This is the simulation code for the paper "Beamforming Design for Large-Scale Antenna Arrays Using Deep Learning". This paper is published on IEEE Wireless Communication Letters.
Arxiv link: https://arxiv.org/abs/1904.03657
I recommand the pre-print version on arxiv.
Also, a Chinese-version blog can be referred to CSDN blog
- Tensorflow-gpu = 1.12.0
Now it supports tf 2.3.0, just run the file train_v2.py
Main revision is that the API batch_dot is different from tensorflow 1
(Tensorflow 1.12.0 is better for debugging, while tensorflow 1.13.0 using cuda10 can run faster)
If you are confused about how to have several different tensorflows and cudas of different versions in one computer, there is a easy guide may help you (in Chinese).
After fork the repo and download the corresponding data sets and trained models, the following performance results can be easily reproduced. (the python codes is only for the blue cerves, and compared cerves should be plot via Matlab codes)
How to use:
- run the train.py to train the model
- run the test.py to test the trained model
- Due to the space limitation of github, we provide two tiny training and testing data sets only for running the example.
Data sets and trained models
- Thanks for the authors of , the simulation codes of the channel is provided here. you can use the direct URL to download it.
- The channel estimation codes can be referred to the website of the first author of .
- We have provided the data sets in our google drive, which can be directly used in our .py files.
- Trained weights, corresponding to the shown results in the paper, is also provided in the google drive.
Due to some readers requiremnets, for Chinese people, the BaiduYun URL is also provided. (password: z9un).
Some coding tricks are used to fit the Keras framework, for example, the loss function is written in an unique way, which is described in the issues and have been questioned many times.
For many readers requirements, I have updated the matlab code for samples generation. Please kindly refer to the gen_samples.m for details. The codes are based on the work .
- Matlab codes of compared algorithms [4，5] can be referred to this repo. In addition, if you are interested in traditional HBF algorithms, you can kindly refer to our previous work Hybrid Beamforming for Millimeter Wave Systems Using the MMSE Criterion, we also provide specific Matlab codes for your reproduction. You can refer to this repo.
- If you have any questions, you can contact the author via
- Hopefully you can star or fork this repo if it is helpful. Thank you for your support in advance.