ONNX Runtime：由微软推出，用于优化和加速机器学习推理和训练，适用于ONNX模型，是一个跨平台推理和训练机器学习加速器(ONNX Runtime is a cross-platform inference and training machine-learning accelerator)，源码地址：https://github.com/microsoft/onnxruntime，最新发布版本为v1.11.1，License为MIT：

      1.ONNX Runtime Inferencing：高性能推理引擎

      (1).可在不同的操作系统上运行，包括Windows、Linux、Mac、Android、iOS等；

      (2).可利用硬件增加性能，包括CUDA、TensorRT、DirectML、OpenVINO等；

      (3).支持PyTorch、TensorFlow等深度学习框架的模型，需先调用相应接口转换为ONNX模型；

      (4).在Python中训练，确可部署到C++/Java等应用程序中。

      2.ONNX Runtime Training：于2021年4月发布，可加快PyTorch对模型训练，可通过CUDA加速，目前多用于Linux平台。

      通过conda命令安装执行：

conda install -c conda-forge onnxruntime

      以下为测试代码：通过ResNet-50对图像进行分类

import numpy as np
import onnxruntime
import onnx
from onnx import numpy_helper
import urllib.request
import os
import tarfile
import json
import cv2

# reference: https://github.com/onnx/onnx-docker/blob/master/onnx-ecosystem/inference_demos/resnet50_modelzoo_onnxruntime_inference.ipynb
def download_onnx_model():
    labels_file_name = "imagenet-simple-labels.json"
    model_tar_name = "resnet50v2.tar.gz"
    model_directory_name = "resnet50v2"

    if os.path.exists(model_tar_name) and os.path.exists(labels_file_name):
        print("files exist, don't need to download")
    else:
        print("files don't exist, need to download ...")

        onnx_model_url = "https://s3.amazonaws.com/onnx-model-zoo/resnet/resnet50v2/resnet50v2.tar.gz"
        imagenet_labels_url = "https://raw.githubusercontent.com/anishathalye/imagenet-simple-labels/master/imagenet-simple-labels.json"

        # retrieve our model from the ONNX model zoo
        urllib.request.urlretrieve(onnx_model_url, filename=model_tar_name)
        urllib.request.urlretrieve(imagenet_labels_url, filename=labels_file_name)

        print("download completed, start decompress ...")
        file = tarfile.open(model_tar_name)
        file.extractall("./")
        file.close()

    return model_directory_name, labels_file_name

def load_labels(path):
    with open(path) as f:
        data = json.load(f)
    return np.asarray(data)

def images_preprocess(images_path, images_name):
    input_data = []

    for name in images_name:
        img = cv2.imread(images_path + name)
        img = cv2.resize(img, (224, 224))
        img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)
        
        data = np.array(img).transpose(2, 0, 1)
        #print(f"name: {name}, opencv image shape(h,w,c): {img.shape}, transpose shape(c,h,w): {data.shape}")
        # convert the input data into the float32 input
        data = data.astype('float32')

        # normalize
        mean_vec = np.array([0.485, 0.456, 0.406])
        stddev_vec = np.array([0.229, 0.224, 0.225])
        norm_data = np.zeros(data.shape).astype('float32')
        for i in range(data.shape[0]):
            norm_data[i,:,:] = (data[i,:,:]/255 - mean_vec[i]) / stddev_vec[i]

        # add batch channel
        norm_data = norm_data.reshape(1, 3, 224, 224).astype('float32')
        input_data.append(norm_data)

    return input_data

def softmax(x):
    x = x.reshape(-1)
    e_x = np.exp(x - np.max(x))
    return e_x / e_x.sum(axis=0)

def postprocess(result):
    return softmax(np.array(result)).tolist()

def inference(onnx_model, labels, input_data, images_name, images_label):
    session = onnxruntime.InferenceSession(onnx_model, None)
    # get the name of the first input of the model
    input_name = session.get_inputs()[0].name
    count = 0
    for data in input_data:
        print(f"{count+1}. image name: {images_name[count]}, actual value: {images_label[count]}")
        count += 1

        raw_result = session.run([], {input_name: data})

        res = postprocess(raw_result)

        idx = np.argmax(res)
        print(f"  result: idx: {idx}, label: {labels[idx]}, percentage: {round(res[idx]*100, 4)}%")

        sort_idx = np.flip(np.squeeze(np.argsort(res)))
        print("  top 5 labels are:", labels[sort_idx[:5]])

def main():
    model_directory_name, labels_file_name = download_onnx_model()

    labels = load_labels(labels_file_name)
    print("the number of categories is:", len(labels)) # 1000

    images_path = "../../data/image/"
    images_name = ["5.jpg", "6.jpg", "7.jpg", "8.jpg", "9.jpg", "10.jpg"]
    images_label = ["goldfish", "hen", "ostrich", "crocodile", "goose", "sheep"]
    if len(images_name) != len(images_label):
        print("Error: images count and labes'length don't match")
        return

    input_data = images_preprocess(images_path, images_name)

    onnx_model = model_directory_name + "/resnet50v2.onnx"
    inference(onnx_model, labels, input_data, images_name, images_label)

    print("test finish")

if __name__ == "__main__":
    main()

      测试图像如下所示：

      执行结果如下所示：

      GitHub： https://github.com/fengbingchun/PyTorch_Test