Research map / Architectures and efficiency / Attention and transformers
FastViT: A Fast Hybrid Vision Transformer Using Structural Reparameterization
Pavan Kumar Anasosalu Vasu, James Gabriel, Jeff Zhu, Oncel Tuzel, Anurag Ranjan
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 changed | Where | Measured | Effect |
|---|---|---|---|
| replace pooling with RepMixer (ablation) | PoolFormer-S12 · ImageNet-1k | Top-1 accuracy / mobile latency: 77.6 -> 78.5; latency 2.25 -> 1.58 ms | Improved |
| factorize dense kxk convolutions (ablation) | FastViT-S12 · ImageNet-1k | Top-1 accuracy: 78.5 -> 78.0 | Hurt |
| add linear train-time overparameterization (ablation) | FastViT-S12 · ImageNet-1k | Top-1 accuracy: 78.0 -> 78.9 | Improved |
| add train-time overparameterization (ablation) | FastViT-SA12 · ImageNet-1k | Top-1 accuracy: 80.0 -> 80.6 | Improved |
| add train-time overparameterization (ablation) | FastViT-SA36 · ImageNet-1k | Top-1 accuracy: 83.3 -> 83.6 | Improved |
| also overparameterize ConvFFN layers (ablation) | FastViT-SA12 · ImageNet-1k | Top-1 accuracy / train time: 80.6 -> 80.6; train time 33.4 -> 40.1 hrs | No clear effect |
| add large kernel conv in FFN (ablation) | FastViT-S12 · ImageNet-1k | Top-1 accuracy: 78.9 -> 79.4 | Improved |
| add large kernel conv in patch embedding (ablation) | FastViT-S12 · ImageNet-1k | Top-1 accuracy: 79.4 -> 79.8 | Improved |
| large kernel convs instead of self-attention in last stage (ablation) | FastViT-S12 · ImageNet-1k | Top-1 accuracy / latency: V5 80.6 at 1.6 ms vs V3 81.0 at 3.7 ms | Mixed |
| increase kernel size beyond 7x7 (ablation) | FastViT-S12 · ImageNet-1k | Top-1 accuracy: 79.8 -> 79.6 (9x9), 79.8 (11x11) | No clear effect |
| replace BatchNorm with LayerNorm (ablation) | FastViT-S12 · ImageNet-1k | Top-1 accuracy / latency: 79.8 -> 79.7; latency 1.4 -> 1.7 ms | Hurt |
| replace GELU with ReLU (ablation) | FastViT-S12 · ImageNet-1k | Top-1 accuracy: 79.8 -> 79.4 | Hurt |
Closest papers by meaning
- EfficientFormer: Vision Transformers at MobileNet Speed (NeurIPS 2022 · similarity 0.90)
- RePaViT: Scalable Vision Transformer Acceleration via Structural Reparameterization on Feedforward Network Layers (ICML 2025 · similarity 0.90)
- MobileViT: Light-weight, General-purpose, and Mobile-friendly Vision Transformer (ICLR 2022 · similarity 0.90)
- RepViT: Revisiting Mobile CNN From ViT Perspective (CVPR 2024 · similarity 0.89)
- FasterViT: Fast Vision Transformers with Hierarchical Attention (ICLR 2024 · similarity 0.89)
- SHViT: Single-Head Vision Transformer with Memory Efficient Macro Design (CVPR 2024 · similarity 0.88)
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.
