NeurIPS 2021 Transformer部署难?北大&华为诺亚提出Vision Transformer的后训练量化方法
详细信息如下:
论文链接:https://arxiv.org/abs/2106.14156
项目链接:未开源
01
02
2.1 Preliminaries
2.2 Optimization for Post-Training Quantization
Similarity-Aware Quantization for Linear Operation
Ranking-Aware Quantization for Self-Attention
Bias Correction
2.3 Mixed-Precision Quantization for Vision Transformer
03
3.1. Results and Analysis
Image classification
Object Detection
3.2. Ablation study
04
END
赞 (0)