Run a model
After you create a worker, use Execute to run the model.
worker.Execute(inputTensor);
You can enable verbose mode when you create a worker. Sentis logs execution to the Console window when you run the model.
worker = WorkerFactory.CreateWorker(BackendType.GPUCompute, runtimeModel, verbose: true);
When you run a model for the first time in the Unity Editor, it might be slow because Sentis needs to compile code and shaders. Later runs are faster.
Refer to the ModelExecution example in the sample scripts for a working example.
Run a model a layer at a time
To run a model a layer at a time, use the StartManualSchedule method of the worker. The method creates an IEnumerator object.
The following code sample runs the model one layer per frame, and executes the rest of the Update method only after the model finishes.
using UnityEngine;
using Unity.Sentis;
using System.Collections;
public class StartManualSchedule : MonoBehaviour
{
public ModelAsset modelAsset;
IWorker worker;
Tensor inputTensor;
public Texture2D inputTexture;
public RenderTexture outputTexture;
IEnumerator modelEnumerator;
bool started = false;
bool hasMoreModelToRun = true;
void Start()
{
Model runtimeModel = ModelLoader.Load(modelAsset);
worker = WorkerFactory.CreateWorker(BackendType.GPUCompute, runtimeModel);
inputTensor = TextureConverter.ToTensor(inputTexture);
}
void Update()
{
if (!started)
{
modelEnumerator = worker.StartManualSchedule(inputTensor);
started = true;
}
// Iterate running the model once per frame
// In each iteration of the do-while loop, use the IEnumerator object to run the next layer of the model
do
{
// Log the progress so far as a float value. 0 means no progress, 1 means complete
Debug.Log(worker.scheduleProgress);
// Use MoveNext() to run the next layer of the model. MoveNext() returns true if there's more work to do
hasMoreModelToRun = modelEnumerator.MoveNext();
} while (hasMoreModelToRun);
// Get the output tensor
var outputTensor = worker.PeekOutput() as TensorFloat;
outputTensor.PrepareCacheForAccess(blocking: true);
float[] outputArray = outputTensor.ToReadOnlyArray();
}
void OnDisable()
{
worker.Dispose();
inputTensor.Dispose();
}
}
Refer to the ModelExecutionInParts example in the sample scripts for a working example.