Dataset information about ptp

[FingerRec]

Hi jaeseokbyun:

Thanks for you attention in our work and carefully check!

Yes, the dataset corpus is from OSCAR and BLIP.
I update the dataset.md, reference for this file for more details.

Thanks for point out the problem in Table6, previous we only use coco image for one times during pertaining.
Later we follow OSCAR, ViLT and use each coco image for five times (each image 5 captions) during pre-training which outperform former significantly. We have alignment this table in camera ready version.

Thanks again.

[jaeseokbyun]

Thanks for sharing the code and pre-training corpus!
In DATASET.md, there are two corpuses

If I want to reproduce your result in the paper,
could I just use the 4M dataset corpus (not 2M dataset) without changing any other corresponding codes or dataset?
(Except path for dataset)

Thanks,