.62M | |
| FISTA-Net | -8.54 | -2.60 | 0.10M | 0.07M | 35.39M |
The results and dataset with low-rank mmWave channel matrix by FISTA-Nets will be add this repository in the future.
If you find our paper and code are helpful for your research or work, please cite our paper.
@ARTICLE{9663378,
author={Guo, Jianhua and Wang, Lei and Li, Feng and Xue, Jiang},
journal={IEEE Communications Letters},
title={CSI Feedback with Model-Driven Deep Learning of Massive MIMO Systems},
year={2021},
volume={},
number={},
pages={1-1},
doi={10.1109/LCOMM.2021.3138927}}