PaddleClas/docs/zh_CN/extension/paddle_quantization.md

6 lines
622 B
Markdown
Raw Blame History

This file contains ambiguous Unicode characters!

This file contains ambiguous Unicode characters that may be confused with others in your current locale. If your use case is intentional and legitimate, you can safely ignore this warning. Use the Escape button to highlight these characters.

# 模型量化
模型量化是 [PaddleSlim](https://github.com/PaddlePaddle/PaddleSlim) 的特色功能之一,支持动态和静态两种量化训练方式,对权重全局量化和 Channel-Wise 量化,同时以兼容 Paddle-Lite 的格式保存模型。
[PaddleClas](https://github.com/PaddlePaddle/PaddleClas) 使用该量化工具量化了78.9%的mobilenet_v3_large_x1_0的蒸馏模型, 量化后SD855上预测速度从19.308ms加速到14.395ms存储大小从21M减小到10M top1识别准确率75.9%。
具体的训练方法可以参见 [PaddleSlim 量化训练](../../../deploy/slim/quant/README.md)。