waxpple / crnet Goto Github PK

This is the PyTorch implementation of paper Multi-resolution CSI Feedback with deep learning in Massive MIMO System.

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

• Python >= 3.7
• PyTorch >= 1.2
• thop

The channel state information (CSI) matrix is generated from COST2100 model. Chao-Kai Wen and Shi Jin group provides a pre-processed version of COST2100 dataset in Google Drive, which is easier to use for the CSI feedback task; You can also download it from Baidu Netdisk.

You can generate your own dataset according to the open source library of COST2100 as well. The details of data pre-processing can be found in our paper.

The model checkpoints should be downloaded if you would like to reproduce our result. All the checkpoints files can be downloaded from Baidu Netdisk or Google Drive

We recommend you to arrange the project tree as follows.

home
├── CRNet  # The cloned CRNet repository
│   ├── dataset
│   ├── models
│   ├── utils
│   ├── main.py
├── COST2100  # The data folder
│   ├── DATA_Htestin.mat
│   ├── ...
├── Experiments
│   ├── checkpoints  # The checkpoints folder
│   │     ├── in_04.pth
│   │     ├── ...
│   ├── run.sh  # The bash script
...

An example of run.sh is listed below. Simply use it with sh run.sh. It will start advanced scheme aided CRNet training from scratch. Change scenario using --scenario and change compression ratio with --cr.

python /home/CRNet/main.py \
  --data-dir '/home/COST2100' \
  --scenario 'in' \
  --epochs 2500 \
  --batch-size 200 \
  --workers 0 \
  --cr 4 \
  --scheduler cosine \
  --gpu 0 \
  2>&1 | tee log.out

The main results reported in our paper are presented as follows. All the listed results can be found in Table1 of our paper. They are achieved from training CRNet with our advanced training scheme (cosine annealing scheduler with warm up for 2500 epochs).

As aforementioned, we provide model checkpoints for all the results. Our code library supports easy inference.

To reproduce all these results, simple add --evaluate to run.sh and pick the corresponding pre-trained model with --pretrained. An example is shown as follows.

python /home/CRNet/main.py \
  --data-dir '/home/COST2100' \
  --scenario 'in' \
  --pretrained './checkpoints/in_04' \
  --evaluate \
  --batch-size 200 \
  --workers 0 \
  --cr 4 \
  --cpu \
  2>&1 | tee log.out

Thank Chao-Kai Wen and Shi Jin group again for providing the pre-processed COST2100 dataset, you can find their related work named CsiNet in Github-Python_CsiNet