Model API is a set of wrapper classes for particular tasks and model architectures, simplifying data preprocess and postprocess as well as routine procedures (model loading, asynchronous execution, etc.). It is aimed at simplifying end-to-end model inference for different deployment scenarios, including local execution and serving. The Model API is based on the OpenVINO inference API.

Model API searches for additional information required for model inference, data, pre/postprocessing, label names, etc. directly in OpenVINO Intermediate Representation. This information is used to prepare the inference data, process and output the inference results in a human-readable format.

Currently, ModelAPI supports models trained in OpenVINO Training Extensions framework. Training Extensions embed all the metadata required for inference into model file. For models coming from other than Training Extensions frameworks metadata generation step is required before using ModelAPI.

• OpenVINO IR
• ONNX

• Python and C++ API
• Synchronous and asynchronous inference
• Local inference and serving through the rest API (Python only)
• Model preprocessing embedding for faster inference

pip install openvino-model-api

• Install dependencies. For installation on Ubuntu, you can use the following script:

  chmod +x src/cpp/install_dependencies.sh
  sudo src/cpp/install_dependencies.sh

• Build library:

  • Create build folder and navigate into it:

  mkdir build && cd build

  • Run cmake:

  cmake ../model_api/cpp -DOpenCV_DIR=<OpenCV cmake dir> -DOpenVINO_DIR=<OpenVINO cmake dir>

OpenCV and OpenVINO locations are optional. In most cases, these dependencies are discovered by cmake without extra guidance.

• Build:

cmake --build . -j

• To build a .tar.gz package with the library, run:

cmake --build . --target package -j

from model_api.models import Model
# Create a model wrapper from a compatible model generated by OpenVINO Training Extensions
# Use URL to work with OVMS-served model, e.g. "localhost:9000/models/ssdlite_mobilenet_v2"
model = Model.create_model("model.xml")
# Run synchronous inference locally
result = model(image)  # image is numpy.ndarray
# Print results in model-specific format
print(f"Inference result: {result}")

In C++ we have to specify model type in advance, let's set it to detection model.

#include <models/detection_model.h>
#include <models/results.h>
// Load the model
auto model = Model::create_model("model.xml");
// Run synchronous inference locally
auto result = model->infer(image); // image is cv::Mat
// Iterate over the vector of DetectedObject with box coordinates, confidence and label string
for (auto& obj : result->objects) {
    std::cout << obj.label << " | " << obj.confidence << " | " << int(obj.x) << " | " << int(obj.y) << " | "
        << int(obj.x + obj.width) << " | " << int(obj.y + obj.height) << std::endl;
}

Model's static method create_model() has two overloads. One constructs the model from a string (a path or a model name) (shown above) and the other takes an already constructed InferenceAdapter.

There are usecases when it is not possible to modify an internal ov::Model and it is hidden behind InferenceAdapter. For example the model can be served using OVMS. create_model() can construct a model from a given InferenceAdapter. That approach assumes that the model in InferenceAdapter was already configured by create_model() called with a string (a path or a model name). It is possible to prepare such model using C++ or Python: C++

auto model = DetectionModel::create_model("~/.cache/omz/public/ssdlite_mobilenet_v2/FP16/ssdlite_mobilenet_v2.xml");
const std::shared_ptr<ov::Model>& ov_model = model->getModel();
ov::serialize(ov_model, "serialized.xml");

Python

model = DetectionModel.create_model("~/.cache/omz/public/ssdlite_mobilenet_v2/FP16/ssdlite_mobilenet_v2.xml")
model.save("serialized.xml")

After that the model can be constructed from InferenceAdapter:

ov::Core core;
std::shared_ptr<ov::Model> ov_model = core.read_model("serialized.xml");
std::shared_ptr<InferenceAdapter> adapter = std::make_shared<OpenVINOInferenceAdapter>();
adapter->loadModel(ov_model, core);
auto model = DetectionModel::create_model(adapter);

For more details please refer to the examples of this project.