search

danfenghong / RSE_Cross-city

Danfeng Hong, Bing Zhang, Hao Li, Yuxuan Li, Jing Yao, Chenyu Li, Martin Werner, Jocelyn Chanussot, Alexander Zipf, Xiao Xiang Zhu. Cross-City Matters: A Multimodal Remote Sensing Benchmark Dataset for Cross-City Semantic Segmentation using High-Resolution Domain Adapation Networks, Remote Sensing of Enviroment, 2023.
49 stars 4 forks source link

can't reproduce the accuracy in the paper #1

Closed xiaoyan07 closed 10 months ago

xiaoyan07 commented 10 months ago

Hi, Prof. Hong:

Thanks for your great work and contribution, I recently run the original code: Train_HighDAN_beijing.py with the default parameter. The changing trend of mIoU is as follows, and I couldn't get the mIoU: 11.92% in the RSE paper. I wonder if some of the parameters in the training file are different from the ones you actually use？Can you share your trained model weights? Thank you so much, and look forward to your reply. miou_score

danfenghong commented 10 months ago

I guess the possible reason is about the version: you can try the torch 2.0.0 with Python 2.8.2. We checked and it works on our server.

Xiaoyan Lu @.***> 于2023年12月24日周日 17:20写道：

Hi, Prof. Hong:

Thanks for your great work and contribution, I recently run the original code: Train_HighDAN_beijing.py with the default parameter. The changing trend of mIoU is as follows, and I couldn't get the mIoU: 11.92% in the RSE paper. I wonder if some of the parameters in the training file are different from the ones you actually use？Can you share your trained model weights? Thank you so much, and look forward to your reply. miou_score.jpg (view on web) https://github.com/danfenghong/RSE_Cross-city/assets/35627108/f07580e4-aa9a-4220-ab16-ec16bbc9121b

— Reply to this email directly, view it on GitHub https://github.com/danfenghong/RSE_Cross-city/issues/1, or unsubscribe https://github.com/notifications/unsubscribe-auth/AFL2GZR7SSCKNTAZJDVPGG3YK7XU7AVCNFSM6AAAAABBBMZCBCVHI2DSMVQWIX3LMV43ASLTON2WKOZSGA2TKMBTHA4DQMY . You are receiving this because you are subscribed to this thread.Message ID: @.***>

Xuaner624 commented 10 months ago

Hi, Prof. Hong:

Thanks for your great work and contribution, I recently run the original code: Train_HighDAN_beijing.py with the default parameter. The changing trend of mIoU is as follows, and I couldn't get the mIoU: 11.92% in the RSE paper. I wonder if some of the parameters in the training file are different from the ones you actually use？Can you share your trained model weights? Thank you so much, and look forward to your reply. miou_score

We have updated the pretraining weights we used in the paper.

xiaoyan07 commented 10 months ago

I guess the possible reason is about the version: you can try the torch 2.0.0 with Python 2.8.2. We checked and it works on our server. Xiaoyan Lu @.> 于2023年12月24日周日 17:20写道： … Hi, Prof. Hong: Thanks for your great work and contribution, I recently run the original code: Train_HighDAN_beijing.py with the default parameter. The changing trend of mIoU is as follows, and I couldn't get the mIoU: 11.92% in the RSE paper. I wonder if some of the parameters in the training file are different from the ones you actually use？Can you share your trained model weights? Thank you so much, and look forward to your reply. miou_score.jpg (view on web) https://github.com/danfenghong/RSE_Cross-city/assets/35627108/f07580e4-aa9a-4220-ab16-ec16bbc9121b — Reply to this email directly, view it on GitHub <#1>, or unsubscribe https://github.com/notifications/unsubscribe-auth/AFL2GZR7SSCKNTAZJDVPGG3YK7XU7AVCNFSM6AAAAABBBMZCBCVHI2DSMVQWIX3LMV43ASLTON2WKOZSGA2TKMBTHA4DQMY . You are receiving this because you are subscribed to this thread.Message ID: @.>

Thanks to Prof. Hong's and Xuaner's advice and patience, after adjusting the batch size to 16 (which is the same as in the paper) and searching for the best weights epoch by epoch, the result is much improved and very close to the paper.