private static void TrainModel(DeviceDescriptor device, PreparedDataInfo preparedData,
NetworkBuildOutput buildOutput, Trainer trainer)
{
const uint minibatchSize = 64;
const int outputFrequencyInMinibatches = 20;
var i = 0;
var epochs = 5;
while (epochs > 0)
{
var minibatchData = preparedData.MinibatchSource.GetNextMinibatch(minibatchSize, device);
var arguments = new Dictionary <Variable, MinibatchData>
{
{ buildOutput.Input, minibatchData[preparedData.FeatureStreamInfo] },
{ buildOutput.Labels, minibatchData[preparedData.LabelStreamInfo] }
};
trainer.TrainMinibatch(arguments, device);
PrintTrainingProgress(trainer, i++, outputFrequencyInMinibatches);
// MinibatchSource is created with MinibatchSource.InfinitelyRepeat.
// Batching will not end. Each time minibatchSource completes an sweep (epoch),
// the last minibatch data will be marked as end of a sweep. We use this flag
// to count number of epochs.
if (MiniBatchDataIsSweepEnd(minibatchData.Values))
{
epochs--;
}
}
}
private static Trainer CreateTrainer(NetworkBuildOutput buildOutput)
{
var parameters = buildOutput.ClassifierOutput.Parameters();
var parameterLearners = CreateLerners(parameters);
var trainer = Trainer.CreateTrainer(buildOutput.ClassifierOutput, buildOutput.TrainingLoss,
buildOutput.Prediction, parameterLearners);
return(trainer);
}
public NetworkBuildOutput Build(DeviceDescriptor device, string featureStreamName, string labelsStreamName,
string classifierName, int numClasses, int[] imageDim)
{
var buildOutput = new NetworkBuildOutput();
buildOutput.Input = CNTKLib.InputVariable(imageDim, DataType.Float, featureStreamName);
buildOutput.ScaledInput = CNTKLib.ElementTimes(Constant.Scalar(0.00390625f, device), buildOutput.Input);
buildOutput.ClassifierOutput = CreateConvolutionNeuralNetwork(buildOutput.ScaledInput,
numClasses, device, classifierName);
buildOutput.Labels = CNTKLib.InputVariable(new[] { numClasses }, DataType.Float, labelsStreamName);
buildOutput.TrainingLoss = CNTKLib.CrossEntropyWithSoftmax(new Variable(buildOutput.ClassifierOutput),
buildOutput.Labels, "lossFunction");
buildOutput.Prediction = CNTKLib.ClassificationError(new Variable(buildOutput.ClassifierOutput),
buildOutput.Labels, "classificationError");
return(buildOutput);
}
private static void SaveModel(NetworkBuildOutput buildOutput, string modelFile)
{
buildOutput.ClassifierOutput.Save(modelFile);
}
