Research map / Architectures and efficiency / Attention and transformers
PaPr: Training-Free One-Step Patch Pruning with Lightweight ConvNets for Faster Inference
Tanvir Mahmud, Burhaneddin Yaman, Chun-Hao Liu, Diana Marculescu
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 PaPr's class weights in the FC layer with uniform weights to build the patch significance map | PaPr · ImageNet-1k | patch localization | Improved |
| use PaPr one-step pruning instead of ToMe at similar FLOPs | AugReg ViT-B-16 · ImageNet-1k val | top-1 accuracy: 80.38 -> 82.11 | Improved |
| add PaPr pruning before ToMe | ToMe on ViT-B-16 · ImageNet-1k val | top-1 accuracy / FLOPs: 82.34 acc at 8.21 GFLOPs vs ToMe 80.38 at 8.78 | Improved |
| use PaPr instead of ToMe on MAE-pretrained models | MAE ViT-B-16 · ImageNet-1k val | top-1 accuracy: 78.82 -> 82.4 | Improved |
| use PaPr on class-token-free ViTs instead of ToMe | ViT-Medium GAP-16 256 · ImageNet-1k val | top-1 accuracy: 80.21 -> 81.8 | Improved |
| use training-free PaPr instead of trained DynamicCNN/DynamicSwin | ConvNeXt Base-1k, Swin B-1k · ImageNet-1k val | top-1 accuracy: 83.08 -> 82.75 (ConvNeXt); 83.18 -> 81.7 (Swin) | Hurt |
| use ImageNet-22k weights with PaPr | ConvNeXt Base · ImageNet-1k val | top-1 accuracy: 82.75 -> 84.27 | Improved |
| swap MobileOne-S0 proposal for ResNet-152 (42x more FLOPs) (ablation) | PaPr on ViT-B · ImageNet-1k val | top-1 accuracy: 0.3% accuracy difference | No clear effect |
| use larger video proposal models (X3d-m, ResNet-101 vs smaller) (ablation) | PaPr on ViT-B MAE · Kinetics-400 val | top-1 accuracy: 81.11 -> 81.19 (X3d-s to X3d-m, z=0.5) | No clear effect |
| use PaPr (z=0.3) on ViT-L MAE video model | ViT-L MAE · Kinetics-400 val | top-1 accuracy: 85.26 -> 84.53 at 598 -> 160 GFLOPs | Mixed |
| replace gradient-based CAM localization with PaPr significance map (ablation) | patch pruning of ViT-B · qualitative samples | prediction confidence | Improved |
| use PaPr for training DeiT-S | DeiT-S · ImageNet-1k | train speed: 1x -> 1.6x | Improved |
Closest papers by meaning
- Synergistic Patch Pruning for Vision Transformer: Unifying Intra- & Inter-Layer Patch Importance (ICLR 2024 · similarity 0.87)
- FlexiViT: One Model for All Patch Sizes (CVPR 2023 · similarity 0.86)
- Memory-efficient Patch-based Inference for Tiny Deep Learning (NeurIPS 2021 · similarity 0.86)
- Patch Reordering: A NovelWay to Achieve Rotation and Translation Invariance in Convolutional Neural Networks (AAAI 2017 · similarity 0.86)
- Preserving Deep Representations in One-Shot Pruning: A Hessian-Free Second-Order Optimization Framework (ICLR 2025 · similarity 0.85)
- Patch Rebirth: Fast and Transferable Model Inversion of Vision Transformers (NeurIPS 2026 · similarity 0.85)
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.
