yoadtew / zero-shot-video-to-text Goto Github PK

I try to use GPT-NEO, rather than GPT-2, to do the video caption. But here is an error:

AssertionError: All appended context options must be of exactly length 3!

Seems the current code can not work with GPT-NEO.

Hey,

Have you tried integrating a stronger VLM such as BLIP instead of CLIP?

Hi,
Super cool paper!
I'd love to know when is the estimated release date of the code?

Thanks!

Hi, I found your work very intersting, and wanted to reproduce the results reported in your paper.

However, it seems that some of the default values in run.py are different from what is reported in your paper (e.g., running rate or sentence iteration).
Also, for some hyperparameters such as repetition penalty or entity penalty, I could not find the exact values you used in your paper.
Could you provide full list of hyperparameters required for reproducing the results in your paper?

Thanks.

It's a great job. But I have a question, the model generates a lot of capion for the video, how to choose a caption to calculate CLIP similarity mumble with the ground-truth tag of this video.

I am confused about the cache in paper, I may ask that if part of K in Prefix-tokens' cache come from Pseudo-tokens?

Hey folks, this looks very exciting, I wanted to test but it seems that this code doesn't run on M1?

Also it seems that maybe the requirements are outdated, they didn't work for me, it couldn't find the right torch librarty