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

where is train.sh? #4

Closed ZHUXUHAN closed 11 months ago

ZHUXUHAN commented 11 months ago

where can i find this train.sh file in 'bash train.sh TRANSGLOVE_novel'

Yuqifan1117 commented 11 months ago

We use Scene-Graph-Benchmark.pytorch's train.sh extended with CaCao, which you can debug on. However, according to the next works, we will sort out this part of the code latter.

ZHUXUHAN commented 11 months ago

We use Scene-Graph-Benchmark.pytorch's train.sh extended with CaCao, which you can debug on. However, according to the next works, we will sort out this part of the code latter.

How do you actually test the "base novel" in OpenWord? I'm still not quite clear. I appreciate your assistance.

base is the 'cacao/VG-SGG-base-EXPANDED-with-attri.h5'? novel is 'open-world/VG-SGG-zs-random-EXPANDED-with-attri.h5'?

and vg+cacao is the ' 'open-world/VG-SGG-zs-random-EXPANDED-with-attri.h5', cacao is the 'cacao/VG-SGG-base-EXPANDED-with-attri.h5'?

I'm very confused about this and would appreciate your help. If possible, could you provide me with the code for evaluating your base novel?

Yuqifan1117 commented 11 months ago

We apologize for your confusion owing to file naming. We split the base and novel predicates randomly and the split result is shown as follows: af88a86781911cd4ff55fbca8189a762 We then use unified embeddings with cross-modal prompts mentioned in the paper to predict unseen predicates. Finally, we compute the performance on the base and novel predicates in turn in the same way as standard recall. By the way, vg+cacao is the ' 'open-world/VG-SGG-zs-random-EXPANDED-with-attri.h5', vg is the 'cacao/VG-SGG-base-zs-random-with-attri.h5', and cacao is the extended data between these two datasets.

ZHUXUHAN commented 11 months ago

yep, this answer helped me a lot, thanks.