help

[shuchao97]

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

train in usa-singapore, test in singapore
@tuanhungvu @maxjaritz @tristanschultz

[xingbw]

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]

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.

[xingbw]

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?

[xingbw]

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]

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