Research map / Architectures and efficiency / Neural architecture search
Once-for-All: Train One Network and Specialize it for Efficient Deployment
Read the paperCheck 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 |
|---|---|---|---|
| Use progressive shrinking rather than training without it (ablation) | OFA · ImageNet, (D=4, W=3, K=3), 224×224 | top-1 accuracy: 71.5 -> 74.8 | Helped |
| Use progressive shrinking rather than training without it (ablation) | OFA · ImageNet, Pixel1 | top-1 accuracy: 72.4 -> 76.0 | Helped |
| Fine-tune specialized sub-networks for 25 epochs (ablation) | OFA w/ PS · ImageNet, Pixel1 | top-1 accuracy: 76.0 -> 76.4 | Helped |
| Fine-tune specialized sub-networks for 75 epochs instead of 25 (ablation) | OFA w/ PS · ImageNet, Pixel1 | top-1 accuracy: 76.4 -> 76.9 | Helped |
| Use OFA instead of EfficientNet-B0 at similar MACs | OFA · ImageNet, 389M MACs | top-1 accuracy: +2.8% | Helped |
| Use OFA instead of EfficientNet-B2 at 143ms Pixel1 latency | OFA · ImageNet, 143ms Pixel1 latency | top-1 accuracy: +0.3% | Helped |
| Use OFA instead of EfficientNet-B2 at 143ms Pixel1 latency | OFA · ImageNet, 143ms Pixel1 latency | latency: 2.6× faster | Helped |
| Use OFA instead of MobileNetV2 0.35 at similar latency | OFA #25 · ImageNet, NVIDIA 1080Ti GPU | top-1 accuracy: 60.3% -> 72.6% (+12.3%) | Helped |
| Train searched architectures from scratch rather than use OFA pretrained weights | OFA specialized architectures · ImageNet | top-1 accuracy | Hurt |
| Search hardware-specific architectures instead of relying on width-multiplier changes alone | OFA · ImageNet, diverse hardware platforms | accuracy-latency trade-off | Helped |
Closest papers by meaning
- Efficient On-Device Training via Gradient Filtering (CVPR 2023 · similarity 0.87)
- CompOFA – Compound Once-For-All Networks for Faster Multi-Platform Deployment (ICLR 2021 · similarity 0.86)
- TinyTrain: Resource-Aware Task-Adaptive Sparse Training of DNNs at the Data-Scarce Edge (ICML 2024 · similarity 0.85)
- Stepping Forward on the Last Mile (NeurIPS 2024 · similarity 0.85)
- Auto-Train-Once: Controller Network Guided Automatic Network Pruning from Scratch (CVPR 2024 · similarity 0.85)
- AtomNet: Designing Tiny Models from Operators Under Extreme MCU Constraints (AAAI 2025 · 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.