AI Paper Atlas

Research map / Architectures and efficiency / Attention and transformers

LeViT: A Vision Transformer in ConvNet's Clothing for Faster Inference

Benjamin Graham, Alaaeldin El-Nouby, Hugo Touvron, Pierre Stock, Armand Joulin, Hervé Jégou, Matthijs Douze

ICCV 2021 · Architectures and efficiency › Attention and transformers · approach: Hybrid and interpretable ViT designs · 1,067 citations (OpenAlex)

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
add multi-resolution pyramid shape (vs straight stack) (ablation)LeViT-128S · ImageNet-1k val (LeViT-128S, 100 epochs)top-1 accuracy (%): 56.5 -> 71.9Improved
add PatchConv (four 3x3 convolutions in patch embedding) (ablation)LeViT-128S · ImageNet-1k val (LeViT-128S, 100 epochs)top-1 accuracy (%): 65.3 -> 71.9Improved
add BatchNorm (instead of LayerNorm) with zero-init residuals (ablation)LeViT-128S · ImageNet-1k val (LeViT-128S, 100 epochs)top-1 accuracy (%): 66.6 -> 71.9Improved
add hard distillation from RegNetY-16GF teacher (ablation)LeViT-128S · ImageNet-1k val (LeViT-128S, 100 epochs)top-1 accuracy (%): 69.7 -> 71.9Improved
add attention bias (replacing positional embedding) (ablation)LeViT-128S · ImageNet-1k val (LeViT-128S, 100 epochs)top-1 accuracy (%): 70.4 -> 71.9Improved
add wider blocks (smaller Q,K, V of 2D, MLP expansion 2) (ablation)LeViT-128S · ImageNet-1k val (LeViT-128S, 100 epochs)top-1 accuracy (%): 70.9 -> 71.9Improved
add Hardswish activation after attention product (ablation)LeViT-128S · ImageNet-1k val (LeViT-128S, 100 epochs)top-1 accuracy (%): 71.1 -> 71.9Improved
graft 2 ResNet-50 stages below 6 DeiT-S layersDeiT-S · ImageNet-1k val (300 epochs)top-1 accuracy (%): 79.9 -> 80.9Improved
graft 3 ResNet-50 stages below 3 DeiT-S layersDeiT-S · ImageNet-1k val (300 epochs)top-1 accuracy (%): 79.9 -> 80.1Mixed
use LeViT-256 instead of EfficientNet B3LeViT-256 · ImageNet-V2 matched frequencytop-1 accuracy (%): 70.6 -> 70.0Hurt
use LeViT-384 instead of DeiT-SmallLeViT-384 · ImageNet-Realtop-1 accuracy (%): 87.8 -> 87.6Hurt
use LeViT-384 instead of DeiT-SmallLeViT-384 · ImageNet-V2 matched frequencytop-1 accuracy (%): 71.7 -> 71.3Hurt

Later papers that compared with this method

5 later accepted papers report LeViT-256 as a baseline in their results tables; 4 of them report a better number on most of the results they share with it.

Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy