search

zilongzhong / SSTN

PyTorch Implementation of SSTNs for hyperspectral image classifications from the IEEE T-GRS paper "Spectral-Spatial Transformer Network for Hyperspectral Image Classification: A FAS Framework."
106 stars 23 forks source link

How to Run the code? Thank you! #2

Open AlanMorningLight opened 3 years ago

AlanMorningLight commented 3 years ago

Dear zilong, I have run the code with the command "python train_UP.py --batch_size 50 --learning_rate 0.001 --gpu 0 --epochs 200 --model SSTN --phi EAAE" in prcharm terminal on win10 system. However, ① I get the Multithreaded error. I modified the num_works=0, The program does not report an error, but no result. ② I also add the "if name == 'main':" in the train_UP.py,Otherwise the program cannot run. ③ Can you give an execution command that can run the program correctly？ I also have some questions which have been sent to your email z26zhong@uwaterloo.ca. Thank you very much.

zilongzhong commented 2 years ago

Dear Wenchao, thanks a lot for your detailed feedback. The repo has been updated with proper explanation and the commands to run codes can be found in the Usage section. Here are my responses to your five questions: 1) Yes, that's one reason. Also, additional experiments are requested by reviewers. 2) The layer-level searching is for FAS. After layer-level searching, we focus on the architecture-level sequential setting. 3) Spectral association is different from spectral attention. We have conducted preliminary experiments that show spectral attention did not work well as expected. Therefore, we further design a spectral association module to improve it. The essence of the spectral association kernel is the kxc spectral kernels, the physical meaning of which is the most prominent spectral features of input feature maps. This step also drastically decreases computational cost. 4) Compared to both the spatial attention module and convolutional module, the spectral association kernels are sparse.
5) The adopted sampling method is without replacement. You understand that Val and Dev sets are generally correct.

AlanMorningLight commented 2 years ago

Dear Wenchao, thanks a lot for your detailed feedback. The repo has been updated with proper explanation and the commands to run codes can be found in the Usage section. Here are my responses to your five questions:

  1. Yes, that's one reason. Also, additional experiments are requested by reviewers.
  2. The layer-level searching is for FAS. After layer-level searching, we focus on the architecture-level sequential setting.
  3. Spectral association is different from spectral attention. We have conducted preliminary experiments that show spectral attention did not work well as expected. Therefore, we further design a spectral association module to improve it. The essence of the spectral association kernel is the kxc spectral kernels, the physical meaning of which is the most prominent spectral features of input feature maps. This step also drastically decreases computational cost.
  4. Compared to both the spatial attention module and convolutional module, the spectral association kernels are sparse.
  5. The adopted sampling method is without replacement. You understand that Val and Dev sets are generally correct.

Thank you very much for your serious reply one by one during your busy schedule. I have benefited a lot from your paper. My contact information is included in the email. If you are convenient, you can also directly add my WeChat to facilitate future academic discussions and exchanges. thanks for your help. Good Luck!!

AlanMorningLight commented 2 years ago

Dear zilong, I think that the FAS codes are not included in the code you are shared. Is it right? Thank you very much. Best wishes.

zilongzhong commented 2 years ago

Dear Wenchao, Yes, since the FAS can be regarded as an independent part of the SSTN model. Best regards, Zilong

On Sun, 9 Jan 2022 at 23:52, Wenchao QI @.***> wrote:

Dear zilong, I think that the FAS codes are not included in the code you are shared. Is it right? Thank you very much. Best wishes.

— Reply to this email directly, view it on GitHub https://github.com/zilongzhong/SSTN/issues/2#issuecomment-1008312244, or unsubscribe https://github.com/notifications/unsubscribe-auth/AD2DHTR7GIRTSHBIFZWLSPDUVGOMPANCNFSM5HPNICRA . Triage notifications on the go with GitHub Mobile for iOS https://apps.apple.com/app/apple-store/id1477376905?ct=notification-email&mt=8&pt=524675 or Android https://play.google.com/store/apps/details?id=com.github.android&referrer=utm_campaign%3Dnotification-email%26utm_medium%3Demail%26utm_source%3Dgithub.

You are receiving this because you commented.Message ID: @.***>