python onnx 快捷安装 onnxruntime 的 gpu 版本如何使用

比眉伴天荒 2022-09-12 07:00 2763阅读 0赞

# 快捷安装cpu版本的onnx #

这样子就能装上cpu版本的：

pip install scipy onnx-simplifier

onnx-simplifier项目地址：  
[https://github.com/daquexian/onnx-simplifier][https_github.com_daquexian_onnx-simplifier]

setup.py中写的：  
![在这里插入图片描述][89440039a1984007a87f423029eb1b0e.png]  
所以正常来说需要这么安装：

pip install onnx onnxoptimizer onnxruntime-gpu protobuf

# 如何使用onnxruntime 的 gpu版本 #

参考：[https://stackoverflow.com/questions/64452013/how-do-you-run-a-onnx-model-on-a-gpu][https_stackoverflow.com_questions_64452013_how-do-you-run-a-onnx-model-on-a-gpu]

Step 1: uninstall your current onnxruntime

pip uninstall onnxruntime

Step 2: install GPU version of onnxruntime environment

pip install onnxruntime-gpu

Step 3: Verify the device support for onnxruntime environment

>> import onnxruntime as rt
    >> rt.get_device()
    'GPU'

Step 4: If you encounter any issue please check with your cuda and CuDNN versions, that must be compatible to each other. Please refer this link here to understand about the version compatibility between cuda and CuDNN.

rt.get\_device()获取到的是安装的onnxruntime版本，但实际使用中需要注意兼容性，这才是重点。如果onnxruntime版本不兼容cuda，那还是会默认使用CPU版本。

# GPU #

代码中加断点，就可以看到自己当前的 onnxruntime session使用的是什么：

>>detector.session.get_providers()
    ['CPUExecutionProvider']

这里是官网的说明：

[https://onnxruntime.ai/docs/execution-providers/CUDA-ExecutionProvider.html][https_onnxruntime.ai_docs_execution-providers_CUDA-ExecutionProvider.html]

可知：

<table> 
 <thead> 
  <tr> 
   <th>ONNX Runtime</th> 
   <th>CUDA</th> 
   <th>cuDNN</th> 
   <th>Notes</th> 
  </tr> 
 </thead> 
 <tbody> 
  <tr> 
   <td>1.8</td> 
   <td>11.0.3</td> 
   <td>8.0.4 (Linux)<br>8.0.2.39 (Windows)</td> 
   <td>libcudart 11.0.221<br>libcufft 10.2.1.245<br>libcurand 10.2.1.245<br>libcublasLt 11.2.0.252<br>libcublas 11.2.0.252<br>libcudnn 8.0.4<br>libcupti.so 2020.1.1</td> 
  </tr> 
  <tr> 
   <td>1.7</td> 
   <td>11.0.3</td> 
   <td>8.0.4 (Linux)<br>8.0.2.39 (Windows)</td> 
   <td>libcudart 11.0.221<br>libcufft 10.2.1.245<br>libcurand 10.2.1.245<br>libcublasLt 11.2.0.252<br>libcublas 11.2.0.252<br>libcudnn 8.0.4</td> 
  </tr> 
  <tr> 
   <td>1.5-1.6</td> 
   <td>10.2</td> 
   <td>8.0.3</td> 
   <td>CUDA 11 can be built from source</td> 
  </tr> 
  <tr> 
   <td>1.2-1.4</td> 
   <td>10.1</td> 
   <td>7.6.5</td> 
   <td>Requires cublas10-10.2.1.243; cublas 10.1.x will not work</td> 
  </tr> 
  <tr> 
   <td>1.0-1.1</td> 
   <td>10.0</td> 
   <td>7.6.4</td> 
   <td>CUDA versions from 9.1 up to 10.1, and cuDNN versions from 7.1 up to 7.4 should also work with Visual Studio 2017</td> 
  </tr> 
 </tbody> 
</table>

所以我执行了：

pip uninstall onnxruntime-gpu
    pip install onnxruntime-gpu==1.4

所以我再次查看：

>>detector.session.get_providers()
    ['CUDAExecutionProvider', 'CPUExecutionProvider']

![在这里插入图片描述][watermark_type_ZHJvaWRzYW5zZmFsbGJhY2s_shadow_50_text_Q1NETiBAUTc0Mjk3MTYzNg_size_20_color_FFFFFF_t_70_g_se_x_16]

# 2080Ti CUDA10.1 的安装方式 #

这样就完美了：

pip install scipy  onnx onnxoptimizer onnxruntime-gpu==1.4 protobuf

[https_github.com_daquexian_onnx-simplifier]: https://github.com/daquexian/onnx-simplifier
[89440039a1984007a87f423029eb1b0e.png]: /images/20220828/60a6dbae86c146c5911e11195b5326b5.png
[https_stackoverflow.com_questions_64452013_how-do-you-run-a-onnx-model-on-a-gpu]: https://stackoverflow.com/questions/64452013/how-do-you-run-a-onnx-model-on-a-gpu
[https_onnxruntime.ai_docs_execution-providers_CUDA-ExecutionProvider.html]: https://onnxruntime.ai/docs/execution-providers/CUDA-ExecutionProvider.html
[watermark_type_ZHJvaWRzYW5zZmFsbGJhY2s_shadow_50_text_Q1NETiBAUTc0Mjk3MTYzNg_size_20_color_FFFFFF_t_70_g_se_x_16]: /images/20220828/28d38ff2c8904b4d90b67bda29784c81.png