Efficient-AI-Backbones icon indicating copy to clipboard operation
Efficient-AI-Backbones copied to clipboard

Published 20 hours ago • verified publisher icon huawei-noah

higher performance of ViG

Open tdzdog opened this issue 2 years ago • 9 comments

I try to train ViG-S on ImageNet and get 80.54% top1 accuracy, which is higher than that in paper, 80.4%. I wonder if 80.4 is the average of multiple trainings? If yes, how many reps do you use?

tdzdog avatar Dec 03 '22 15:12 tdzdog

80.4 is a single training result. A slight fluctuation of accuracy is normal.

iamhankai avatar Dec 04 '22 05:12 iamhankai

@tdzdog on what resolution did u train the model ? is it by default 224x224

abhigoku10 avatar Dec 09 '22 05:12 abhigoku10

224x224

iamhankai avatar Dec 09 '22 07:12 iamhankai

@iamhankai can we train on higher resolutions ? Do we have segmentation based models

abhigoku10 avatar Dec 09 '22 07:12 abhigoku10

We have trained on COCO dataset whose resolution is much higher.

iamhankai avatar Dec 09 '22 07:12 iamhankai

@iamhankai can you share the weight files or point to where it is available ??

abhigoku10 avatar Dec 09 '22 08:12 abhigoku10

Like this: https://github.com/huawei-noah/Efficient-AI-Backbones/issues/114

iamhankai avatar Dec 09 '22 09:12 iamhankai

@iamhankai Thought the #params of pvig-b is more compared to pvig-m y is the metrics very nearer any explanation on this

abhigoku10 avatar Dec 09 '22 17:12 abhigoku10

Like this: #114

Thnaks for sharing the reference is it possible to share the already trained model from u on google drive or one drive

abhigoku10 avatar Dec 12 '22 04:12 abhigoku10