Research map / Detection and segmentation / CNN features
Rethinking Channel Dimensions for Efficient Model Design
Check your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 |
|---|---|---|---|
| Replace stage-wise channel dimensions with a linear channel configuration by block index | MobileNetV2-based lightweight network · ImageNet | Top-1 accuracy: 77.3% -> 77.9% | Helped |
| Use a linear channel configuration instead of EfficientNet-B1's channel design | ReXNet (×1.3) vs EfficientNet-B1 · ImageNet | Top-1 accuracy: 79.2% -> 79.5% | Helped |
| Use ReXNet as the SSDLite backbone rather than EfficientNet-B0 | SSDLite · COCO testdev2017 | Bbox AP: 23.5 -> 25.3 | Helped |
| Use ReXNet instead of ResNet50 as Faster R-CNN FPN backbone | Faster R-CNN with FPN · COCO val2017 | Bbox AP: 37.6 -> 41.5 | Helped |
| Use ReXNet instead of EfficientNet-B0 for fine-tuning | ImageNet-pretrained model · Food-101 | Top-1 accuracy: 87.5% -> 88.4% | Helped |
| Use ReXNet instead of ResNet50 as Mask R-CNN FPN backbone | Mask R-CNN with FPN · COCO val2017 | Mask AP: 34.6 -> 37.8 | Helped |
| Increase the first 1×1 convolution's dimension ratio from 1/20 to 1/6 (ablation) | Two-IB CIFAR-100 model · CIFAR-100 | Top-1 accuracy: 48.0% -> 52.1% | Helped |
| Increase every IB's dimension ratio from 0.22 to 0.8 (ablation) | Two-IB CIFAR-100 model · CIFAR-100 | Top-1 accuracy: 52.1% -> 53.8% | Helped |
| Replace the first ReLU6 in each IB with SiLU (ablation) | Two-IB CIFAR-100 model · CIFAR-100 | Top-1 accuracy: 53.8% -> 54.6% | Helped |
| Replace SiLU with ReLU6 after the first 1×1 convolution in each IB (ablation) | ReXNet (×1.0) · ImageNet | Top-1 accuracy: 77.9 -> 77.3 | Hurt |
| Replace SiLU with ELU after the first 1×1 convolution in each IB (ablation) | ReXNet (×1.0) · ImageNet | Top-1 accuracy: 77.9 -> 77.6 | Hurt |
| Use a linear channel configuration rather than conventional stage-wise channels | Searched lightweight networks · CIFAR-100 | Top-1 accuracy: 63.1±0.1 vs 62.2±0.0 | Helped |
Closest papers by meaning
- ChannelNets: Compact and Efficient Convolutional Neural Networks via Channel-Wise Convolutions (NeurIPS 2018 · similarity 0.87)
- CT-Net: Channel Tensorization Network for Video Classification (ICLR 2021 · similarity 0.86)
- Learning Features with Parameter-Free Layers (ICLR 2022 · similarity 0.86)
- Do Better ImageNet Models Transfer Better? (CVPR 2019 · similarity 0.85)
- The Heterogeneity Hypothesis: Finding Layer-Wise Differentiated Network Architectures (CVPR 2021 · similarity 0.85)
- Channel Vision Transformers: An Image Is Worth 1 x 16 x 16 Words (ICLR 2024 · 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.