Bit-shrinking: Limiting Instantaneous Sharpness for Improving Post-training Quantization

Lin, Chen; Peng, Bo; Li, Zheyang; Tan, Wenming; Ren, Ye; Xiao, Jun; Pu, Shiliang

doi:10.1109/cvpr52729.2023.01554

Cited by 4 publications

(1 citation statement)

References 15 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…It can be divided into posttraining quantization (PTQ) and quantization-aware training (QAT). PTQ compresses a pretrained full-precision model into a low-bit model without retraining or fine-tuning [33][34][35]. However, PTQ often suffers from significant accuracy degradation due to the lack of weightaware training.…”

Section: Model Quantizationmentioning

confidence: 99%

Quantized Graph Neural Networks for Image Classification

Xu,

Ma,

Zeng

et al. 2023

Mathematics

View full text Add to dashboard Cite

Researchers have resorted to model quantization to compress and accelerate graph neural networks (GNNs). Nevertheless, several challenges remain: (1) quantization functions overlook outliers in the distribution, leading to increased quantization errors; (2) the reliance on full-precision teacher models results in higher computational and memory overhead. To address these issues, this study introduces a novel framework called quantized graph neural networks for image classification (QGNN-IC), which incorporates a novel quantization function, Pauta quantization (PQ), and two innovative self-distillation methods, attention quantization distillation (AQD) and stochastic quantization distillation (SQD). Specifically, PQ utilizes the statistical characteristics of distribution to effectively eliminate outliers, thereby promoting fine-grained quantization and reducing quantization errors. AQD enhances the semantic information extraction capability by learning from beneficial channels via attention. SQD enhances the quantization robustness through stochastic quantization. AQD and SQD significantly improve the performance of the quantized model with minimal overhead. Extensive experiments show that QGNN-IC not only surpasses existing state-of-the-art quantization methods but also demonstrates robust generalizability.

show abstract

Section: Model Quantizationmentioning

confidence: 99%