search

zjfheart / Geometry-aware-Instance-reweighted-Adversarial-Training

the paper "Geometry-aware Instance-reweighted Adversarial Training" ICLR 2021 oral
57 stars 8 forks source link

train with unlabeled data #6

Open silvercherry opened 2 years ago

silvercherry commented 2 years ago

I also have an questions about the 'entropy_weight' in learning from additional unlabeled data. I notice the 'entropy_weight' is '0'; I would like to ask, why do you want to set it is ‘0’？why not subtract the unlabeled data loss ？

Best wishes

ZFancy commented 2 years ago

Hi silvercherry,

Thanks for your question!

For the entropy_weight, we keep the same with the original code [https://github.com/yaircarmon/semisup-adv/blob/master/robust_self_training.py] in Carmon et.al 2019. The unique part we changed is the KL loss terms, you can try to set different weight values to test the performance.

Best regards, Jianing