void NDMask::MarkSectionAs(const std::vector<size_t>& sectionOffset, const NDShape& sectionShape, MaskKind maskKind)
{
// TODO: Implement batching of masking operation for masks residing on GPUs to avoid making
// GPU invocations for each MaskSection call.
if (sectionOffset.size() > m_maskShape.Rank())
LogicError("NDMask::MaskSection: The sectionOffset cannot have dimensionality higher than the rank of 'this' mask");
if (sectionShape.Rank() > m_maskShape.Rank())
LogicError("NDMask::MaskSection: The section shape cannot have an axes count higher than the rank of 'this' mask");
std::vector<size_t> offset(m_maskShape.Rank(), 0);
for (size_t i = 0; i < sectionOffset.size(); ++i)
offset[i] = sectionOffset[i];
NDShape shape = sectionShape.AppendShape(NDShape(m_maskShape.Rank() - sectionShape.Rank(), NDShape::InferredDimension));
auto maskMatrix = GetMatrix();
size_t rowOffset = offset[0];
size_t colOffset = offset[1];
size_t sliceRowLength = (shape[0] != NDShape::InferredDimension) ? shape[0] : (maskMatrix->GetNumRows() - rowOffset);
size_t sliceColLength = (shape[1] != NDShape::InferredDimension) ? shape[1] : (maskMatrix->GetNumCols() - colOffset);
if ((rowOffset == 0) && (sliceRowLength == maskMatrix->GetNumRows()))
maskMatrix->ColumnSlice(colOffset, sliceColLength).SetValue((char)maskKind);
else
{
// Since Matrix does not support strides in the row dimension, we will need to create separate slices for each column
for (size_t i = colOffset; i < (colOffset + sliceColLength); ++i)
{
auto column = maskMatrix->ColumnSlice(i, 1);
column.Reshape(1, maskMatrix->GetNumRows());
column.ColumnSlice(rowOffset, sliceRowLength).SetValue((char)maskKind);
}
}
}
void CheckValue(const ValuePtr testValue, const size_t dimension, const vector<vector<size_t>>& expectedData, const vector<size_t>& seqLenList, const vector<bool>& seqStartFlags = {})
{
// Check parameters
BOOST_TEST(expectedData.size() == seqLenList.size(), "Parameter error: the sequence number in the exepected data and sequence list does not match.");
for (size_t i = 0; i < expectedData.size(); i++)
{
if (expectedData[i].size() != seqLenList[i])
{
ReportFailure("Parameter erroe: the number of data for sequence %" PRIu64 " in the expected data does not match. Expected: %" PRIu64 ", actual: %" PRIu64 ".",
i, seqLenList[i], expectedData[i].size());
}
}
// Check shape
NDShape shape = testValue->Shape();
size_t valueRank = shape.Rank();
if (valueRank < 2 || valueRank > 3 || shape[0] != dimension)
{
ReportFailure("The shape of the value does not match\n");
}
size_t numOfSequences = valueRank == 2 ? 1 : shape[2];
if (numOfSequences != expectedData.size())
{
ReportFailure("The sequence number in the Value does not match. Expected: %" PRIu64 ", actual: %" PRIu64 ".", expectedData.size(), numOfSequences);
}
CheckMask(testValue, seqLenList, seqStartFlags);
// Get data from Value
vector<ElementType> outputData(shape.TotalSize());
NDArrayViewPtr arrayOutput = MakeSharedObject<NDArrayView>(shape, outputData, false);
arrayOutput->CopyFrom(*testValue->Data());
size_t maxSeqLen = *max_element(seqLenList.begin(), seqLenList.end());
size_t oIndex = 0;
for (size_t seq = 0; seq < seqLenList.size(); seq++)
{
size_t seqLen = seqLenList[seq];
for (size_t sample = 0; sample < seqLen; sample++)
{
for (size_t c = 0; c < dimension; c++, oIndex++)
{
if (outputData[oIndex] != 0)
{
if (outputData[oIndex] != 1)
{
ReportFailure("OneHot vector contains value other than 0 and 1 at seqNo=%" PRIu64 " sampleNo=%" PRIu64 " position=%" PRIu64 "\n", seq, sample, c);
}
if (c != expectedData[seq][sample])
{
ReportFailure("OneHot Index does match at seqNo=%" PRIu64 ", sampleNo=%" PRIu64 ", expected: %" PRIu64 ", actual: %" PRIu64 "\n", seq, sample, expectedData[seq][sample], c);
}
}
}
}
// Skip mask data
oIndex += (maxSeqLen - seqLen) * dimension;
}
}
NDMask::NDMask(const NDShape& shape, const DeviceDescriptor& device)
: NDMask(shape, AllocateMatrix(shape, device))
{
if (shape.Rank() > 2)
LogicError("NDMask instances with more than 2 axes are currently unsupported");
Clear();
}
/*static*/ NDArrayViewPtr Variable::CreateValueFromParameterInitializer(const NDShape& shape, const ParameterInitializer& initConfig, const DeviceDescriptor& device)
{
auto dataType = AsDataType<ElementType>();
auto value = MakeSharedObject<NDArrayView>(dataType, shape, device);
auto valueMatrix = value->template GetWritableMatrix<ElementType>();
auto initializerType = initConfig[InitializerTypeAttributeName].Value<std::wstring>();
if (initializerType == Microsoft::MSR::CNTK::ConstantInitializerTypeName)
{
auto constantInitValue = initConfig[ValueAttributeName].Value<double>();
valueMatrix->SetValue((ElementType)constantInitValue);
}
else if (initializerType == Microsoft::MSR::CNTK::BilinearInitializerTypeName)
{
auto kernelWidth = initConfig[KernelWidthAttributeName].Value<size_t>();
auto kernelHeight = initConfig[KernelHeightAttributeName].Value<size_t>();
Microsoft::MSR::CNTK::LearnableParameter<ElementType>::InitBilinear(*valueMatrix, AsTensorShape(shape), kernelWidth, kernelHeight, AsCNTKImplDeviceId(device));
}
else
{
auto randomSeed = (unsigned long)initConfig[RandomSeedAttributeName].Value<size_t>();
if (randomSeed == SentinelValueForAutoSelectRandomSeed)
randomSeed = s_currentRandomSeed++;
auto scale = initConfig[ScaleAttributeName].Value<double>();
int outputRank = DefaultParamInitOutputRank, filterRank = DefaultParamInitFilterRank;
if (initializerType != Microsoft::MSR::CNTK::UniformInitializerTypeName)
{
outputRank = initConfig[OutputRankAttributeName].Value<int>();
filterRank = initConfig[FilterRankAttributeName].Value<int>();
if (outputRank == SentinelValueForInferParamInitRank)
outputRank = DefaultParamInitOutputRank;
if (filterRank == SentinelValueForInferParamInitRank)
filterRank = DefaultParamInitFilterRank;
if ((filterRank + outputRank) > shape.Rank())
InvalidArgument("Sum of filter rank (%d) and output rank (%d) of the parameter initializer cannot exceed the Parameter's rank(%d)", filterRank, outputRank, (int)shape.Rank());
}
Microsoft::MSR::CNTK::LearnableParameter<ElementType>::InitRandom(*valueMatrix, AsTensorShape(shape), initializerType, randomSeed, (ElementType)scale,
filterRank, outputRank, /*initOnCPUOnly=*/true,
AsCNTKImplDeviceId(device));
}
return value;
}
void CheckValue(const ValuePtr testValue, const NDShape& sampleShape, const vector<vector<ElementType>>& expectedData, const vector<size_t>& seqLenList, const vector<bool>& seqStartFlags = {})
{
size_t sampleSize = sampleShape.TotalSize();
// Check parameters
BOOST_TEST(expectedData.size() == seqLenList.size(), "Parameter error: the sequence number in the exepected data and sequence list does not match.");
for (size_t i = 0; i < expectedData.size(); i++)
{
if (expectedData[i].size() != seqLenList[i] * sampleSize)
{
ReportFailure("Parameter erroe: the number of data for sequence %" PRIu64 " in the expected data does not match. Expected: %" PRIu64 ", actual: %" PRIu64 ".",
i, seqLenList[i] * sampleSize, expectedData[i].size());
}
}
// Check shape
auto valueRank = testValue->Shape().Rank();
auto sampleRank = sampleShape.Rank();
auto shapeIsCorrect = !((valueRank < sampleRank + 1) || (valueRank > sampleRank + 2) || (sampleShape != testValue->Shape().SubShape(0, sampleRank)));
BOOST_TEST(shapeIsCorrect, "The Value does not have the expected shape.");
size_t numOfSequences;
if (valueRank == sampleShape.Rank() + 1)
{
// no batch axis, only sequence axis
numOfSequences = 1;
}
else
{
assert(valueRank == sampleShape.Rank() + 2);
numOfSequences = testValue->Shape()[valueRank - 1];
}
if (numOfSequences != expectedData.size())
{
ReportFailure("The sequence number in the Value does not match. Expected: %" PRIu64 ", actual: %" PRIu64 ".", expectedData.size(), numOfSequences);
}
CheckMask(testValue, seqLenList, seqStartFlags);
// Get data from Value
vector<ElementType> outputData(testValue->Shape().TotalSize());
NDArrayViewPtr arrayOutput = MakeSharedObject<NDArrayView>(testValue->Shape(), outputData, false);
arrayOutput->CopyFrom(*testValue->Data());
size_t maxSeqLen = *max_element(seqLenList.begin(), seqLenList.end());
size_t oIndex = 0;
for (size_t seq = 0; seq < seqLenList.size(); seq++)
{
size_t seqLen = seqLenList[seq];
for (size_t sIndex = 0; sIndex < seqLen * sampleSize; sIndex++, oIndex++)
{
if (expectedData[seq][sIndex] != outputData[oIndex])
{
ReportFailure("Data does match at position %" PRIu64 ", expected: %f, actual: %f\n", oIndex, expectedData[seq][sIndex], outputData[oIndex]);
}
}
// Skip mask data
oIndex += (maxSeqLen - seqLen) * sampleSize;
}
}