search

Cocofeat / EyeMoSt

【MICCAI 2023 Early Accept & MedIA submission】EyeMost "Reliable Multimodality Eye Disease Screening via Mixture of Student's t Distributions"
16 stars 0 forks source link

A question about loss and results #4

Open HYC01 opened 1 day ago

HYC01 commented 1 day ago

Thanks for your impressive work EyeMoSt and EyeMoSt+ in MICCAI2023 and MIA 2024. When I reproduced the work based on this code, I found that the calculated loss would be negative. image And according to the method of model selection, the final weights are saved, and the results of the inference are different from the experimental results.

Cocofeat commented 1 day ago

Yes, loss will become negative because of the NIG distribution, which will cause a negative situation. You can refer to the following two papers:

  1. Deep evidential regression
  2. Trustworthy multimodal regression with mixture of normal-inverse gamma distributions

By the way, what did you test the results on?