Quantization without Tears (CVPR 2025)

Implementation of QwT, a simple, fast, and general approach to network quantization that “adds no tears” to your workflow. QwT augments any PTQ model with lightweight linear compensation layers to recover information lost during quantization .

---

📖 Paper

Minghao Fu, Hao Yu, Jie Shao, Junjie Zhou, Ke Zhu & Jianxin Wu
Quantization without Tears, the Proceedings of the Computer Vision and Pattern Recognition Conference (CVPR)
PDF • CVPR 2025 Version

---

🚀 Features

• Fast: closed‐form compensation parameters can be computed on a small calibration set in under 2 minutes.
• Accurate: outperforms standard PTQ methods without any back‐prop.
• Simple: zero task‐specific hyperparameters; only a few linear layers (W, b) per block.
• General: applies to CNNs (ResNet), Transformers (ViT, Swin), detection (Mask R-CNN, DETR), segmentation, multi‐modal (CLIP), generative (DiT) and LLMs (LLaMA).
• Practical: integrates seamlessly with TensorRT or any existing INT8/PTQ pipeline.

---

📌 Changelog

• 2025-09-30: Release code implementing QwT + RepQ-ViT for CLIP-based multimodal recognition.
• 2025-07-22: Release code implementing QwT with pytorch‑percentile, along with latency‑testing scripts to reproduce the results in Table 2 of our paper.
• 2025-06-23: Release code implementing QwT + ResNet, evaluated on ImageNet.
• 2025-06-09: Release code showcasing QwT integrated with the baseline PTQ method RepQ-ViT for both classification and detection tasks.
• 2025-02-27: QwT is accepted by CVPR 2025 link.
• 2024-11-21: QwT preprint published on (arXiv:2411.13918).

---

📦 Installation

• To install QwT and develop locally:

  git clone https://github.com/wujx2001/qwt.git
  cd qwt

🛠️ Usage

For detailed reproduction instructions, please refer to:

• QwT-CLIP-Classification (RepQ‑ViT) README — Reproduce QwT zero-shot classification results using RepQ‑ViT on CLIP model.
• QwT-Classification (PyTorch‑Percentile) README — Reproduce QwT ImageNet classification results using the PyTorch‑Percentile method and measure inference latency.
• QwT-Classification (RepQ‑ViT) README — Reproduce QwT ImageNet classification results using RepQ‑ViT (vit & swin) and the Percentile (resnet) baseline.
• QwT-Detection (RepQ‑ViT) README — Reproduce QwT COCO detection results using the RepQ‑ViT baseline.

---

🙏 Acknowledgements

This implementation builds on code from RepQ-ViT and leverages the timm library.

🎓 Citation

We would greatly appreciate it if you could cite our paper if you find our implementation helpful in your work.

@InProceedings{Fu_2025_CVPR,
    author    = {Fu, Minghao and Yu, Hao and Shao, Jie and Zhou, Junjie and Zhu, Ke and Wu, Jianxin},
    title     = {Quantization without Tears},
    booktitle = {Proceedings of the Computer Vision and Pattern Recognition Conference (CVPR)},
    month     = {June},
    year      = {2025},
    pages     = {4462-4472}
}