潇湘竹 發表於 2019-7-3 22:20:00

(原)Ubuntu安装TensorRT

<p>转载请注明出处:</p>
<p>https://www.cnblogs.com/darkknightzh/p/11129472.html</p>
<p>参考网址:</p>
<p>https://docs.nvidia.com/deeplearning/sdk/tensorrt-install-guide/index.html</p>
<p>https://arleyzhang.github.io/articles/7f4b25ce/</p>
<p>https://docs.nvidia.com/deeplearning/sdk/cudnn-install/index.html</p>
<h1><span style="font-size: 18px"><strong>1 说明</strong></span></h1>
<p>1.1 直接使用deb安装的。不过不记得之前cuda驱动是如何安装的了。网上说cuda驱动和TensorRTF都是deb安装的才行。。。</p>
<p>1.2 安装环境是ubuntu 16.04.1+anaconda(python3.6.8)+cuda9.0+cudnn7.5</p>
<h1><span style="font-size: 18px"><strong>2 安装步骤</strong></span></h1>
<p>2.1 sudo dpkg -i nv-tensorrt-repo-ubuntu1604-cuda9.0-trt5.1.5.0-ga-20190427_1-1_amd64.deb</p>
<p>2.2 sudo apt-key -add /var/nv-tensorrt-repo-cuda9.0-trt5.1.5.0-ga-20190427/7fa2af80.pub</p>
<p><img src="https://img2018.cnblogs.com/blog/682463/201907/682463-20190703221421508-2113183792.png" alt=""></p>
<p>2.3 sudo apt-get update</p>
<p>2.4 sudo apt-get install tensorrt</p>
<p>至此,TensorRTF安装完成。</p>
<p>2.5 通过下面命令检查一下是否安装成功:</p>
<div class="cnblogs_code">
<pre>dpkg -l | <span style="color: rgba(0, 0, 255, 1)">grep</span> TensorRT</pre>
</div>
<p><img src="https://img2018.cnblogs.com/blog/682463/201907/682463-20190703221440856-1267528097.png" alt=""></p>
<p>2.6 通过下面命令安装(不知道是啥,反正安装就对了):</p>
<div class="cnblogs_code">
<pre><span style="color: rgba(0, 0, 255, 1)">sudo</span> apt-get <span style="color: rgba(0, 0, 255, 1)">install</span> python3-libnvinfer-dev</pre>
</div>
<p><img src="https://img2018.cnblogs.com/blog/682463/201907/682463-20190703221452317-2079461425.png" alt=""></p>
<p>2.7 到上面,python里面无法import tensorrt。直接使用ubuntu自带的archive manager打开nv-tensorrt-repo-ubuntu1604-cuda9.0-trt5.1.5.0-ga-20190427_1-1_amd64.deb文件,可以看到里面有python3.6的XXX。通过下面命令安装该whl文件。</p>
<div class="cnblogs_code">
<pre>pip <span style="color: rgba(0, 0, 255, 1)">install</span> tensorrt-<span style="color: rgba(128, 0, 128, 1)">5.1</span>.<span style="color: rgba(128, 0, 128, 1)">5.0</span>-cp36-none-linux_x86_64.whl</pre>
</div>
<p><img src="https://img2018.cnblogs.com/blog/682463/201907/682463-20190703221509892-98220158.png" alt=""></p>
<p><img src="https://img2018.cnblogs.com/blog/682463/201907/682463-20190703221525904-1028622323.png" alt=""></p>
<p>2.8 验证python能否导入tensorrt(此处成功):</p>
<p><img src="https://img2018.cnblogs.com/blog/682463/201907/682463-20190703221545199-1088529574.png" alt=""></p>
<p>2.9 安装PyCUDA。</p>
<p>PyCUDA允许python直接访问CUDA的API。</p>
<p>首先保证nvcc在PATH路径中。而后输入:</p>
<div class="cnblogs_code">
<pre>pip <span style="color: rgba(0, 0, 255, 1)">install</span> <span style="color: rgba(128, 0, 0, 1)">'</span><span style="color: rgba(128, 0, 0, 1)">pycuda&gt;=2017.1.1</span><span style="color: rgba(128, 0, 0, 1)">'</span></pre>
</div>
<p>2.10 验证程序能否运行(见下面网址)</p>
<p>https://arleyzhang.github.io/articles/7f4b25ce/</p>
<p>可以把 tensorrt 文件夹拷贝到用户目录下,方便自己修改测试例程中的代码。</p>
<p>进入 samples 文件夹直接 make,会在 bin 目录中生成可执行文件,可以一一进行测试学习。</p>
<p>运行了sample_mnist,结果如下:</p>
<p><img src="https://img2018.cnblogs.com/blog/682463/201907/682463-20190703221617426-2060555616.png" alt=""></p>
<h1><span style="font-size: 18px"><strong>3 卸载</strong></span></h1>
<p>具体见网址:</p>
<p>https://docs.nvidia.com/deeplearning/sdk/tensorrt-install-guide/index.html#uninstalling</p>
<div class="cnblogs_code">
<pre><span style="color: rgba(0, 0, 255, 1)">sudo</span> apt-get purge <span style="color: rgba(128, 0, 0, 1)">"</span><span style="color: rgba(128, 0, 0, 1)">libnvinfer*</span><span style="color: rgba(128, 0, 0, 1)">"</span><span style="color: rgba(0, 0, 0, 1)">
pip uninstall tensorrt</span></pre>
</div>
<h1><span style="font-size: 18px"><strong>4 其他</strong></span></h1>
<p>4.1 cudnn使用deb安装:</p>
<p>说明:不建议使用这种方式安装。</p>
<p>Navigate to your &lt;cudnnpath&gt; directory containing cuDNN Debian file.</p>
<p>Install the runtime library, for example:</p>
<div class="cnblogs_code">
<pre><span style="color: rgba(0, 0, 255, 1)">sudo</span> dpkg -i libcudnn7_7.<span style="color: rgba(128, 0, 128, 1)">0.3</span>.<span style="color: rgba(128, 0, 128, 1)">11</span>-<span style="color: rgba(128, 0, 128, 1)">1</span>+cuda9.0_amd64.deb</pre>
</div>
<p>Install the developer library, for example:</p>
<div class="cnblogs_code">
<pre><span style="color: rgba(0, 0, 255, 1)">sudo</span> dpkg -i libcudnn7-devel_7.<span style="color: rgba(128, 0, 128, 1)">0.3</span>.<span style="color: rgba(128, 0, 128, 1)">11</span>-<span style="color: rgba(128, 0, 128, 1)">1</span>+cuda9.0_amd64.deb</pre>
</div>
<p>Install the code samples and the cuDNN Library User Guide, for example:</p>
<div class="cnblogs_code">
<pre><span style="color: rgba(0, 0, 255, 1)">sudo</span> dpkg -i libcudnn7-doc_7.<span style="color: rgba(128, 0, 128, 1)">0.3</span>.<span style="color: rgba(128, 0, 128, 1)">11</span>-<span style="color: rgba(128, 0, 128, 1)">1</span>+cuda9.0_amd64.deb</pre>
</div>
<p>4.2 使用tar文件安装&nbsp;</p>
<p>Navigate to your &lt;cudnnpath&gt; directory containing the cuDNN Tar file.</p>
<p>Unzip the cuDNN package.</p>
<div class="cnblogs_code">
<pre><span style="color: rgba(0, 0, 255, 1)">tar</span> -xzvf cudnn-<span style="color: rgba(128, 0, 128, 1)">9.0</span>-linux-x64-v7.tgz</pre>
</div>
<p>Copy the following files into the CUDA Toolkit directory, and change the file permissions.</p>
<div class="cnblogs_code">
<pre><span style="color: rgba(0, 0, 255, 1)">sudo</span> <span style="color: rgba(0, 0, 255, 1)">cp</span> cuda/include/cudnn.h /usr/local/cuda/<span style="color: rgba(0, 0, 0, 1)">include
</span><span style="color: rgba(0, 0, 255, 1)">sudo</span> <span style="color: rgba(0, 0, 255, 1)">cp</span> cuda/lib64/libcudnn* /usr/local/cuda/<span style="color: rgba(0, 0, 0, 1)">lib64
</span><span style="color: rgba(0, 0, 255, 1)">sudo</span> <span style="color: rgba(0, 0, 255, 1)">chmod</span> a+r /usr/local/cuda/include/cudnn.h /usr/local/cuda/lib64/libcudnn*</pre>
</div>
<p>4.3 测试cudnn是否安装成功</p>
<p>1)/usr/src/cudnn_samples_v7放到home文件夹下。</p>
<p>2)cd mnistCUDNN/</p>
<p>3)make</p>
<p><img src="https://img2018.cnblogs.com/blog/682463/201907/682463-20190703221852003-1181697583.png" alt=""></p>
<p>4)./mnistCUDNN</p>
<p><img src="https://img2018.cnblogs.com/blog/682463/201907/682463-20190703221909650-1874400492.png" alt=""></p>
<p>出错了。。。</p>
<p>原因是之前/usr/local/cuda中cudnn用的是7.4.2版本的。替换为7.5之后,运行成功。(实际上在/usr/lib/x86_64-linux-gnu中有刚刚使用deb装上的7.5版本的cudnn,但是程序未能找到。不知道目前有两套相同版本的cudnn,以后会不会出问题吧。。。所以不建议使用deb安装,位置不可控)</p>
<p><img src="https://img2018.cnblogs.com/blog/682463/201907/682463-20190703221933390-897046453.png" alt=""></p><br><br>
来源:https://www.cnblogs.com/darkknightzh/p/11129472.html
頁: [1]
查看完整版本: (原)Ubuntu安装TensorRT