my results is not same your results in your paper

[TranNgocHuyen]

when I retrain the model risnet , my results is higher your results in your paper, and gain good results at 10 iteration. I cannot explain my results.

[bilepeng]

Hi, thank you for your interest. I cannot explain it without many details but I am happy that you reach higher results.

Bile

On 30. May 2024, at 12:29, TranNgocHuyen-3463 @.***> wrote:

when I retrain the model risnet , my results is higher your results in your paper, and gain good results at 10 iteration. I cannot explain my results.

— Reply to this email directly, view it on GitHub https://github.com/bilepeng/risnet/issues/1, or unsubscribe https://github.com/notifications/unsubscribe-auth/ABFKK45A4ONCPPHLNCKSQLDZE35QNAVCNFSM6AAAAABIQUUFRSVHI2DSMVQWIX3LMV43ASLTON2WKOZSGMZDKMZRGE3TCMA. You are receiving this because you are subscribed to this thread.

[TranNgocHuyen]

OMG, thanks for your reply TvT This is my results I cannot explain my results is higher your results in your paper when my teacher ask me. I train in my server, it is gpu RTX 3090 And I have some questions :

1. In the model, why in final layer,r = self.conv8(r) * np.pi and I know r= ψn but ψn∈ [0, 2π), How???
2. IN your paper, In the final layer (i = L), the output of the RISnet is computed as fL+1 = ReLU() but I don't see it
3. Would you like to explain about WMMSE precoding, I don't see in your paper, it is difficult :((((( Please help me

[bilepeng]

1. We do not have constraints on the phase since it is periodical anyway.
2. Sorry that was a mistake. We do not have relu in the final layer.
3. It is complicated to explain… You can read the paper https://ieeexplore.ieee.org/document/5756489.

On 30. May 2024, at 13:58, TranNgocHuyen-3463 @.***> wrote:

OMG, thanks for your reply TvT Figure_1.png (view on web) https://github.com/bilepeng/risnet/assets/146775773/81095fd1-cfbc-4908-a771-ad23afa627ba This is my results. and I have some questions :

In the model, why in final layer, r = self.conv8(r) * np.pi and I know r= ψn but ψn∈ [0, 2π), How??? IN your paper, In the final layer (i = L), the output of the RISnet is computed as fL+1 = ReLU() but I don't see it Would you like to explain about WMMSE precoding, I don't see in your paper, it is difficult :((((( Please help me — Reply to this email directly, view it on GitHub https://github.com/bilepeng/risnet/issues/1#issuecomment-2139397648, or unsubscribe https://github.com/notifications/unsubscribe-auth/ABFKK46USTZ73PBFPI5HO2TZE4H5PAVCNFSM6AAAAABIQUUFRSVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZDCMZZGM4TONRUHA. You are receiving this because you commented.

[TranNgocHuyen]

Thanks so much for your answer.

I continually search for the reasons why my results is higher when i don't have any edit.

• Would you like to give me some information about environment requirements and your training computer?
• And I wonder about above things:
  1. When I train , have some warning:

     /mnt/2E287FDD287FA28F/teamAI/Huyen/Risnet_env/lib/python3.10/site-packages/scipy/optimize/_minpack_py.py:177: RuntimeWarning: The iteration is not making good progress, as measured by the 
     improvement from the last ten iterations.
     warnings.warn(msg, RuntimeWarning)
     WMMSE round: 12/400, Epoch: 0/1, WSR = 1.069804072380066
     WMMSE round: 13/400, Epoch: 0/1, WSR = 1.1867262125015259
     /mnt/2E287FDD287FA28F/teamAI/Huyen/Risnet_env/lib/python3.10/site-packages/scipy/optimize/_minpack_py.py:177: RuntimeWarning: The iteration is not making good progress, as measured by the 
     improvement from the last five Jacobian evaluations.
     warnings.warn(msg, RuntimeWarning)

     but my training also finish.

  2. My results gain high suddenly and very soon in 10 iteration but in your results, gain high in 50th iteration my results: Your results:

[bilepeng]

1. I used Pytorch 2.1 on the server of the university: https://www.tu-braunschweig.de/it/hpc
2. The warning is because the search for mu (Eq. 18 in https://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=5756489) is not numerically stable. I think it is a problem of WMMSE (or my implementation), but the simulation can finish.
3. I do not know the reason... have you changed any code?

[TranNgocHuyen]

I'm sorry, but I haven't changed any code. That's exactly what I'm worried about, I can't explain this difference. Is the results picture in your paper the results of this code? Did you change any code, such as WMMSE precoding, because I see many version precoding code in util.py? And how did you collect dataset?

[bilepeng]

Me neither… maybe because of the update of PyTorch?

I think it is fine since your results are better than mine. A good improvement without an unknown reason :-)

On 4. Jun 2024, at 14:46, TranNgocHuyen-3463 @.***> wrote:

I'm sorry, but I haven't changed any code. That's exactly what I'm worried about, I can't explain this difference. Is the results picture in your paper the results of this code? Did you change any code, such as WMMSE precoding, because I see many version precoding code in util.py?

— Reply to this email directly, view it on GitHub https://github.com/bilepeng/risnet/issues/1#issuecomment-2147449192, or unsubscribe https://github.com/notifications/unsubscribe-auth/ABFKK464VKOYQNDF42OHCUTZFWZLLAVCNFSM6AAAAABIQUUFRSVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZDCNBXGQ2DSMJZGI. You are receiving this because you commented.