search

WangYueFt / dgcnn

MIT License
1.62k stars 420 forks source link

Serious problem. About label smoothing in loss #49

Closed a18700 closed 4 years ago

a18700 commented 4 years ago

While examining provided codes,

I found out label smoothing term is applied for cross entroy loss.

However, there is no description about that in the paper.

I doubt for the effects of this term.. Is there any official results without label smoothing term

a18700 commented 4 years ago

I've experimented without label smoothing, and found out performance is hugely deterior to paper's. I think this should be clarified...

Maybe i think it's possible that performance of reported dgcnn is from effects of label smoothing, not a virtue of architectural innovations.

WangYueFt commented 4 years ago

I've experimented without label smoothing, and found out performance is hugely deterior to paper's. I think this should be clarified...

Maybe i think it's possible that performance of reported dgcnn is from effects of label smoothing, not a virtue of architectural innovations.

Hi,

Are you running the pytorch code? I re-run the code without label smoothing and the result is very close.