【模型压缩推全计划】为六大套件新增模型压缩功能 #12736
Replies: 8 comments
-
队伍名:风清扬 |
Beta Was this translation helpful? Give feedback.
-
队伍名:风清扬 |
Beta Was this translation helpful? Give feedback.
-
队伍名:飞桨小队 |
Beta Was this translation helpful? Give feedback.
-
队伍名:德布罗意波 |
Beta Was this translation helpful? Give feedback.
-
队伍名:Dec20b |
Beta Was this translation helpful? Give feedback.
-
队伍名:晓飞队 |
Beta Was this translation helpful? Give feedback.
-
队伍名:Zheng-Bicheng |
Beta Was this translation helpful? Give feedback.
-
队伍名:Zheng-Bicheng |
Beta Was this translation helpful? Give feedback.
-
背景
目前各套件的模型压缩能力参差不齐,而模型压缩作为部署之前的一步,可以在不损害或者少量损害模型精度的情况下,对模型的能耗,速度、大小都有显著的改善。因此为了对各套件的模型压缩进行推全,我们提出了基于PaddleSlim的ACT为各大套件新增模型压缩功能的计划。
解决步骤:
查看Paddleslim的ACT文档,学习ACT,熟悉基于不同任务进行ACT中量化压缩的主要步骤。
为某一任务接入ACT中的量化训练压缩功能:在适配paddle2.5,paddleslim2.5和套件develop版本的前提下,仿照pp-Liteseg接入ACT的PR,在各大套件的部署文件夹下(例如分割为PaddleSeg/deploy/slim/act) 接入ACT。下面以分割为例说明需要包括的新增文件内容:
参考paddleslim的ACT使用文档,对量化训练模型进行导出,并验证不同模型在量化蒸馏训练前后的精度和速度对比,验证通过后,撰写量化训练使用文档,增加相关Readme在 PaddleSeg/deploy/slim/act/readme.md。
提交PR到对应模型的仓库中。
注:不同套件的部署目录如下,将上述分割的目录进行相应替换即可:
模型列表
同一个任务仅有配置文件差异,一个任务需要完成对应的模型配置,领取通用目标检测任务的开发者将认领四个检测模型。
Beta Was this translation helpful? Give feedback.
All reactions