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.

Read output from a model asynchronously

After you schedule a model and access an output tensor from PeekOutput, the following are true:

Sentis might not have finished calculating the final tensor data, so there's pending scheduled work.
If you use a graphics processing unit (GPU) backend, the calculated tensor data might be on the GPU. This requires a read back to copy the data to the central processing unit (CPU) in a readable format.

If either of these conditions is true, ReadbackAndClone or CompleteAllPendingOperations methods block the main thread until the operations are complete.

To avoid this, follow these two methods to use asynchronous readback:

Use the awaitable ReadbackAndCloneAsync method. Sentis returns a CPU copy of the input tensor in a non blocking way.

using Unity.Sentis;
using UnityEngine;

public class AsyncReadbackCompute : MonoBehaviour
{
    [SerializeField]
    ModelAsset modelAsset;

    Tensor<float> m_Input;
    Worker m_Worker;

    async void OnEnable()
    {
        var model = ModelLoader.Load(modelAsset);
        m_Input = new Tensor<float>(new TensorShape(1, 1), new[] { 43.0f });
        m_Worker = new Worker(model, BackendType.GPUCompute);
        m_Worker.Schedule(m_Input);

        // Peek the value from Sentis, without taking ownership of the tensor
        var outputTensor = m_Worker.PeekOutput() as Tensor<float>;
        var cpuCopyTensor = await outputTensor.ReadbackAndCloneAsync();

        Debug.Assert(cpuCopyTensor[0] == 42);
        Debug.Log($"Output tensor value {cpuCopyTensor[0]}");
        cpuCopyTensor.Dispose();
    }

    void OnDisable()
    {
        m_Input.Dispose();
        m_Worker.Dispose();
    }
}

Use a polling mechanism with the ReadbackRequest and Tensor.IsReadbackRequestDone methods.

bool inferencePending = false;
Tensor<float> outputTensor;

void OnUpdate()
{
    if (!inferencePending)
    {
        m_Worker.Schedule(m_Input);
        outputTensor = m_Worker.PeekOutput() as Tensor<float>;

        // Trigger a non-blocking readback request
        outputTensor.ReadbackRequest();
        inferencePending = true;
    }
    else if (inferencePending && outputTensor.IsReadbackRequestDone())
    {
        // m_Output is now downloaded to the cpu. Using ReadbackAndClone or ToReadOnlyArray will not be blocking
        var array = outputTensor.DownloadToArray();
        inferencePending = false;
    }
}

Use an awaitable with a callback.

bool inferencePending = false;

void Update()
{
    if (!inferencePending)
    {
        m_Worker.Schedule(m_Input);
        var outputTensor = m_Worker.PeekOutput() as Tensor<float>;
        inferencePending = true;

        var awaiter = outputTensor.ReadbackAndCloneAsync().GetAwaiter();
        awaiter.OnCompleted(() =>
        {
            var tensorOut = awaiter.GetResult();
            inferencePending = false;
            tensorOut.Dispose();
        });
    }
}

Note

To avoid a Tensor data mutation to a CPU tensor from calling ReadbackAndClone, call tensor.dataOnBackend.Download to get the data directly. This keeps the tensor.dataOnBackend on the given backend while providing a CPU copy. Be cautious with synchronization issues: if you re-schedule a worker, make a new download request.

For an example, refer to the Read output asynchronously example in the sample scripts.

Note

Read output from a model asynchronously

Note

Additional resources