Research map / Architectures and efficiency / Neural architecture search
CompOFA – Compound Once-For-All Networks for Faster Multi-Platform 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 |
|---|---|---|---|
| Constrain the architecture family by coupling depth and width and fixing kernel size | OFA with MobileNetV3 · ImageNet training | GPU hours: 978 GPU hours -> 493 GPU hours | Helped |
| Constrain the architecture family by coupling depth and width while retaining elastic kernels | OFA with MobileNetV3 · ImageNet training | GPU hours: 978h 18m -> 677h 48m | Helped |
| Couple depth and width, fix kernel size, and memoize measured latencies | OFA evolutionary model search · ImageNet model search | average search time: 4.5 hours -> 75 seconds | Helped |
| Constrain the architecture family by coupling depth and width and fixing kernel size | OFA with MobileNetV3 · ImageNet training | training cost: $2.4k -> $1.2k | Helped |
| Constrain the architecture family by coupling depth and width and fixing kernel size | OFA with MobileNetV3 · ImageNet training | CO2 emission: 277 lbs -> 138 lbs | Helped |
| Couple depth and width rather than sampling them independently | OFA with MobileNetV3 · ImageNet validation, latency-constrained models | accuracy | No clear effect |
| Couple depth and width rather than sampling them independently | OFA with MobileNetV3 · ImageNet validation, sampled models per 5ms latency bucket | average accuracy | Helped |
| Remove progressive shrinking (ablation) | OFA · ImageNet validation, common configurations | top-1 accuracy: up to 3.7% | Hurt |
| Remove progressive shrinking (ablation) | CompOFA · ImageNet validation, common configurations | accuracy | No clear effect |
| Apply depth-width coupling and fixed kernels to ProxylessNAS | OFA with ProxylessNAS · ImageNet validation, common configurations | accuracy | No clear effect |
| Reduce evolutionary search iterations to 50 for CompOFA Fixed-Kernel (ablation) | CompOFA Fixed-Kernel · ImageNet model search | Pareto optimality: N=50 | No clear effect |
| Reduce evolutionary search iterations to 300 for CompOFA Elastic-Kernel (ablation) | CompOFA Elastic-Kernel · ImageNet model search | Pareto optimality: N=300 | No clear effect |
Closest papers by meaning
- Once-for-All: Train One Network and Specialize it for Efficient Deployment (ICLR 2020 · similarity 0.86)
- Model LEGO: Creating Models Like Disassembling and Assembling Building Blocks (NeurIPS 2024 · similarity 0.82)
- COGNATE: Acceleration of Sparse Tensor Programs on Emerging Hardware using Transfer Learning (ICML 2025 · similarity 0.82)
- ModHiFi: Identifying High Fidelity predictive components for Model Modification (NeurIPS 2025 · similarity 0.82)
- FlashMoE: Fast Distributed MoE in a Single Kernel (NeurIPS 2025 · similarity 0.81)
- Accessing Vision Foundation Models via ImageNet-1K (ICLR 2025 · similarity 0.81)
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.