Note

Sentis is now called Inference Engine. The documentation has moved to https://docs.unity3d.com/Packages/com.unity.ai.inference@latest. Refer to the new location for the latest updates and guidance. Make sure to update your bookmarks and references accordingly.

Create a new model

You can create a new runtime model without an ONNX file using Sentis. For example if you want to do a series of tensor operations without weights, or if you want to build your own model serialization from another model format.

Using the functional API

  1. Create a FunctionalGraph object.
  2. Add the inputs to the graph with their data type and shape getting the respective FunctionalTensor for each input.
  3. Apply a series of functional API methods to the functional tensors objects to create the desired output functional tensors.
  4. Compile the model using the Compile method with the output functional tensors.

The following example shows the creation of a simple model to calculate the dot product of two vectors.

using System;
using Unity.Sentis;
using UnityEngine;

public class CreateNewModel : MonoBehaviour
{
    Model model;

    void Start()
    {
        // Create the functional graph.
        FunctionalGraph graph = new FunctionalGraph();

        // Add two inputs to the graph with data types and shapes.
        // Our dot product operates on two vector tensors of the same size `6`.
        FunctionalTensor x = graph.AddInput<float>(new TensorShape(6));
        FunctionalTensor y = graph.AddInput<float>(new TensorShape(6));

        // Calculate the elementwise product of the input `FunctionalTensor`s using an operator.
        FunctionalTensor prod = x * y;

        // Sum the product along the first axis flattening the summed dimension.
        FunctionalTensor reduce = Functional.ReduceSum(prod, dim: 0, keepdim: false);

        // Build the model using the `Compile` method.
        model = graph.Compile(reduce, prod);

        // The final model has two inputs `x` and `y`, which are denoted as `input_0` and `input_1` respectively
        // and two outputs `reduce` and `prod`, which are denoted as `output_0` and `output_1` respectively.
    }
}

To debug your code, use ToString to retrieve information, such as its shape and datatype.

You can then create an engine to run a model.

Model inputs and outputs

Compile returns a model with the number of inputs created with AddInput. In our example x is the first input and y the second. The inputs have the names input_0, input_1 ... in the compiled model.

Compile returns a model with the number of outputs that are provided to the Compile method. In our example reduce is the first output and prod the second. The outputs have the names output_0, output_1 ... in the compiled model.

Additional resources