Research map / Architectures and efficiency / Attention and transformers
DS-Net++: Dynamic Weight Slicing for Efficient Inference in CNNs and Vision Transformers
Changlin Li, Guangrun Wang, Bing Wang, Xiaodan Liang, Zhihui Li, Xiaojun Chang
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 generalized residual learning instead of naive weight sharing (ablation) | dynamic depth supernet · MobileNet V1, 2 paths | accuracy psi_2: 50.0 -> 53.2 | Improved |
| use generalized residual learning instead of naive weight sharing (ablation) | dynamic depth supernet · MobileNet V1, 2 paths | accuracy psi_1: 46.1 -> 48.3 | Improved |
| use weight slicing instead of masking (ablation) | dynamic ResNet-50 25% channels · latency RTX 2080 Ti | latency: 12.4ms -> 7.9 ms | Improved |
| use weight slicing instead of indexing (ablation) | dynamic ResNet-50 25% channels · latency RTX 2080 Ti | latency: 16.6 ms -> 7.9 ms | Improved |
| replace LV-ViT stem with depthwise separable stem (ablation) | DS-ViT++ patch embedding · ViT stem complexity | MAdds: 6.8G -> 5.6G | Improved |
| transfer the trained routing gate to CIFAR-10 (ablation) | DS-ResNet · CIFAR-10 | accuracy: +0.4% | Improved |
| use DS-MBNet-M++ instead of DS-MBNet-M (++ improvements) | DS-Net · ImageNet MobileNet 300M | top-1: 72.8 -> 73.2 | Improved |
| use DS-MBNet-M++ over static AutoSlim | DS-MBNet · ImageNet MobileNet 300M | top-1: 71.5 -> 73.2 | Improved |
| use DS-ViT-Ti++ over statically scaled DeiT-Ti | DS-ViT++ · ImageNet | top-1: 72.1 -> 78.2 | Improved |
Closest papers by meaning
- Dynamic Slimmable Network (CVPR 2021 · similarity 0.93)
- Latency-Aware Unified Dynamic Networks for Efficient Image Recognition (TPAMI 2024 · similarity 0.88)
- Dynamic Sparse Graph for Efficient Deep Learning (ICLR 2019 · similarity 0.87)
- KernelWarehouse: Rethinking the Design of Dynamic Convolution (ICML 2024 · similarity 0.87)
- Slicing Vision Transformer for Flexible Inference (NeurIPS 2024 · similarity 0.86)
- Advancing Dynamic Sparse Training by Exploring Optimization Opportunities (ICML 2024 · similarity 0.86)
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.
