Jaa

Core ML 2 in Xamarin.iOS

  • Artikkeli

Core ML is a machine learning technology available on iOS, macOS, tvOS, and watchOS. It allows apps to make predictions based on machine learning models.

In iOS 12, Core ML includes a batch processing API. This API makes Core ML more efficient and provides performance improvements in scenarios where a model is used to make a sequence of predictions.

Generate sample data

In ViewController, the sample app's ViewDidLoad method calls LoadMLModel, which loads the included Core ML model:

void LoadMLModel()
{
    var assetPath = NSBundle.MainBundle.GetUrlForResource("CoreMLModel/MarsHabitatPricer", "mlmodelc");
    model = MLModel.Create(assetPath, out NSError mlErr);
}

Then, the sample app creates 100,000 MarsHabitatPricerInput objects to use as input for sequential Core ML predictions. Each generated sample has a random value set for the number of solar panels, the number of greenhouses, and the number of acres:

async void CreateInputs(int num)
{
    // ...
    Random r = new Random();
    await Task.Run(() =>
    {
        for (int i = 0; i < num; i++)
        {
            double solarPanels = r.NextDouble() * MaxSolarPanels;
            double greenHouses = r.NextDouble() * MaxGreenHouses;
            double acres = r.NextDouble() * MaxAcres;
            inputs[i] = new MarsHabitatPricerInput(solarPanels, greenHouses, acres);
        }
    });
    // ...
}

Tapping any of the app's three buttons executes two sequences of predictions: one using a for loop, and another using the new batch GetPredictions method introduced in iOS 12:

async void RunTest(int num)
{
    // ...
    await FetchNonBatchResults(num);
    // ...
    await FetchBatchResults(num);
    // ...
}

for loop

The for loop version of the test naively iterates over the specified number of inputs, calling GetPrediction for each and discarding the result. The method times how long it takes to make the predictions:

async Task FetchNonBatchResults(int num)
{
    Stopwatch stopWatch = Stopwatch.StartNew();
    await Task.Run(() =>
    {
        for (int i = 0; i < num; i++)
        {
            IMLFeatureProvider output = model.GetPrediction(inputs[i], out NSError error);
        }
    });
    stopWatch.Stop();
    nonBatchMilliseconds = stopWatch.ElapsedMilliseconds;
}

GetPredictions (new batch API)

The batch version of the test creates an MLArrayBatchProvider object from the input array (since this is a required input parameter for the GetPredictions method), creates an MLPredictionOptions object that prevents prediction computations from being restricted to the CPU, and uses the GetPredictions API to fetch the predictions, again discarding the result:

async Task FetchBatchResults(int num)
{
    var batch = new MLArrayBatchProvider(inputs.Take(num).ToArray());
    var options = new MLPredictionOptions()
    {
        UsesCpuOnly = false
    };

    Stopwatch stopWatch = Stopwatch.StartNew();
    await Task.Run(() =>
    {
        model.GetPredictions(batch, options, out NSError error);
    });
    stopWatch.Stop();
    batchMilliseconds = stopWatch.ElapsedMilliseconds;
}

Results

On both simulator and device, GetPredictions finishes more quickly than the loop-based Core ML predictions.