search

MendelXu / SAN

Open-vocabulary Semantic Segmentation
https://mendelxu.github.io/SAN/
MIT License
295 stars 27 forks source link

question about clip-aware and clip-unaware #9

Closed jainie-max closed 1 year ago

jainie-max commented 1 year ago

Hi, great work! I have some questions about clip aware and clip unaware in the paper. Since the clip model is locked, why the gradients can pass through clip in an e2e training manner and block in a two-stage manner training in paper fig.5?

MendelXu commented 1 year ago

The fact that the model is locked only means the parameters of the model will not be updated during training. The gradient of the activation still exists as the input to the deepr layer requires gradient.

jainie-max commented 1 year ago

get it！Tanks