【环境搭建】onnxruntime

1，介绍
    onnxruntime是一个用于onnx模型推理的引擎。

2，安装
2.1 cuda,cudnn
2.2 cmake，版本>=3.13.0

sudo apt-get install libssl-dev
sudo apt-get autoremove cmake # 卸载
wget https://cmake.org/files/v3.23/cmake-3.23.4.tar.gz
tar -xf cmake-3.23.4
cd cmake-3.23.4
./bootstrap --prefix=/usr
make -j 8
sudo make install

cmake -version
cmake version 3.23.4
CMake suite maintained and supported by Kitware (kitware.com/cmake).

2.3 tensorrt
2.4 onnxruntime

conda activate py36 # 切换虚拟环境
git clone https://github.com/microsoft/onnxruntime.git
cd onnxruntime
git submodule sync
git submodule update --init --recursive 
./build.sh \
	--use_cuda \
	--cuda_version=11.0 \
	--cuda_home=/usr/local/cuda \
	--cudnn_home=/usr/local/cuda \
	--use_tensorrt --tensorrt_home=$TENSORRT_ROOT \
	--build_shared_lib --enable_pybind \
	--build_wheel --update --build
pip build/Linux/Debug/dist/onnxruntime_gpu_tensorrt-1.6.0-cp36-cp36m-linux_x86_64.whl

# 获取指定版本
git clone -b v1.6.0 https://github.com/microsoft/onnxruntime.git
cd onnxruntime
git checkout -b v1.6.0
git submodule sync
git submodule update --init --recursive 

3,pip 安装

pip install onnxruntime-gpu==1.6.0 onnx==1.9.0 onnxconverter_common==1.6.0 # cuda 10.2
pip install onnxruntime-gpu==1.8.1 onnx==1.9.0 onnxconverter_common==1.8.1 # cuda 11.0

import onnxruntime
onnxruntime.get_available_providers()
sess = onnxruntime.InferenceSession(onnx_path)
sess.set_providers(['CUDAExecutionProvider'], [ {'device_id': 0}])
result = sess.run([output_name], {input_name:x})  # x is input

3，onnxruntime and pytorch inference

# -*- coding: utf-8 -*-
import torch
from torchvision import models
import onnxruntime
import numpy as np

model = models.resnet18(pretrained=True)
model.eval().cuda()

## pytorch
x = torch.rand(2,3,224,224).cuda()
out_pt = model(x)
print(out_pt.size())

# onnx
onnx_path = "resnet18.onnx"
dynamic_axes = {'input': {0: 'batch_size'}}
torch.onnx.export(model,
                  x,
                  onnx_path,
                  export_params=True,
                  opset_version=11,
                  do_constant_folding=True,
                  input_names=['input'],
                  dynamic_axes=dynamic_axes)

sess = onnxruntime.InferenceSession(onnx_path)
sess.set_providers(['CUDAExecutionProvider'], [ {'device_id': 0}])
input_name = sess.get_inputs()[0].name
output_name = sess.get_outputs()[0].name

# onnxruntime inference
xx = x.cpu().numpy().astype(np.float32)
result = sess.run([output_name], {input_name:xx})  # x is input
print(result[0].shape)

# MSE
mse = np.mean((out_pt.data.cpu().numpy() - result[0]) ** 2)
print(mse)

输出结果

torch.Size([2, 1000])
(2, 1000)
9.275762e-13