search

ChenRocks / UNITER

Research code for ECCV 2020 paper "UNITER: UNiversal Image-TExt Representation Learning"
https://arxiv.org/abs/1909.11740
784 stars 109 forks source link

Retrieval time #45

Open antoine77340 opened 4 years ago

antoine77340 commented 4 years ago

Thank you so much for providing the code for UNITER, great work! The README mentions the following "NOTE: Image-Text Retrieval is computationally heavy, especially on COCO."

Do you have a rough estimation in mind of how many hours (or days) does it take to run the full text-to-image retrieval on COCO on 1 GPU (eg. Tesla V100).

Thanks for your help!

RenShuhuai-Andy commented 3 years ago

The size of val and test set on COCO is very large, conduct one evaluation with 8 GPUs (V100) needs 2~3hours. If you only evaluate on val and test set once after training, it will take more than 8 hours with 8 GPUs.