33 ms·
ViT gives you 90% top-1 accuracy on ImageNet: https://paperswithcode.com/sota/image-classification-on-imagenet https://paperswithcode.com/sota/image-classificat
by thomasahle 4y ago
ViT gives you 90% top-1 accuracy on ImageNet: https://paperswithcode.com/sota/image-classification-on-imagenet https://paperswithcode.com/sota/image-classification-on-imag... I don't know where they get the 77.9% number from.
75% is pretty bad. Similar to the 2015 VGG net as the authors also admit.
- thomasahle 4y agoNevermind, I guess it's "ImageNet-1K trained models" on which ViT gets 79.9% and the 90% is only when pretraining with ImageNet-22K. There are other non-attention based networks that get 90% too though: https://arxiv.org/pdf/2212.11696v3.pdf https://arxiv.org/pdf/2212.11696v3.pdf