search

google-research / scenic

Scenic: A Jax Library for Computer Vision Research and Beyond
Apache License 2.0
3.26k stars 428 forks source link

Are there checkpoint and config for pre-training vid2seq models based on YT-Temporal-1B？ #735

Open JaylanMei opened 1 year ago

JaylanMei commented 1 year ago

in scenic/projects/vid2seq/configs/youcook2.py

config.init_from.checkpoint_path = 'path_to_checkpoint_pretrained_on_yt_temporal_1bn' config.init_from.model_config = 'path_to_yt_temporal_1bn_config' config.init_from.step = 200000

Are there checkpoint and config for pre-training vid2seq models based on YT-Temporal-1B？

antoyang commented 1 year ago

The release of model checkpoints for Vid2Seq is planned as written in the readme. It is currently pending internal approval and will be done as quickly as possible.

sachit-menon commented 1 year ago

Hi, any rough timeline for when it might be released? Sorry to be a pain, I was just interested in using it for a project but am trying to gauge if the timings would work out.

antoyang commented 1 year ago

Hi, this is still pending internal approval. This is a process I am not used to so I can hardly predict any timeline. I really hope to get it done before the CVPR conference.

sachit-menon commented 1 year ago

Thanks for the quick update!

anilbatra2185 commented 1 year ago

hi @antoyang,

Thanks for sharing the code of a great work.

I am wondering if it is possible to share the final validation results of YouCook2 dataset (to compute caption and segmentation metrics) earlier until the checkpoints are available.

Thanks

antoyang commented 12 months ago

Hi, I do not have access to the prediction files as this was an internship work. However, there are now checkpoints available as well as an updated PyTorch implementation here: https://github.com/antoyang/VidChapters.