search

valeoai / xmuda

Cross-Modal Unsupervised Domain Adaptationfor 3D Semantic Segmentation
Other
194 stars 36 forks source link

help #21

Open shuchao97 opened 2 years ago

shuchao97 commented 2 years ago

Hello author, when I implemented your code recently, I found that the training results of the model have some large deviations from the results of your paper. The training environment I used is 3060 、 pytorch1.10、cuda11.3、cudnn8 .2 1 train in usa-singapore, test in singapore @tuanhungvu @maxjaritz @tristanschultz

xingbw commented 2 years ago

hello, I got the same results with you, showing large deviations with the result in paper. Have you find anything wrong in training, or can you reproduce the result in paper now? thank you!

shuchao97 commented 2 years ago

hello, i have trained the model again, i find the same results with the article. Now, I have finished the traning process with xmuda without pl. there is test result. xmuda-new-singapore

xingbw commented 2 years ago

Thanks for your reply! I have also trained the model several times, and each time the 2d+3d IOU is only near 60(as follows), not so good as the article says and not so good as your result. I wonder how do you reproduce the result. Have you changed anything in code or in the config? Should I just try more times? image

xingbw commented 2 years ago

I noticed that you leaved an issue in DsCML repo, and I also meet the same problem like you in that repo. Is it convenient to add friend with you? my Wechat number is xing_bowei. Looking forward to have further discussion with you.

DouGuangjian commented 1 year ago

hello ，I used this loss training my network to encounter loss nan。How to solve it？