Skip to content

Python专题
Python专题

⁢い寻常づ / Python专题
与 Fork 源项目一致

Fork自 GitCode官方 / Python专题

通知 3
Star 0
Fork 0
Python专题
Python专题

前往新版Gitcode，体验更适合开发者的 AI 搜索 >>

提交 68ba74f3 编写于 作者: M MaoXianxin
浏览文件
操作

2021/06/11-2021/06/18项目关注数增长排行榜

上级 191d026e
隐藏空白更改
内联 并排
Showing with 99 addition and 108 deletion
README.md
浏览文件 @ 68ba74f3
# CODE CHINA项目推荐｜基于强化学习的自动化剪枝模型
CODE CHINA上最新开源的一个基于强化学习的自动化剪枝模型，本模型在图像识别的实验证明了能够有效减少计算量，同时还能提高模型的精度。
今天为大家介绍一个CODE CHINA上最新开源的一个**基于强化学习的自动化剪枝模型**，本模型在图像识别的实验证明了能够有效减少计算量，同时还能提高模型的精度。
项目地址：[https://codechina.csdn.net/greenplum/cv-automatic-pruning-transformer](https://codechina.csdn.net/greenplum/cv-automatic-pruning-transformer)
## **介绍**
目前的强化学习工作很多集中在利用外部环境的反馈训练agent，忽略了模型本身就是一种能够获得反馈的环境。本项目的核心思想是：将模型视为环境，构建附生于模型的 agent ，以辅助模型进一步拟合真实样本。
![](https://maoxianxin1996.oss-accelerate.aliyuncs.com/codechina1/20210618103018.png)
大多数领域的模型都可以采用这种方式来优化，如cv/多模态等。它至少能够以三种方式工作：
**1.过滤噪音信息，如删减语音或图像特征；**
**2.进一步丰富表征信息，如高效引用外部信息；**
**3.实现记忆、联想、推理等复杂工作，如构建重要信息的记忆池。**
这里推出一款早期完成的裁剪机制transformer版本(后面称为APT)，实现了一种更高效的训练模式，能够优化模型指标；此外，可以使用动态图丢弃大量的不必要单元，在指标基本不变的情况下，大幅降低计算量。
该项目希望为大家抛砖引玉。
![](https://maoxianxin1996.oss-accelerate.aliyuncs.com/codechina1/20210618103136.png)
## **为什么要做自动剪枝**
在具体任务中，往往存在大量毫无价值的信息和过渡性信息，有时不但对任务无益，还会成为噪声。比如：表述会存在冗余/无关片段以及过渡性信息；动物图像识别中，有时候背景无益于辨别动物主体，即使是动物部分图像，也仅有小部分是关键的特征。
![](https://maoxianxin1996.oss-accelerate.aliyuncs.com/codechina1/20210618103220.png)
以transformer为例，在进行self-attention计算时其复杂度与序列长度平方成正比。长度为10，复杂度为100；长度为9，复杂度为81。
利用强化学习构建agent，能够精准且自动化地动态裁剪已丧失意义部分，甚至能将长序列信息压缩到50-100之内（实验中有从500+的序列长度压缩到个位数的示例），以大幅减少计算量。
实验中，发现与裁剪agent联合训练的模型比普通方法训练的模型效果要更好。
## **模型介绍及实验**
**模型主体**
基于transformer的视觉预训练模型ViT是本项目的模型主体，具体细节可以查看论文:《An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale》
![](https://maoxianxin1996.oss-accelerate.aliyuncs.com/codechina1/20210618103254.png)
**自动化裁剪的智能体**
对于强化学习agent来说，最关键的问题之一是如何衡量动作带来的反馈。为了评估单次动作所带来的影响，使用了以下三步骤：
1、使用一个普通模型（无裁剪模块）进行预测；
2、使用一个带裁剪器的模型（执行一次裁剪动作）进行预测；
3、对比两次预测的结果，若裁剪后损失相对更小，则说明该裁剪动作帮助了模型进一步拟合真实状况，应该得到奖励；反之，应该受到惩罚。
但是在实际预测过程中，模型是同时裁剪多个单元的，这或将因为多个裁剪的连锁反应而导致模型失效。训练过程中需要构建一个带裁剪器的模型（可执行多次裁剪动作），以减小该问题所带来的影响。
综上，本模型使用的是三通道模式进行训练。
![](https://maoxianxin1996.oss-accelerate.aliyuncs.com/codechina1/20210618103327.png)
关于裁剪器的模型结构设计，本模型中认为如何衡量一个信息单元是否对模型有意义，建立于其自身的信息及它与任务的相关性上。
因此以信息单元本身及它与CLS单元的交互作为agent的输入信息。
![](https://maoxianxin1996.oss-accelerate.aliyuncs.com/codechina1/20210618103351.png)
**实验**
| 数据集 | ViT | APT(pruning) | APT(no pruning) |
| --------- | ----- | ------------ | --------------- |
| CIFAR-100 | 92.3 | 92.6 | 93.03 |
| CIFAR-10 | 99.08 | 98.93 | 98.92 |
以上加载的均为ViT-B_16，resolution为224*224。
## **使用说明**
**环境**
![](https://maoxianxin1996.oss-accelerate.aliyuncs.com/codechina1/20210618103423.png)
下载经过预先训练的模型（来自Google官方）
本项目使用的型号：ViT-B_16（您也可以选择其它型号进行测试）
![](https://maoxianxin1996.oss-accelerate.aliyuncs.com/codechina1/20210618103444.png)
**训练与推理**
下载好预训练模型就可以跑了。
![](https://maoxianxin1996.oss-accelerate.aliyuncs.com/codechina1/20210618103505.png)
CIFAR-10和CIFAR-100会自动下载和培训。如果使用其他数据集，您需要自定义data_utils.py。
在裁剪模式的推理过程中，预期您将看到如下格式的输出。
![](https://maoxianxin1996.oss-accelerate.aliyuncs.com/codechina1/20210618103525.png)
默认的batch size为72、gradient_accumulation_steps为3。当GPU内存不足时，您可以通过它们来进行训练。
注：相较于原始的ViT，APT(Automatic pruning transformer)的训练步数、训练耗时都会上升。原因是使用pruning agent的模型由于总会丢失部分信息，使得收敛速度变慢，同时为了训练pruning agent，也需要多次的观测、行动、反馈。
![](https://maoxianxin1996.oss-accelerate.aliyuncs.com/codechina/20210615093836.png)
\ No newline at end of file
# 2021/06/11-2021/06/18项目关注数排行榜
| 项目 | 链接 |
| ---------------------------------------------------------- | ------------------------------------------------------------ |
| hihell/python120 | https://codechina.csdn.net/hihell/python120 |
| mirrors/ultralytics/yolov5 | https://codechina.csdn.net/mirrors/ultralytics/yolov5 |
| mirrors/vuejs/vue-devtools | https://codechina.csdn.net/mirrors/vuejs/vue-devtools |
| csdn_codechina/enterprise_technology | https://codechina.csdn.net/csdn_codechina/enterprise_technology |
| EricLee/handpose_x | https://codechina.csdn.net/EricLee/handpose_x |
| csdn_codechina/deep_learning | https://codechina.csdn.net/csdn_codechina/deep_learning |
| csdn_codechina/dataset | https://codechina.csdn.net/csdn_codechina/dataset |
| mirrors/mobz/elasticsearch-head | https://codechina.csdn.net/mirrors/mobz/elasticsearch-head |
| mirrors/vuejs-templates/webpack | https://codechina.csdn.net/mirrors/vuejs-templates/webpack |
| geeknews/CSDN-Developer-Weekly | https://codechina.csdn.net/geeknews/CSDN-Developer-Weekly |
| mirrors/LiuKay/WareHouseManagSys | https://codechina.csdn.net/mirrors/LiuKay/WareHouseManagSys |
| mirrors/PanJiaChen/vue-element-admin | https://codechina.csdn.net/mirrors/PanJiaChen/vue-element-admin |
| mirrors/alibaba/easyexcel | https://codechina.csdn.net/mirrors/alibaba/easyexcel |
| csdn_codechina/frequency_command | https://codechina.csdn.net/csdn_codechina/frequency_command |
| mirrors/lutzroeder/netron | https://codechina.csdn.net/mirrors/lutzroeder/netron |
| dcloud/uni-starter | https://codechina.csdn.net/dcloud/uni-starter |
| MiddlewareDesign/doc | https://codechina.csdn.net/MiddlewareDesign/doc |
| mirrors/apache/incubator-echarts | https://codechina.csdn.net/mirrors/apache/incubator-echarts |
| mirrors/shaochen2015821426/music-player | https://codechina.csdn.net/mirrors/shaochen2015821426/music-player |
| COPU/OCOW-2021 | https://codechina.csdn.net/COPU/OCOW-2021 |
| EricLee/dpcas | https://codechina.csdn.net/EricLee/dpcas |
| mirrors/bubbliiiing/yolo3-pytorch | https://codechina.csdn.net/mirrors/bubbliiiing/yolo3-pytorch |
| mirrors/qwerty472123/wxappUnpacker | https://codechina.csdn.net/mirrors/qwerty472123/wxappUnpacker |
| mirrors/tensorflow/tensorflow | https://codechina.csdn.net/mirrors/tensorflow/tensorflow |
| mirrors/antabot/white-jotter | https://codechina.csdn.net/mirrors/antabot/white-jotter |
| chasedream1129/SleepyCD | https://codechina.csdn.net/chasedream1129/SleepyCD |
| mirrors/bubbliiiing/yolov4-pytorch | https://codechina.csdn.net/mirrors/bubbliiiing/yolov4-pytorch |
| mirrors/gildas-lormeau/jsonview-for-chrome | https://codechina.csdn.net/mirrors/gildas-lormeau/jsonview-for-chrome |
| mirrors/medcl/elasticsearch-analysis-ik | https://codechina.csdn.net/mirrors/medcl/elasticsearch-analysis-ik |
| mirrors/google-research/bert | https://codechina.csdn.net/mirrors/google-research/bert |
| mirrors/microsoftarchive/redis | https://codechina.csdn.net/mirrors/microsoftarchive/redis |
| mirrors/tesseract-ocr/tessdata | https://codechina.csdn.net/mirrors/tesseract-ocr/tessdata |
| mirrors/ultralytics/yolov3 | https://codechina.csdn.net/mirrors/ultralytics/yolov3 |
| qing_gee/csdn-book | https://codechina.csdn.net/qing_gee/csdn-book |
| EricLee/yolo_v3 | https://codechina.csdn.net/EricLee/yolo_v3 |
| codechina/developer-roadmap/python/intro | https://codechina.csdn.net/codechina/developer-roadmap/python/intro |
| mirrors/ecomfe/echarts-for-weixin | https://codechina.csdn.net/mirrors/ecomfe/echarts-for-weixin |
| mirrors/audi-1/sqli-labs | https://codechina.csdn.net/mirrors/audi-1/sqli-labs |
| mirrors/cheungchingyin/studentmanagementsystembyjsp | https://codechina.csdn.net/mirrors/cheungchingyin/studentmanagementsystembyjsp |
| mirrors/lenve/vhr | https://codechina.csdn.net/mirrors/lenve/vhr |
| mirrors/philferriere/cocoapi | https://codechina.csdn.net/mirrors/philferriere/cocoapi |
| mirrors/qqwweee/keras-yolo3 | https://codechina.csdn.net/mirrors/qqwweee/keras-yolo3 |
| mirrors/steveloughran/winutils | https://codechina.csdn.net/mirrors/steveloughran/winutils |
| mirrors/tensorflow/models | https://codechina.csdn.net/mirrors/tensorflow/models |
| mirrors/wkentaro/labelme | https://codechina.csdn.net/mirrors/wkentaro/labelme |
| csdn_codechina/opencv_technology | https://codechina.csdn.net/csdn_codechina/opencv_technology |
| lyndon_li/go_web | https://codechina.csdn.net/lyndon_li/go_web |
| mirrors/52JDK/Source-Collection | https://codechina.csdn.net/mirrors/52JDK/Source-Collection |
| mirrors/alibaba/canal | https://codechina.csdn.net/mirrors/alibaba/canal |
| mirrors/alibaba/nacos | https://codechina.csdn.net/mirrors/alibaba/nacos |
| mirrors/hanhanwanghahaha/Confession | https://codechina.csdn.net/mirrors/hanhanwanghahaha/Confession |
| mirrors/nltk/nltk_data | https://codechina.csdn.net/mirrors/nltk/nltk_data |
| mirrors/alibaba/fastjson | https://codechina.csdn.net/mirrors/alibaba/fastjson |
| mirrors/best-bo-cai/books_management | https://codechina.csdn.net/mirrors/best-bo-cai/books_management |
| mirrors/bigkoo/android-pickerview | https://codechina.csdn.net/mirrors/bigkoo/android-pickerview |
| mirrors/shanglianlm0525/pytorch-networks | https://codechina.csdn.net/mirrors/shanglianlm0525/pytorch-networks |
| mirrors/square/okhttp | https://codechina.csdn.net/mirrors/square/okhttp |
| mirrors/4ttty/winutils | https://codechina.csdn.net/mirrors/4ttty/winutils |
| mirrors/alibaba/p3c | https://codechina.csdn.net/mirrors/alibaba/p3c |
| mirrors/bitcarmanlee/easy-algorithm-interview-and-practice | https://codechina.csdn.net/mirrors/bitcarmanlee/easy-algorithm-interview-and-practice |
| mirrors/cartucho/map | https://codechina.csdn.net/mirrors/cartucho/map |
| mirrors/ctripcorp/apollo | https://codechina.csdn.net/mirrors/ctripcorp/apollo |
| mirrors/jikeMisma/java_datapase_project | https://codechina.csdn.net/mirrors/jikeMisma/java_datapase_project |
| mirrors/pytorch/vision | https://codechina.csdn.net/mirrors/pytorch/vision |
| mirrors/scwang90/SmartRefreshLayout | https://codechina.csdn.net/mirrors/scwang90/SmartRefreshLayout |
| mirrors/spring-projects/spring-framework | https://codechina.csdn.net/mirrors/spring-projects/spring-framework |
| mirrors/srccodes/hadoop-common-2.2.0-bin | https://codechina.csdn.net/mirrors/srccodes/hadoop-common-2.2.0-bin |
| mirrors/wszqkzqk/deepin-wine-ubuntu | https://codechina.csdn.net/mirrors/wszqkzqk/deepin-wine-ubuntu |
| mirrors/youth5201314/banner | https://codechina.csdn.net/mirrors/youth5201314/banner |
| paddlepaddle/Paddle | https://codechina.csdn.net/paddlepaddle/Paddle |
| linxinfa/UnityTile3D | https://codechina.csdn.net/linxinfa/UnityTile3D |
| mirrors/arut/nginx-rtmp-module | https://codechina.csdn.net/mirrors/arut/nginx-rtmp-module |
| mirrors/espressif/arduino-esp32 | https://codechina.csdn.net/mirrors/espressif/arduino-esp32 |
| mirrors/jesolem/pcv | https://codechina.csdn.net/mirrors/jesolem/pcv |
| mirrors/litepalframework/litepal | https://codechina.csdn.net/mirrors/litepalframework/litepal |
| mirrors/macrozheng/mall | https://codechina.csdn.net/mirrors/macrozheng/mall |
| mirrors/tianxiaomo/pytorch-yolov4 | https://codechina.csdn.net/mirrors/tianxiaomo/pytorch-yolov4 |
| mirrors/zxing/zxing | https://codechina.csdn.net/mirrors/zxing/zxing |
| mirrors/Awesome-HarmonyOS/HarmonyOS | https://codechina.csdn.net/mirrors/Awesome-HarmonyOS/HarmonyOS |
| mirrors/bilibili/ijkplayer | https://codechina.csdn.net/mirrors/bilibili/ijkplayer |
| mirrors/harisiqbal88/plotneuralnet | https://codechina.csdn.net/mirrors/harisiqbal88/plotneuralnet |
| mirrors/jessyancoding/androidautosize | https://codechina.csdn.net/mirrors/jessyancoding/androidautosize |
| mirrors/youngyangyang04/leetcode-master | https://codechina.csdn.net/mirrors/youngyangyang04/leetcode-master |
| mirrors/zhihu/matisse | https://codechina.csdn.net/mirrors/zhihu/matisse |
| EricLee/classification | https://codechina.csdn.net/EricLee/classification |
| csdn_codechina/it_information | https://codechina.csdn.net/csdn_codechina/it_information |
| mirrors/bubbliiiing/unet-pytorch | https://codechina.csdn.net/mirrors/bubbliiiing/unet-pytorch |
| mirrors/detectrecog/ccpd | https://codechina.csdn.net/mirrors/detectrecog/ccpd |
| mirrors/fo40225/tensorflow-windows-wheel | https://codechina.csdn.net/mirrors/fo40225/tensorflow-windows-wheel |
| mirrors/lin-xin/vue-manage-system | https://codechina.csdn.net/mirrors/lin-xin/vue-manage-system |
| mirrors/littlecurl/appprojects | https://codechina.csdn.net/mirrors/littlecurl/appprojects |
| mirrors/nothings/stb | https://codechina.csdn.net/mirrors/nothings/stb |
| mirrors/whyalwaysmea/spring-security | https://codechina.csdn.net/mirrors/whyalwaysmea/spring-security |
| mirrors/zxhjames/phpstudentmanagesystem | https://codechina.csdn.net/mirrors/zxhjames/phpstudentmanagesystem |
| zorrosoft/pluginok | https://codechina.csdn.net/zorrosoft/pluginok |
\ No newline at end of file
Markdown is supported
0% 或 .
You are about to add 0 people to the discussion. Proceed with caution.
先完成此消息的编辑！
想要评论请 注册 或