search

zjunlp / MKG_Analogy

[ICLR 2023] Multimodal Analogical Reasoning over Knowledge Graphs
https://zjunlp.github.io/project/MKG_Analogy/
MIT License
99 stars 11 forks source link

Lower results #15

Closed sana-mine closed 9 months ago

sana-mine commented 9 months ago

Nice work! But I reproduce the program's results based on pretrained checkpoints, which are lower than those mentioned in the paper. I would like to know which part of the reason is causing this? And when I saw the image encoding features, the code used a random sampling process. Will it affect the results? Thanks! image

flow3rdown commented 9 months ago

Thanks for your attention! The hyper-parameters and the image sampling strategy may effect the results, you can refer to issue #9 and issue #8.

zxlzr commented 9 months ago

Hi, do you have any further questions?