search

TianLin0509 / BF-design-with-DL

Beamforming design with deep learning.
258 stars 105 forks source link

How to calculate the loss for imperfect CSI case? #2

Open SharonWxy opened 5 years ago

SharonWxy commented 5 years ago

First I’d like to thank you very much for providing the code online for further learning. Then, if you get some time, may I ask some questions regarding your code? I put my question here as following:

  1. For imperfect CSI case, to calculate the loss, do you use the true H, or the estimated H? I think it should be the true H, right?
  2. If you use the real H, how do you pass the value of true H to the self-defined loss function?

I really appreciate your reply, and thank you very much in advance.

TianLin0509 commented 5 years ago
  1. For imperfect CSI case, we calculate the loss with the true H during the offline training. (Noticed when deployed online, it is not needed).
  2. In this code, we use the input layer of Keras to input the true H, which is directly used to compute the loss. (Only used to compute loss and did not through the neural network, so when deployed online, i.e., the testing procedure, only imperfect CSI is needed to fed into the BFNN to compute the analog beamformer.
  3. Thank you for your approval of this work.
SharonWxy commented 5 years ago

Thank you very much.~

AlirezaJahanbakhshi commented 4 years ago

Dear Mr. Lin, I also have question about calculation of loss function. would you please tell where did you exactly calculate loss function in your code? I appreciate your kind cooperation

TianLin0509 commented 4 years ago

The calculation of loss function, is discussed in the issues of the repo. You can refer to it, because some other people have the similar question, so I answered it already. The channels is generated via the codes provided via the author of "Reliable Beamspace Channel Estimation for Millimeter-Wave Massive MIMO Systems with Lens Antenna Array", you can find it in the authors'own website. I just used his codes.

You can also try other channel models, hopefully my BFNN can also work.

-----原始邮件----- 发件人:AlirezaJahanbakhshi notifications@github.com 发送时间:2020-02-25 17:48:57 (星期二) 收件人: TianLin0509/BF-design-with-DL BF-design-with-DL@noreply.github.com 抄送: "Lin Tian" 17210720082@fudan.edu.cn, Comment comment@noreply.github.com 主题: Re: [TianLin0509/BF-design-with-DL] How to calculate the loss for imperfect CSI case? (#2)

Dear Mr. Lin, I also have question about calculation of loss function. would you please tell where did you exactly calculate loss function in your code? I appreciate your kind cooperation

— You are receiving this because you commented. Reply to this email directly, view it on GitHub, or unsubscribe.

AlirezaJahanbakhshi commented 4 years ago

Thank you very much for your answer.