mirrors
/
PaddleClas
mirror of https://github.com/PaddlePaddle/PaddleClas.git
1
0
Fork 0
Code Issues Projects Releases Wiki Activity

add quantize

pull/160/head
WuHaobo 2020-06-11 01:23:46 +08:00
parent 6fcbfdc785
commit 48014f0973
1 changed files with 10 additions and 4 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

14
docs/en/extension/paddle_quantization.md
View File

@ -1,7 +1,13 @@
# Model Quantifization
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](https://paddlepaddle.github.io/PaddleSlim/quick_start/quant_aware_tutorial.html)。
Int8 quantization is one of the key features in [PaddleSlim](https://github.com/PaddlePaddle/PaddleSlim).
It supports two kinds of training aware, **Dynamic strategy** and **Static strategy**,
layer-wise and channel-wise quantization,
and using PaddleLite to deploy models generated by PaddleSlim.
By using this toolkit, [PaddleClas](https://github.com/PaddlePaddle/PaddleClas) quantized the mobilenet_v3_large_x1_0 model whose accuracy is 78.9% after distilled.
After quantized, the prediction speed is accelerated from 19.308ms to 14.395ms on SD855.
The storage size is reduced from 21M to 10M.
The top1 recognition accuracy rate is 75.9%.
For specific training methods, please refer to [PaddleSlim quant aware](https://paddlepaddle.github.io/PaddleSlim/quick_start/quant_aware_tutorial.html)。