Research map / Architectures and efficiency / CNN design
Extreme Network Compression via Filter Group Approximation
Bo Peng, Wenming Tan, Zheyang Li, Shun Zhang, Di Xie, Shiliang Pu
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 |
|---|---|---|---|
| compress with filter group approximation instead of asymmetric low-rank decomposition | VGG16(S) · CIFAR100 (88% FLOPs reduction) | top-1 accuracy loss: 4.47% -> 1.93% | Improved |
| compress with filter group approximation instead of asymmetric low-rank decomposition | VGG16 · ImageNet (about 7x speed-up) | top-5 accuracy loss: 5.14% -> 2.03% | Improved |
| compress with filter group approximation instead of asymmetric low-rank decomposition | ResNet34 · ImageNet (84.8% FLOPs reduction) | top-5 accuracy loss: 8.27% -> 1.22% | Improved |
| compress with filter group approximation instead of Li et al. filter pruning | ResNet34 · ImageNet (84.84% FLOPs reduction) | top-1 accuracy loss: 13.27% -> 3.03% | Improved |
| add spatial decomposition (3D) to the Asym. baseline | Asym. · ImageNet ResNet34 | top-5 accuracy loss: 8.27% -> 2.78% | Improved |
| use Quarter compression schedule (n ratio 1:4 between stages) instead of Constant or Half (ablation) | ResNet34 compression · ImageNet | top-5 accuracy | Improved |
| compress the network with filter group approximation instead of training same architecture from scratch (ablation) | VGG16 compressed 57% · ImageNet | accuracy | Improved |
| initialize with filter group approximation instead of MSRA or orthogonal initialization (ablation) | Ours-Res34/A · ImageNet | accuracy loss | Improved |
| use filter group approximation instead of asymmetric decomposition for CPU latency | VGG16 (85.80% FLOPs reduction) · ImageNet CPU time | ms per image: 2078 -> 628 | Improved |
| run filter group convolution on Caffe GPU | VGG16 compressed · TITAN-XP GPU time | ms per image: 2.91 -> 2.17 | Hurt |
Closest papers by meaning
- Group Sparsity: The Hinge Between Filter Pruning and Decomposition for Network Compression (CVPR 2020 · similarity 0.89)
- Compressing Neural Networks: Towards Determining the Optimal Layer-wise Decomposition (NeurIPS 2021 · similarity 0.89)
- On Compressing Deep Models by Low Rank and Sparse Decomposition (CVPR 2017 · similarity 0.88)
- Towards Compact CNNs via Collaborative Compression (CVPR 2021 · similarity 0.87)
- Model Preserving Compression for Neural Networks (NeurIPS 2022 · similarity 0.87)
- Coreset-Based Neural Network Compression (ECCV 2018 · similarity 0.87)
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.
