About input text token max_length

[lyakaap]

Great work!

I noticed max_length is set to 25 here, which is much smaller than 77 of CLIP.
How do you derive this setting?

[viyjy]

Hi, thanks for you interest in our paper.
This setting is determined by (1) the average sentence length of the pre-training dataset and (2) computational efficiency.
Feel free to let me know if you might have any other questions/suggestions. Thanks.

[lyakaap]

Thank you very much for your answer!