Research map / Architectures and efficiency / Quantization
Towards Accurate Post-training Network Quantization via Bit-Split and Stitching
Peisong Wang, Qiang Chen, Xiangyu He, Jian Cheng
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 |
|---|---|---|---|
| Bit-split weight quantization (vs TF-Lite) at 3-bit, weights only | ResNet-18/50/101, VGG-16-BN weights · ImageNet classification | top-5 accuracy: ResNet-18 87.45 vs 2.68 for TF-Lite | Improved |
| Bit-split weight quantization vs TF-Lite at 4-bit and higher, weights only | weight quantization · ImageNet classification | top-1 accuracy: gap grows as bit-width decreases | Improved |
| Per-channel activation quantization vs per-layer (ResNet-18, 8-bit weights) (ablation) | activation quantization · ImageNet ResNet-18 | top-1 accuracy: 3-bit: 65.91 vs 64.77 | Improved |
| ECAQ (error compensated per-layer activation quantization) vs direct per-layer quantization (ablation) | activation quantization · ImageNet ResNet-18, 3-bit and 4-bit activations | top-1 accuracy: +1.44% at 3-bit, +0.45% at 4-bit | Improved |
| ECAQ vs direct per-layer quantization at 8-bit to 5-bit activations (ablation) | activation quantization · ImageNet ResNet-18 | top-1 accuracy: differences less than 0.1% at higher bit widths | No clear effect |
| Bit-split 8-bit activation quantization added to 8-bit-weight... (A8 per-layer) on top of Bit-split weights (ablation) | ResNet-18 activations · ImageNet classification, weights 3-8 bit | top-1 accuracy: differences under 0.1% | No clear effect |
| Bit-split + ECAQ joint quantization of weights and activations at 4-bit vs TF-Lite | ResNet-18, ResNet-50, VGG-16-BN · ImageNet classification W4A4 | top-5 accuracy: ResNet-18 87.76 vs 36.33 | Improved |
| Bit-split with 4-bit weights and 8-bit activations (per-layer) vs ACIQ-Mix (mixed precision, per-channel, dynamic) | ResNet-18 A8W4 · ImageNet classification | top-1 accuracy: 69.1 vs 68.3 (Bit-split higher on ResNet-18); ResNet-50 75.6 vs 75.3, but ResNet-101 76.9 tie and VGG-16-BN 73.0 vs 72.4 | Mixed |
| Per-channel activation quantization added to Bit-split (W4A4) (ablation) | Bit-split W4A4 · ImageNet classification | top-1 accuracy: ResNet-18 68.1 vs 67.6 (per-layer) | Improved |
| Bit-split 4-bit weights with 8-bit activations applied to detection/segmentation | RetinaNet and Mask R-CNN (ResNet-50) · COCO minival | AP0.5:0.95: about 0.8%~1.2% mAP degradation | No clear effect |
Closest papers by meaning
- Optimization-Based Post-Training Quantization With Bit-Split and Stitching (TPAMI 2022 · similarity 0.94)
- Bit-Shrinking: Limiting Instantaneous Sharpness for Improving Post-Training Quantization (CVPR 2023 · similarity 0.89)
- Once Quantization-Aware Training: High Performance Extremely Low-Bit Architecture Search (ICCV 2021 · similarity 0.88)
- Reg-PTQ: Regression-specialized Post-training Quantization for Fully Quantized Object Detector (CVPR 2024 · similarity 0.86)
- AIQViT: Architecture-Informed Post-Training Quantization for Vision Transformers (AAAI 2025 · similarity 0.86)
- ParetoQ: Improving Scaling Laws in Extremely Low-bit LLM Quantization (NeurIPS 2025 · similarity 0.85)
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.
