Openvino2022 dev tools installation and use

Click on the blue words above to follow us

WeChat official account ：OpenCV School Focus on getting more computer vision and deep learning knowledge

introduction

OpenVINO2022 After the release , Make a clear division of the functions of previous versions , Which still passes exe The way to install the program is runtime Inference package , Support ONNX、IR、PADDLE Wait for model reading and reasoning . But model optimization transformation 、 Other functions, such as the model library download function, are divided into one called Dev Tool Part of . This part can be passed through pip Way to install directly , Then direct the line through the command line , Complete the model transformation , Download and other operations , Compared with previous versions, the usability has been greatly improved ！ Make a comparison as follows ：

Dev Tools Installation and use

Dev Tools Easy to install , You can choose to install directly through the command line of the official script , The only thing to note is the choice of model framework support , I chose ONNX/Pytorch Format conversion support , The installed command line is as follows ：

pip install openvino-dev[onnx,pytorch]==2022.1.0

Copy

After executing this command line, the installation is completed , As long as the network is not connected, the installation can be successful ！

installation is complete , Change one ONNX The format model is IR Format (xml/bin) file , With Pytorch Of ResNet18 For example , First convert to ONNX, The code is as follows ：

model = models.resnet18(pretrained=True) model.eval() model.cpu()
 dummy_input1 = torch.randn(1, 3, 224, 224) torch.onnx.export(model, (dummy_input1), "resnet_model.onnx", verbose=True)

Copy

Then run the command line directly to convert IR Format file , The screenshot is as follows ：

Model download

installation is complete Dev Tools after , Download model , Just execute the command line ： Examples are as follows ：

omz_downloader --name person-detection-0200

Copy

Means downloading the model  person-detection-0200  It is a lightweight face detection model .

omz_downloader  Supported parameters ：

--all Means to download all models , Don't do this ！--name  Download one or more models with the specified names , Recommended ！--precisions  Represents the downloaded model accuracy parameters , Support FP32/FP18/INT8

Copy

Python SDK Use

Before comparison Python Version of SDK, It works a lot , The most obvious feeling is that there is no need to read input and output , Then a bunch of settings , For a single I / O network , The call is especially simple and convenient ！ It's easier for developers to get started ！ I have successfully converted one above pytorch The image classification model is IR Format , You can use it now , be based on OpenVINO2022 The latest version Python SDK Deployment call , The implementation code is as follows :

#  Load tag data 
with open('imagenet_classes.txt') as f:
    labels = [line.strip() for line in f.readlines()]


def resnet_demo():
    ie = Core()
    # model = ie.read_model(model="resnet_model.onnx")
    model = ie.read_model(model="resnet_model.xml")
    compiled_model = ie.compile_model(model=model, device_name="CPU")

    output_layer = compiled_model.output(0)

    means = np.zeros((224, 224, 3), dtype=np.float32)
    means[: ,:] = (0.485, 0.456, 0.406)
    dev = np.zeros((224, 224, 3), dtype=np.float32)
    dev[: ,:] = (0.229, 0.224, 0.225)

    image = cv.imread("D:/images/space_shuttle.jpg")
    rgb = cv.cvtColor(image, code=cv.COLOR_BGR2RGB)

    # resize to MobileNet image shape
    input_image = cv.resize(src=rgb, dsize=(224, 224))
    blob_img = np.float32(input_image) / 255.0
    input_x = (blob_img - means) / dev
    input_x = input_x.transpose((2, 0, 1))
    input_x = np.expand_dims(input_x, 0)
    print(input_x.shape)

    result_infer = compiled_model([input_x])[output_layer]
    result_index = np.argmax(result_infer)
    cv.putText(image, labels[result_index], (20, 50), cv.FONT_HERSHEY_SIMPLEX, 1.0, (0, 0, 255), 2)

    cv.imshow("OpenVINO2022 + Pythorch ResNet18", image)
    cv.waitKey(0)
    cv.destroyAllWindows()


if __name__ == "__main__":
    resnet_demo()

Copy

Scan the code to check OpenCV+OpenVIO+Pytorch Systematic learning roadmap

  Recommended reading  

CV The full stack developer said - From traditional algorithm to deep learning, how to practice

2022 Pit depth learning , I choose Pytorch frame ！

Pytorch Easily achieve classic visual tasks

Tutorial recommendation | Pytorch frame CV Development - From introduction to practice

OpenCV4 C++ Study Essential basic grammar knowledge 3

OpenCV4 C++ Study Essential basic grammar knowledge 2

OpenCV4.5.4 Face detection + Five o'clock landmark New functional testing

OpenCV4.5.4 Detailed explanation of face recognition and code demonstration

OpenCV Of binary image analysis Blob Analyze and find the circle

OpenCV4.5.x DNN + YOLOv5 C++ Reasoning

OpenCV4.5.4 Direct support for YOLOv5 6.1 Version model reasoning

OpenVINO2021.4+YOLOX Target detection model deployment test

Than YOLOv5 Not bad YOLOX coming , Official support OpenVINO Reasoning