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. Define the forward method of the model which returns the outputs as FunctionalTensor from the inputs.
  2. Define an InputDef for each input to the model with the data type and shape.
  3. Compile the model using the Compile method.

In this example a simple model is created to calculate the dot product of two vectors.

using System;
using Unity.Sentis;
using UnityEngine;

public class CreateNewModel : MonoBehaviour
{
    Model model;

    void Start()
    {
        // Define the forward method of the model.
        // In this example a single `FunctionalTensor` output is calculated from two `FunctionalTensor` inputs.
        FunctionalTensor DotProduct(FunctionalTensor x, FunctionalTensor y)
        {
            // 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.
            return Functional.ReduceSum(prod, dim: 0, keepdim: false);
        }

        // Set up the input data types and shapes for the model as `InputDef`s.
        // Our dot product operates on two vector tensors of the same size `6`.
        (InputDef, InputDef) inputDefs =
        (
            new InputDef(DataType.Float, new TensorShape(6)),
            new InputDef(DataType.Float, new TensorShape(6))
        );

        // Build the model using the `Compile` method.
        model = Functional.Compile(DotProduct, inputDefs);
    }
}

You can then create an engine to run a model.

Additional resources