Research map / Architectures and efficiency / Pruning and efficient tuning
Channel Pruning for Accelerating Very Deep Neural Networks
Yihui He, Xiangyu Zhang, Jian Sun
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 fine-tuning after channel pruning (ablation) | channel pruning · ImageNet VGG-16 4x | increase of top-5 error (%): 7.9 -> 1.0 | Improved |
| add fine-tuning after channel pruning (ablation) | channel pruning · ImageNet VGG-16 5x | increase of top-5 error (%): 22.0 -> 1.7 | Improved |
| add fine-tuning after channel pruning (ablation) | channel pruning · ImageNet VGG-16 2x | increase of top-5 error (%): 2.7 -> 0 | Improved |
| add multi-branch enhancement and fine-tuning (ablation) | channel pruning · ImageNet ResNet-50 2x | increase of top-5 error (%): 8.0 -> 4.0 -> 1.4 | Improved |
| prune with reconstruction-based channel selection instead of first k or max response (ablation) | channel pruning · VGG-16 single layer | increase of top-5 error | Improved |
| select channels by max response (filter weight magnitude) instead of first k (ablation) | channel pruning · VGG-16 single layer | increase of top-5 error | No clear effect |
| combine channel pruning with spatial and channel factorization (3C) | VGG-16 · ImageNet 4x | increase of top-5 error (%): 1.0 -> 0.0 | Improved |
| prune channels instead of train same architecture from scratch | VGG-16 4x · ImageNet | increase of top-5 error (%): 1.8 -> 1.0 | Improved |
| prune channels instead of uniformly halving filters and training from scratch | VGG-16 4x · ImageNet | increase of top-5 error (%): 2.4 -> 1.0 | Improved |
| channel pruning vs Filter pruning (Li et al.) | VGG-16 · ImageNet 5x | increase of top-5 error (%): 14.6 -> 1.7 | Improved |
| channel pruning on Faster R-CNN backbone 2x | VGG-16 · VOC07 | mAP: 68.7 -> 68.3 | No clear effect |
| channel pruning on Faster R-CNN backbone 4x | VGG-16 · VOC07 | mAP: 68.7 -> 66.9 | Hurt |
Closest papers by meaning
- Operation-Aware Soft Channel Pruning using Differentiable Masks (ICML 2020 · similarity 0.88)
- Collaborative Channel Pruning for Deep Networks (ICML 2019 · similarity 0.88)
- Rethinking the Smaller-Norm-Less-Informative Assumption in Channel Pruning of Convolution Layers (ICLR 2018 · similarity 0.88)
- Storage Efficient and Dynamic Flexible Runtime Channel Pruning via Deep Reinforcement Learning (NeurIPS 2020 · similarity 0.88)
- REPrune: Channel Pruning via Kernel Representative Selection (AAAI 2024 · similarity 0.87)
- BilevelPruning: Unified Dynamic and Static Channel Pruning for Convolutional Neural Networks (CVPR 2024 · 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.
