寻找最新模型并转为TensorRT模型进行并行加速的方法

1.寻找最新开源的ONNX模型

https://github.com/PINTO0309/PINTO_model_zoo/tree/main
这个大佬把2023年新开源的模型都转成ONNX了，很好用。

2.安装ONNX2TensorRT

git clone --recurse-submodules https://github.com/onnx/onnx-tensorrt.git

# Build from source
cd onnx-tensorrt
mkdir build && cd build
cmake ..
make -j$(nproc)
sudo make install

(1).注意适配版本号

注意这里有个坑，git clone默认下载的是最新版的，onnx-tensorrt，有可能和你的tensorrt版本不合。
解决方法是，我的tensorrt版本为8.4,然后就在这个网址里找打onnx2trt的8.4 版本下载即可

https://github.com/onnx/onnx-tensorrt/releases

 dpkg -l | grep TensorRT

这个是查看TensorRT版本号的代码
一般出现这样的错误就是tensorrt版本不合

/home/zhangqi/Documents/Library/onnx-tensorrt/TensorOrWeights.hpp:115:38: error: ‘kUINT8’ is not a member of ‘nvinfer1::DataType’; did you mean ‘kINT8’?
  115 |             case nvinfer1::DataType::kUINT8: return "UINT8";
      |                                      ^~~~~~
      |                                      kINT8
/home/zhangqi/Documents/Library/onnx-tensorrt/TensorOrWeights.hpp:118:38: error: ‘kFP8’ is not a member of ‘nvinfer1::DataType’
  118 |             case nvinfer1::DataType::kFP8: return "FP8";
      |                                      ^~~~
In file included from /home/zhangqi/Documents/Library/onnx-tensorrt/trt_utils.hpp:8,
                 from /home/zhangqi/Documents/Library/onnx-tensorrt/onnx2trt_utils.hpp:10,
                 from /home/zhangqi/Documents/Library/onnx-tensorrt/onnx2trt_utils.cpp:5:
/home/zhangqi/Documents/Library/onnx-tensorrt/TensorOrWeights.hpp: In member function ‘std::string onnx2trt::TensorOrWeights::getType() const’:
/home/zhangqi/Documents/Library/onnx-tensorrt/TensorOrWeights.hpp:115:38: error: ‘kUINT8’ is not a member of ‘nvinfer1::DataType’; did you mean ‘kINT8’?
  115 |             case nvinfer1::DataType::kUINT8: return "UINT8";

(2).注意适配合适的cuda

fatal error: crt/host_defines.h: No such file or directory
  147 | #include "crt/host_defines.h"
      |          ^~~~~~~~~~~~~~~~~~~~

出现这种状况，就直接搜索全计算机有没有这个文件，我的这个文件在cuda11.1文件夹里而不是默认的cuda文件夹。于是我把onnx2trt里面CmakeList里的

if (NOT CUDA_TOOLKIT_ROOT_DIR)
  set(CUDA_TOOLKIT_ROOT_DIR /usr/local/cuda)

改为

if (NOT CUDA_TOOLKIT_ROOT_DIR)
  set(CUDA_TOOLKIT_ROOT_DIR /usr/local/cuda-11.1)

3.转换开始！！！

cd build
/onnx2trt ../decoder.onnx -o decoder.trt