On an ACCV-2020 paper's claims pertaining to this repo #454
Unanswered
vinayprabhu
asked this question in
General
Replies: 1 comment 1 reply
-
@vinayprabhu there are no models released to the public that were directly trained with JFT-300M, or even directly pretrained and then fine-tuned (to my knowledge), Google would not allow that. However, the Noisy-Student Efficientnets relased by Google did use JFT-300M as an unlabeled dataset as per the noisy student algorithm. See https://github.com/tensorflow/tpu/tree/master/models/official/efficientnet#2-using-pretrained-efficientnet-checkpoints and https://arxiv.org/abs/1911.04252 .. I converted the weights from the official repository above and they are in the |
Beta Was this translation helpful? Give feedback.
1 reply
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
-
Hi Ross,
Thank you for the wonderful service to the community.
I came across this paper titled "Compensating for the Lack of Extra Training Data by Learning Extra Representation" which is about " ... a novel framework, Extra Representation (ExRep), to surmount the problem of not having access to the JFT-300M data by instead using ImageNet and the publicly available model that has been pre-trained on JFT-300M".
In section-3, they make this specific claim: " Our choice of EfficientNet-B1 as the backbone is attributed to its relatively small number of parameters (7.8M), which enables efficient computation. The weights of the model pre-trained on JFT-300M
in PyTorch are available here [3]" : [3]: https://github.com/rwightman/pytorch-image-models
I wanted to check with you if this claim is correct. To the best of my knowledge, your collection does not have JFT-300M pre-trained models correct?
A reply would be utmost appreciated!
Kind regards,
Beta Was this translation helpful? Give feedback.
All reactions