search

google-research / big_transfer

Official repository for the "Big Transfer (BiT): General Visual Representation Learning" paper.
https://arxiv.org/abs/1912.11370
Apache License 2.0
1.5k stars 175 forks source link

Plan to release weights pretrained on JFT-300M dataset? #4

Closed nlml closed 4 years ago

nlml commented 4 years ago

Thanks for releasing this - when I saw the paper I really hoped it would be open sourced.

It seems the biggest gains can come from using the models pretrained on your internal JFT-300M dataset. Are there plans to release weights from the models pretrained on this dataset?

Cheers!

lucasb-eyer commented 4 years ago

Unfortunately, this is an internal dataset and we are not allowed to release these weights. ImageNet-21k was in anticipation for this, and we are actively looking at public datasets larger than ImageNet-21k in order to be able to release even better pre-trained models.

Happy to hear recommendations for large, public datasets we could try!

nlml commented 4 years ago

Bummer!

No alternative datasets of that scale spring immediately to mind.

ghost commented 4 years ago

Unfortunately, this is an internal dataset and we are not allowed to release these weights. ImageNet-21k was in anticipation for this, and we are actively looking at public datasets larger than ImageNet-21k in order to be able to release even better pre-trained models.

Happy to hear recommendations for large, public datasets we could try!

Surely, there must be some subset for which you may be able to train and release without any licensing hassle. Or maybe the permissions your team needs is few mails down the line, if you and everyone here feels that releasing the model will be a great service to humanity.