Research map / Architectures and efficiency / Quantization
Unified Data-Free Compression: Pruning and Quantization without Fine-Tuning
Shipeng Bai, Jun Chen, Xintian Shen, Yixuan Qian, Yong Liu
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 linear-combination channel reconstruction after pruning (UDFC vs plain pruning) | structured pruning · ImageNet pruning 30% | Top-1 accuracy (%): 10.10 -> 65.76 (ResNet-101, l2-norm, with 6-bit) | Improved |
| replace one-to-one Neuron Merging compensation with many-to-one linear combination reconstruction | data-free pruning · CIFAR-10 VGG-16 80% pruning | accuracy (%): 40.26 -> 91.79 | Improved |
| data-free UDFC weight quantization instead of data-driven PTQ (DSG, SQuant) | ResNet-18 W6/A6 · ImageNet | Top-1 accuracy (%): 70.74 -> 72.76 (vs SQuant) | Improved |
| increase pruning ratio (ablation) | UDFC · ResNet-56 CIFAR-10 | accuracy: 30%: 90.33; 40%: 86.99; 50%: 81.90 | Hurt |
| reduce weight bit-width to 4-bit (ablation) | UDFC · ResNet-56 CIFAR-10 | accuracy | Mixed |
| add bias-term weighting alpha1 in reconstruction error (0 -> 0.01, then 0.04) (ablation) | UDFC pruning (50% ResNet-56) · CIFAR-10 | accuracy | Mixed |
| tune alpha2 for 2-bit quantization (ablation) | UDFC quantization (ResNet-56) · CIFAR-10 | accuracy | Mixed |
| use UDFC at very low bit-widths (3-bit, 2-bit) | UDFC · ResNet-56 CIFAR-10 | accuracy | Hurt |
| replace data-free-synthetic-sample quantization with data-free closed-form quantization | ResNet-50 quantization · ImageNet | quantization time: 29 seconds (ZeroQ, 8 V100) -> 2 seconds (UDFC, 1 RTX 1080Ti) | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- SQuant — SQuant: On-the-Fly Data-Free Quantization via Diagonal Hessian Approximation (ICLR 2022) · better on 3 of 3 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- CLIP-Q: Deep Network Compression Learning by In-Parallel Pruning-Quantization (CVPR 2018 · similarity 0.90)
- Automatic Neural Network Compression by Sparsity-Quantization Joint Learning: A Constrained Optimization-Based Approach (CVPR 2020 · similarity 0.89)
- OPQ: Compressing Deep Neural Networks with One-shot Pruning-Quantization (AAAI 2021 · similarity 0.89)
- Differentiable Joint Pruning and Quantization for Hardware Efficiency (ECCV 2020 · similarity 0.89)
- Automatic Joint Structured Pruning and Quantization for Efficient Neural Network Training and Compression (CVPR 2025 · similarity 0.88)
- Data-Free Quantization via Pseudo-label Filtering (CVPR 2024 · 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.
