Research map / Architectures and efficiency / Pruning and efficient tuning
CHEX: CHannel EXploration for CNN Model Compression
Check your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 periodic pruning and regrowing instead of one-shot early-stage pruning (ablation) | CHEX on ResNet-50 · ImageNet | top-1 accuracy: 73.9% -> 74.8% | Helped |
| add dynamic sub-model structure exploration (ablation) | CHEX on ResNet-50 · ImageNet | top-1 accuracy: 74.8% -> 75.6% | Helped |
| add importance sampling-based regrowing instead of uniform sampling (ablation) | CHEX on ResNet-50 · ImageNet | top-1 accuracy: 75.6% -> 76.0% | Helped |
| use CSS rather than BN pruning criterion (ablation) | CHEX on ResNet-50 · ImageNet | top-1 accuracy: 75.3% -> 76.0% | Helped |
| use CSS rather than magnitude pruning criterion (ablation) | CHEX on ResNet-50 · ImageNet | top-1 accuracy: 75.7% -> 76.0% | Helped |
| restore MRU weights rather than initialize regrown channels to zero (ablation) | CHEX on ResNet-50 · ImageNet | top-1 accuracy: 74.1% -> 76.0% | Helped |
| restore MRU weights rather than initialize regrown channels randomly (ablation) | CHEX on ResNet-50 · ImageNet | top-1 accuracy: 74.5% -> 76.0% | Helped |
| restore MRU weights rather than use EMA weights for regrown channels (ablation) | CHEX on ResNet-50 · ImageNet | top-1 accuracy: 75.5% -> 76.0% | Helped |
| increase the initial regrowing factor from δ0=0.1 to δ0=0.3 (ablation) | CHEX on ResNet-50 · ImageNet | top-1 accuracy: 75.2% -> 76.0% | Helped |
| increase the initial regrowing factor from δ0=0.3 to δ0=0.4 (ablation) | CHEX on ResNet-50 · ImageNet | top-1 accuracy: 76.0% -> 75.9% | Hurt |
| use cosine decay rather than constant regrowing (ablation) | CHEX on ResNet-50 · ImageNet | top-1 accuracy: 75.3% -> 76.0 | Helped |
| use cosine decay rather than linear decay for regrowing (ablation) | CHEX on ResNet-50 · ImageNet | top-1 accuracy: 75.6% -> 76.0 | Helped |
Closest papers by meaning
- Rethinking the Smaller-Norm-Less-Informative Assumption in Channel Pruning of Convolution Layers (ICLR 2018 · similarity 0.91)
- Discrimination-aware Channel Pruning for Deep Neural Networks (NeurIPS 2018 · similarity 0.90)
- REPrune: Channel Pruning via Kernel Representative Selection (AAAI 2024 · similarity 0.89)
- Revisiting Random Channel Pruning for Neural Network Compression (CVPR 2022 · similarity 0.89)
- BilevelPruning: Unified Dynamic and Static Channel Pruning for Convolutional Neural Networks (CVPR 2024 · similarity 0.89)
- Reborn Filters: Pruning Convolutional Neural Networks with Limited Data (AAAI 2020 · 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.