Research map / Architectures and efficiency / CNN design
1xN Pattern for Pruning Convolutional Neural Networks
Mingbao Lin, Yuxin Zhang, Yuchao Li, Bohong Chen, Fei Chao, Mengdi Wang, Shen Li, Yonghong Tian, Rongrong Ji
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 |
|---|---|---|---|
| add filter rearrangement before 1xN pruning (N=16) (ablation) | 1xN pruning, MobileNet-V2 · ImageNet-1k val, p=50% | top-1 accuracy: 68.971 -> 69.352 | Improved |
| add filter rearrangement before 1xN pruning (N=2) (ablation) | 1xN pruning, MobileNet-V2 · ImageNet-1k val, p=50% | top-1 accuracy: 69.900 -> 70.233 | Improved |
| add filter rearrangement before 1xN pruning (N=4) (ablation) | 1xN pruning, MobileNet-V2 · ImageNet-1k val, p=50% | top-1 accuracy: 69.521 -> 69.579 | Improved |
| add filter rearrangement before 1xN pruning (N=32) (ablation) | 1xN pruning, MobileNet-V2 · ImageNet-1k val, p=50% | top-1 accuracy: 68.431 -> 68.762 | Improved |
| remove blocks of consecutive kernels by l1 norm instead of random kernels (ablation) | Kernel-wise pruning, ResNet-50 · ImageNet-1k val, p=50% | top-1 accuracy: 74.834 -> 75.370 | Improved |
| use 1xN block-level magnitude (N=4) instead of kernel-level l1 magnitude (ablation) | Kernel-wise pruning, ResNet-50 · ImageNet-1k val, p=50% | top-1 accuracy: 75.370 -> 76.506 | Improved |
| use 1xN block-level magnitude (N=4) instead of kernel-level l1 magnitude (ablation) | Kernel-wise pruning, MobileNet-V2 · ImageNet-1k val, p=50% | top-1 accuracy: 69.514 -> 69.706 | Improved |
| use 1x4 pattern instead of filter pruning | Pruned MobileNet-V2 · ImageNet-1k val, p=50% | top-1 accuracy: 66.730 -> 69.706 | Improved |
| use 1x4 pattern instead of weight pruning | Pruned MobileNet-V2 · ImageNet-1k val, p=50% | top-1 accuracy: 71.146 -> 69.706 | Hurt |
| increase block size N from 2 to 32 | 1xN pruning, MobileNet-V2 · ImageNet-1k val, p=50% | top-1 accuracy: 70.233 -> 68.762 | Hurt |
| increase block size N from 2 to 32 | 1xN pruning, ResNet-50 · ImageNet-1k val, p=50% | top-1 accuracy: 76.654 -> 75.960 | Hurt |
| use 1x4 pattern instead of weight pruning for latency | Pruned MobileNet-V2 on Cortex-A7 · Inference latency, p=50% | latency savings: 56.04ms inference savings | Improved |
Closest papers by meaning
- Pruning Filters for Efficient ConvNets (ICLR 2017 · similarity 0.92)
- Rethinking the Pruning Criteria for Convolutional Neural Network (NeurIPS 2021 · similarity 0.92)
- Structured Pruning for Deep Convolutional Neural Networks: A Survey (TPAMI 2023 · similarity 0.92)
- Convolutional Neural Network Pruning With Structural Redundancy Reduction (CVPR 2021 · similarity 0.92)
- Dynamic Structure Pruning for Compressing CNNs (AAAI 2023 · similarity 0.91)
- A Survey on Deep Neural Network Pruning: Taxonomy, Comparison, Analysis, and Recommendations (TPAMI 2024 · similarity 0.91)
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.
