# FBNetV3: Joint Architecture-Recipe Search using Predictor Pretraining

arXiv: https://arxiv.org/abs/2006.02049 · HTML: https://mlsanity.com/paper/2006.02049/

## Key facts

- Cited by 35 papers in the map (data through 14 Oct 2026).
- First posted on arXiv 3 Jun 2020.

## Built on by (most cited first, 30 of 35)

- [Early Convolutions Help Transformers See Better](https://mlsanity.com/paper/2106.14881/) (2021)
- [You Only Look at One Sequence: Rethinking Transformer in Vision through Object Detection](https://mlsanity.com/paper/2106.00666/) (2021)
- [Fast and Accurate Model Scaling](https://mlsanity.com/paper/2103.06877/) (2021)
- [ViTCoD: Vision Transformer Acceleration via Dedicated Algorithm and Accelerator Co-Design](https://mlsanity.com/paper/2210.09573/) (2022)
- [AlphaNet: Improved Training of Supernets with Alpha-Divergence](https://mlsanity.com/paper/2102.07954/) (2021)
- [Data Efficient Language-supervised Zero-shot Recognition with Optimal Transport Distillation](https://mlsanity.com/paper/2112.09445/) (2021)
- [ViTAS: Vision Transformer Architecture Search](https://mlsanity.com/paper/2106.13700/) (2021)
- [Stronger NAS with Weaker Predictors](https://mlsanity.com/paper/2102.10490/) (2021)
- [Castling-ViT: Compressing Self-Attention via Switching Towards Linear-Angular Attention at Vision Transformer Inference](https://mlsanity.com/paper/2211.10526/) (2022)
- [UniNet: Unified Architecture Search with Convolution, Transformer, and MLP](https://mlsanity.com/paper/2110.04035/) (2021)
- [AutoLoss-Zero: Searching Loss Functions from Scratch for Generic Tasks](https://mlsanity.com/paper/2103.14026/) (2021)
- [Searching by Generating: Flexible and Efficient One-Shot NAS with Architecture Generator](https://mlsanity.com/paper/2103.07289/) (2021)
- [FBNetV5: Neural Architecture Search for Multiple Tasks in One Run](https://mlsanity.com/paper/2111.10007/) (2021)
- [SplitNets: Designing Neural Architectures for Efficient Distributed Computing on Head-Mounted Systems](https://mlsanity.com/paper/2204.04705/) (2022)
- [Automated Deep Learning: Neural Architecture Search Is Not the End](https://mlsanity.com/paper/2112.09245/) (2021)
- [Generative Adversarial Neural Architecture Search](https://mlsanity.com/paper/2105.09356/) (2021)
- [BasisNet: Two-stage Model Synthesis for Efficient Inference](https://mlsanity.com/paper/2105.03014/) (2021)
- [FOX-NAS: Fast, On-device and Explainable Neural Architecture Search](https://mlsanity.com/paper/2108.08189/) (2021)
- [DHA: End-to-End Joint Optimization of Data Augmentation Policy, Hyper-parameter and Architecture](https://mlsanity.com/paper/2109.05765/) (2021)
- [Speeding up NAS with Adaptive Subset Selection](https://mlsanity.com/paper/2211.01454/) (2022)
- [AutoSpace: Neural Architecture Search with Less Human Interference](https://mlsanity.com/paper/2103.11833/) (2021)
- [Collaboration of Experts: Achieving 80% Top-1 Accuracy on ImageNet with 100M FLOPs](https://mlsanity.com/paper/2107.03815/) (2021)
- [DetOFA: Efficient Training of Once-for-All Networks for Object Detection Using Path Filter](https://mlsanity.com/paper/2303.13121/) (2023)
- [Profiling Neural Blocks and Design Spaces for Mobile Neural Architecture Search](https://mlsanity.com/paper/2109.12426/) (2021)
- [DAAS: Differentiable Architecture and Augmentation Policy Search](https://mlsanity.com/paper/2109.15273/) (2021)
- [Privacy-preserving Online AutoML for Domain-Specific Face Detection](https://mlsanity.com/paper/2203.08399/) (2022)
- [DISGO: Automatic End-to-End Evaluation for Scene Text OCR](https://mlsanity.com/paper/2308.13173/) (2023)
- [Pruning Compact ConvNets for Efficient Inference](https://mlsanity.com/paper/2301.04502/) (2023)
- [ANNETTE: Accurate Neural Network Execution Time Estimation with Stacked Models](https://mlsanity.com/paper/2105.03176/) (2021)
- [Meta-Learning of NAS for Few-shot Learning in Medical Image Applications](https://mlsanity.com/paper/2203.08951/) (2022)

## Abstract (arXiv)

Neural Architecture Search (NAS) yields state-of-the-art neural networks that outperform their best manually-designed counterparts. However, previous NAS methods search for architectures under one set of training hyper-parameters (i.e., a training recipe), overlooking superior architecture-recipe combinations. To address this, we present Neural Architecture-Recipe Search (NARS) to search both (a) architectures and (b) their corresponding training recipes, simultaneously. NARS utilizes an accuracy predictor that scores architecture and training recipes jointly, guiding both sample selection and ranking. Furthermore, to compensate for the enlarged search space, we leverage "free" architecture statistics (e.g., FLOP count) to pretrain the predictor, significantly improving its sample efficiency and prediction reliability. After training the predictor via constrained iterative optimization, we run fast evolutionary searches in just CPU minutes to generate architecture-recipe pairs for a variety of resource constraints, called FBNetV3. FBNetV3 makes up a family of state-of-the-art compact neural networks that outperform both automatically and manually-designed competitors. For example, FBNetV3 matches both EfficientNet and ResNeSt accuracy on ImageNet with up to 2.0x and 7.1x fewer FLOPs, respectively. Furthermore, FBNetV3 yields significant performance gains for downstream object detection tasks, improving mAP despite 18% fewer FLOPs and 34% fewer parameters than EfficientNet-based equivalents.