Research map / Architectures and efficiency / CNN design
GhostNets on Heterogeneous Devices via Cheap Operations
Kai Han, Yunhe Wang, Chang Xu, Jianyuan Guo, Chunjing Xu, Enhua Wu, Qi Chuan Tian
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 C-Ghost bottlenecks instead of inverted residual with SE (Net-A) (ablation) | C-GhostNet 0.5x · ImageNet val | top-1 accuracy (%): 64.2 -> 66.2 | Improved |
| use C-Ghost bottlenecks instead of lower resolution inverted residual (Net-B) (ablation) | C-GhostNet 0.5x · ImageNet val | top-1 accuracy (%): 65.7 -> 66.2 | Improved |
| use C-Ghost bottlenecks instead of deeper inverted residual network (Net-C) (ablation) | C-GhostNet 0.5x · ImageNet val | top-1 accuracy (%): 62.3 -> 66.2 | Improved |
| use C-Ghost bottlenecks instead of deeper inverted residual network (Net-D) (ablation) | C-GhostNet 0.5x · ImageNet val | top-1 accuracy (%): 61.0 -> 66.2 | Improved |
| use depthwise kernel d=3 instead of d=1 in C-Ghost module (ablation) | C-Ghost VGG16 · CIFAR-10 (VGG16) | accuracy (%): 93.5 -> 93.7 | Improved |
| increase depthwise kernel size from d=3 to d=7 (ablation) | C-Ghost VGG16 · CIFAR-10 (VGG16) | accuracy (%): 93.7 -> 93.1 | Hurt |
| increase ratio s from 2 to 5 (ablation) | C-Ghost VGG16 · CIFAR-10 (VGG16) | accuracy (%): 93.7 -> 92.9 | Hurt |
| add mix operation (intrinsic feature aggregation) to G-Ghost stage (ablation) | G-Ghost-ResNet56 · CIFAR-10 (ResNet56) | accuracy (%): 93.7 -> 93.9 | Improved |
| add mix operation to G-Ghost stage (ablation) | G-Ghost-RegNetX · ImageNet val (RegNetX-3.2GF) | top-1 accuracy (%): 76.8 -> 77.8 | Improved |
| add mix operation to G-Ghost stage (ablation) | G-Ghost-ResNet34 · ImageNet val (ResNet34) | top-1 accuracy (%): 72.8 -> 73.1 | Improved |
| use 1x1 conv cheap operation instead of identity mapping (ablation) | G-Ghost-ResNet56 · CIFAR-10 (ResNet56) | accuracy (%): 93.3 -> 93.7 | Improved |
| use 1x1 conv cheap operation instead of removing it (ablation) | G-Ghost-ResNet56 · CIFAR-10 (ResNet56) | accuracy (%): 93.2 -> 93.7 | Improved |
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.
- HRank — HRank: Filter Pruning Using High-Rank Feature Map (CVPR 2020) · better on 2 of 2 shared results
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
- GhostNet: More Features From Cheap Operations (CVPR 2020 · similarity 0.93)
- GhostNetV2: Enhance Cheap Operation with Long-Range Attention (NeurIPS 2022 · similarity 0.87)
- Faster CNNs with Direct Sparse Convolutions and Guided Pruning (ICLR 2017 · similarity 0.83)
- Multi-Granularity Pruning for Model Acceleration on Mobile Devices (ECCV 2022 · similarity 0.83)
- XiNet: Efficient Neural Networks for tinyML (ICCV 2023 · similarity 0.83)
- Once-for-All: Train One Network and Specialize it for Efficient Deployment (ICLR 2020 · similarity 0.82)
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.
