Very low accuracy on ImageNet-LT

val-iisc / DeiT-LT

[CVPR 2024] Code for our Paper "DeiT-LT: Distillation Strikes Back for Vision Transformer training on Long-Tailed Datasets"

https://rangwani-harsh.github.io/DeiT-LT/

MIT License

33 stars 3 forks source link

Open AlvinHan123 opened 1 month ago

AlvinHan123 commented 1 month ago

Hi,

Thanks for your great job! Here is a problem a problem when I attempt reproduce your work.

I got very low accuracy on Imagenet-LT dataset. Could you please have a look on my problem?

The bash command and result have been attached. 460255922_887940873261359_1623481579448230716_n 461609504_891035332453376_221387602586693602_n

Cheers, Alvin.

fantastice commented 3 weeks ago

Could you please tell me if the model you used on the ImageNet dataset is the resnet.py file located in the teacher_models folder?Thank you!

pradipto111 commented 3 weeks ago

Hi @AlvinHan123

Thanks for your interest in our work.

Please find the recent commit for code and shell script changes. This should help you resolve the issue.

Regards, Pradipto

pradipto111 commented 3 weeks ago

Hi @fantastice

You can find the resnet file used for ImageNet-LT and iNaturalist-2018 here - teacher_models/resnet_imagenet_paco.py, added in the latest commit.

Regards, Pradipto