jasonkyuyim / multiflow

https://arxiv.org/abs/2402.04997
MIT License
106 stars 5 forks source link

Question about loss weight #3

Closed Wangchentong closed 3 months ago

Wangchentong commented 5 months ago

Hi Jason,

Nice work!

I notice that the auxiliary loss weight in base.yaml is set to 0, which is different from the FrameDiff and Frameflow, do you observe a better performance by removing the auxiliary loss?

jasonkyuyim commented 5 months ago

Hi, we noticed the auxiliary losses are not necessary anymore. It simplified training so we dropped it. There wasn't a loss in performance from my experiments.