Merge branch 'master' into qiwye/multiverso
This commit is contained in:
Qiwei Ye
Коммит
c7bfebe740
2
Makefile
2
Makefile
|
|
@ -423,7 +423,7 @@ ALL += $(IMAGEREADER)
|
|||
SRC+=$(IMAGEREADER_SRC)
|
||||
|
||||
INCLUDEPATH += $(OPENCV_PATH)/include
|
||||
LIBPATH += $(OPENCV_PATH)/release/lib
|
||||
LIBPATH += $(OPENCV_PATH)/lib
|
||||
|
||||
$(IMAGEREADER): $(IMAGEREADER_OBJ) | $(CNTKMATH_LIB)
|
||||
@echo $(SEPARATOR)
|
||||
|
|
|
|||
|
|
@ -501,6 +501,22 @@ public:
|
|||
return (size_t) location;
|
||||
}
|
||||
|
||||
// get begin and end location (first offset after last element), for validation purposes
|
||||
pair<ptrdiff_t, ptrdiff_t> GetLocationRange() const
|
||||
{
|
||||
auto result = make_pair(m_offset, m_offset);
|
||||
for (size_t k = 0; k < size(); k++)
|
||||
{
|
||||
ptrdiff_t step = (ptrdiff_t)(m_dims[k] - 1) * m_strides[k];
|
||||
if (m_strides[k] > 0) // strides may be negative
|
||||
result.second += step;
|
||||
else
|
||||
result.first += step;
|
||||
}
|
||||
result.second++; // max --> end
|
||||
return result;
|
||||
}
|
||||
|
||||
// helpers for tensor operations
|
||||
bool CanFlatten(size_t k) const // can dims k and k-1 be flattened into a single vector? (do they form a matrix without stride)
|
||||
{
|
||||
|
|
|
|||
|
|
@ -413,10 +413,33 @@ void fprintfOrDie(FILE* f, const char* fmt, ...)
|
|||
void fflushOrDie(FILE* f)
|
||||
{
|
||||
int rc = fflush(f);
|
||||
|
||||
if (rc != 0)
|
||||
{
|
||||
RuntimeError("error flushing to file: %s", strerror(errno));
|
||||
}
|
||||
|
||||
int fd = fileno(f);
|
||||
|
||||
if (fd == -1)
|
||||
{
|
||||
RuntimeError("unable to convert file handle to file descriptor: %s", strerror(errno));
|
||||
}
|
||||
|
||||
// Ensure that all data is synced before returning from this function
|
||||
#ifdef _WIN32
|
||||
if (!FlushFileBuffers((HANDLE)_get_osfhandle(fd)))
|
||||
{
|
||||
RuntimeError("error syncing to file: %d", (int) ::GetLastError());
|
||||
}
|
||||
#else
|
||||
rc = fsync(fd);
|
||||
|
||||
if (rc != 0)
|
||||
{
|
||||
RuntimeError("error syncing to file: %s", strerror(errno));
|
||||
}
|
||||
#endif
|
||||
}
|
||||
|
||||
// ----------------------------------------------------------------------------
|
||||
|
|
|
|||
|
|
@ -185,6 +185,7 @@ void ComputationNode<ElemType>::ValidateInferInputDimsFrom(const TensorShape& ot
|
|||
|
||||
// determine the sample tensor dimension to use for operations based on output and all inputs
|
||||
// 'Sample tensor' means we only consider single samples. If we have an MBLayout, that is the sample layout of a single matrix column.
|
||||
// TODO: Turn rank into a member variable, and call this method once in validation (currently called for every single ForwardProp/BackpropTo()).
|
||||
size_t ComputationNodeBase::DetermineElementwiseTensorRank() const
|
||||
{
|
||||
// determine largest tensor dimension amongst the sample shapes of output and the selected inputs
|
||||
|
|
|
|||
|
|
@ -81,7 +81,7 @@ struct /*interface*/ IComputationNode
|
|||
virtual void BackpropTo(const size_t inputIndex, const FrameRange&) = 0; // backprop gradient into one of the inputs
|
||||
virtual void EndBackprop() = 0; // called after last iteration step of ComputeGradient()
|
||||
|
||||
// --- these are meant to be overridden by ControlFlowNodes
|
||||
// --- this is meant to be overridden by ControlFlowNodes
|
||||
|
||||
virtual void Backprop(const FrameRange& fr, bool childrenInThisLoop, bool childrenInOuterLoop) = 0;
|
||||
|
||||
|
|
@ -491,10 +491,11 @@ public:
|
|||
protected:
|
||||
|
||||
size_t DetermineElementwiseTensorRank() const; // determine tensor rank when considering all inputs with padding
|
||||
TensorShape GetTensorSliceFor(size_t rank, const FrameRange& fr) const; // form tensor shape of the slice referenced by FrameRange
|
||||
|
||||
public:
|
||||
|
||||
TensorShape GetTensorSliceFor(size_t rank, const FrameRange& fr) const; // form tensor shape of the slice referenced by FrameRange. Public since nodes may call it for their inputs.
|
||||
|
||||
// -----------------------------------------------------------------------
|
||||
// inputs
|
||||
// -----------------------------------------------------------------------
|
||||
|
|
|
|||
|
|
@ -209,21 +209,6 @@ public:
|
|||
{
|
||||
}
|
||||
|
||||
virtual void /*ComputationNode::*/ BackpropTo(const size_t /*inputIndex*/, const FrameRange& fr) override
|
||||
{
|
||||
Input(0)->GradientFor(fr.WithLayout(Input(0)->GetMBLayout())) += GradientFor(fr);
|
||||
// TODO: Once we do in-place, the above must include a copy-to-self check (pay special attention to adding vs. copying).
|
||||
}
|
||||
|
||||
virtual bool OutputUsedInComputingInputNodesGradients() const override
|
||||
{
|
||||
return false;
|
||||
}
|
||||
virtual bool InputUsedInComputingInputNodesGradients(size_t /*childIndex*/) const override
|
||||
{
|
||||
return false;
|
||||
}
|
||||
|
||||
virtual void /*ComputationNode::*/ ForwardProp(const FrameRange& fr) override
|
||||
{
|
||||
// enforce compatibility of 'dataInput' with 'layoutInput'
|
||||
|
|
@ -239,6 +224,15 @@ public:
|
|||
// TODO: Once we do in-place, the above must include a copy-to-self check (either here or inside the matrix lib).
|
||||
}
|
||||
|
||||
virtual void /*ComputationNode::*/ BackpropTo(const size_t /*inputIndex*/, const FrameRange& fr) override
|
||||
{
|
||||
Input(0)->GradientFor(fr.WithLayout(Input(0)->GetMBLayout())) += GradientFor(fr);
|
||||
// TODO: Once we do in-place, the above must include a copy-to-self check (pay special attention to adding vs. copying).
|
||||
}
|
||||
|
||||
virtual bool OutputUsedInComputingInputNodesGradients() const override { return false; }
|
||||
virtual bool InputUsedInComputingInputNodesGradients(size_t /*childIndex*/) const override { return false; }
|
||||
|
||||
virtual void /*ComputationNodeBase::*/ Validate(bool isFinalValidationPass) override
|
||||
{
|
||||
Base::Validate(isFinalValidationPass);
|
||||
|
|
@ -256,8 +250,7 @@ template class ReconcileMBLayoutNode<double>;
|
|||
|
||||
// -----------------------------------------------------------------------
|
||||
// RowSliceNode (input)
|
||||
// this node extracts part of the input by rows as the output
|
||||
// it has to be continuous segments of rows since each column is treated as one sample
|
||||
// This node extracts a slice of the first tensor dimension (row).
|
||||
// -----------------------------------------------------------------------
|
||||
|
||||
template <class ElemType>
|
||||
|
|
@ -277,6 +270,7 @@ public:
|
|||
m_sliceHeight(numRows)
|
||||
{
|
||||
}
|
||||
|
||||
RowSliceNode(const ScriptableObjects::IConfigRecordPtr configp)
|
||||
: RowSliceNode(configp->Get(L"deviceId"), L"<placeholder>", configp->Get(L"startIndex"), configp->Get(L"numRows"))
|
||||
{
|
||||
|
|
@ -292,58 +286,62 @@ public:
|
|||
node->m_sliceHeight = m_sliceHeight;
|
||||
}
|
||||
|
||||
virtual void Save(File& fstream) const override
|
||||
{
|
||||
Base::Save(fstream);
|
||||
fstream << m_startIndex << m_sliceHeight;
|
||||
}
|
||||
|
||||
virtual void Load(File& fstream, size_t modelVersion) override
|
||||
{
|
||||
Base::Load(fstream, modelVersion);
|
||||
fstream >> m_startIndex >> m_sliceHeight;
|
||||
}
|
||||
|
||||
virtual void /*ComputationNode::*/ BackpropTo(const size_t /*inputIndex*/, const FrameRange& fr) override
|
||||
virtual void Save(File& fstream) const override
|
||||
{
|
||||
Input(0)->GradientFor(fr).AddToRowSliceValuesOf(GradientFor(fr), m_startIndex, m_sliceHeight);
|
||||
Base::Save(fstream);
|
||||
fstream << m_startIndex << m_sliceHeight;
|
||||
}
|
||||
|
||||
virtual bool OutputUsedInComputingInputNodesGradients() const override
|
||||
private:
|
||||
|
||||
// determine the tensor shape that represents slice of the input that we are taking
|
||||
TensorShape GetInputSlice(size_t rank, const FrameRange & fr) const
|
||||
{
|
||||
// The RowSliceNode does not require its output value for computing
|
||||
// the gradients of its input nodes
|
||||
return false;
|
||||
auto inputSlice = Input(0)->GetTensorSliceFor(rank, fr); // input must be narrowed down
|
||||
inputSlice.NarrowTo(0, m_startIndex, m_startIndex + m_sliceHeight);
|
||||
return inputSlice;
|
||||
}
|
||||
|
||||
virtual bool InputUsedInComputingInputNodesGradients(size_t childIndex) const override
|
||||
{
|
||||
// The RowSliceNode does not require any of it's input's values for computing
|
||||
// the gradients of its input nodes
|
||||
UNREFERENCED_PARAMETER(childIndex);
|
||||
return false;
|
||||
}
|
||||
public:
|
||||
|
||||
virtual void /*ComputationNode::*/ ForwardProp(const FrameRange& fr) override
|
||||
{
|
||||
ValueFor(fr).AssignRowSliceValuesOf(Input(0)->ValueFor(fr), m_startIndex, m_sliceHeight);
|
||||
size_t rank = DetermineElementwiseTensorRank();
|
||||
auto output = ValueTensorFor(rank, fr);
|
||||
let input = TensorView<ElemType>(Input(0)->Value(), GetInputSlice(rank, fr.AllowBroadcast()));
|
||||
output.AssignCopyOf(input);
|
||||
}
|
||||
|
||||
virtual void /*ComputationNode::*/ BackpropTo(const size_t /*inputIndex*/, const FrameRange& fr) override
|
||||
{
|
||||
size_t rank = DetermineElementwiseTensorRank();
|
||||
let outputGrad = GradientTensorFor(rank, fr);
|
||||
auto inputGrad = TensorView<ElemType>(Input(0)->Gradient(), GetInputSlice(rank, fr));
|
||||
inputGrad.AddCopyOf(outputGrad);
|
||||
}
|
||||
|
||||
virtual bool OutputUsedInComputingInputNodesGradients() const override { return false; }
|
||||
virtual bool InputUsedInComputingInputNodesGradients(size_t /*childIndex*/) const override { return false; }
|
||||
|
||||
virtual void /*ComputationNodeBase::*/ Validate(bool isFinalValidationPass) override
|
||||
{
|
||||
Base::Validate(isFinalValidationPass);
|
||||
InferMBLayoutFromInputsForStandardCase();
|
||||
|
||||
if (isFinalValidationPass && Input(0)->GetSampleMatrixNumRows() < m_startIndex + m_sliceHeight)
|
||||
RuntimeError("%ls %ls operation: m_startIndex + m_sliceHeight exceeds number of rows in the input.", NodeName().c_str(), OperationName().c_str());
|
||||
auto sampleLayout = Input(0)->GetSampleLayout();
|
||||
if (isFinalValidationPass && sampleLayout[0] < m_startIndex + m_sliceHeight)
|
||||
RuntimeError("%ls %ls operation: m_startIndex + m_sliceHeight (%d) exceeds number of rows in the input ([%s]).", NodeName().c_str(), OperationName().c_str(), (int)(m_startIndex + m_sliceHeight), string(sampleLayout).c_str());
|
||||
|
||||
// RowSlice cannot slice tensors.
|
||||
// TODO: Create a TensorSlice operation, or just Slice.
|
||||
if (isFinalValidationPass && !Input(0)->GetSampleLayout().IsColumnVector()
|
||||
&& !Input(0)->GetSampleLayout().IsVectorStoredAsImage() // legacy
|
||||
)
|
||||
RuntimeError("%ls %ls operation: Input must be a vector, tensor shape [%s] not allowed.", NodeName().c_str(), OperationName().c_str(), string(Input(0)->GetSampleLayout()).c_str());
|
||||
SetDims(TensorShape(m_sliceHeight), HasMBLayout());
|
||||
if (sampleLayout[0] >= m_startIndex + m_sliceHeight) // (this guards against failing an out-of-bounds error if not isFinalValidationPass)
|
||||
sampleLayout.NarrowTo(0, m_startIndex, m_startIndex + m_sliceHeight);
|
||||
|
||||
SetDims(TensorShape(sampleLayout.GetDims()), HasMBLayout());
|
||||
}
|
||||
|
||||
private:
|
||||
|
|
@ -604,24 +602,6 @@ public:
|
|||
{
|
||||
}
|
||||
|
||||
virtual void Validate(bool isFinalValidationPass) override
|
||||
{
|
||||
Base::Validate(isFinalValidationPass);
|
||||
m_pMBLayout = nullptr;
|
||||
|
||||
if (isFinalValidationPass && Input(0)->HasMBLayout())
|
||||
InvalidArgument("%ls %ls operation cannot operate on minibatch data (which have a layout)", NodeName().c_str(), OperationName().c_str());
|
||||
|
||||
size_t dim = Input(0)->GetAsMatrixNumCols();
|
||||
if (isFinalValidationPass && dim != Input(0)->GetAsMatrixNumRows())
|
||||
InvalidArgument("%ls %ls operation requires a square matrix as its input.", NodeName().c_str(), OperationName().c_str());
|
||||
|
||||
if (Input(0)->HasSampleLayout())
|
||||
fprintf(stderr, "WARNING: Diagonal operation cannot inherit image size information from its child. Image size info is lost.\n");
|
||||
|
||||
SetDims(TensorShape(1, dim), false);
|
||||
}
|
||||
|
||||
virtual void /*ComputationNodeNonLooping::*/ ForwardPropNonLooping() override
|
||||
{
|
||||
Input(0)->ValueAsMatrix().AssignDiagonalValuesTo(ValueAsMatrix()); // TODO: use tensor lib; this is a stride operation
|
||||
|
|
@ -646,19 +626,25 @@ public:
|
|||
inputGradientValues.SetDiagonalValue(diag);
|
||||
}
|
||||
|
||||
virtual bool OutputUsedInComputingInputNodesGradients() const override
|
||||
{
|
||||
// The DiagonalNode does not require its output value for computing
|
||||
// the gradients of its input nodes
|
||||
return false;
|
||||
}
|
||||
virtual bool OutputUsedInComputingInputNodesGradients() const override { return false; }
|
||||
virtual bool InputUsedInComputingInputNodesGradients(size_t /*childIndex*/) const override { return false; }
|
||||
|
||||
virtual bool InputUsedInComputingInputNodesGradients(size_t childIndex) const override
|
||||
virtual void Validate(bool isFinalValidationPass) override
|
||||
{
|
||||
// The DiagonalNode does not require any of it's input's values for computing
|
||||
// the gradients of its input nodes
|
||||
UNREFERENCED_PARAMETER(childIndex);
|
||||
return false;
|
||||
Base::Validate(isFinalValidationPass);
|
||||
m_pMBLayout = nullptr;
|
||||
|
||||
if (isFinalValidationPass && Input(0)->HasMBLayout())
|
||||
InvalidArgument("%ls %ls operation cannot operate on minibatch data (which have a layout)", NodeName().c_str(), OperationName().c_str());
|
||||
|
||||
size_t dim = Input(0)->GetAsMatrixNumCols();
|
||||
if (isFinalValidationPass && dim != Input(0)->GetAsMatrixNumRows())
|
||||
InvalidArgument("%ls %ls operation requires a square matrix as its input.", NodeName().c_str(), OperationName().c_str());
|
||||
|
||||
if (Input(0)->HasSampleLayout())
|
||||
fprintf(stderr, "WARNING: Diagonal operation cannot inherit image size information from its child. Image size info is lost.\n");
|
||||
|
||||
SetDims(TensorShape(1, dim), false);
|
||||
}
|
||||
};
|
||||
|
||||
|
|
@ -839,13 +825,6 @@ public:
|
|||
}
|
||||
}
|
||||
|
||||
virtual void Save(File& fstream) const override
|
||||
{
|
||||
Base::Save(fstream);
|
||||
fstream << m_numTargetRows;
|
||||
m_targetImageLayout.Save(fstream);
|
||||
}
|
||||
|
||||
virtual void Load(File& fstream, size_t modelVersion) override
|
||||
{
|
||||
Base::Load(fstream, modelVersion);
|
||||
|
|
@ -853,6 +832,13 @@ public:
|
|||
m_targetImageLayout.Load(fstream, /*acceptLegacyFormat=*/true);
|
||||
}
|
||||
|
||||
virtual void Save(File& fstream) const override
|
||||
{
|
||||
Base::Save(fstream);
|
||||
fstream << m_numTargetRows;
|
||||
m_targetImageLayout.Save(fstream);
|
||||
}
|
||||
|
||||
virtual void /*IComputationNode::*/ PrintSelfBeforeValidation() const override
|
||||
{
|
||||
fprintf(stderr, "\nValidating --> %ls = %ls", NodeName().c_str(), OperationName().c_str());
|
||||
|
|
@ -871,56 +857,6 @@ public:
|
|||
// BUGBUG: This interpretaion as image dims is only correct for the 'legacy format, not for cudnn.
|
||||
}
|
||||
|
||||
virtual void /*ComputationNodeBase::*/ Validate(bool isFinalValidationPass) override
|
||||
{
|
||||
Base::Validate(isFinalValidationPass);
|
||||
if (factor() == 1) // canonical case: keeps the MBLayout(e.g. only changing the TensorShape)
|
||||
m_pMBLayout = Input(0)->GetMBLayout();
|
||||
else if (Input(0)->HasMBLayout())
|
||||
{
|
||||
if (!m_pMBLayout)
|
||||
m_pMBLayout = make_shared<MBLayout>(); // mini-batch data: this generates a new layout
|
||||
}
|
||||
else
|
||||
assert(!m_pMBLayout); // reshaping non-mini-batch data
|
||||
|
||||
size_t newCols = 1; // dummy
|
||||
if (!m_pMBLayout)
|
||||
{
|
||||
size_t rows = Input(0)->GetAsMatrixNumRows(), cols = Input(0)->GetAsMatrixNumCols();
|
||||
newCols = cols * rows / m_numTargetRows;
|
||||
if (isFinalValidationPass)
|
||||
{
|
||||
if ((m_numTargetRows > rows && m_numTargetRows % rows != 0) || // grouping columns
|
||||
(m_numTargetRows < rows && rows % m_numTargetRows != 0)) // splitting columns
|
||||
InvalidArgument("%ls %ls operation: output row dimension %d is not an integer multiple or divisor of input dimension %d", NodeName().c_str(), OperationName().c_str(), (int) m_numTargetRows, (int) rows);
|
||||
if (rows * cols != m_numTargetRows * newCols)
|
||||
LogicError("%ls %ls operation: unexpected dimension mismatch", NodeName().c_str(), OperationName().c_str());
|
||||
}
|
||||
}
|
||||
|
||||
// patch up m_targetImageLayout, which was originally a construction parameter
|
||||
InferTargetSampleLayout();
|
||||
|
||||
// setting any dimension to 0 means lose the tensor, flatten to vector
|
||||
if (m_targetImageLayout.GetNumElements() == 0)
|
||||
{
|
||||
if (Input(0)->HasSampleLayout())
|
||||
fprintf(stderr, "WARNING: Reshape operation cannot inherit image size information from its child. Image size info is lost.\n");
|
||||
// TODO: We need to decide what reshaping means in presence of a tensor.
|
||||
if (HasMBLayout())
|
||||
SetDims(TensorShape(m_numTargetRows), true);
|
||||
else
|
||||
SetDims(TensorShape(m_numTargetRows, newCols), false);
|
||||
}
|
||||
else
|
||||
{
|
||||
if (m_numTargetRows != m_targetImageLayout.GetNumElements())
|
||||
LogicError("LegacyReshapeNode: InferTargetSampleLayout() computed a sample layout [%s] that mismatches m_numTargetRows %d.", string(m_targetImageLayout).c_str(), (int) m_numTargetRows);
|
||||
SetDims(m_targetImageLayout, HasMBLayout());
|
||||
}
|
||||
}
|
||||
|
||||
// TODO: Clarify/resolve the semantic overlap between BeginForwardProp() and UpdateFunctionMBSize().
|
||||
virtual void /*IComputationNode::*/ BeginForwardProp() override
|
||||
{
|
||||
|
|
@ -1002,19 +938,57 @@ public:
|
|||
}
|
||||
}
|
||||
|
||||
virtual bool OutputUsedInComputingInputNodesGradients() const override
|
||||
virtual bool OutputUsedInComputingInputNodesGradients() const override { return false; }
|
||||
virtual bool InputUsedInComputingInputNodesGradients(size_t /*childIndex*/) const override { return false; }
|
||||
|
||||
virtual void /*ComputationNodeBase::*/ Validate(bool isFinalValidationPass) override
|
||||
{
|
||||
// The LegacyReshapeNode does not require its output value for computing
|
||||
// the gradients of its input nodes
|
||||
return false;
|
||||
Base::Validate(isFinalValidationPass);
|
||||
if (factor() == 1) // canonical case: keeps the MBLayout(e.g. only changing the TensorShape)
|
||||
m_pMBLayout = Input(0)->GetMBLayout();
|
||||
else if (Input(0)->HasMBLayout())
|
||||
{
|
||||
if (!m_pMBLayout)
|
||||
m_pMBLayout = make_shared<MBLayout>(); // mini-batch data: this generates a new layout
|
||||
}
|
||||
else
|
||||
assert(!m_pMBLayout); // reshaping non-mini-batch data
|
||||
|
||||
size_t newCols = 1; // dummy
|
||||
if (!m_pMBLayout)
|
||||
{
|
||||
size_t rows = Input(0)->GetAsMatrixNumRows(), cols = Input(0)->GetAsMatrixNumCols();
|
||||
newCols = cols * rows / m_numTargetRows;
|
||||
if (isFinalValidationPass)
|
||||
{
|
||||
if ((m_numTargetRows > rows && m_numTargetRows % rows != 0) || // grouping columns
|
||||
(m_numTargetRows < rows && rows % m_numTargetRows != 0)) // splitting columns
|
||||
InvalidArgument("%ls %ls operation: output row dimension %d is not an integer multiple or divisor of input dimension %d", NodeName().c_str(), OperationName().c_str(), (int) m_numTargetRows, (int) rows);
|
||||
if (rows * cols != m_numTargetRows * newCols)
|
||||
LogicError("%ls %ls operation: unexpected dimension mismatch", NodeName().c_str(), OperationName().c_str());
|
||||
}
|
||||
}
|
||||
|
||||
virtual bool InputUsedInComputingInputNodesGradients(size_t childIndex) const override
|
||||
// patch up m_targetImageLayout, which was originally a construction parameter
|
||||
InferTargetSampleLayout();
|
||||
|
||||
// setting any dimension to 0 means lose the tensor, flatten to vector
|
||||
if (m_targetImageLayout.GetNumElements() == 0)
|
||||
{
|
||||
// The LegacyReshapeNode does not require any of it's input's values for computing
|
||||
// the gradients of its input nodes
|
||||
UNREFERENCED_PARAMETER(childIndex);
|
||||
return false;
|
||||
if (Input(0)->HasSampleLayout())
|
||||
fprintf(stderr, "WARNING: Reshape operation cannot inherit image size information from its child. Image size info is lost.\n");
|
||||
// TODO: We need to decide what reshaping means in presence of a tensor.
|
||||
if (HasMBLayout())
|
||||
SetDims(TensorShape(m_numTargetRows), true);
|
||||
else
|
||||
SetDims(TensorShape(m_numTargetRows, newCols), false);
|
||||
}
|
||||
else
|
||||
{
|
||||
if (m_numTargetRows != m_targetImageLayout.GetNumElements())
|
||||
LogicError("LegacyReshapeNode: InferTargetSampleLayout() computed a sample layout [%s] that mismatches m_numTargetRows %d.", string(m_targetImageLayout).c_str(), (int) m_numTargetRows);
|
||||
SetDims(m_targetImageLayout, HasMBLayout());
|
||||
}
|
||||
}
|
||||
|
||||
private:
|
||||
|
|
|
|||
|
|
@ -312,13 +312,10 @@ public:
|
|||
|
||||
protected:
|
||||
void Clear()
|
||||
{
|
||||
if (m_matrixName != nullptr)
|
||||
{
|
||||
delete[] m_matrixName;
|
||||
m_matrixName = nullptr;
|
||||
}
|
||||
}
|
||||
|
||||
protected:
|
||||
size_t m_numRows;
|
||||
|
|
@ -330,6 +327,7 @@ protected:
|
|||
ElemType* m_pArray;
|
||||
mutable DEVICEID_TYPE m_computeDevice; // current GPU device Id or CPUDEVICE
|
||||
size_t m_nz; // Number of non-zero elements for sparse matrices (unused in other formats)
|
||||
wchar_t* m_matrixName;
|
||||
wchar_t* m_matrixName; // TODO: Use std::wstring?
|
||||
};
|
||||
|
||||
} } }
|
||||
|
|
|
|||
|
|
@ -94,10 +94,23 @@ const char* CudaErrString<cudaError_t>(cudaError_t x)
|
|||
return cudaGetErrorString(x);
|
||||
}
|
||||
template <>
|
||||
const char* CudaErrString<cublasStatus_t>(cublasStatus_t)
|
||||
const char* CudaErrString<cublasStatus_t>(cublasStatus_t e)
|
||||
{
|
||||
cudaDeviceSynchronize();
|
||||
return "(see cublas_api.h & look for cublasStatus_t or CUBLAS_STATUS_xxx)";
|
||||
switch (e)
|
||||
{
|
||||
case CUBLAS_STATUS_SUCCESS: return "CUBLAS_STATUS_SUCCESS";
|
||||
case CUBLAS_STATUS_NOT_INITIALIZED: return "CUBLAS_STATUS_NOT_INITIALIZED";
|
||||
case CUBLAS_STATUS_ALLOC_FAILED: return "CUBLAS_STATUS_ALLOC_FAILED";
|
||||
case CUBLAS_STATUS_INVALID_VALUE: return "CUBLAS_STATUS_INVALID_VALUE";
|
||||
case CUBLAS_STATUS_ARCH_MISMATCH: return "CUBLAS_STATUS_ARCH_MISMATCH";
|
||||
case CUBLAS_STATUS_MAPPING_ERROR: return "CUBLAS_STATUS_MAPPING_ERROR";
|
||||
case CUBLAS_STATUS_EXECUTION_FAILED: return "CUBLAS_STATUS_EXECUTION_FAILED";
|
||||
case CUBLAS_STATUS_INTERNAL_ERROR: return "CUBLAS_STATUS_INTERNAL_ERROR";
|
||||
case CUBLAS_STATUS_NOT_SUPPORTED: return "CUBLAS_STATUS_NOT_SUPPORTED";
|
||||
case CUBLAS_STATUS_LICENSE_ERROR: return "CUBLAS_STATUS_LICENSE_ERROR";
|
||||
default: return "(look for CUBLAS_STATUS_xxx in cublas_api.h)";
|
||||
}
|
||||
}
|
||||
template <>
|
||||
const char* CudaErrString<curandStatus>(curandStatus)
|
||||
|
|
|
|||
|
|
@ -524,9 +524,8 @@ public:
|
|||
};
|
||||
|
||||
typedef GPUMatrix<float> GPUSingleMatrix;
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
}}}
|
||||
|
||||
// Error handling
|
||||
template <typename ERRTYPE>
|
||||
|
|
|
|||
|
|
@ -36,17 +36,19 @@ using namespace std;
|
|||
// construction
|
||||
// -------------------------------------------------------------------
|
||||
|
||||
// cast a matrix as a TensorView
|
||||
// main constructor (all constructors except the default one route through this)
|
||||
template <class ElemType>
|
||||
TensorView<ElemType>::TensorView(const Matrix<ElemType>& sob)
|
||||
: m_sob(sob.AsReference()), m_shape(TensorShape(array<size_t, 2>{sob.GetNumRows(), sob.GetNumCols()}))
|
||||
{
|
||||
}
|
||||
// reshape a TensorView
|
||||
template <class ElemType>
|
||||
TensorView<ElemType>::TensorView(const TensorView<ElemType>& other, const TensorShape& shape)
|
||||
: m_sob(other.m_sob.AsReference()), m_shape(shape)
|
||||
TensorView<ElemType>::TensorView(const Matrix<ElemType>& sob, const TensorShape& shape)
|
||||
: m_sob(sob.AsReference()), m_shape(shape)
|
||||
{
|
||||
#ifdef _DEBUG
|
||||
// check bounds of TensorShape against underlying storage object
|
||||
// This is useful to detect errors like passing a matrix from the wrong input.
|
||||
const auto r = shape.GetLocationRange();
|
||||
const auto n = m_sob.GetNumElements();
|
||||
if (r.first < 0 || (size_t)r.second > n)
|
||||
LogicError("TensorView: Shape bounds [%d,%d) exceed bounds of underlying storage object [0,%d).", (int) r.first, (int) r.second, (int) n);
|
||||
#endif
|
||||
}
|
||||
|
||||
// -------------------------------------------------------------------
|
||||
|
|
|
|||
|
|
@ -25,13 +25,16 @@ public:
|
|||
// construction
|
||||
// -------------------------------------------------------------------
|
||||
|
||||
// cast a matrix storage object (SOB) as a TensorView (without shape change)
|
||||
TensorView(const Matrix<ElemType>& sob);
|
||||
// reinterpret a matrix storage object (SOB) as a TensorView with a given TensorShape --this is the main constructor
|
||||
TensorView(const Matrix<ElemType>& sob, const TensorShape& shape);
|
||||
// cast a Matrix as a 2D TensorView (without shape change)
|
||||
TensorView(const Matrix<ElemType>& sob)
|
||||
: m_sob(sob.AsReference()), m_shape(TensorShape(array<size_t, 2>{sob.GetNumRows(), sob.GetNumCols()}))
|
||||
{
|
||||
}
|
||||
// reshape a TensorView
|
||||
TensorView(const TensorView<ElemType>& other, const TensorShape& shape);
|
||||
// reinterpret a SOB as a TensorView with a given TensorShape
|
||||
TensorView(const Matrix<ElemType>& sob, const TensorShape& shape)
|
||||
: TensorView(TensorView(sob) /*cast as a TensorView*/, shape /*with a shape*/)
|
||||
TensorView(const TensorView<ElemType>& other, const TensorShape& shape)
|
||||
: m_sob(other.m_sob.AsReference()), m_shape(shape)
|
||||
{
|
||||
}
|
||||
// empty constructor
|
||||
|
|
|
|||
|
|
@ -37,6 +37,10 @@ Using full BrainScript configuration
|
|||
|
||||
COMMAND: --cd $(SolutionDir)Tests\EndToEndTests\Speech\Data -f $(SolutionDir)Tests\EndToEndTests\Speech\LSTM\lstm.bs -D stderr='$(SolutionDir)Tests\EndToEndTests\Speech\RunDir\LSTM\FullUtterance\models\cntkSpeech.dnn.log' -D RunDir='$(SolutionDir)Tests\EndToEndTests\Speech\RunDir\LSTM\FullUtterance' -D NdlDir='$(SolutionDir)Tests\EndToEndTests\Speech\LSTM' -D DataDir='.' -D DeviceId='Auto' -D Truncated=false -D speechTrain=[reader=[nbruttsineachrecurrentiter=1];SGD=[epochSize=2560;maxEpochs=2;numMBsToShowResult=1]] -D makeMode=false
|
||||
|
||||
--- Speech\AN4:
|
||||
|
||||
COMMAND: configFile=$(SolutionDir)Examples\Speech\AN4\Config\LSTM-NDL.config currentDirectory=$(SolutionDir)Examples\Speech\AN4\Data RunDir=$(SolutionDir)Examples\RunDir\Speech\AN4 DataDir=$(SolutionDir)Examples\Speech\AN4\Data ConfigDir=$(SolutionDir)Examples\Speech\AN4\Config OutputDir=$(SolutionDir)Examples\RunDir\Speech\AN4 stderr=$(SolutionDir)Examples\RunDir\Speech\AN4\cntkSpeech.dnn.log DeviceId=auto speechTrain=[SGD=[maxEpochs=1]] speechTrain=[SGD=[epochSize=64]] parallelTrain=false makeMode=false
|
||||
|
||||
--- Speech\DiscriminativePreTraining: --currently fails with MEL error 'Parameter name could not be resolved 'HL2.y'
|
||||
|
||||
COMMAND: currentDirectory=$(SolutionDir)Tests\EndToEndTests\Speech\Data configFile=..\DNN\DiscriminativePreTraining\cntk_dpt.config stderr=$(SolutionDir)Tests\EndToEndTests\Speech\RunDir\DNN\DiscriminativePreTraining\models\cntkSpeech.dnn.log ConfigDir=$(SolutionDir)Tests\EndToEndTests\Speech\DNN\DiscriminativePreTraining RunDir=..\RunDir\DNN\DiscriminativePreTraining DataDir=. DeviceId=auto makeMode=false
|
||||
|
|
|
|||
Загрузка…
Ссылка в новой задаче