Research map / Architectures and efficiency / Attention and transformers
Isomorphic Pruning for Vision Models
Gongfan Fang, Xinyin Ma, Michael Bi Mi, Xinchao Wang
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 |
|---|---|---|---|
| prune DeiT-S-size model from DeiT-B instead of training DeiT-T from scratch (isomorphic pruning) | DeiT-T · ImageNet-1K val | top-1 accuracy (%): 74.52 -> 77.50 | Improved |
| isomorphic pruning ConvNext-B to ConvNext-T size | ConvNext-T · ImageNet-1K val | top-1 accuracy (%): 82.06 -> 82.19 | Improved |
| group sub-structures by isomorphism and rank within groups instead of global ranking (ablation) | Taylor, Hessian, L1 criteria on ResNet-50 · ImageNet-1K val without finetuning | accuracy and loss | Improved |
| use isomorphic pruning on ResNet-101 for 4.5G and 3.8G budgets | ResNet-101 · ImageNet-1K val | top-1 accuracy drop: +0.16 and +0.05 | Improved |
| isomorphic pruning on lightweight MobileNet-v2 to 0.15G MACs | MobileNet-v2 · ImageNet-1K val | top-1 accuracy drop: 71.88 -> 68.91 | Mixed |
| global ranking without grouping on DeiT-B at 50% pruning (ablation) | naive global pruning · DeiT-B importance distribution | pruning balance | Hurt |
| non-uniform pruned DeiT-S instead of uniform DeiT-S | DeiT-S · peak memory | peak memory (MB): 1363 -> 1547 | Hurt |
| halve MACs of ResNet-50 by pruning | ResNet-50 · GPU latency | speedup: 1.33x speedup | Mixed |
| transfer pruned DeiT-S to CIFAR-100 | DeiT-S · CIFAR-100 | accuracy (%): 87.07 -> 90.09 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 2 of 2 on most shared results.
- GOHSP — GOHSP: A Unified Framework of Graph and Optimization-Based Heterogeneous Structured Pruning for Vision Transformer (AAAI 2023) · better on 1 of 1 shared result
- X-Pruner — X-Pruner: eXplainable Pruning for Vision Transformers (CVPR 2023) · better on 1 of 1 shared result
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
- Structured Pruning for Deep Convolutional Neural Networks: A Survey (TPAMI 2023 · similarity 0.91)
- MDP: Multidimensional Vision Model Pruning with Latency Constraint (CVPR 2025 · similarity 0.90)
- UPDP: A Unified Progressive Depth Pruner for CNN and Vision Transformer (AAAI 2024 · similarity 0.90)
- GOHSP: A Unified Framework of Graph and Optimization-Based Heterogeneous Structured Pruning for Vision Transformer (AAAI 2023 · similarity 0.90)
- Balanced Token Pruning: Accelerating Vision Language Models Beyond Local Optimization (NeurIPS 2025 · similarity 0.89)
- OSSCAR: One-Shot Structured Pruning in Vision and Language Models with Combinatorial Optimization (ICML 2024 · similarity 0.89)
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.
