Research map / Architectures and efficiency / Quantization
Make RepVGG Greater Again: A Quantization-Aware Approach
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Diagnoses why RepVGG collapses under INT8 post-training quantization and proposes QARepVGG, a reparameterizable block redesign (no BN on identity/1x1 branches, extra post-addition BN, normal L2 weight decay) that keeps FP32 accuracy while recovering most INT8 accuracy.
Problem
RepVGG-style reparameterized networks suffer severe accuracy collapse under INT8 post-training quantization (e.g., RepVGG-A0 drops from ~72% to ~50% top-1), blocking deployment.
Approach
Attribute the quantization failure to the custom L2 weight decay (which inflates activation variance) and to the identity-branch BN creating outlier fused weights and same-mean branch sums; remove BN from identity and 1x1 branches, use normal L2, and add a post-addition BN so the fused kernel stays deploy-equivalent while being quantization friendly.
Key results
- QARepVGG-A0 reaches 72.2% FP32 and 70.4% INT8 top-1 on ImageNet versus RepVGG-A0's 72.2% FP32 and 50.3% INT8 under the same PTQ setup.
- QARepVGG-B0 keeps INT8 drop within 2% (74.8% FP32 to 72.9% INT8), while RepVGG-B0 falls from 75.1% to 40.2%.
- On YOLOv6s, QARepVGG INT8 mAP is 41.0% (1.3 drop) versus 35.0% (7.4 drop) for RepVGG.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Replace custom L2 with normal L2 (ablation) | RepVGG-A0 · ImageNet validation | INT8 top-1 accuracy: 50.3 -> 61.6 | Helped |
| Remove the denominator from custom L2 (ablation) | RepVGG-A0 · ImageNet validation | INT8 top-1 accuracy: 50.3 -> 61.2 | Helped |
| Remove BN from the identity branch (ablation) | RepVGG-A0 with M1 · ImageNet validation | INT8 top-1 accuracy: 61.6 -> 62.5 | Helped |
| Remove BN from the 1×1 branch (ablation) | RepVGG-A0 with M1 and M2 · ImageNet validation | INT8 top-1 accuracy: 62.5 -> 69.5 | Helped |
| Add a post-addition BN (ablation) | RepVGG-A0 with M1, M2, and M3 · ImageNet validation | INT8 top-1 accuracy: 69.5 -> 70.4 | Helped |
| Apply M1–M4 together (QARepVGG structure) (ablation) | RepVGG-A0 · ImageNet validation | INT8 top-1 accuracy: 50.3 -> 70.4 | Helped |
| Use only the post-addition BN modification (ablation) | RepVGG-A0 · ImageNet validation | INT8 top-1 accuracy: 50.3 -> 57.0 | Helped |
| Set 1×1 branch BN affine=False (ablation) | RepVGG-A0 without identity-branch BN · ImageNet validation | INT8 top-1 accuracy: 62.5 -> 65.5 | Helped |
| Remove BN from the 3×3 branch (ablation) | RepVGG-A0 without identity-branch BN · ImageNet validation | FP32 top-1 accuracy: 0.1 | Hurt |
| Remove BN from both 3×3 and 1×1 branches (ablation) | RepVGG-A0 without identity-branch BN · ImageNet validation | FP32 top-1 accuracy: 0.1 | Hurt |
| Replace RepVGG with QARepVGG structure | RepVGG-A0 · ImageNet validation | INT8 top-1 accuracy: 50.3 -> 70.4 | Helped |
| Replace RepVGG with QARepVGG structure | YOLOv6s · COCO 2017 | INT8 mAP: 35.0 -> 41.0 | Helped |
Surprising findings
- Removing the denominator from the custom L2 loss (or using normal L2) raises INT8 accuracy from 50.3% to about 61% while FP32 barely changes, indicating the weight-decay design itself harms quantization.
- Removing BN from the 1x1 branch lowers INT8 drop substantially while a plain VGG-BN-A0 also quantizes relatively well (70.1% INT8 vs 70.4% FP32).
Limitations
- Evaluation focuses on post-training quantization; QAT results are limited to a few models and a short 10-epoch schedule.
- RepOpt-VGG comparison depends on hyperparameter search that the authors found hard to reproduce for A0/B0.
- Analysis relies on symmetric uniform PyTorch-Quantization toolkit settings; other quantization schemes are not explored.
Experimental setup
Datasets: ImageNet COCO 2017 Cityscapes CIFAR-100
Benchmarks: ImageNet COCO 2017 Cityscapes
Metrics: top-1 accuracy INT8 accuracy mAP mIoU FPS
Baselines: RepVGG RepOpt-VGG OREPA-RepVGG VGG-BN YOLOv6-RepVGG YOLOv6-RepOpt
Builds on: RepVGG RepOpt-VGG OREPA
Closest papers by meaning
- RePaViT: Scalable Vision Transformer Acceleration via Structural Reparameterization on Feedforward Network Layers (ICML 2025 · similarity 0.87)
- Zero-Shot Quantization via Weight-Space Arithmetic (NeurIPS 2026 · similarity 0.85)
- Amortized-Precision Quantization for Early-Exit Vision Transformers (NeurIPS 2026 · similarity 0.85)
- Reg-PTQ: Regression-specialized Post-training Quantization for Fully Quantized Object Detector (CVPR 2024 · similarity 0.85)
- SageAttention: Accurate 8-Bit Attention for Plug-and-play Inference Acceleration (ICLR 2025 · similarity 0.85)
- Quantization without Tears (CVPR 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.