-
Notifications
You must be signed in to change notification settings - Fork 94
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
Token Labeling #8
Comments
Hi, token labeling can improve volo_d1 by 0.4% (from 83.8% to 84.2%). For larger volo models, token labeling can achieve improvement from 0.0 to 0.4%. |
Thank you, I think these numbers are probably worth adding to the paper as well for context. |
Could you please also put a copy of token labeling data and pretrained models on Baidu Netdisk? |
Hi,thank you for sharing this work. If I want to use VOLO to finetune on my own dataset, how can I create my own token labeling? |
Hi @Ree1s , You can refer to zihangJiang/TokenLabeling#7 (comment) for generating token label data for your own dataset. |
If you only want to 'finetune' VOLO on your own dataset, one option is to directly use the 'class token' or mean the feature tokens without using token labeling, thus you don't need to generate your own token labels. |
Thanks, and I will try on my dataset through this way mentioned. |
I'll have a try. Thanks a lot! |
Hi, when you say class token, do you mean configure |
Hi, thank you for the paper.
Do you have any numbers for how the networks perform without token labeling, only using MixToken and other augmentations?
The text was updated successfully, but these errors were encountered: