(c) 2018 Wang-Ting Shih and Chao-Kai Wen e-mail: [email protected] and [email protected]

This repository contains the original models described in Chao-Kai Wen, Wan-Ting Shih, and Shi Jin, “Deep learning for massive MIMO CSI feedback,” IEEE Wireless Communications Letters, 2018. [Online]. Available: https://ieeexplore.ieee.org/document/8322184/

• Python 3.5 (or 3.6)
• Keras (>=2.1.1)
• Tensorflow (>=1.4)
• Numpy

There are two models in the paper:

• CsiNet: CSI sensing (or encoder) and recovery (or decoder) network
• CS-CsiNet: Only learns to recover CSI from CS random linear measurements

We provide two types of code:

• xxx_onlytest: This type of code is used to reproduce the results in our paper based on our training weights. The model and weights we trained are put in folder 'saved_model'.
• xxx_train: This type of code provide a procedure to train the weights yourself.

Download the data from https://www.dropbox.com/scl/fo/tqhriijik2p76j7kfp9jl/h?rlkey=4r1zvjpv4lh5h4fpt7lbpus8c&dl=0. After you got the data, put the data as shown below.

*.py
saved_model/
  *.h5
  *.json
data/
  *.mat

Now, you are ready to run any *.py to get the results (i.e., CS-CsiNet and CsiNet in Table I of our paper).

The following results are reproduced from Table I of our paper:

1. The file DATA_HtestFin_all.mat is mainly used to calculate $\rho$. According to the definition in the article, we should use all subcarriers to calculate to get the average. However, due to the limitation of the computer, we only compared the first 125 sub-carriers. In other words, the dimension of the matrix in DATA_HtestFin_all.mat, 4000, is the result of 125 (subcarriers) *32 (antenna).
2. The source code of the CsiNet-LSTM can be found in the Book "Intelligent communication: physical layer design based on deep learning".