microsoft
/
CNTK
зеркало из
1
0
Форкнуть 0
Код Задачи Пакеты Проекты Релизы Вики Активность

full tensor support for Mean and InvStdDev operations

This commit is contained in:
Frank Seide 2016-03-03 23:46:17 -08:00
Родитель e0be5a1c58
Коммит 359d90ab09
8 изменённых файлов: 131 добавлений и 179 удалений
Показать статистику Скачать Patch файл Скачать Diff файл Показать все файлы Свернуть все файлы

24
Source/ComputationNetworkLib/LinearAlgebraNodes.h
Просмотреть файл

@ -47,7 +47,7 @@ public:
virtual void /*ComputationNode::*/ ForwardProp(const FrameRange& fr) override
{
size_t rank = DetermineElementwiseTensorRank();
auto result = ValueTensorFor(rank, fr);
auto result = ValueTensorFor(rank, fr);
auto input0 = Input(0)->ValueTensorFor(rank, fr.AllowBroadcast());
auto input1 = Input(1)->ValueTensorFor(rank, fr.AllowBroadcast());
result.AssignSumOf(input0, input1);
@ -56,7 +56,7 @@ public:
virtual void /*ComputationNode::*/ BackpropTo(const size_t inputIndex, const FrameRange& fr) override
{
size_t rank = DetermineElementwiseTensorRank();
auto gradient = GradientTensorFor(rank, fr);
auto gradient = GradientTensorFor(rank, fr);
auto inputGradient = Input(inputIndex)->GradientTensorFor(rank, fr.AllowBroadcast());
// if reduction then mask the respective input(s) (zero out the gaps)
@ -77,12 +77,8 @@ template class PlusNode<double>;
template <class ElemType>
class MinusNode : public BinaryElementWiseNode<ElemType>
{
typedef BinaryElementWiseNode<ElemType> Base;
UsingBinaryElementwiseNodeBaseMembers;
static const std::wstring TypeName()
{
return L"Minus";
}
typedef BinaryElementWiseNode<ElemType> Base; UsingBinaryElementwiseNodeBaseMembers;
static const std::wstring TypeName() { return L"Minus"; }
public:
DeclareConstructorFromConfigWithNumInputs(MinusNode);
@ -95,7 +91,7 @@ public:
{
ElemType sign = inputIndex == 0 ? 1.0f : -1.0f;
size_t rank = DetermineElementwiseTensorRank();
auto gradient = GradientTensorFor(rank, fr);
auto gradient = GradientTensorFor(rank, fr);
auto inputGradient = Input(inputIndex)->GradientTensorFor(rank, fr.AllowBroadcast());
// if reduction then mask the respective input(s) (zero out the gaps)
@ -108,7 +104,7 @@ public:
virtual void /*ComputationNode::*/ ForwardProp(const FrameRange& fr) override
{
size_t rank = DetermineElementwiseTensorRank();
auto result = ValueTensorFor(rank, fr);
auto result = ValueTensorFor(rank, fr);
auto input0 = Input(0)->ValueTensorFor(rank, fr.AllowBroadcast());
auto input1 = Input(1)->ValueTensorFor(rank, fr.AllowBroadcast());
result.AssignDifferenceOf(input0, input1);
@ -126,12 +122,8 @@ template class MinusNode<double>;
template <class ElemType>
class NegateNode : public ComputationNode<ElemType>, public NumInputs<1>
{
typedef ComputationNode<ElemType> Base;
UsingComputationNodeMembersBoilerplate;
static const std::wstring TypeName()
{
return L"Negate";
}
typedef ComputationNode<ElemType> Base; UsingComputationNodeMembersBoilerplate;
static const std::wstring TypeName() { return L"Negate"; }
public:
DeclareConstructorFromConfigWithNumInputs(NegateNode);

260
Source/ComputationNetworkLib/PreComputeNodes.h
Просмотреть файл

@ -138,8 +138,7 @@ public:
template <class ElemType>
class MeanInvStdDevNodeBase : public PreComputedNodeBase<ElemType>, public NumInputs<1>
{
typedef PreComputedNodeBase<ElemType> Base;
UsingPreComputedNodeMembers;
typedef PreComputedNodeBase<ElemType> Base; UsingPreComputedNodeMembers;
// static const std::wstring TypeName() { return L"MeanInvStdDev (base)"; }
public:
// DeclareConstructorFromConfigWithNumInputs(MeanInvStdDevNodeBase);
@ -219,12 +218,8 @@ protected:
template <class ElemType>
class MeanNode : public MeanInvStdDevNodeBase<ElemType>
{
typedef MeanInvStdDevNodeBase<ElemType> Base;
UsingMeanInvStdDevNodeBaseNodeMembers;
static const std::wstring TypeName()
{
return L"Mean";
}
typedef MeanInvStdDevNodeBase<ElemType> Base; UsingMeanInvStdDevNodeBaseNodeMembers;
static const std::wstring TypeName() { return L"Mean"; }
public:
DeclareConstructorFromConfigWithNumInputs(MeanNode);
@ -232,11 +227,11 @@ public:
: Base(deviceId, name)
{
}
MeanNode(DEVICEID_TYPE deviceId, const wstring& name, size_t)
: Base(deviceId, name)
{
}
virtual void /*PreComputedNodeBase::*/ MarkComputed(const bool hasComputed)
{
Base::MarkComputed(hasComputed);
@ -260,19 +255,27 @@ public:
// set gaps to zero, since we are reducing in time
Input(0)->MaskMissingValueColumnsToZero(fr);
auto& samples = Input(0)->Value();
auto& avg = Value();
#if NANCHECK
samples.HasNan("Mean-Samples");
#endif
size_t numNewSamples = Input(0)->GetMBLayout()->GetActualNumSamples();
size_t totalNumSamples = m_numSamples + numNewSamples;
if (totalNumSamples == 0)
totalNumSamples = 1; // 0/0=1 in this context
Matrix<ElemType>::MultiplyAndWeightedAdd(1.0f / totalNumSamples, samples, false,
ElemType alpha = 1.0f / totalNumSamples;
ElemType beta = (ElemType)m_numSamples / totalNumSamples;
#if 1
size_t rank = DetermineElementwiseTensorRank();
auto mean = ValueTensorFor(rank, FrameRange()); // mean is formed directly in our m_value
auto input = Input(0)->ValueTensorFor(rank, fr);
mean.DoCopyOf(beta, input, alpha);
// Note: We leverage that TensorView allows "broadcasting" the output,
// which really means a reduction.
#else
auto& samples = Input(0)->Value();
auto& avg = Value();
Matrix<ElemType>::MultiplyAndWeightedAdd(alpha, samples, false,
ConstOnes(Input(0)->Value().GetNumCols(), 1, samples.GetDeviceId()),
false, (ElemType) m_numSamples / totalNumSamples, avg);
false, beta, avg);
#endif
#if NANCHECK
avg.HasNan("Mean-avg");
#endif
@ -292,12 +295,8 @@ template class MeanNode<double>;
template <class ElemType>
class InvStdDevNode : public MeanInvStdDevNodeBase<ElemType>
{
typedef MeanInvStdDevNodeBase<ElemType> Base;
UsingMeanInvStdDevNodeBaseNodeMembers;
static const std::wstring TypeName()
{
return L"InvStdDev";
}
typedef MeanInvStdDevNodeBase<ElemType> Base; UsingMeanInvStdDevNodeBaseNodeMembers;
static const std::wstring TypeName() { return L"InvStdDev"; }
public:
DeclareConstructorFromConfigWithNumInputs(InvStdDevNode);
@ -316,31 +315,21 @@ public:
if (!m_hasComputed) // initialize
{
// reset accumulators
size_t inputDim = Input(0)->GetSampleMatrixNumRows();
m_mean.Resize(inputDim, 1);
m_var.Resize(inputDim, 1);
m_mean.SetValue(0);
m_var.SetValue(0);
UpdateFunctionValuesSize();
Value().SetValue(0); // also set this because not doing it may flag during debugging; avoids special-casing this
Value().SetValue(0); // Note: We must do this here already because dimensions are verified at places.
m_mean.Resize(Value()); // mean accumulator normalized by #samples in it
m_var .Resize(Value()); // likewise the variance
m_temp.Resize(Value()); // and a temp
m_mean.SetValue(0); // reset the mean and var accumulators
m_var .SetValue(0);
}
else // finalize
{
// m_value <- 1/stddev
ElemType sqrtFloor = 1e-10f;
m_var.InplaceTruncateBottom(sqrtFloor); // prevent too small variance (and negative square roots due to numeric inaccuracy)
#if NANCHECK
m_var.HasNan("MarkComputed-InplaceTruncateBottom");
#endif
m_var.InplaceSqrt();
#if NANCHECK
m_var.HasNan("MarkComputed-InplaceSqrt");
#endif
m_var.ElementInverse();
#if NANCHECK
m_var.HasNan("MarkComputed-ElementInverse()");
#endif
Value().SetValue(m_var);
}
}
@ -357,29 +346,55 @@ public:
// set gaps to zero, since we are reducing in time
Input(0)->MaskMissingValueColumnsToZero(fr);
auto& samples = Input(0)->Value();
#if NANCHECK
samples.HasNan("InvStdDev-Samples");
#endif
m_temp.SetValue(m_mean);
//m_temp.SetValue(m_mean); // old mean
size_t numNewSamples = Input(0)->GetMBLayout()->GetActualNumSamples();
size_t totalNumSamples = m_numSamples + numNewSamples;
if (totalNumSamples == 0)
totalNumSamples = 1; // 0/0=1 in this context
Matrix<ElemType>::MultiplyAndWeightedAdd(1.0f / totalNumSamples, samples, false,
ConstOnes(Input(0)->Value().GetNumCols(), 1, samples.GetDeviceId()),
false, (ElemType) m_numSamples / totalNumSamples, m_mean);
ElemType alpha = 1.0f / totalNumSamples;
ElemType beta = (ElemType)m_numSamples / totalNumSamples;
#if 1
size_t rank = DetermineElementwiseTensorRank();
auto input = Input(0)->ValueTensorFor( rank, fr);
auto mean = DataTensorFor(m_mean, rank, FrameRange());
auto temp = DataTensorFor(m_temp, rank, FrameRange());
auto var = DataTensorFor(m_var, rank, FrameRange());
// preserve the old mean value for the next step
temp.AssignCopyOf(mean);
// accumulate the mean
mean.DoCopyOf(beta, input, alpha); // Note: This reduces over samples.
#else
auto& samples = Input(0)->Value();
Matrix<ElemType>::MultiplyAndWeightedAdd(alpha, samples, false,
ConstOnes(Input(0)->Value().GetNumCols(), 1, samples.GetDeviceId()),
false, beta, m_mean);
#endif
// compute the correction term
#if 1
// var += (oldMean - newMean)^2
temp.DoCopyOf(1.0f, mean, -1.0f); // subtract new 'mean' from the old one
var .DoSqrOf (1.0f, temp, 1.0f); // add the square
// var += (input - mean)^2
auto& temp2 = temp; // another temp variable, for which we can reuse the first one
temp2.AssignDifferenceOf(input, mean); // Note: This also reduces over samples.
var.DoSqrOf(beta, temp2, alpha);
#else
// var += (oldMean - newMean)^2
m_temp -= m_mean;
m_temp.AssignElementPowerOf(m_temp, 2);
m_var += m_temp;
m_var += m_temp;
m_temp.AssignDifferenceOf(samples, m_mean);
m_temp.AssignDifferenceOf(Input(0)->Value(), m_mean);
m_temp.AssignElementPowerOf(m_temp, 2);
Matrix<ElemType>::MultiplyAndWeightedAdd(1.0f / totalNumSamples, m_temp, false,
Matrix<ElemType>::MultiplyAndWeightedAdd(alpha, m_temp, false,
ConstOnes(Input(0)->Value().GetNumCols(), 1, samples.GetDeviceId()),
false, (ElemType) m_numSamples / totalNumSamples, m_var);
false, beta, m_var);
#endif
#if NANCHECK
m_var.HasNan("InvStdDev-m_var");
@ -415,6 +430,11 @@ template class InvStdDevNode<double>;
// -----------------------------------------------------------------------
// PerDimMeanVarNormalizationNode (feature, mean, invStdDev)
// Computes
// output = (feature - mean) .* invStdDev
// where mean and invStdDev are meant to be single elements while features
// is minibatch data.
// TODO: Why do we need this? Why not use Plus and ElementTimes?
// -----------------------------------------------------------------------
template <class ElemType>
@ -441,34 +461,24 @@ public:
virtual void /*ComputationNode::*/ ForwardProp(const FrameRange& fr) override
{
#if 1
size_t rank = DetermineElementwiseTensorRank();
auto output = ValueTensorFor(rank, fr);
auto input = Input(0)->ValueTensorFor(rank, fr);
auto mean = Input(1)->ValueTensorFor(rank, fr.AllowBroadcast());
auto invStdDev = Input(2)->ValueTensorFor(rank, fr.AllowBroadcast());
output.AssignDifferenceOf(input, mean); // output = input - mean
output.AssignElementwiseProductOf(output, invStdDev); // output *= invStdDev
#else
// only feature (input0) and output needs to be sliced
auto sliceInput0Value = Input(0)->ValueFor(fr);
auto sliceOutputValue = ValueFor(fr);
auto functionValues = Input(0)->ValueFor(fr);
auto input0 = ValueFor(fr);
const auto& input1 = Input(1)->Value(); // mean
const auto& input2 = Input(2)->Value(); // inv stddev
ForwardPropS(sliceOutputValue, sliceInput0Value, Input(1)->Value(), Input(2)->Value());
}
/*TODO: merge with call site*/ void ForwardPropS(Matrix<ElemType>& functionValues, const Matrix<ElemType>& input0,
const Matrix<ElemType>& input1, const Matrix<ElemType>& input2)
{
#if DUMPOUTPUT
//input0.Print("PerDimMeanVarNormalization-input0");
//input1.Print("PerDimMeanVarNormalization-input1");
//input2.Print("PerDimMeanVarNormalization-input2");
#endif
#if NANCHECK
input0.HasNan("PerDimMeanVarNormalization-input0");
input1.HasNan("PerDimMeanVarNormalization-input1");
input2.HasNan("PerDimMeanVarNormalization-input2");
#endif
functionValues.AssignDifferenceOf(input0, input1);
functionValues.ColumnElementMultiplyWith(input2);
#if NANCHECK
functionValues.HasNan("PerDimMeanVarNormalization");
#endif
#if DUMPOUTPUT
functionValues.Print("PerDimMeanVarNormalizationNode");
#endif
}
@ -477,31 +487,11 @@ public:
Base::Validate(isFinalValidationPass);
InferMBLayoutFromInputsForStandardCase();
if (Input(0)->RequiresPreCompute())
{
LogicError(
"PerDimMeanVarNormalizationNode criterion forbids first input from being a pre-compute node. "
"The first input should be the node whose output should be normalized, and the second and third inputs "
"should be LearnableParameter type or (Mean, InvStdDev) so that the values will be saved.");
}
if (!(Input(1)->OperationName() == OperationNameOf(LearnableParameter) &&
Input(2)->OperationName() == OperationNameOf(LearnableParameter)) &&
!(Input(1)->OperationName() == OperationNameOf(MeanNode) &&
Input(2)->OperationName() == OperationNameOf(InvStdDevNode)))
{
LogicError(
"PerDimMeanVarNormalizationNode criterion requires the last two inputs to be LearnableParameter "
"type or (Mean, InvStdDev) so that the values will be saved.");
}
Input(1)->ValidateInferInputDimsFrom(Input(0)->GetSampleLayout());
Input(2)->ValidateInferInputDimsFrom(Input(0)->GetSampleLayout());
if (isFinalValidationPass)
{
if (!Input(0)->HasMBLayout() || Input(1)->HasMBLayout() || Input(2)->HasMBLayout())
InvalidArgument("PerDimMeanVarNormalizationNode: Inputs must be data, while mean and InvStdDev must be column vectors.");
if (!Input(0)->GetSampleLayout().IsElementwiseCompatibleWith(Input(1)->GetSampleLayout()) || !Input(0)->GetSampleLayout().IsElementwiseCompatibleWith(Input(2)->GetSampleLayout()))
InvalidArgument("PerDimMeanVarNormalizationNode: All inputs should have same sample layout.");
}
@ -515,17 +505,17 @@ template class PerDimMeanVarNormalizationNode<double>;
// -----------------------------------------------------------------------
// PerDimMeanVarDeNormalizationNode (feature, mean, invStdDev)
// Computes
// output = feature ./ invStdDev + mean
// with parameters the same as PerDimMeanVarNormalizationNode.
// TODO: Why do we need this? Why not use Plus and ElementDividedBy?
// -----------------------------------------------------------------------
template <class ElemType>
class PerDimMeanVarDeNormalizationNode : public ComputationNode<ElemType>, public NumInputs<3>
{
typedef ComputationNode<ElemType> Base;
UsingComputationNodeMembersBoilerplate;
static const std::wstring TypeName()
{
return L"PerDimMeanVarDeNormalization";
}
typedef ComputationNode<ElemType> Base; UsingComputationNodeMembersBoilerplate;
static const std::wstring TypeName() { return L"PerDimMeanVarDeNormalization"; }
public:
DeclareConstructorFromConfigWithNumInputs(PerDimMeanVarDeNormalizationNode);
@ -539,44 +529,28 @@ public:
InvalidArgument("PerDimMeanVarDeNormalizationNode should only be called in the evaluation stage. Is any of its descendents a learnable parameter that requires gradient?");
}
// (feature-mean).*InvStdDev
// feature ./ invStdDev + mean
virtual void /*ComputationNode::*/ ForwardProp(const FrameRange& fr) override
{
#if 1
size_t rank = DetermineElementwiseTensorRank();
auto output = ValueTensorFor(rank, fr);
auto input = Input(0)->ValueTensorFor(rank, fr);
auto mean = Input(1)->ValueTensorFor(rank, fr.AllowBroadcast());
auto invStdDev = Input(2)->ValueTensorFor(rank, fr.AllowBroadcast());
output.AssignElementwiseQuotientOf(input, invStdDev); // output = input / invStdDev
output.AssignDifferenceOf(output, mean); // output += mean
#else
// only feature (input0) and output needs to be sliced
auto sliceInput0Value = Input(0)->ValueFor(fr);
auto sliceOutputValue = ValueFor(fr);
auto functionValues = Input(0)->ValueFor(fr);
auto input0 = ValueFor(fr);
const auto& input1 = Input(1)->Value(); // mean
const auto& input2 = Input(2)->Value(); // inv stddev
ForwardPropS(sliceOutputValue, sliceInput0Value, Input(1)->Value(), Input(2)->Value());
}
/*TODO: merge with call site*/ void ForwardPropS(Matrix<ElemType>& functionValues, const Matrix<ElemType>& input0,
const Matrix<ElemType>& input1, const Matrix<ElemType>& input2)
{
#if DUMPOUTPUT
//input0.Print("PerDimMeanVarDeNormalization-input0");
//input1.Print("PerDimMeanVarDeNormalization-input1");
//input2.Print("PerDimMeanVarDeNormalization-input2");
#endif
#if NANCHECK
input0.HasNan("PerDimMeanVarDeNormalization-input0");
input1.HasNan("PerDimMeanVarDeNormalization-input1");
input2.HasNan("PerDimMeanVarDeNormalization-input2");
#endif
// functionValues.AssignDifferenceOf(input0, input1);
// functionValues.ColumnElementMultiplyWith(input2);
// functionValues.AssignDifferenceOf(input0, input0);
// functionValues += input2;
// functionValues.ElementInverse();
// functionValues.ElementMultiplyWith(input0);
functionValues.SetValue(input0);
functionValues.ColumnElementDivideBy(input2);
functionValues += input1;
#if NANCHECK
functionValues.HasNan("PerDimMeanVarDeNormalization");
#endif
#if DUMPOUTPUT
functionValues.Print("PerDimMeanVarDeNormalizationNode");
#endif
}
@ -585,31 +559,11 @@ public:
Base::Validate(isFinalValidationPass);
InferMBLayoutFromInputsForStandardCase();
if (Input(0)->RequiresPreCompute())
{
LogicError(
"PerDimMeanVarDeNormalizationNode criterion forbids first input from being a pre-compute node. "
"The first input should be the node whose output should be de-normalized, and the second and third inputs "
"should be LearnableParameter type or (Mean, InvStdDev) so that the values will be saved.");
}
if (!(Input(1)->OperationName() == OperationNameOf(LearnableParameter) &&
Input(2)->OperationName() == OperationNameOf(LearnableParameter)) &&
!(Input(1)->OperationName() == OperationNameOf(MeanNode) &&
Input(2)->OperationName() == OperationNameOf(InvStdDevNode)))
{
LogicError(
"PerDimMeanVarDeNormalizationNode criterion requires the last two inputs to be "
"LearnableParameter type or (Mean, InvStdDev) so that the values will be saved.");
}
Input(1)->ValidateInferInputDimsFrom(Input(0)->GetSampleLayout());
Input(2)->ValidateInferInputDimsFrom(Input(0)->GetSampleLayout());
if (isFinalValidationPass)
{
if (!Input(0)->HasMBLayout() || Input(1)->HasMBLayout() || Input(2)->HasMBLayout())
InvalidArgument("PerDimMeanVarDeNormalizationNode: Inputs must be data, while mean and InvStdDev must be column vectors.");
if (!Input(0)->GetSampleLayout().IsElementwiseCompatibleWith(Input(1)->GetSampleLayout()) || !Input(0)->GetSampleLayout().IsElementwiseCompatibleWith(Input(2)->GetSampleLayout()))
InvalidArgument("PerDimMeanVarDeNormalizationNode: All inputs should have same sample layout.");
}

2
Source/Math/CommonMatrix.h
Просмотреть файл

@ -78,6 +78,7 @@ enum ElementWiseOperator
opAbs,
opSigmoid,
opTanh,
opSqr,
opSqrt,
opExp,
opLog,
@ -129,6 +130,7 @@ enum ElementWiseOperator
Macro(Abs); \
Macro(Sigmoid); \
Macro(Tanh); \
Macro(Sqr); \
Macro(Sqrt); \
Macro(Exp); \
Macro(Log); \

2
Source/Math/Matrix.h
Просмотреть файл

@ -357,6 +357,8 @@ public:
Matrix<ElemType>& InplaceAbs();
Matrix<ElemType>& AssignAbsOf(const Matrix<ElemType>& a);
// TODO: rename these to InPlaceFloor() and -Ceil() (I never know what it means to truncate a bottom)
// And also document and implement that sparse matrices can only truncate towards 0.
Matrix<ElemType>& InplaceTruncateBottom(const ElemType threshold);
Matrix<ElemType>& AssignTruncateBottomOf(const Matrix<ElemType>& a, const ElemType threshold);
Matrix<ElemType>& InplaceTruncateTop(const ElemType threshold);

13
Source/Math/TensorOps.h
Просмотреть файл

@ -96,6 +96,12 @@ DECL ElemType LinearRectifierDerivative(ElemType z)
return z > 0 ? (ElemType) 1 : 0;
}
template <class ElemType>
DECL ElemType Sqr(ElemType z)
{
return z * z;
}
template <class ElemType>
DECL ElemType Sqrt(ElemType z)
{
@ -143,12 +149,6 @@ DECL ElemType LogAdd(ElemType x, ElemType y)
}
}
template <class ElemType>
DECL ElemType Sqr(ElemType z)
{
return z * z;
}
// IndexElement reindexes a tensor along one dimension.
// For the indexed dimension, the tensor op is prepared by setting 'a' to be broadcasting along the indexed dimension.
// I.e. pa = &a points to the first element (as if index == 0).
@ -188,6 +188,7 @@ DefUnaryOp(Not, !a);
DefUnaryOp(Abs, fabs_(a));
DefUnaryOp(Sigmoid, Sigmoid(a));
DefUnaryOp(Tanh, tanh_(a));
DefUnaryOp(Sqr, Sqr(a));
DefUnaryOp(Sqrt, Sqrt(a));
DefUnaryOp(Exp, exp_(a));
DefUnaryOp(Log, ClippedLog(a));

2
Source/Math/TensorView.h
Просмотреть файл

@ -56,7 +56,7 @@ public:
// c.AssignDiffOf(c,a) means c -= a,
// and c.AddElementwiseProductOf(a, b, 1) means c += a .* b.
// All operators support elementwise in-place operations, i.e. a, b, and c
// may all reference the same underlying SOB, with onee exception:
// may all reference the same underlying SOB, with one exception:
// The output cannot be in-place and inverse-broadcasting at the same time.
// E.g. with c=[10] and a=[10 x 20], c.AssignDiffOf(c,a) will fail.
// In that case, you can use c.AddCopyOf(a,-1).

5
Source/Readers/LMSequenceReader/SequenceReader.cpp
Просмотреть файл

@ -1747,8 +1747,9 @@ size_t BatchSequenceReader<ElemType>::DetermineSequencesToProcess()
// and count tokens
numTokens += m_parser.mSentenceIndex2SentenceInfo[seq].sLen;
}
// if all are already done, we will return sln=0
fprintf(stderr, "DetermineSequencesToProcess: %d sequences of len %d, %d tokens\n", (int) mToProcess.size(), (int) sln, (int) numTokens);
// if all were already done, we will get here with sln=0 and return that
//fprintf(stderr, "DetermineSequencesToProcess: %d sequences of len %d, %d tokens\n", (int) mToProcess.size(), (int) sln, (int) numTokens);
return sln;
}

2
Tests/EndToEndTests/Speech/README_Windows_Debug_commands.txt
Просмотреть файл

@ -54,7 +54,7 @@ COMMAND: currentDirectory=\\storage.ccp.philly.selfhost.corp.microsoft.com\pu
COMMAND: configFile=$(SolutionDir)Examples/Image/MNIST/Config/01_OneHidden.cntk currentDirectory=$(SolutionDir)Tests/EndToEndTests/Image/Data RunDir=$(SolutionDir)Tests/EndToEndTests/RunDir/Image/MNIST_01_OneHidden DataDir=$(SolutionDir)Tests/EndToEndTests/Image/Data ConfigDir=$(SolutionDir)Examples/Image/MNIST/Config OutputDir=$(SolutionDir)Tests/EndToEndTests/RunDir/Image/MNIST_01_OneHidden DeviceId=0 MNISTtrain=[reader=[file=$(SolutionDir)Tests/EndToEndTests/Image/Data/Train.txt]] MNISTtest=[reader=[file=$(SolutionDir)Tests/EndToEndTests/Image/Data/Test.txt]] MNISTtrain=[SGD=[maxEpochs=1]] MNISTtrain=[SGD=[epochSize=100]] MNISTtrain=[reader=[randomize=none]] imageLayout="cudnn" makeMode=false
COMMAND: configFile=$(SolutionDir)Examples/Image/MNIST/Config/02_Convolution.cntk currentDirectory=$(SolutionDir)Tests/EndToEndTests/Image/Data RunDir=$(SolutionDir)Tests/EndToEndTests/RunDir/Image/MNIST_02_Convolution DataDir=$(SolutionDir)Tests/EndToEndTests/Image/Data ConfigDir=$(SolutionDir)Examples/Image/MNIST/Config OutputDir=$(SolutionDir)Tests/EndToEndTests/RunDir/Image/MNIST_02_Convolution DeviceId=0 MNISTtrain=[reader=[file=$(SolutionDir)Tests/EndToEndTests/Image/Data/Train.txt]] MNISTtest=[reader=[file=$(SolutionDir)Tests/EndToEndTests/Image/Data/Test.txt]] MNISTtrain=[SGD=[maxEpochs=1]] MNISTtrain=[SGD=[epochSize=100]] MNISTtrain=[reader=[randomize=none]] imageLayout="cudnn" makeMode=false
COMMAND: configFile=$(SolutionDir)Examples/Image/MNIST/Config/02_Convolution.cntk currentDirectory=$(SolutionDir)Tests/EndToEndTests/Image/Data RunDir=$(SolutionDir)Tests/EndToEndTests/RunDir/Image/MNIST_02_Convolution DataDir=$(SolutionDir)Tests/EndToEndTests/Image/Data ConfigDir=$(SolutionDir)Examples/Image/MNIST/Config OutputDir=$(SolutionDir)Tests/EndToEndTests/RunDir/Image/MNIST_02_Convolution DeviceId=0 train=[reader=[file=$(SolutionDir)Tests/EndToEndTests/Image/Data/Train.txt]] MNISTtest=[reader=[file=$(SolutionDir)Tests/EndToEndTests/Image/Data/Test.txt]] train=[SGD=[maxEpochs=1]] train=[SGD=[epochSize=100]] train=[reader=[randomize=none]] imageLayout="cudnn" makeMode=false
TODO out-of-date:
COMMAND: currentDirectory=$(SolutionDir)ExampleSetups\Image\MNIST configFile=02_Conv.cntk configName=02_Conv