Research map / Architectures and efficiency / CNN design
Latency-Aware Unified Dynamic Networks for Efficient Image Recognition
Yizeng Han, Zeyu Liu, Zhihang Yuan, Yifan Pu, Chaofei Wang, Shiji Song, Gao Ming Huang
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 |
|---|---|---|---|
| fuse masker with the first convolution (ablation) | LAUD^s-ResNet-50 block · Operator fusion, V100 | latency (us): 162.4 -> 135.1 | Improved |
| also fuse gather with convolution (ablation) | LAUD^s-ResNet-50 block · Operator fusion, V100 | latency (us): 135.1 -> 131.7 | Improved |
| also fuse scatter with add (ablation) | LAUD^s-ResNet-50 block · Operator fusion, V100 | latency (us): 131.7 -> 118.3 | Improved |
| add full operator fusion (ablation) | LAUD^s-ResNet-50 block · Operator fusion, TX2 | latency (us): 1084.5 -> 859.7 | Improved |
| use coarse spatial granularity (S>1) instead of pixel-level (S=1) (ablation) | spatially dynamic convolution · latency ratio on V100 and TX2 | latency ratio r_l: S=1 does not consistently give speedup; S>1 attains r_l<1 at larger activation rates | Improved |
| increase S from 7 to 14 in second stage (ablation) | LAUD-RegNetY-800MF · TX2 | latency: detrimentally impacts efficiency | Hurt |
| raise channel granularity G from 1 to 2 (ablation) | LAUD^c-ResNet-50 · ImageNet latency | latency: speedup for ResNet-50 | Improved |
| raise channel granularity G from 1 to 2 (ablation) | LAUD^c-RegNetY-800M · ImageNet latency | latency: comparable performance | No clear effect |
| use layer skipping instead of spatial-wise dynamic convolution | LAUD-ResNet-101 · V100 ImageNet latency | latency: ~53% latency reduction at t=0.4 | Improved |
| use layer skipping instead of channel skipping (ablation) | LAUD-T2T-ViT · TX2, RTX3060, V100 | accuracy-latency trade-off: consistently outperforms head skipping | Improved |
| combine all three paradigms (ablation) | LAUD-T2T-ViT · ImageNet | accuracy-efficiency trade-off: further enhances trade-off | Improved |
| use LAUD^s-ResNet-101 (t=0.5) backbone | Mask R-CNN · COCO instance segmentation | AP mask: 36.1 -> 37.0 | Improved |
Closest papers by meaning
- Latency-aware Spatial-wise Dynamic Networks (NeurIPS 2022 · similarity 0.91)
- DynamicDet: A Unified Dynamic Architecture for Object Detection (CVPR 2023 · similarity 0.89)
- Dynamic Neural Networks: A Survey (TPAMI 2021 · similarity 0.88)
- DS-Net++: Dynamic Weight Slicing for Efficient Inference in CNNs and Vision Transformers (TPAMI 2022 · similarity 0.88)
- Dynamic Network Quantization for Efficient Video Inference (ICCV 2021 · similarity 0.87)
- KernelWarehouse: Rethinking the Design of Dynamic Convolution (ICML 2024 · similarity 0.85)
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.
