Skip to content

This issue was moved to a discussion.

You can continue the conversation there. Go to discussion →

New issue

Have a question about this project? # for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “#”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? # to your account

【模型压缩推全计划】为六大套件新增模型压缩功能 #10657

Closed
shiyutang opened this issue Aug 16, 2023 · 8 comments
Closed
Assignees
Labels
Code PR is needed This issue could inspire a code PR good first issue Good for newcomers

Comments

@shiyutang
Copy link
Collaborator

shiyutang commented Aug 16, 2023

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:

  1. 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。

  2. 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:

    • 增加ACT启动脚本 run_seg.py 在 PaddleSeg/deploy/slim/act/run_seg.py
    • 增加量化训练配置文件在 PaddleSeg/deploy/slim/act/configs/ppliteseg_qat.py
    • 增加用于验证精度和速度的推理脚本在PaddleSeg/deploy/slim/act/test_seg.py
  3. 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。

  4. 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:

  • PaddleClas: PaddleClas/deploy/slim/act
  • PaddleDetection: PaddleDetection/deploy/auto_compression(该目录下的文件为之前接入的文件,可以复用)
  • Paddle3D:Paddle3D/deploy/slim_act
  • PaddleOCR: PaddleOCR/deploy/slim/act
  • PaddleSeg: PaddleSeg/deploy/slim/act/

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。

任务名称 模型参数名称 报名开发者
通用目标检测 DINO_r50_4scale @MINGtoMING
通用目标检测 RT-DETR-L @MINGtoMING
通用目标检测 PP-YOLOE+_crn_l_80e @MINGtoMING
通用目标检测 PP-PicoDet_s_320_LCNet @MINGtoMING
通用图像分类 SwinTransformer_base_patch4_window7_224 @vax521
通用图像分类 PP-LCNet_x1_0 @vax521
通用图像分类 CLIP_vit_base_patch16_224 @vax521
通用图像分类 PP-HGNet_small @vax521
通用图像分类 ResNet50 @vax521
通用图像分类 MobileNetV3_small_x1_0 @vax521
通用文本检测 PP-OCRv4-server-det @xu-peng-7
通用文本检测 PP-OCRv4-mobile-det @xu-peng-7
通用文本识别 PP-OCRv4-server-rec
通用文本识别 PP-OCRv4-mobile-rec
通用图像分割 OCRNet_HRNetW48 @marshall-dteach
通用图像分割 SegFormer-B0 @marshall-dteach
通用图像分割 PP-LiteSeg-Tiny @marshall-dteach
通用图像分割 PP-MobileSeg-Base @marshall-dteach
3D目标检测 CenterPoint_pillars_02voxel
3D目标检测 PETRv2_v99
3D目标检测 PETRv1_v99
3D目标检测 CaDDN_ocrnet_hrnet_w18
通用主体检测 PP-ShiTuv2-det @Zheng-Bicheng
通用检索特征 PP-ShiTuv2-rec @Zheng-Bicheng
通用版面分析 PP-Structurev2-layout
通用表格识别 PP-Structurev2-SLANet
通用关键信息抽取 PP-Structurev2-vi-layoutxlm
@WilliamQf-AI
Copy link
Contributor

WilliamQf-AI commented Aug 16, 2023

队伍名:风清扬
队伍成员微信昵称:WilliamQf-AI
功能描述:通用版面分析

@shiyutang shiyutang added the Code PR is needed This issue could inspire a code PR label Aug 16, 2023
@WilliamQf-AI
Copy link
Contributor

队伍名:风清扬
队伍成员微信昵称:WilliamQf-AI
功能描述:通用表格识别

@ToddBear
Copy link
Collaborator

队伍名:飞桨小队
队伍成员微信昵称:zejia_su
功能描述:通用文本识别

@marshall-dteach
Copy link

队伍名:德布罗意波
队伍成员微信昵称:TB1250859983
功能描述:通用图像分割

@xu-peng-7
Copy link
Contributor

队伍名:Dec20b
队伍成员微信昵称:muted
功能描述:通用文本检测

@vax521
Copy link

vax521 commented Aug 30, 2023

队伍名:晓飞队
队伍成员微信昵称:晓飞
功能描述:通用图像分类

@Zheng-Bicheng
Copy link

队伍名:Zheng-Bicheng
队伍成员微信昵称:Zheng-Bicheng
功能描述:通用主体检测

@Zheng-Bicheng
Copy link

队伍名:Zheng-Bicheng
队伍成员微信昵称:Zheng-Bicheng
功能描述:通用检索特征

@PaddlePaddle PaddlePaddle locked and limited conversation to collaborators Jun 5, 2024
@SWHL SWHL converted this issue into discussion #12736 Jun 5, 2024

This issue was moved to a discussion.

You can continue the conversation there. Go to discussion →

Labels
Code PR is needed This issue could inspire a code PR good first issue Good for newcomers
Projects
None yet
Development

No branches or pull requests

8 participants