BatsResearch / menghini-neurips23-code

Exploring prompt tuning with pseudolabels for multiple modalities, learning settings, and training strategies.
https://openreview.net/pdf?id=2b9aY2NgXE
42 stars 2 forks source link

quesiton in TRZSL setting #4

Closed jihwanp closed 1 week ago

jihwanp commented 2 weeks ago

Hi im new to this field and i have a minor question about TRZSL setting. At first, I thought that in TRZSL, only images corresponding to unseen classes could be accessed during training. But after reading the paper in detail, pseudo labeling seems done for the unseen classes. Is my understanding correct? Thanks

stephenbach commented 1 week ago

Yup, that's right. The "unseen" classes are pseudolabeled in this setting. They are unseen in the sense that we never have any labeled examples of them. The "transductive" part of the TRZSL name is what indicates this.