Research map / Architectures and efficiency / Pruning and efficient tuning
UPSCALE: Unconstrained Channel Pruning
Alvin Wan, Hanxiang Hao, Kaushik Patnaik, Yueyang Xu, Omer Hadad, David Güera, Zhile Ren, Qi Shan
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 |
|---|---|---|---|
| remove channel-pruning constraints (unconstrained pruning) and export with UPSCALE | post-training pruned models, averaged across heuristics and sparsity levels · ImageNet | top-1 accuracy: +2.1 points on average | Improved |
| remove constraints (unconstrained pruning) | DenseNet121 post-training pruned models · ImageNet | top-1 accuracy: +16.9 points on average | Improved |
| remove constraints on EfficientNetV2-Small with LAMP at higher sparsity | EfficientNetV2-Small LAMP · ImageNet post-training | top-1 accuracy: 78.7% vs 11.4% at 10%; 3.8% vs 0.1% at 30% | Improved |
| remove constraints (unconstrained pruning) with ResNet18 smaller model | ResNet18 post-training pruning · ImageNet | top-1 accuracy: -0.8 points | Hurt |
| export unconstrained-pruned models with UPSCALE reordering instead of baseline memory-copy export (ablation) | exported model latency across heuristics and architectures · ImageNet models latency benchmark | latency reduction: 8.6% average; up to 52.8% at a specific sparsity (ResNet18, FPGM) | Improved |
| prune multi-branch segments in addition to single-branch segments (using UPSCALE export) (ablation) | MobileNetV3-Small with fine-tuning · ImageNet FLOPs-accuracy tradeoff | top-1 accuracy at matched compression: 59.32 vs 55.33 at 10% | Improved |
| export unconstrained-pruned models without UPSCALE (baseline memory copies) on DenseNet121 HRank (ablation) | DenseNet121 latency · ImageNet latency | latency (ms): 5.93 vs 5.44 ms at 1% sparsity for UPSCALE; baseline latency above unpruned ~1.1ms claim is for MobileNetV3-Small | Hurt |
Closest papers by meaning
- Influence Function Based Second-Order Channel Pruning: Evaluating True Loss Changes for Pruning is Possible Without Retraining (TPAMI 2024 · similarity 0.90)
- Rethinking the Smaller-Norm-Less-Informative Assumption in Channel Pruning of Convolution Layers (ICLR 2018 · similarity 0.90)
- Revisiting Random Channel Pruning for Neural Network Compression (CVPR 2022 · similarity 0.89)
- Network Pruning via Performance Maximization (CVPR 2021 · similarity 0.89)
- Targeted Low-rank Refinement: Enhancing Sparse Language Models with Precision (ICML 2025 · similarity 0.88)
- Sparsest Models Elude Pruning: An Exposé of Pruning’s Current Capabilities (ICML 2024 · similarity 0.87)
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.
