search

Yuqifan1117 / CaCao

This is the official repository for the paper "Visually-Prompted Language Model for Fine-Grained Scene Graph Generation in an Open World" (Accepted by ICCV 2023)
40 stars 5 forks source link

How to train openworld? #5

Closed ZHUXUHAN closed 11 months ago

ZHUXUHAN commented 11 months ago

I just use the openword/files to train the baseline motifs model, and just test the model (588 classes) with evulating the base and novel classes' Recall, is it right?

Yuqifan1117 commented 11 months ago

yes, it seems right, however, you should predict novel classes by unified embedding mapping (Open-World Predicate SGG Backbone).

ZHUXUHAN commented 11 months ago

i don't understand what is the unified embedding mapping? if i train the model as what i say, is the output embedding of the model is the ''unified embedding mapping'?

Yuqifan1117 commented 11 months ago

yes

ZHUXUHAN commented 11 months ago

yes

ok, thanks for your answer.