Research map / Architectures and efficiency / Attention and transformers
EfficientViT: Memory Efficient Vision Transformer With Cascaded Group Attention
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 |
|---|---|---|---|
| Use the sandwich layout instead of a Swin block (ablation) | EfficientViT-M4 · ImageNet-1K | Top-1 accuracy: 68.3 -> 71.3 | Helped |
| Use N=1 rather than N=2 FFNs around attention (ablation) | EfficientViT-M4 · ImageNet-1K | Top-1 accuracy: 70.2 -> 71.3 | Helped |
| Use N=1 rather than N=3 FFNs around attention (ablation) | EfficientViT-M4 · ImageNet-1K | Top-1 accuracy: 65.7 -> 71.3 | Helped |
| Use cascaded group attention instead of MHSA (ablation) | EfficientViT-M4 · ImageNet-1K | Top-1 accuracy: 70.2 -> 71.3 | Helped |
| Use cascaded group attention instead of MHSA (ablation) | EfficientViT-M4 · ImageNet-1K | ONNX throughput: 102.2 -> 108.6 | Helped |
| Cascade attention-head outputs (ablation) | EfficientViT-M4 · ImageNet-1K | Top-1 accuracy: 69.8 -> 71.3 | Helped |
| Reallocate channel dimensions across Q, K, and V (ablation) | EfficientViT-M4 · ImageNet-1K | Top-1 accuracy: 69.9 -> 71.3 | Helped |
| Reduce the FFN expansion ratio from 4 to 2 (ablation) | EfficientViT-M4 · ImageNet-1K | Top-1 accuracy: 69.8 -> 71.3 | Helped |
| Add DWConv token interaction (ablation) | EfficientViT-M4 · ImageNet-1K | Top-1 accuracy: 69.9 -> 71.3 | Helped |
| Use BatchNorm rather than LayerNorm (ablation) | EfficientViT-M4 · ImageNet-1K | Top-1 accuracy: 70.4 -> 71.3 | Helped |
| Use ReLU rather than HardSwish (ablation) | EfficientViT-M4 · ImageNet-1K | ONNX throughput: 87.5 -> 108.6 | Helped |
| Finetune the model at higher image resolution | EfficientViT-M5 · ImageNet-1K | Top-1 accuracy: 77.1 -> 79.8; 77.1 -> 80.8 | Helped |
Closest papers by meaning
- SHViT: Single-Head Vision Transformer with Memory Efficient Macro Design (CVPR 2024 · similarity 0.91)
- FasterViT: Fast Vision Transformers with Hierarchical Attention (ICLR 2024 · similarity 0.90)
- You Only Need Less Attention at Each Stage in Vision Transformers (CVPR 2024 · similarity 0.89)
- Efficient Modulation for Vision Networks (ICLR 2024 · similarity 0.88)
- Real-time Core-Periphery Guided ViT with Smart Data Layout Selection on Mobile Devices (NeurIPS 2024 · similarity 0.88)
- Skip-Attention: Improving Vision Transformers by Paying Less Attention (ICLR 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.