question in base-to-novel setting

zhengli97 / PromptKD

[CVPR 2024] Official PyTorch Code for "PromptKD: Unsupervised Prompt Distillation for Vision-Language Models"

https://zhengli97.github.io/PromptKD/

Apache License 2.0

221 stars 3 forks source link

question in base-to-novel setting #27

Open jihwanp opened 3 days ago

jihwanp commented 3 days ago

Hi i have a question about base-to-novel setting. When training PromptKD in this setting, do you also use image samples for unseen classes, or do you only use those for the base classes? It isn’t clear because the paper specifies only cross-dataset as transductive. Thanks

zhengli97 commented 3 days ago

The base-to-novel setting is also transductive ZSL. We mentioned this approach in Section 4.1 Base-to-novel Generalization.