Research map / Architectures and efficiency / Pruning and efficient tuning
Soft Masking for Cost-Constrained Channel Pruning
Ryan Humble, Maying Shen, Jorge Albericio Latorre, Eric Darve, Jose Alvarez
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 batch normalization scaling (ablation) | SMCP ResNet50 (EagleEye baseline) · ImageNet val, high pruning | top-1 accuracy and FPS | Improved |
| add soft input channel masking instead of permanent pruning (ablation) | SMCP ResNet50 (EagleEye baseline) · ImageNet val, high pruning | top-1 accuracy and FPS | Improved |
| allow pruning entire layers (ablation) | SMCP ResNet50 (EagleEye baseline) · ImageNet val | top-1 accuracy vs FPS | Hurt |
| vary target schedule (Kw,Kt,Kc) (ablation) | SMCP ResNet50 70% reduction · ImageNet val | top-1 accuracy: 0.07% | No clear effect |
| change mask recompute interval r (ablation) | SMCP ResNet50 70% reduction · ImageNet val | top-1 accuracy and cost | Hurt |
| solve MCK with meet-in-the-middle instead of DP (s=1000) (ablation) | SMCP resource allocation · ResNet50 solve time | solve time: ~70 second -> <1 second | Improved |
| SMCP vs HALP at 30% latency | HALP · ImageNet val ResNet50 EagleEye | top-1 accuracy: 74.5 -> 75.1 | Improved |
| SMCP vs HALP at 80% latency (PyTorch baseline) | HALP · ImageNet val ResNet50 | top-1 accuracy: 77.2 -> 77.1 | No clear effect |
| SMCP vs HALP at 60% latency | HALP · ImageNet val ResNet101 | top-1 accuracy: 78.3 -> 78.1 | Hurt |
| SMCP vs HALP on MobileNet-V1 | HALP · ImageNet val MobileNet-V1 | top-1 accuracy: 68.3 -> 68.3 | No clear effect |
| SMCP vs HALP at 30% latency FPS | HALP · ImageNet val ResNet50 | FPS: 2755 -> 2947 | Improved |
| prune SSD512-ResNet50 with SMCP | SSD512-RN50 · PASCAL VOC 07+12 | mAP and FPS: 2.63 point higher mAP, 12% higher FPS | Improved |
Closest papers by meaning
- Rethinking the Smaller-Norm-Less-Informative Assumption in Channel Pruning of Convolution Layers (ICLR 2018 · similarity 0.92)
- Operation-Aware Soft Channel Pruning using Differentiable Masks (ICML 2020 · similarity 0.91)
- Structured Pruning of Neural Networks With Budget-Aware Regularization (CVPR 2019 · similarity 0.90)
- BilevelPruning: Unified Dynamic and Static Channel Pruning for Convolutional Neural Networks (CVPR 2024 · similarity 0.89)
- REPrune: Channel Pruning via Kernel Representative Selection (AAAI 2024 · similarity 0.88)
- UPDP: A Unified Progressive Depth Pruner for CNN and Vision Transformer (AAAI 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.
