Research map / Architectures and efficiency / Pruning and efficient tuning
SUBP: Soft Uniform Block Pruning for 1$\times$N Sparse CNNs Multithreading Acceleration
Jingyang Xiang, Siqi Li, Jun Chen, Guang Dai, Shipeng Bai, Yukai Ma, Yong Liu
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 |
|---|---|---|---|
| use BPAR instead of l1 norm as block pruning criterion (1x16) (ablation) | 1xN pruning of MobileNetV1 · ImageNet, p=50% | top-1 accuracy: 69.56 -> 70.30 | Improved |
| use BPAR instead of l1 norm (1x8) (ablation) | 1xN pruning of MobileNetV1 · ImageNet, p=50% | top-1 accuracy: 69.91 -> 70.55 | Improved |
| use BPAR instead of l1 norm (1x32) (ablation) | 1xN pruning of MobileNetV1 · ImageNet, p=50% | top-1 accuracy: 69.54 -> 70.03 | Improved |
| use BPAR instead of l1 norm (1x8) (ablation) | 1xN pruning of ResNet-50 · ImageNet, p=50% | top-1 accuracy: 76.15 -> 76.70 | Improved |
| use BPAR instead of l1 norm (1x16) (ablation) | 1xN pruning of ResNet-50 · ImageNet, p=50% | top-1 accuracy: 76.25 -> 76.52 | Improved |
| use BPAR instead of l1 norm (1x32) (ablation) | 1xN pruning of ResNet-50 · ImageNet, p=50% | top-1 accuracy: 75.96 -> 76.31 | Improved |
| increase block size N from 2 to 32 with l1 norm (ablation) | 1xN pruning of MobileNetV1 · ImageNet, p=50% | top-1 accuracy: 70.28 -> 69.54 | Hurt |
| increase block size N from 16 to 32 | SUBP on ResNet-50 (1.0G) · ImageNet val | top-1 accuracy: 76.3% -> 76.0% | Hurt |
| use 1x32 instead of 1x16 blocks | SUBP on MobileNetV1 (279M) · ImageNet val | top-1 accuracy: 70.8% -> 71.1% | Improved |
| filter pruning instead of weight pruning | pre-trained MobileNetV1 at p=50% · ImageNet | top-1 accuracy: 70.76 -> 65.35 | Hurt |
| initialize SUBP from pre-trained weights (90+120 epochs) instead of scratch (250 epochs) (ablation) | SUBP(1x16) ResNet-18 · ImageNet val | top-1 accuracy: 69.9% -> 69.5% | Hurt |
| sparse 2.0G ResNet-50 backbone vs dense backbone | Faster R-CNN · COCO | mAP: 37.4 -> 38.5 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- CHEX — CHEX: CHannel EXploration for CNN Model Compression (CVPR 2022) · better on 5 of 5 shared results
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
- Connectivity Matters: Neural Network Pruning Through the Lens of Effective Sparsity (JMLR 2023 · similarity 0.90)
- Accelerated Sparse Neural Training: A Provable and Efficient Method to Find N:M Transposable Masks (NeurIPS 2021 · similarity 0.90)
- Pruning via Sparsity-indexed ODE: a Continuous Sparsity Viewpoint (ICML 2023 · similarity 0.90)
- DPFPS: Dynamic and Progressive Filter Pruning for Compressing Convolutional Neural Networks from Scratch (AAAI 2021 · similarity 0.90)
- SequentialAttention++ for Block Sparsification: Differentiable Pruning Meets Combinatorial Optimization (NeurIPS 2024 · similarity 0.89)
- Sparsest Models Elude Pruning: An Exposé of Pruning’s Current Capabilities (ICML 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.
