Research map / Architectures and efficiency / Quantization
Transform Quantization for CNN Compression
Sean I. Young, Zhe Wang, David Taubman, Bernd Girod
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 the end-to-end learned transform (row-ELT) instead of the KLT (ablation) | transform quantization without retraining · ImageNet val (ResNet-18, 3.0 bits) | top-1 accuracy (%): 67.2 -> 67.9 | Improved |
| use row-ELT instead of row-KLT (ablation) | transform quantization without retraining · ImageNet val (AlexNet, 1.0 bits) | top-1 accuracy (%): 53.1 -> 53.4 | Improved |
| use row-ELT instead of row-KLT (ablation) | transform quantization without retraining · ImageNet val (ResNet-50, about 3 bits) | top-1 accuracy (%): 73.6 -> 73.6 | No clear effect |
| add 3 epochs of retraining | transform quantization · ImageNet val (ResNet-18, 3.0 bits) | top-1 accuracy (%): 67.9 -> 69.8 | Improved |
| add 3 epochs of retraining | transform quantization · ImageNet val (ResNet-50, about 3 bits) | top-1 accuracy (%): 73.6 -> 76.0 | Improved |
| add 3 epochs of retraining | transform quantization · ImageNet val (AlexNet, 1.0 bits) | top-1 accuracy (%): 53.4 -> 55.3 | Improved |
| increase bit-rate from 1.0 to 3.0 bits with retraining | transform quantization · ImageNet val (ResNet-18) | top-1 accuracy (%): 64.5 -> 69.8 | Improved |
| increase EDSR bit-depth from 1 to 2 bits | transform quantization · Set5 4x super-resolution | PSNR (dB): 30.31 -> 32.28 | Improved |
| increase EDSR bit-depth from 3 to 4 bits | transform quantization · Set5 4x super-resolution | PSNR (dB): 32.51 -> 32.60 | Improved |
| lower MobileNet-v2 bit-rate from 7.3 to 5.5 bits | transform quantization · ImageNet val (MobileNet-v2) | top-1 accuracy (%): 71.3 -> 64.8 | Hurt |
| lower MobileNet-v2 bit-rate from 7.3 to 5.5 bits | transform quantization · MobileNet-v2 on TPU | inference rate (img/s): 2197 -> 2407 | Improved |
| use transform quantization instead of HAQ at similar bit-rate | MobileNet-v2 · ImageNet val | top-1 accuracy (%): 71.2 -> 71.3 | No clear effect |
Closest papers by meaning
- Focused Quantization for Sparse CNNs (NeurIPS 2019 · similarity 0.92)
- And the Bit Goes Down: Revisiting the Quantization of Neural Networks (ICLR 2020 · similarity 0.91)
- Bit-Shrinking: Limiting Instantaneous Sharpness for Improving Post-Training Quantization (CVPR 2023 · similarity 0.90)
- Post-Training Quantization for Vision Transformer (NeurIPS 2021 · similarity 0.90)
- Optimization-Based Post-Training Quantization With Bit-Split and Stitching (TPAMI 2022 · similarity 0.90)
- Quantization without Tears (CVPR 2025 · 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.
