Research map / Image generation / Super-resolution
Learning Filter Basis for Convolutional Neural Network Compression
Yawei Li, Shuhang Gu, Luc Van Gool, Radu Timofte
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 |
|---|---|---|---|
| add basis sharing between the two convolutions in a residual block (m=16) (ablation) | Basis-compressed EDSR · Set5 x4 | PSNR (dB): 32.14 -> 32.16 | No clear effect |
| add basis sharing within residual block (m=64) (ablation) | Basis-compressed EDSR · Set5 x4 | PSNR (dB): 32.33 -> 32.30 | No clear effect |
| add basis sharing within residual block (parameters) (ablation) | Basis-compressed EDSR (m=64) · EDSR one residual block | #Params: 106k -> 70k | Improved |
| increase the number of basis from 16 to 64 (ablation) | Basis-compressed EDSR · Set5 x4 | PSNR (dB): 32.14 -> 32.33 | Improved |
| increase the number of basis from 24 to 32 (ablation) | Basis-compressed DenseNet-12-40 · CIFAR-10 | top-1 error (%): 5.69% to 5.57% | Improved |
| add compression of the 1x1 transition layers (6 splits) and use more basis (ablation) | Basis-compressed DenseNet-12-40 (M32 -> M36T6) · CIFAR-10 | top-1 error (%): 5.57 -> 5.32 | Improved |
| increase transition-block splits to 12 with 38 basis (ablation) | Basis-compressed DenseNet-12-40 (M36T6 -> M38T12) · CIFAR-10 | top-1 error (%): 5.32 -> 5.56 | Hurt |
| implement split-wise decomposition with s shared-weight 2D convolutions instead of a 3D convolution | Compressed EDSR · training iteration, batch 16, 48x48 patches | run time (ms): 121 ms -> 62 ms | Improved |
| use basis learning instead of Factor | VGG-16 compression · CIFAR-10 | top-1 error (%): 7.12 -> 6.18 | Improved |
| use basis learning instead of Factor | EDSR compression · Urban100 x4 | PSNR (dB): 25.98 -> 26.36 | Improved |
| add basis sharing (Basis-S instead of Basis) (ablation) | EDSR compression · Set5 x4 | PSNR (dB): 32.39 -> 32.29 | Mixed |
| use basis learning instead of Factor-SIC3 | SRResNet compression · Set5 x4 | PSNR (dB): 31.86 -> 31.90 | Improved |
Closest papers by meaning
- Compressing Convolutional Neural Networks via Factorized Convolutional Filters (CVPR 2019 · similarity 0.90)
- Compressing Deep Convolutional Neural Networks by Stacking Low-dimensional Binary Convolution Filters (AAAI 2021 · similarity 0.89)
- FSNet: Compression of Deep Convolutional Neural Networks by Filter Summary (ICLR 2020 · similarity 0.89)
- Packing Convolutional Neural Networks in the Frequency Domain (TPAMI 2018 · similarity 0.89)
- CNNpack: Packing Convolutional Neural Networks in the Frequency Domain (NeurIPS 2016 · similarity 0.88)
- Learning Versatile Filters for Efficient Convolutional Neural Networks (NeurIPS 2018 · similarity 0.88)
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.
