Image Text Retrieval Loss Discussion

ChenRocks / UNITER

Research code for ECCV 2020 paper "UNITER: UNiversal Image-TExt Representation Learning"

https://arxiv.org/abs/1909.11740

777 stars 109 forks source link

Image Text Retrieval Loss Discussion #68

Open jipson7 opened 3 years ago

jipson7 commented 3 years ago

Hi.

Why did you decide to use a BCE loss on the ITM pretraining text and a ranking loss on the ITM downstream task? Is there any intuition behind this? Why not use a ranking loss on both?

Thanks for the great work.

-Caleb