# OpenPose採坑紀錄
###### tags: `其他`
官方github:https://github.com/CMU-Perceptual-Computing-Lab/openpose
使用30系顯示卡仍存在cuDNN問題[（參考踩坑记录）](https://blog.csdn.net/weixin_41693877/article/details/123272174?spm=1001.2101.3001.6661.1&utm_medium=distribute.pc_relevant_t0.none-task-blog-2%7Edefault%7ECTRLIST%7Edefault-1-123272174-blog-106665336.pc_relevant_sortByAnswer&depth_1-utm_source=distribute.pc_relevant_t0.none-task-blog-2%7Edefault%7ECTRLIST%7Edefault-1-123272174-blog-106665336.pc_relevant_sortByAnswer&utm_relevant_index=1)與[Ubuntu20.04运行Openpose出现已放弃(核心已转储)的原因和解决方案](https://www.zhihu.com/question/475846925)
## 系統配置
* 顯示卡:Nvidia FE RTX3070
* os: Ubuntu 20.04 LTS
* Anaconda installed(懶得解安裝)
* CUDA version: 11.7
* cudnn-linux-x86_64-8.4.1.50_cuda11.6-archive.tar.xz
* cuDNN:8.4.1.50
* protobuf:2.6.1
* cmake:3.12.2
* gcc/g++:5.3.1

詳細安裝教學跳轉至<a href="#instalation">安裝</a>

## 安裝 <div id="instalation"></div>
1. 安裝Nvidia driver/CUDA/cuDNN
參考其他網站安裝教學：
* [Ubuntu 安裝 CUDA cuDNN pytorch tensorflow mxnet](https://mikethreeacer.medium.com/ubuntu-18-04-%E5%AE%89%E8%A3%9D-cuda-cudnn-anaconda-pytorch-1f170b3326a4)
* [深度学习三件套:Ubuntu 20.04 安装 NVIDIA 驱动/CUDA/cuDNN全流程](https://cyfeng.science/2020/05/02/ubuntu-install-nvidia-driver-cuda-cudnn-suits/)

2. CUDA：除參考以上連結外，其中安裝cuda時請根據[官方cuda toolkit對應系統版本](https://developer.nvidia.com/cuda-downloads?target_os=Linux&target_arch=x86_64&Distribution=Ubuntu&target_version=20.04&target_type=runfile_local)下載；若遇到wget下載大檔案卡死問題請參考[這篇](https://blog.csdn.net/Carina_Cao/article/details/99647165)

3. cuDNN：NVIDIA針對DNN的加速lib
* 主要參考[此篇](https://mikethreeacer.medium.com/ubuntu-18-04-%E5%AE%89%E8%A3%9D-cuda-cudnn-anaconda-pytorch-1f170b3326a4) & [安裝參考](https://mikethreeacer.medium.com/ubuntu-18-04-%E5%AE%89%E8%A3%9D-cuda-cudnn-anaconda-pytorch-1f170b3326a4)
* [通过软链接同时维护和使用多个版本的CUDA库](https://cyfeng.science/2020/05/02/ubuntu-install-nvidia-driver-cuda-cudnn-suits/) 因此檔案應該cp到usr/../cuda-11.x 這部份可能要在研究差別

4. 安裝openpose
* <span style="color:red;">主要參考</span>[Ubuntu20.04编译安装openpose使用pythonAPI](https://blog.csdn.net/weixin_44379605/article/details/121980761)
* protobuf版本過新會造成cmake設置完make all時錯誤，先參考[protobuf 深坑 - 版本冲突 - 彻底解决](https://my.oschina.net/u/4298883/blog/4272253)這篇**完整解安裝protobuf**
* 問題
* python 使用protobuf出错：protoc: error while loading shared libraries: libprotoc.so.9: cannot open shared object file:No such...

* 解决方法：linux 敲击命令：export LD_LIBRARY_PATH=/usr/local/lib
* ref：[用protobuf编译时报错：protoc: error while loading shared libraries: libprotoc.so.9: cannot open shared obje...](https://blog.csdn.net/qq_27563511/article/details/80696403)
* 安装protobuf 2.6.1
* https://blog.csdn.net/lwplwf/article/details/76532804
* cmake-gui成功cofiguration且generate後，在buildy資料夾執行
* ```linux=
make all

    * 編譯成功時顯示下圖，可以開始使用
        * ![](https://i.imgur.com/QI4F5HD.png)
## Openpose操作
[ubuntu下OpenPose的安裝、使用、初步介紹](https://www.twblogs.net/a/5c0a8d4fbd9eee6fb37bc32f)
[參數說明](https://blog.csdn.net/zziahgf/article/details/84668319)

顯示卡內存過小，執行時加入

--net_resolution 320x176

* ref:https://www.codeleading.com/article/48545516446/

* webCam
    * 若加入--face --hand參數，顯存不足(3070)
    * ref:
        * [TX2入门教程软件篇-安装openpose(jetpack3.3)](https://www.ncnynl.com/archives/201910/3380.html )
        * [ubuntu下OpenPose的安裝、使用、初步介紹](https://www.twblogs.net/a/5c0a8d4fbd9eee6fb37bc32f) 
        * [[Windows]OpenPose安装问题小结](https://blog.csdn.net/jd_1230/article/details/104096159)
        * openCV[支援webcam解析度](https://coder.gay/1922383/python-opencv-access-webcam-maximum-resolution)：
            * 160.0 x 120.0
            * 176.0 x 144.0
            * 320.0 x 240.0
            * 352.0 x 288.0
            * 640.0 x 480.0
            * 1024.0 x 768.0
            * 1280.0 x 1024.0
        * [openpose demo运行报错Check failed: error == cudaSuccess (2 vs. 0) out of memory](https://blog.csdn.net/AshleyXM/article/details/115863739)
        * [Ubuntu20.04 RTX3060 安装openpose CUDA11.5.2 CUDNN](https://blog.csdn.net/weixin_43454032/article/details/125611686?spm=1001.2101.3001.6650.1&utm_medium=distribute.pc_relevant.none-task-blog-2%7Edefault%7ECTRLIST%7Edefault-1-125611686-blog-119780910.pc_relevant_aa&depth_1-utm_source=distribute.pc_relevant.none-task-blog-2%7Edefault%7ECTRLIST%7Edefault-1-125611686-blog-119780910.pc_relevant_aa&utm_relevant_index=1)

./build/examples/openpose/openpose.bin -camera_resolution 320x240 -net_resolution 128x96

## Openpose python API
先到build/python中執行：

sudo make install

使用過多python版本時，改用**Cmake設定之python path位置**執行：

/usr/bin/python3.x 01_body_from_image.py

## 模型研究
[欸你在幹嘛?跟一眼瞬間就懂你的行為辨識技術初次見面 ](https://ithelp.ithome.com.tw/users/20129150/ironman/2969?page=3)