Create input for a model

A model requires input tensors with certain shapes and data types. Use this information to find model inputs and create input tensors for your model.

Understand the required input

Before you can create input tensors for a model, you must first understand the shape and data types of the model inputs. For more information, refer to Model inputs.

The TensorShape of the Tensor you create must be compatible with the DynamicTensorShape, which defines the shape of the model input.

Convert an array to a tensor

To create a CPU tensor from a one-dimensional data array, follow these steps:

  1. Create a TensorShape object that has the length of each axis.
  2. Create a Tensor<T> object with the shape and the data array.

For example, the following code creates a tensor for a model that takes an input tensor of shape 3 × 1 × 3.

using UnityEngine;
using Unity.Sentis;

public class ConvertArrayToTensor : MonoBehaviour
{
    void Start()
    {
        // Create a data array with 9 values
        float[] data = new float[] { 1f, 2f, 3f, 4f, 5f, 6f, 7f, 8f, 9f };

        // Create a 3D tensor shape with size 3 × 1 × 3
        TensorShape shape = new TensorShape(3, 1, 3);

        // Create a new tensor from the array
        Tensor<float> tensor = new Tensor<float>(shape, data);
    }
}

Create an empty tensor

You can create a CPU tensor with its memory 0-initialized as follows:

var tensor = new Tensor<int>(new TensorShape(1), clearOnInit: true);

The parameter clearOnInit determines whether the resulting tensor memory should be 0-initialized. If it is unimportant what the initial data of the tensor is, set clearOnInit to false to avoid this additional operation.

Pass inputs to a worker

If a model needs multiple input tensors, do one of the following:

  • Call SetInput on the worker for each input, and then call Schedule on the worker with no arguments.
  • Call Schedule on the worker with an array of the desired inputs.
worker.SetInput("x", xTensor);
worker.SetInput("y", yTensor);
worker.Schedule();

worker.Schedule(xTensor, yTensor);

To avoid garbage collection due to the params array creation, you can pre-allocate the input array.

var inputs = new [] { xTensor, yTensor };
//...
worker.Schedule(inputs);

Edit a model

Use the functional API to add operations to your model inputs. For more information, refer to Edit a model.

Additional resources