Research map / Architectures and efficiency / Neural architecture search
NPAS: A Compiler-Aware Framework of Unified Network Pruning and Architecture Search for Beyond Real-Time Mobile Acceleration
Zhengang Li, Geng Yuan, Wei Niu, Pu Zhao, Yanyu Li, Yuxuan Cai, Xuan Shen, Zheng Zhan, Zhenglun Kong, Qing Jin, Zhiyu Chen, Sijia Liu and 4 more
Check 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 |
|---|---|---|---|
| Use NPAS with compiler-aware joint architecture search and pruning | ImageNet classification models · ImageNet | Top-1 accuracy: 78.2 vs. 75.2 | Improved |
| Use NPAS with compiler-aware joint architecture search and pruning | ImageNet classification models · ImageNet on Galaxy S10 GPU | latency: 6.7ms | Improved |
| Use the proposed compiler optimizations instead of MNN | MobileNet-V3 · MobileNet-V3 inference on mobile CPU and GPU | inference speedup: 46% and 141% | Improved |
| Use block-punched pruning rather than unstructured pruning (ablation) | ResNet-50 · ImageNet, uniform 6× pruning | latency | Improved |
| Use unstructured pruning rather than block-punched pruning (ablation) | ResNet-50 · ImageNet, uniform 6× pruning | accuracy | Improved |
| Replace the original ResNet-50 with a narrower model having twice as many layers at the same MACs (ablation) | ResNet-50 · Mobile GPU inference | latency: 1.22× slower (44ms vs. 36ms) | Hurt |
| Replace mobile-unfriendly activations with hard-sigmoid and hard-swish (ablation) | Pretrained DNN models · ImageNet | accuracy | No clear effect |
| Choose 3×3 convolution filters rather than 1×1 filters at equal computation (ablation) | Convolution layers · Mobile CPU, input feature map 56×56 | latency | Improved |
| Use pattern-based or block-punched fine-grained pruning rather than unstructured pruning (ablation) | 3×3 convolution layer · Mobile CPU, pruning rates below 5× | speedup | Improved |
| Use pattern-based or block-punched fine-grained pruning rather than coarse-grained structured pruning (ablation) | 3×3 convolution layer · Mobile CPU, pruning rates below 5× | speedup | Mixed |
Closest papers by meaning
- Multi-Granularity Pruning for Model Acceleration on Mobile Devices (ECCV 2022 · similarity 0.90)
- GRIM: A General, Real-Time Deep Learning Inference Framework for Mobile Devices Based on Fine-Grained Structured Weight Sparsity (TPAMI 2021 · similarity 0.88)
- PCONV: The Missing but Desirable Sparsity in DNN Weight Pruning for Real-Time Execution on Mobile Devices (AAAI 2020 · similarity 0.88)
- CPrune: Compiler-Informed Model Pruning for Efficient Target-Aware DNN Execution (ECCV 2022 · similarity 0.87)
- MobileNetV4: Universal Models for the Mobile Ecosystem (ECCV 2024 · similarity 0.87)
- DPP-Net: Device-aware Progressive Search for Pareto-optimal Neural Architectures (ECCV 2018 · 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.
