Kylin9511/GsmEFBNet

Overview

This is the PyTorch implementation of paper <Deep Learning for Hybrid Beamforming with Finite Feedback in GSM Aided mmWave MIMO Systems>, which has been submitted to IEEE for possible publication. The general spatial modulation scheme is first considered into the joint optimized pipeline of channel estimation, CSI feedback, and beamforming. The test script and trained models are listed here and the key results can be reproduced as a validation of our work.

Requirements

To use this project, you need to ensure the following requirements are installed.

Python >= 3.8
PyTorch >= 1.10.1
fvcore

Project Preparation

A. Data Preparation

The channel state information (CSI) matrix is generated according to the influential clustered Saleh Valenzuela (SV) model. The test dataset is provided Baidu Netdisk(passwd: 842t), which is easy for you to download and reproduce the experiment results. You can also generate your own dataset according to the SV channel model. The details of data pre-processing can be found in our paper.

B. Checkpoints Downloading

The model checkpoints should be downloaded if you would like to reproduce our result. All the checkpoints files can be downloaded from Baidu Netdisk(passwd: wh9v).

C. Project Tree Arrangement

We recommend you to arrange the project tree as follows.

home
├── GsmEFBNet  # The cloned GsmEFBNet repository
│   ├── dataset
│   ├── model
│   ├── utils
│   ├── main.py
├── data_files # Test data files
│   ├── Data_Nt16Nr4Nrf2Ncl2Nray8_2000.mat
├── Nt16Nr4Nrf2Nk4      # Checkpoints under the scenario (Nt=16,Nr=4,Nrf=2,Nk=4)
│   ├── AblationOnB  # Checkpoints for the ablation study vs. B (SNR=10dB)
│   │     ├── model_B3.pth
│   │     ├── ...
│   ├── AblationOnSNRWithB6 # Checkpoints for the ablation study vs. SNR (B=6)
│   │     ├── model_SNR-3.pth
│   │     ├── ...
│   ├── AblationOnSNRWithB36 # Checkpoints for the ablation study vs. SNR (B=36)
│   │     ├── model_SNR-3.pth
│   │     ├── ...
├── evaluate.sh  # The test script
...

Results and Reproduction

The main results of the deep learning method reported in our paper are presented in the following tables. All the listed results are marked in Fig. 3 and Fig. 4 in our paper. Our proposed GsmEFBNet first introduces the general spatial modulation (GSM) scheme into the end-to-end jointly optimization pipeline of channel estimation, CSI feedback and beamforming.

The performance and corresponding checkpoints of Fig. 3 in the paper is given as follows.

Feedback Bits B	Sum Rate (Bits/s/Hz)	Checkpoint
3	8.5679	Nt16Nr4Nrf2Nk4/AblationOnB/model_B3.pth
10	8.8024	Nt16Nr4Nrf2Nk4/AblationOnB/model_B10.pth
20	9.5025	Nt16Nr4Nrf2Nk4/AblationOnB/model_B20.pth
30	10.0396	Nt16Nr4Nrf2Nk4/AblationOnB/model_B30.pth
40	10.0978	Nt16Nr4Nrf2Nk4/AblationOnB/model_B40.pth
50	10.1071	Nt16Nr4Nrf2Nk4/AblationOnB/model_B50.pth

The performance and corresponding checkpoints of Fig. 4 in the paper is given as follows.

Feedback Bits B	SNR (dB)	Sum Rate (Bits/s/Hz)	Checkpoint
6	-10	0.5604	Nt16Nr4Nrf2Nk4/AblationOnSNRWithB6/model_SNR-10.pth
6	-7	1.0139	Nt16Nr4Nrf2Nk4/AblationOnSNRWithB6/model_SNR-7.pth
6	-5	1.4414	Nt16Nr4Nrf2Nk4/AblationOnSNRWithB6/model_SNR-5.pth
6	-3	1.9514	Nt16Nr4Nrf2Nk4/AblationOnSNRWithB6/model_SNR-3.pth
6	0	3.2281	Nt16Nr4Nrf2Nk4/AblationOnSNRWithB6/model_SNR0.pth
6	3	4.4839	Nt16Nr4Nrf2Nk4/AblationOnSNRWithB6/model_SNR3.pth
6	5	5.5980	Nt16Nr4Nrf2Nk4/AblationOnSNRWithB6/model_SNR5.pth
6	7	6.9675	Nt16Nr4Nrf2Nk4/AblationOnSNRWithB6/model_SNR7.pth
6	10	8.6310	Nt16Nr4Nrf2Nk4/AblationOnSNRWithB6/model_SNR10.pth
6	13	10.8503	Nt16Nr4Nrf2Nk4/AblationOnSNRWithB6/model_SNR13.pth
6	15	12.3268	Nt16Nr4Nrf2Nk4/AblationOnSNRWithB6/model_SNR15.pth
36	-10	0.6634	Nt16Nr4Nrf2Nk4/AblationOnSNRWithB36/model_SNR-10.pth
36	-7	1.2979	Nt16Nr4Nrf2Nk4/AblationOnSNRWithB36/model_SNR-7.pth
36	-5	1.9451	Nt16Nr4Nrf2Nk4/AblationOnSNRWithB36/model_SNR-5.pth
36	-3	2.6870	Nt16Nr4Nrf2Nk4/AblationOnSNRWithB36/model_SNR-3.pth
36	0	4.1003	Nt16Nr4Nrf2Nk4/AblationOnSNRWithB36/model_SNR0.pth
36	3	5.6763	Nt16Nr4Nrf2Nk4/AblationOnSNRWithB36/model_SNR3.pth
36	5	6.8352	Nt16Nr4Nrf2Nk4/AblationOnSNRWithB36/model_SNR5.pth
36	7	8.1042	Nt16Nr4Nrf2Nk4/AblationOnSNRWithB36/model_SNR7.pth
36	10	10.0434	Nt16Nr4Nrf2Nk4/AblationOnSNRWithB36/model_SNR10.pth
36	13	12.1352	Nt16Nr4Nrf2Nk4/AblationOnSNRWithB36/model_SNR13.pth
36	15	13.5398	Nt16Nr4Nrf2Nk4/AblationOnSNRWithB36/model_SNR15.pth

As aforementioned, we provide model checkpoints for all the deep learning-based results. Our code library supports easy inference. It is worth mentioning that the inference results have a certain degree of randomness brought by the random Gaussian noise in the SubArrayGSMPilotNet.

To reproduce all these results, you need to download the given dataset and corresponding checkpoints. Also, you should arrange your projects tree as instructed. An example of evaluate.sh is shown as follows. Change the SNR and feedback bits with SNR and B, respectively. Change the number of transmit antennas, receive antennas, RF chains and users with Nt, Nr, Nrf and Nk, respectively. Change the pilot length with L.

python3 home/GsmEFBNet/main.py \
    --evaluate \
    --pretrained home/Nt8Nr2Nrf1Nk4/AblationOnB/model_B30.pth \
    --gpu 0 \
    --test-data-dir home/data_files/Data_Nt8Nr2Nrf1Ncl2Nray8_2000.mat \
    --model GsmEFBNet \
    -SNR 10 \
    -B 30 \
    -Nt 8 \
    -Nr 2 \
    -Nrf 1 \
    -Nk 4 \
    -TP 1 \
    -L 4 \
    2>&1 | tee log_inference.txt

Acknowledgment

This repository is constructed referring to DL-DSC-FDD. Thank Foad Sohrabi, Kareem M. Attiah, and Wei Yu for their excellent work and you can find it in detail from this paper.

Kylin9511 / GsmEFBNet