public static Function OptimizedRNNStack(Variable input, int hiddenSize, int layerSize = 1, bool bidirectional = false, string cellType = "lstm", string name = "")
{
try
{
NodeGroup.EnterNewGroup(name);
var dim = input.Shape.Dimensions[0];
var weightSize = (dim - 1) * 4 * hiddenSize;
weightSize += (layerSize - 1) * (8 * hiddenSize * hiddenSize + 8 * hiddenSize);
weightSize += 4 * hiddenSize * hiddenSize + 12 * hiddenSize;
var w = new Parameter(new int[] { weightSize }, DataType.Float, CNTKLib.GlorotUniformInitializer(), DeviceDescriptor.UseDefaultDevice(), name + "_w");
Register(w);
var rnn = CNTKLib.OptimizedRNNStack(input, w, (uint)hiddenSize, (uint)layerSize, bidirectional, cellType, name + "_rnn");
Register(rnn);
var output = CNTKLib.SequenceLast(rnn);
Register(output);
output.RootFunction.SetName(name);
return(output);
}
finally
{
NodeGroup.LeaveGroup();
}
}
public static Function BatchNormalization(Variable input, bool spatial, double initScale, double normalizationTimeConstant, double blendTimeConstant, double epsilon, bool useCuDNNEngine, bool disableRegularization, string name)
{
try
{
NodeGroup.EnterNewGroup(name);
var normShape = new int[] { CNTK.NDShape.InferredDimension };
var scale = new Parameter(normShape, DataType.Float, initScale, DeviceDescriptor.UseDefaultDevice(), name + "/scale");
Register(scale);
var bias = new Parameter(normShape, DataType.Float, 0, DeviceDescriptor.UseDefaultDevice(), name + "/bias");
Register(bias);
var runningMean = new Constant(normShape, 0.0f, DeviceDescriptor.UseDefaultDevice());
Register(runningMean);
var runningInvStd = new Constant(normShape, 0.0f, DeviceDescriptor.UseDefaultDevice());
Register(runningInvStd);
var runningCount = Constant.Scalar(0.0f, DeviceDescriptor.UseDefaultDevice());
Register(runningCount);
var output = CNTKLib.BatchNormalization(
input, // CNTK.Variable operand
scale, // CNTK.Variable scale
bias, // CNTK.Variable bias
runningMean, // CNTK.Variable runningMean
runningInvStd, // CNTK.Variable runningInvStd
runningCount, // CNTK.Variable runningCount
spatial, // bool spatial
normalizationTimeConstant, // double normalizationTimeConstant
blendTimeConstant, // double blendTimeConstant
epsilon, // double epsilon
useCuDNNEngine, // bool useCuDNNEngine
disableRegularization, // bool disableRegularization
"" // string name
);
Register(output);
output.RootFunction.SetName(name);
return(output);
}
finally
{
NodeGroup.LeaveGroup();
}
}
// Assume input shape is such as (x [, y [, z]], channels)
public static Function ConvolutionTranspose(Variable input, int[] filterShape, int numFilters, string activation, CNTKDictionary initializer, bool[] padding, int[] strides, bool useBias, CNTKDictionary biasInitializer, int[] outputShape, int[] dilation, int reductionRank, int maxTempMemSizeInSamples, string name)
{
try
{
NodeGroup.EnterNewGroup(name);
// Initializers
if (initializer == null)
{
initializer = CNTKLib.GlorotUniformInitializer();
}
if (useBias && biasInitializer == null)
{
biasInitializer = CNTKLib.ConstantInitializer(0);
}
// Convolution map
// (kernelWidth, kernelHeight, featureMapCount, kernelChannel)
var convDims = new int[filterShape.Length + 2];
filterShape.CopyTo(convDims, 0);
convDims[convDims.Length - 2] = numFilters;
convDims[convDims.Length - 1] = input.Shape.Dimensions[filterShape.Length]; // input channel
var convolutionMap = new Parameter(convDims, DataType.Float, initializer, DeviceDescriptor.UseDefaultDevice(), name + "/weight");
Register(convolutionMap);
var conv = CNTKLib.ConvolutionTranspose(
convolutionMap, // CNTK.Variable convolutionMap
input, // CNTK.Variable operand
strides, // CNTK.NDShape strides
new BoolVector(new bool[] { true }), // CNTK.BoolVector sharing
new BoolVector(padding), // CNTK.BoolVector autoPadding
outputShape, // CNTK.NDShape outputShape
dilation, // CNTK.NDShape dilation
(uint)reductionRank, // uint reductionRank
(uint)maxTempMemSizeInSamples, // uint maxTempMemSizeInSamples
"" // string name
);
Register(conv);
if (useBias)
{
var bias = new Parameter(conv.Output.Shape, DataType.Float, biasInitializer, DeviceDescriptor.UseDefaultDevice(), name + "/bias");
Register(bias);
conv = CNTKLib.Plus(conv, bias);
Register(conv);
}
conv = ApplyActivation(conv, activation);
conv.RootFunction.SetName(name);
return(conv);
}
finally
{
NodeGroup.LeaveGroup();
}
}
public static Function Dense(Variable input, int[] outputDimensions, CNTKDictionary initializer, bool useBias, CNTKDictionary biasInitializer, bool stabilize, double steepness, string activation, DeviceDescriptor device, string name)
{
try
{
NodeGroup.EnterNewGroup(name);
if (outputDimensions == null)
{
outputDimensions = new Shape(input.Shape.Dimensions.ToArray());
}
if (initializer == null)
{
initializer = CNTKLib.GlorotUniformInitializer();
}
if (useBias && biasInitializer == null)
{
biasInitializer = CNTKLib.ConstantInitializer(0);
}
if (input.Shape.Rank > 1)
{
int newDim = input.Shape.Dimensions.Aggregate((d1, d2) => d1 * d2);
input = CNTKLib.Reshape(input, new int[] { newDim });
Register(input);
}
var inputDimensions = input.Shape.Dimensions[0];
int hiddenSize = outputDimensions.Aggregate((d1, d2) => d1 * d2);
var weight = new Parameter(new int[] { hiddenSize, inputDimensions }, DataType.Float, initializer, device, name + "/weight");
Register(weight);
Parameter bias = null;
if (useBias)
{
bias = new Parameter(new int[] { hiddenSize }, DataType.Float, biasInitializer, device, name + "/bias");
Register(bias);
}
if (stabilize)
{
input = Stabilize(input, steepness, device, name + "/stabilizer");
}
var output = GetAffine(input, weight, bias);
if (outputDimensions.Length > 1)
{
output = CNTKLib.Reshape(output, outputDimensions);
Register(output);
}
output = ApplyActivation(output, activation);
output.RootFunction.SetName(name);
return(output);
}
finally
{
NodeGroup.LeaveGroup();
}
}