search

Gank0078 / FineSSL

Pytorch implementation for "Erasing the Bias: Fine-Tuning Foundation Models for Semi-Supervised Learning" (ICML 2024)
13 stars 0 forks source link

self.betabase is not updated. #1

Open DevinCheung opened 1 month ago

DevinCheung commented 1 month ago

Thanks for the great work. I am confused with self.betabase in trainer.py. Does it refer to ∆^{t}_{y} in Eq. (3) in the main paper? Also it seems that it is initialized but not updated during training. Coud you please help explain this? Thanks!

Gank0078 commented 1 month ago

Thank you for your interest in our work! self.betabase in our code is applied to constrcut margins for balanced margin softmax, so it is unnecessary to update it. alpha[targetsx] * self.betabase is refer to ∆^{t}{y} in Eq. (3).