/// <summary>
/// Helper functoin to create gerantor;
/// </summary>
/// <param name="inputNoiseSize"></param>
/// <param name="inputConditionSize"></param>
/// <param name="outputSize"></param>
/// <param name="generatorLayerSize"></param>
/// <param name="generatorLayerCount"></param>
/// <param name="device"></param>
protected void CreateGenerator(int inputNoiseSize, int inputConditionSize, int outputSize, int generatorLayerSize, int generatorLayerCount, DeviceDescriptor device)
{
//create generator
Variable concatenatedInput;
if (inputNoiseSize > 0 && inputConditionSize > 0)
{
InputNoiseGenerator = CNTKLib.InputVariable(new int[] { inputNoiseSize }, DataType.Float);
InputConditionGenerator = CNTKLib.InputVariable(new int[] { inputConditionSize }, DataType.Float);
var vsgenerator = new VariableVector();
vsgenerator.Add(InputNoiseGenerator);
vsgenerator.Add(InputConditionGenerator);
concatenatedInput = CNTKLib.Splice(vsgenerator, new Axis(0));
}
else if (inputNoiseSize > 0)
{
InputNoiseGenerator = CNTKLib.InputVariable(new int[] { inputNoiseSize }, DataType.Float);
InputConditionGenerator = null;
concatenatedInput = InputNoiseGenerator;
}
else
{
InputNoiseGenerator = null;
InputConditionGenerator = CNTKLib.InputVariable(new int[] { inputConditionSize }, DataType.Float);
concatenatedInput = InputConditionGenerator;
}
var inputG = new InputLayerCNTKVar(concatenatedInput);
var outputLayerG = new OutputLayerDense(outputSize, null, OutputLayerDense.LossFunction.Square);
GeneratorSequentialModel = new SequentialNetworkDense(inputG, LayerDefineHelper.DenseLayers(generatorLayerCount, generatorLayerSize, true, NormalizationMethod.None), outputLayerG, device);
GeneratorOutput = GeneratorSequentialModel.OutputLayer.GetOutputVariable();
InputTargetGenerator = GeneratorSequentialModel.OutputLayer.GetTargetInputVariable();
}
/// <summary>
/// Helper functio to create discriminators
/// </summary>
/// <param name="fakeDataFromGenerator"></param>
/// <param name="inputConditionSize"></param>
/// <param name="outputSize"></param>
/// <param name=""></param>
/// <param name="discriminatorLayerSize"></param>
/// <param name="discriminatorLayerCount"></param>
/// <param name="device"></param>
protected void CreateDiscriminators(Variable fakeDataFromGenerator, int inputConditionSize, int outputSize, int discriminatorLayerSize, int discriminatorLayerCount, DeviceDescriptor device)
{
//create discriminator
Variable concatenatedInput = null;
//create input based on whether it is a conditional gan
if (inputConditionSize > 0)
{
InputDataDiscriminatorReal = CNTKLib.InputVariable(new int[] { outputSize }, DataType.Float);
InputConditionDiscriminatorReal = CNTKLib.InputVariable(new int[] { inputConditionSize }, DataType.Float);
InputConditionDiscriminatorFake = CNTKLib.InputVariable(new int[] { inputConditionSize }, DataType.Float);
var vsDiscriminator = new VariableVector();
vsDiscriminator.Add(InputDataDiscriminatorReal);
vsDiscriminator.Add(InputConditionDiscriminatorReal);
concatenatedInput = CNTKLib.Splice(vsDiscriminator, new Axis(0));
}
else
{
InputDataDiscriminatorReal = CNTKLib.InputVariable(new int[] { outputSize }, DataType.Float);
InputConditionDiscriminatorReal = null;
InputConditionDiscriminatorFake = null;
concatenatedInput = InputDataDiscriminatorReal;
}
var inputD = new InputLayerCNTKVar(concatenatedInput);
var outputLayerD = new OutputLayerDense(1, new SigmoidDef(), OutputLayerDense.LossFunction.Square);
//create the discriminator sequential model
DiscriminatorSequentialModel = new SequentialNetworkDense(inputD, LayerDefineHelper.DenseLayers(discriminatorLayerCount, discriminatorLayerSize, true, NormalizationMethod.None), outputLayerD, device);
//real discriminator output
DiscriminatorRealOutput = DiscriminatorSequentialModel.OutputLayer.GetOutputVariable();
//clone the discriminator with shared parameters
if (inputConditionSize > 0)
{
DiscriminatorFakeOutput = ((Function)DiscriminatorRealOutput).Clone(ParameterCloningMethod.Share,
new Dictionary <Variable, Variable>()
{
{ InputDataDiscriminatorReal, fakeDataFromGenerator }, { InputConditionDiscriminatorReal, InputConditionDiscriminatorFake }
});
}
else
{
DiscriminatorFakeOutput = ((Function)DiscriminatorRealOutput).Clone(ParameterCloningMethod.Share,
new Dictionary <Variable, Variable>()
{
{ InputDataDiscriminatorReal, fakeDataFromGenerator }
});
}
DiscriminatorMerged = Function.Combine(new List <Variable>()
{
DiscriminatorRealOutput, DiscriminatorFakeOutput
});
}
public TrainerSimpleNN(SequentialNetworkDense net, LearnerDefs.LearnerDef trainingLearner, DeviceDescriptor device, int maxDataBufferCount = 50000)
{
Device = device;
networkRef = net;
learners = new List <Learner>();
var paramsToTrain = net.CNTKFunction.Parameters();
Learner learner = trainingLearner.Create(paramsToTrain);
learners.Add(learner);
trainer = Trainer.CreateTrainer(networkRef.OutputLayer.GetOutputVariable(), networkRef.OutputLayer.GetTrainingLossVariable(),
networkRef.OutputLayer.GetTrainingLossVariable(),//use training loss for eval error for now
learners);
//get the input shape for creating the buffer
var inputDims = networkRef.InputLayer.InputVariable.Shape.Dimensions;
int inputSize = 1;
foreach (var d in inputDims)
{
inputSize *= d;
}
var targetDims = networkRef.OutputLayer.GetTargetInputVariable().Shape.Dimensions;
int targetSize = 1;
foreach (var d in targetDims)
{
targetSize *= d;
}
//create databuffer
dataBuffer = new DataBuffer(maxDataBufferCount,
new DataBuffer.DataInfo("Input", DataBuffer.DataType.Float, inputSize),
new DataBuffer.DataInfo("Target", DataBuffer.DataType.Float, targetSize)
);
}
