Finspire13 / CMCS-Temporal-Action-Localization

Completeness Modeling and Context Separation for Weakly Supervised Temporal Action Localization (CVPR2019)
MIT License
150 stars 17 forks source link

Regarding the exact value of converged div loss and norm loss #24

Open June01 opened 4 years ago

June01 commented 4 years ago

Hi,

Thanks for releasing the code! I tried only your div loss and norm loss in my own dataset, but it seems the losses doesn't behave as expected. So if I may ask for help, the questions are:

  1. I use $\alpha L{div} + \beta L{norm}$, while $\alpha$ and $\beta$ are set to be 0.2 as in your paper. During the training process, $L{norm}$ increased a bit from 0 and then stopped changing at around 0.04 or 0.05, to make branches to be different; at the same time, $L{div}$ decreased at the beginning from 0.36 and stopped at 0.05. Are the losses reasonable according to your experience? If not, could you think of any mistake that I could make to cause it?
  2. As to the $L_{mil}$, I am a little bit confusing, did you use the mean of all branches to serve as the final T-CAM for loss? Or you calculate the classification loss in each branch?

Thanks for reading. Looking forward to your response!

June