Low Complexity Multi-User MIMO Precoding Design

Descriptions

This repository contains the code for our TWC work "A Deep Learning-Based Framework for Low Complexity Multi-User MIMO Precoding Design", available at https://ieeexplore.ieee.org/document/9834153?denied=

For any repreduce, further research or development, please kindly cite our TWC journal paper:

M. Zhang, J. Gao and C. Zhong, "A Deep Learning-Based Framework for Low Complexity Multiuser MIMO Precoding Design," in IEEE Transactions on Wireless Communications, vol. 21, no. 12, pp. 11193-11206, Dec. 2022, doi: 10.1109/TWC.2022.3190435.

Requirements

torch==1.8.0
tensorflow==2.3.0
python==3.8.0
hdf5storage
sklearn
numpy

Structures

There are several folders: ''differ_ds', '''DUU_MISO_pytorch',''execution_time_test',''imperfect_CSI',''learn_UW',''multi_RB',''single_RB',''zero_patch_new'

Generate dataset

The folder ''generate_dataset'' contains the code for dataset generation.

generate_channel.py: generate the MIMO channels

python3 generate_dataset/generate_channel.py

generate_dataset.py: generate the labels for learning

python3 generate_dataset/generate_dataset.py --Nt 64 --Nr 4 --K 10 --dk 2 --B 1 --SNR 0 --SNR_channel 100 --gpu 0 --mode gpu --batch_size 200 --epoch 1000 --factor 1

Learning based precoding for single RB MIMO system

The folder ''single_RB'' contains the code for Fig. 2.

data_preprocess.py: some functions for dataset preprocessing, including transforming method in Section III-B
train_main.py: main training functions for the method presented in Fig 2. Some baseline schemes (including ZF, WMMSE) are also provided.

python3 -u single_RB/train_main.py --Nt 64 --Nr 4 --K 10 --dk 2 --B 1 --SNR 0 --SNR_channel 100 --gpu 0 --mode gpu --batch_size 200 --epoch 1000 --factor 1

Learning based low complexity precoding design for OFDM-MIMO system

The folder ''multi_RB'' contains the code for the proposed method in Fig. 4.

data_preprocess.py: some functions for dataset preprocessing, including transforming method in Section V
train_main.py: main training functions for the method presented in Fig 4. Some baseline schemes (including ZF, WMMSE) are also provided.

python3 -u multi_RB/train_main.py --Nt 64 --Nr 4 --K 10 --dk 2 --B 4 --SNR 0 --SNR_channel 100 --gpu 0 --mode gpu --batch_size 200 --epoch 1000 --factor 1

Varying Number of User Streams

The folder ''differ_ds'' contains the code for evaluating the performance when $d_k \neq d_j$.

python3 -u differ_ds/learn_from_bar_merge_rb.py --Nt 64 --Nr 4 --K 8 --dk 2 --B 4 --SNR 0 --SNR_channel 100 --gpu 0 --mode gpu --batch_size 200 --epoch 1000 --factor 1

Varying Number of Users

The folder ''zero_patch'' contains the code for evaluating the performance when user number varies.

python3 -u zero_patch/train.py  --Nt 64 --Nr 4 --dk 2 --K 12 --SNR 0 --B 1 --SNR_channel 100 --gpu 0 --mode gpu --batch_size 200 --epoch 1000 --factor 2

Imperfect CSI

The folder ''imperfect_CSI'' contains the code for evaluating the performance when user number varies.

python3 -u zero_patch/train_main.py  --Nt 64 --Nr 4 --K 12 --dk 2 --B 1 --SNR 0 --SNR_channel 10 --gpu 0 --mode gpu --batch_size 200 --epoch 1000 --factor 1

Network Pruning

The folder ''DUU_MISO_pytorch'' contains the code for evaluating the performance when user number varies.

python3 -u DUU_MISO_pytorch/prune_naive.py --Nt 64 --Nr 4 --dk 2 --K 16 --SNR 0 --SNR_channel 100 --gpu 0 --mode gpu  --batch_size 200 --epoch 1000 --factor 2

MauroZMJ/Deep_Learning_based_Low_complexity_MIMO_Precoding_Design