RoyZry98 / MoFME-Pytorch

[AAAI 2024] Official code for Efficient Deweather Mixture-of-Experts with Uncertainty-aware Feature-wise Linear Modulation
16 stars 1 forks source link

about balance loss and uncertainty loss #1

Open azcdk opened 4 months ago

azcdk commented 4 months ago

It seems that load balance loss and uncertainty loss are not being used in the MoFME. It does have an l_aux, maybe an auxiliary loss, but it is initialized with 'None'.

syrGitHub commented 4 months ago

I also have the same question, can the author answer it?

RoyZry98 commented 3 months ago

Thank you for your interest in our work. In subsequent experiments, we found that MoFME can outperform MoE without the need for these two loss functions. To minimize computational load and enhance model efficiency, we have removed these loss functions from the code. We'll also make further exploration to find the optimal experimental setting to boost the effectiveness of the load balance loss and uncertainty loss for the best model performance in the future.

Yuki2L0ve commented 2 months ago

Could the author please indicate which part of the code reflects Uncertainty-aware?