Research map / Architectures and efficiency / Pruning and efficient tuning
Flexible Group Count Enables Hassle-Free Structured Pruning
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes LeanFlex-GKP, a one-shot, post-train, data-agnostic grouped kernel pruning method that makes the convolution group count flexible and uses simple k-Means++ filter grouping and L2/geometric-median kernel pruning to produce dense pruned CNNs.
Problem
Structured pruning that keeps models densely structured while achieving fine granularity; existing grouped kernel pruning methods rely on expensive clustering, dynamic pruning rates, or custom architecture support, and are hard to use.
Approach
Instead of introducing dynamic clustering schemes or irregular group sizes, make Conv2d group count a flexible per-layer choice selected by a post-prune evaluation (inter-group vs intra-group geometric-median distance), keeping equal-sized groups and standard grouped convolution.
Key results
- On ResNet56 CIFAR100, LeanFlex-GKP reaches 72.11% accuracy (+0.58%) with 43.22% MACs drop, versus TMI-GKP at 71.11% (+0.26%).
- On VGG16 CIFAR10 with 43.15% MACs drop, LeanFlex-GKP achieves 94.15% (+0.21%), above TMI-GKP's 94.07%.
- Of 21 reported LeanFlex-GKP results, 18 showed accuracy improvements after pruning, per the authors.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use LeanFlex-GKP instead of CC | VGG16 · CIFAR10 | accuracy: 94.14 -> 94.15 | Helped |
| Use LeanFlex-GKP instead of CC | ResNet32 · CIFAR10 | accuracy: 92.39 -> 92.40 | Helped |
| Use LeanFlex-GKP instead of CC | ResNet110 · CIFAR10 | accuracy: 94.29 -> 94.35 | Helped |
| Use LeanFlex-GKP instead of NPPM | ResNet56 · CIFAR100 | accuracy: 71.57 -> 72.11 | Helped |
| Use LeanFlex-GKP instead of LRF | ResNet110 · CIFAR100 | accuracy: 73.58 -> 73.63 | Helped |
| Use LeanFlex-GKP instead of LRF | ResNet56 · Tiny-ImageNet | accuracy: 55.95 -> 55.67 | Hurt |
| Use LeanFlex-GKP instead of TMI-GKP | ResNet50 · ImageNet-1K | accuracy: 75.53 -> 75.62 | Helped |
Surprising findings
- Pruned models often exceed the unpruned baseline accuracy, which the authors attribute to a generalization benefit of pruning.
- Standard silhouette scores for clustering quality were found to have little bearing in the pruning context, motivating a geometric-median-based evaluation.
Limitations
- Limited to CNNs; not as general as architecture-agnostic methods like OTOv2 or DepGraph.
- Some comparisons are not perfectly aligned in baseline accuracy or pruning rates, as acknowledged by the authors.
- Accuracy gap to the best competing method is sometimes small or negative on specific settings.
Experimental setup
Datasets: CIFAR10 CIFAR100 Tiny-ImageNet ImageNet-1K
Benchmarks: CIFAR10 CIFAR100 Tiny-ImageNet ImageNet-1K
Metrics: top-1 accuracy accuracy change after pruning MACs reduction parameter reduction wall-clock pruning time
Baselines: TMI-GKP DSP KPGP CC HRank L1Norm FPGM NPPM SFP LRF OTOv2 ThiNet
Builds on: Grouped Kernel Pruning (TMI-GKP) FPGM Dynamic Structure Pruning (DSP) OTOv2
Closest papers by meaning
- GDeR: Safeguarding Efficiency, Balancing, and Robustness via Prototypical Graph Pruning (NeurIPS 2024 · similarity 0.86)
- Sparsest Models Elude Pruning: An Exposé of Pruning’s Current Capabilities (ICML 2024 · similarity 0.85)
- SequentialAttention++ for Block Sparsification: Differentiable Pruning Meets Combinatorial Optimization (NeurIPS 2024 · similarity 0.84)
- Exploring Learning Complexity for Efficient Downstream Dataset Pruning (ICLR 2025 · similarity 0.84)
- MDP: Multidimensional Vision Model Pruning with Latency Constraint (CVPR 2025 · similarity 0.83)
- Free Lunch in the Forest: Functionally-Identical Pruning of Boosted Tree Ensembles (AAAI 2025 · similarity 0.83)
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.