Model Representation in OpenVINO™ Runtime

In OpenVINO™ Runtime, a model is represented by the ov::Model class.

The ov::Model object stores shared pointers to ov::op::v0::Parameter, ov::op::v0::Result, and ov::op::Sink operations, which are inputs, outputs, and sinks of the graph. Sinks of the graph have no consumers and are not included in the results vector. All other operations hold each other via shared pointers, in which a child operation holds its parent via a hard link. If an operation has no consumers and is neither the Result nor the Sink operation whose shared pointer counter is zero, the operation will be destructed and not be accessible anymore.

Each operation in ov::Model has the std::shared_ptr<ov::Node> type.

How OpenVINO Runtime Works with Models

OpenVINO™ Runtime enables you to use different approaches to work with model inputs/outputs:

  • The ov::Model::inputs() / ov::Model::outputs() methods are used to get vectors of all input/output ports.

        inputs = model.inputs
    outputs = model.outputs

    /* Take information about all topology inputs */
auto inputs = model->inputs();
/* Take information about all topology outputs */
auto outputs = model->outputs();

  • For a model that has only one input or output, you can use the ov::Model::input() or ov::Model::output() methods without any arguments to get input or output port respectively.

  • The ov::Model::input() and ov::Model::output() methods can be used with the index of inputs or outputs from the framework model to get specific ports by index.

    ov_model_input = model.input(index)
ov_model_output = model.output(index)

    auto ov_model_input = ov_model->input(index);
auto ov_model_output = ov_model->output(ind

  • You can use the tensor name of input or output from the original framework model together with the ov::Model::input() or ov::Model::output() methods to get specific ports. It means that you do not need to have any additional mapping of names from framework to OpenVINO as it was before. OpenVINO Runtime allows the usage of native framework tensor names, for example:

Warning

All inputs/outputs of ov::Model are numbered, so the preferred way to retrieve them is to use indices.

Using tensor names can potentially be a less reliable approach, since the mandatory presence of tensor names for inputs and outputs is not guaranteed in the original frameworks. Therefore ov::Model may contain empty list of tensor_names for inputs/outputs.

To get all tensor names which are associated with the corresponding input/output, OpenVINO Runtime has get_names method. To get some name from all names associated with a given input/output, the get_any_name method was introduced. These methods may return empty names list/empty name if the names are not present.

ov_model_input = model.input(original_fw_in_tensor_name)
ov_model_output = model.output(original_fw_out_tensor_name)

auto ov_model_input = ov_model->input(original_fw_in_tensor_name);
auto ov_model_output = ov_model->output(original_fw_out_tensor_name);

For details on how to build a model in OpenVINO™ Runtime, see the Build a Model in OpenVINO Runtime section.

OpenVINO™ Runtime model representation uses special classes to work with model data types and shapes. The ov::element::Type is used for data types.

ov_input.get_element_type()

ov_input->get_element_type();

Representation of Shapes

OpenVINO™ Runtime provides two types for shape representation:

  • ov::Shape - Represents static (fully defined) shapes.

  • ov::PartialShape - Represents dynamic shapes. This means that the rank or some of dimensions are dynamic (dimension defines an interval or undefined).

ov::PartialShape can be converted to ov::Shape by using the get_shape() method if all dimensions are static; otherwise, the conversion will throw an exception. For example:

    partial_shape = node.output(0).get_partial_shape() # get zero output partial shape
    if not partial_shape.is_dynamic: # or partial_shape.is_static
        static_shape = partial_shape.get_shape()

    ov::Shape static_shape;
    ov::PartialShape partial_shape = node->output(0).get_partial_shape(); // get zero output partial shape
    if (!partial_shape.is_dynamic() /* or partial_shape.is_static() */) {
        static_shape = partial_shape.get_shape();
    }

However, in most cases, before getting static shape using the get_shape() method, you need to check if that shape is static.

Representation of Operations

The ov::Op class represents any abstract operation in the model representation. Use this class to create custom operations.

Representation of Operation Sets

An operation set (opset) is a collection of operations that can be used to construct a model. The ov::OpSet class provides the functionality to work with operation sets. For each operation set, OpenVINO™ Runtime provides a separate namespace, for example opset8.

Each OpenVINO™ Release introduces new operations and adds them to new operation sets, within which the new operations would change the behavior of previous operations. Using operation sets helps you avoid changing your application when new operations are introduced. For a complete list of operation sets supported in OpenVINO™ toolkit, see the Available Operations Sets. To add the support for custom operations, see OpenVINO Extensibility Mechanism.

Building a Model in OpenVINO™ Runtime

You can create a model from source. This section illustrates how to construct a model composed of operations from an available operation set.

Operation set opsetX integrates a list of pre-compiled operations that work for this purpose. In other words, opsetX defines a set of operations for building a graph.

To build an ov::Model instance from opset8 operations, include the following files:

import openvino as ov

#include <openvino/core/model.hpp>
#include <openvino/opsets/opset8.hpp>

The following code demonstrates how to create a simple model:

def create_simple_model():
    # This example shows how to create ov::Function
    #
    # Parameter--->Multiply--->Add--->Result
    #    Constant---'          /
    #              Constant---'
    data = ops.parameter([3, 1, 2], ov.Type.f32)
    mul_constant = ops.constant([1.5], ov.Type.f32)
    mul = ops.multiply(data, mul_constant)
    add_constant = ops.constant([0.5], ov.Type.f32)
    add = ops.add(mul, add_constant)
    res = ops.result(add)
    return ov.Model([res], [data], "model")

std::shared_ptr<ov::Model> create_simple_model() {
    // This example shows how to create ov::Model
    //
    // Parameter--->Multiply--->Add--->Result
    //    Constant---'          /
    //              Constant---'

    // Create opset8::Parameter operation with static shape
    auto data = std::make_shared<ov::opset8::Parameter>(ov::element::f32, ov::Shape{3, 1, 2});

    auto mul_constant = ov::opset8::Constant::create(ov::element::f32, ov::Shape{1}, {1.5});
    auto mul = std::make_shared<ov::opset8::Multiply>(data, mul_constant);

    auto add_constant = ov::opset8::Constant::create(ov::element::f32, ov::Shape{1}, {0.5});
    auto add = std::make_shared<ov::opset8::Add>(mul, add_constant);

    // Create opset8::Result operation
    auto res = std::make_shared<ov::opset8::Result>(mul);

    // Create nGraph function
    return std::make_shared<ov::Model>(ov::ResultVector{res}, ov::ParameterVector{data});
}

The following code creates a model with several outputs:

def create_advanced_model():
    # Advanced example with multi output operation
    #
    # Parameter->Split---0-->Result
    #               | `--1-->Relu-->Result
    #               `----2-->Result
    data = ops.parameter(ov.Shape([1, 3, 64, 64]), ov.Type.f32)
    # Create Constant for axis value
    axis_const = ops.constant(1, dtype=ov.Type.i64)

    # Create opset12::Split operation that splits input to three slices across 1st dimension
    split = ops.split(data, axis_const, 3)

    # Create opset12::Relu operation that takes 1st Split output as input
    relu = ops.relu(split.output(1))

    # Results operations will be created automatically based on provided OutputVector
    return ov.Model([split.output(0), relu.output(0), split.output(2)], [data], "model")

std::shared_ptr<ov::Model> create_advanced_model() {
    // Advanced example with multi output operation
    //
    // Parameter->Split---0-->Result
    //               | `--1-->Relu-->Result
    //               `----2-->Result

    auto data = std::make_shared<ov::opset8::Parameter>(ov::element::f32, ov::Shape{1, 3, 64, 64});

    // Create Constant for axis value
    auto axis_const = ov::opset8::Constant::create(ov::element::i64, ov::Shape{} /*scalar shape*/, {1});

    // Create opset8::Split operation that splits input to three slices across 1st dimension
    auto split = std::make_shared<ov::opset8::Split>(data, axis_const, 3);

    // Create opset8::Relu operation that takes 1st Split output as input
    auto relu = std::make_shared<ov::opset8::Relu>(split->output(1) /*specify explicit output*/);

    // Results operations will be created automatically based on provided OutputVector
    return std::make_shared<ov::Model>(ov::OutputVector{split->output(0), relu, split->output(2)},
                                       ov::ParameterVector{data});
}

Model Debugging Capabilities

OpenVINO™ provides several debug capabilities:

  • To receive additional messages about applied model modifications, rebuild the OpenVINO™ Runtime library with the -DENABLE_OPENVINO_DEBUG=ON option.

  • Model can be visualized to image from the xDot format:

def visualize_example(m : ov.Model):
    # Need import:
    # * import openvino.runtime.passes as passes
    pass_manager = passes.Manager()
    pass_manager.register_pass(passes.VisualizeTree(file_name='image.svg'))
    pass_manager.run_passes(m)

void visualize_example(const std::shared_ptr<ov::Model>& m) {
    // Need include:
    // * openvino/pass/manager.hpp
    // * openvino/pass/visualize_tree.hpp
    ov::pass::Manager manager;

    // Serialize ov::Model to before.svg file before transformation
    manager.register_pass<ov::pass::VisualizeTree>("image.svg");

    manager.run_passes(m);
}

`ov::pass::VisualizeTree` can be parametrized via environment variables:

OV_VISUALIZE_TREE_OUTPUT_SHAPES=1       - visualize shapes

OV_VISUALIZE_TREE_OUTPUT_TYPES=1        - visualize types

OV_VISUALIZE_TREE_MIN_MAX_DENORMAL=1    - pretty denormal values

OV_VISUALIZE_TREE_RUNTIME_INFO=1        - print runtime information

OV_VISUALIZE_TREE_IO=1                  - print I/O ports

OV_VISUALIZE_TREE_MEMBERS_NAME=1        - print member names

  • Also model can be serialized to IR:

def serialize_example(m : ov.Model):
    ov.serialize(m, xml_path='model.xml', bin_path='model.bin')

void serialize_example(const std::shared_ptr<ov::Model>& model) {
    ov::serialize(model, "/path/to/file/model.xml", "/path/to/file/model.bin");
}