search

hytseng0509 / CrossDomainFewShot

Cross-Domain Few-Shot Classification via Learned Feature-Wise Transformation (ICLR 2020 spotlight)
329 stars 62 forks source link

Auxiliary Loss #9

Open Franklin-Yao opened 4 years ago

Franklin-Yao commented 4 years ago

Hi,

I saw you updated your code and added auxiliary classifier. Why can it stabilize the training? Where did you get this idea?

hytseng0509 commented 4 years ago

The training in the initial stage is not stable and may harm the model performance. We use the auxiliary training to solve the problem and decay the weight of the auxiliary training loss for later epochs.

Franklin-Yao commented 4 years ago

Why does this auxiliary training help stabilize the training? The loss may let your model fit the val dataset and give worse acc for test dataset.