Research map / Architectures and efficiency / Neural architecture search
ShuffleNet V2: Practical Guidelines for Efficient CNN Architecture Design
Ningning Ma, Xiangyu Zhang, Hai-Tao Zheng, Jian Sun
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 |
|---|---|---|---|
| use equal input/output channel width (1:1) instead of 1:12 (ablation) | pointwise-conv benchmark network · Guideline 1 benchmark (GPU, x1) | GPU speed (batches/sec): 748 -> 1480 | Improved |
| use 1 group instead of 8 groups in pointwise group convolution (ablation) | group-conv benchmark network · Guideline 2 benchmark (GPU, x1) | GPU speed (batches/sec): 634 -> 2451 | Improved |
| use 1-fragment instead of 4-fragment-series block (ablation) | fragmentation benchmark · Guideline 3 benchmark (GPU, c=128) | GPU speed (batches/sec): 752 -> 2446 | Improved |
| remove ReLU from bottleneck unit (ablation) | bottleneck unit benchmark · Guideline 4 benchmark (GPU, c=32) | GPU speed (batches/sec): 2427 -> 2672 | Improved |
| remove shortcut from bottleneck unit (ablation) | bottleneck unit benchmark · Guideline 4 benchmark (GPU, c=32) | GPU speed (batches/sec): 2427 -> 2647 | Improved |
| remove both ReLU and shortcut from bottleneck unit (ablation) | bottleneck unit benchmark · Guideline 4 benchmark (GPU, c=32) | GPU speed (batches/sec): 2427 -> 2842 | Improved |
| add squeeze-and-excitation module | ShuffleNet v2 2x · ImageNet val (~500 MFLOPs) | top-1 error (%): 25.1 -> 24.6 | Mixed |
| use ShuffleNet v2 instead of ShuffleNet v1 at 50 layers | 50-layer network · ImageNet val (large models) | top-1 error (%): 25.2 -> 22.8 | Improved |
| add extra 3x3 depthwise conv to enlarge receptive field (ShuffleNet v2*) | ShuffleNet v2 · COCO minival detection (40M FLOPs backbone) | mmAP (%): 22.5 -> 23.7 | Improved |
| add extra 3x3 depthwise conv to enlarge receptive field (ShuffleNet v2*) | ShuffleNet v2 · COCO minival detection (500M FLOPs backbone) | mmAP (%): 33.3 -> 34.2 | Improved |
| use ShuffleNet v2 backbone instead of ShuffleNet v1 | Light-Head RCNN · COCO minival detection (500M FLOPs backbone) | mmAP (%): 32.9 -> 33.3 | Improved |
| use ShuffleNet v2 instead of MobileNet v2 at 40M FLOPs | small-model comparison · ImageNet val (~40 MFLOPs) | top-1 error (%): 43.4 -> 39.7 | Improved |
Later papers that compared with this method
5 later accepted papers report ShuffleNetV2 as a baseline in their results tables; 3 of them report a better number on most of the results they share with it.
- Stripe Observation Guided Inference Cost-free Attention Mechanism (ECCV 2024 · better on 0 of 1 shared result)
- Channel Regeneration: Improving Channel Utilization for Compact DNNs (AAAI 2023 · better on 3 of 3 shared results)
- Dynamic Convolutions: Exploiting Spatial Sparsity for Faster Inference (CVPR 2020 · better on 2 of 2 shared results)
- Fast and Practical Neural Architecture Search (ICCV 2019 · better on 2 of 2 shared results)
- Exploring Randomly Wired Neural Networks for Image Recognition (ICCV 2019 · better on 0 of 1 shared result)
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- Xception — Xception: Deep Learning With Depthwise Separable Convolutions (CVPR 2017) · better on 1 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- ShuffleNet: An Extremely Efficient Convolutional Neural Network for Mobile Devices (CVPR 2018 · similarity 0.85)
- Butterfly Transform: An Efficient FFT Based Neural Architecture Design (CVPR 2020 · similarity 0.84)
- Run, Don't Walk: Chasing Higher FLOPS for Faster Neural Networks (CVPR 2023 · similarity 0.84)
- Neural Routing by Memory (NeurIPS 2021 · similarity 0.84)
- Efficient Architecture Search by Network Transformation (AAAI 2018 · similarity 0.84)
- LotteryCodec: Searching the Implicit Representation in a Random Network for Low-Complexity Image Compression (ICML 2025 · similarity 0.83)
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.
