Research map / Architectures and efficiency / Quantization
AdaLog: Post-Training Quantization for Vision Transformers with Adaptive Logarithm Quantizer
Zhuguanyu Wu, Jiaxin Chen, Hanwen Zhong, Di Huang, Yunhong Wang
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 AdaLog quantizer on post-Softmax and post-GELU activations (ablation) | ViT-S PTQ · ImageNet W4/A4 | top-1 accuracy: 62.20 -> 72.01 | Improved |
| add FPCS hyperparameter search (ablation) | ViT-B PTQ · ImageNet W3/A3 | top-1 accuracy: 9.68 -> 15.50 | Improved |
| combine AdaLog with FPCS (ablation) | ViT-B PTQ · ImageNet W3/A3 | top-1 accuracy: 9.68 -> 37.91 | Improved |
| add FPCS on top of AdaLog for DeiT-S (ablation) | DeiT-S PTQ · ImageNet W3/A3 | top-1 accuracy: 22.81 -> 24.47 | No clear effect |
| replace RepQ-ViT quantizers with AdaLog at 4-bit | RepQ-ViT · ImageNet W4/A4 average | top-1 accuracy: average +5.13% | Improved |
| use AdaLog at 3-bit instead of RepQ-ViT | RepQ-ViT · ImageNet W3/A3 ViT-S | top-1 accuracy: 0.10 -> 13.88 | Improved |
| replace log-sqrt2 quantizer with AdaLog (lookup tables, integer-only) | DeiT-T W4/A4 · inference cost | FixOPs: 0.613B -> 0.539B | Improved |
| train activation scaling with LSQ in BRECQ with AdaLog (ablation) | BRECQ ViT-S · ImageNet W4/A4 | top-1 accuracy: 76.48 -> 77.25 | Improved |
| train activation scaling with LSQ in BRECQ without AdaLog (ablation) | BRECQ ViT-S · ImageNet W4/A4 | top-1 accuracy: 69.21 -> 1.93 | Hurt |
| replace alternating search with FPCS (ablation) | AdaLog calibration · DeiT-T W3A3 | top-1 accuracy: 28.41 -> 31.56 | Improved |
| replace brute force search with FPCS (ablation) | AdaLog calibration · DeiT-T W3A3 | GPU minutes: 183 -> 4.1 | Improved |
Later papers that compared with this method
2 later accepted papers report AdaLog as a baseline in their results tables; 2 of them report a better number on most of the results they share with it.
- LampQ: Towards Accurate Layer-wise Mixed Precision Quantization for Vision Transformers (AAAI 2026 · better on 8 of 8 shared results)
- FIMA-Q: Post-Training Quantization for Vision Transformers by Fisher Information Matrix Approximation (CVPR 2025 · better on 1 of 1 shared result)
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
- AIQViT: Architecture-Informed Post-Training Quantization for Vision Transformers (AAAI 2025 · similarity 0.93)
- GPLQ: A General, Practical, and Lightning QAT Method for Vision Transformers (NeurIPS 2025 · similarity 0.90)
- Amortized-Precision Quantization for Early-Exit Vision Transformers (NeurIPS 2026 · similarity 0.90)
- APHQ-ViT: Post-Training Quantization with Average Perturbation Hessian Based Reconstruction for Vision Transformers (CVPR 2025 · similarity 0.90)
- FIMA-Q: Post-Training Quantization for Vision Transformers by Fisher Information Matrix Approximation (CVPR 2025 · similarity 0.90)
- ERQ: Error Reduction for Post-Training Quantization of Vision Transformers (ICML 2024 · similarity 0.90)
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.
