Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

《Fast and Accurate Model Scaling》 reproducte for EfficientNet #171

Closed
dalin477 opened this issue Dec 6, 2021 · 4 comments
Closed

《Fast and Accurate Model Scaling》 reproducte for EfficientNet #171

dalin477 opened this issue Dec 6, 2021 · 4 comments

Comments

@dalin477
Copy link

dalin477 commented Dec 6, 2021

how can i get the same reproductive accuracy for efficientnet in 《Fast and Accurate Model Scaling》?
Sincere thanks

@dalin477
Copy link
Author

dalin477 commented Dec 6, 2021

I change the efficientnet configs file and add the following parameters, and i get the efficientnet-B1 err top1 is 23.246 ,but the paper reproduction about 21.7 err top1 for 100 epochs. how can i get the same acc for efficientnet in the code? thanks!
DROPOUT_RATIO: 0.2
DC_RATIO: 0.2
AUGMENT: AutoAugment

@pdollar
Copy link
Member

pdollar commented Dec 21, 2021

Details of the recipe are in the paper, they are described there. Note that we do not use dropout, but other things are used. Please see the paper for details. Eventually we'd like to release configs time permitting. Thanks!

@pdollar pdollar closed this as completed Dec 21, 2021
@dalin477
Copy link
Author

The stochastic weight averaging(SWA) optimization setup is not provided in the code? Could you'd like to offer the configs of the EfficientNet to me? My email is 1905663409@qq.com. Very thanks!

@dalin477
Copy link
Author

I want to reproductive the accuracy of the EfficientNet and RegNet in this paper. If you could provide the YAML configs of the related Net(EfficientNet & RegNet). Very thanks!

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
None yet
Projects
None yet
Development

No branches or pull requests

2 participants