Why it matters

ViT serving powers vision AI. Foundation of modern models.

Advertisement

The architecture

Patch tokenization.

Transformer forward.

Feature or class output.

Vision transformer servingPatches16x16 tokensTransformerself-attentionOutputfeatures or classDosovitskiy et al. 2020; ViT-B/16, ViT-L/14 common; huge scale ViT-H
Vision transformer serving.
Advertisement

How it works end to end

Dosovitskiy 2020.

ViT-B/L/H sizes.

Patch-based tokenization.