const std::string ComputationNodeBase::ShapeDescription() const
{
return msra::strfun::strprintf("[%s%s%ls]",
string(m_sampleLayout).c_str(),
HasMBLayout() ? " x " : "",
HasMBLayout() ? GetMBLayout()->GetAxisName() : L"");
}
/*virtual*/ void GatherPackedNode<ElemType>::Validate(bool isFinalValidationPass) /*override*/
{
ComputationNodeBase::Validate(isFinalValidationPass);
// inherit MBLayout from indexData
m_pMBLayout = Input(INDEXDATA)->GetMBLayout();
if (isFinalValidationPass && (!Input(INDEXDATA)->HasMBLayout()))
LogicError("%ls requires first argument (index data) to have a time dimension.", NodeDescription().c_str());
bool sourceHasTimeDimension = Input(SOURCEDATA)->HasMBLayout();
if (isFinalValidationPass && Input(INDEXDATA)->GetSampleLayout().GetNumElements() != 1)
InvalidArgument("%ls requires the first argument (index data) to be a scalar time sequence.", NodeDescription().c_str());
// inherit tensor dimension from sourceData, minus the last (column or time) dimension. TODO this needs to become simpler...
if (sourceHasTimeDimension)
SetDims(Input(SOURCEDATA)->GetSampleLayout(), HasMBLayout());
else
{
SmallVector<size_t> layout = { 1 }; // Scalar
if (Input(SOURCEDATA)->GetSampleLayout().GetRank() > 1)
{
auto srcLayout = Input(SOURCEDATA)->GetSampleLayout().GetDims();
layout.assign(srcLayout.begin(), srcLayout.end() - 1);
}
SetDims(TensorShape(layout), HasMBLayout());
}
}
// binary zip operation, e.g. Plus
// If allowBroadcast then one can be a sub-dimension of the other (if layout then only for rows, otherwise for cols, too).
// This also helpfully resizes the children if not yet sized.
void ComputationNodeBase::ValidateBinaryZip(bool isFinalValidationPass, bool allowBroadcast)
{
assert(m_inputs.size() == 2);
ComputationNodeBase::Validate(isFinalValidationPass);
InferMBLayoutFromInputsForStandardCase(isFinalValidationPass);
ValidateInferBinaryInputDims();
if (isFinalValidationPass)
ValidateMBLayout(Input(0), Input(1));
// result has tensor shape with dimensions being the max over both
let shape0 = GetInputSampleLayout(0);
let shape1 = GetInputSampleLayout(1);
SmallVector<size_t> dims = shape0.GetDims();
if (shape1.GetRank() > dims.size())
dims.resize(shape1.GetRank(), 1); // pad with ones
// If rank of [0] is higher than we only need to take max over rank [1].
// If rank of [1] is higher then we have padded to equal lentgh.
for (size_t k = 0; k < shape1.GetRank(); k++)
{
size_t dim1 = shape1[k];
// BUGBUG: We must consider the allowBroadcast flag here.
if (dims[k] <= 1 && dim1 != 0) // is [0] broadcasting (1) or unspecified (0)?
dims[k] = dim1; // then use dimension we broadcast to
else if (dim1 <= 1 && dims[k] != 0) // if [1] is broadcasting or unspecified
; // then dims is already correct
else if (isFinalValidationPass && dim1 != dims[k]) // no broadcasting or unspecified: they must match
InvalidArgument("%ls: Input dimensions [%s] and [%s] are not compatible.",
NodeDescription().c_str(), string(shape0).c_str(), string(shape1).c_str());
}
SetDims(TensorShape(dims), HasMBLayout());
}
// same as GetTensorSliceFor() except that 'fr' refers to a single column, and result will not have seq/time axes
// This is needed by TimesNode when the left argument has to be broken up into individual matrices/GEMM calls.
// To enable its first argument to have an MBLayout, it needs to un-pad if we have an MBLayout but only refer to a single sequence and time step.
TensorShape ComputationNodeBase::GetOneSampleTensorSliceFor(size_t rank, const FrameRange& fr) const
{
TensorShape result = GetTensorSliceFor(rank, fr);
// undo the adding of (seq, time) axes that was done by GetTensorShape()
if (!fr.IsOneColumnWrt(GetMBLayout()))
LogicError("GetOneSampleTensorSliceFor: Requires 'fr' to refer to a single sample.");
if (HasMBLayout())
result.TrimRankInPlace(rank); // Note: This function will verify once again that the extra dimensions have been reduced to [1 x 1]
return result;
}
// form the actual tensor that describes the full object
TensorShape ComputationNodeBase::GetTensorShape(size_t rank) const
{
// If we have an MB layout then add the necessary sequence and time axes. If we have none, then absorb the column dimension.
TensorShape tensorShape = GetSampleLayout(); // TODO: Do we need to expect this tensor to have arbitrary strides? In case it came out of a Slice, Reshape, or Transpose op in-place?
if (HasMBLayout())
{
size_t i = (rank != SIZE_MAX) ? rank : tensorShape.GetRank();
tensorShape.AppendInPlace(i++, GetMBLayout()->GetNumParallelSequences());
tensorShape.AppendInPlace(i++, GetMBLayout()->GetNumTimeSteps());
}
return tensorShape;
}
/*virtual*/ void PackedIndexNode<ElemType>::Validate(bool isFinalValidationPass) /*override*/
{
ComputationNodeBase::Validate(isFinalValidationPass);
// inherit both MBLayout and sample dimension (scalar) from indexData
// Because we map (per-seq) index sequence to (packed) index sequence. Target is only for index calculation.
m_pMBLayout = Input(INDEXDATA)->GetMBLayout();
if (isFinalValidationPass && (!Input(INDEXDATA)->HasMBLayout() || !Input(SOURCEDATA)->HasMBLayout()))
LogicError("%ls %ls operation requires both inputs to be minibatch data (must have MBLayouts).", NodeName().c_str(), OperationName().c_str());
if (isFinalValidationPass && Input(INDEXDATA)->GetSampleLayout().GetNumElements() != 1)
InvalidArgument("%ls %ls operation requires the second argument (indexData) to be a scalar sequence.", NodeName().c_str(), OperationName().c_str());
SetDims(Input(INDEXDATA)->GetSampleLayout(), HasMBLayout());
}
/*virtual*/ void GatherPackedNode<ElemType>::Validate(bool isFinalValidationPass) /*override*/
{
ComputationNodeBase::Validate(isFinalValidationPass);
// inherit MBLayout from indexData
m_pMBLayout = Input(INDEXDATA)->GetMBLayout();
if (isFinalValidationPass && (!Input(INDEXDATA)->HasMBLayout() || !Input(SOURCEDATA)->HasMBLayout()))
LogicError("%ls %ls operation requires both inputs to be minibatch data (must have MBLayouts).", NodeName().c_str(), OperationName().c_str());
if (isFinalValidationPass && Input(INDEXDATA)->GetSampleLayout().GetNumElements() != 1)
InvalidArgument("%ls %ls operation requires the first argument (indexData) to be a scalar sequence.", NodeName().c_str(), OperationName().c_str());
// inherit tensor dimension from sourceData
SetDims(Input(SOURCEDATA)->GetSampleLayout(), HasMBLayout());
}
/*virtual*/ void ScatterPackedNode<ElemType>::Validate(bool isFinalValidationPass) /*override*/
{
ComputationNodeBase::Validate(isFinalValidationPass);
// inherit MBLayout from layoutData (that's the only thing we use it for)
m_pMBLayout = Input(LAYOUTDATA)->GetMBLayout();
if (isFinalValidationPass && (!Input(LAYOUTDATA)->HasMBLayout() || !Input(INDEXDATA)->HasMBLayout() || !Input(SOURCEDATA)->HasMBLayout()))
LogicError("%ls %ls operation requires all inputs to be minibatch data (must have MBLayouts).", NodeName().c_str(), OperationName().c_str());
if (isFinalValidationPass && Input(INDEXDATA)->GetSampleLayout().GetNumElements() != 1)
InvalidArgument("%ls %ls operation requires the second argument (indexData) to be a scalar sequence.", NodeName().c_str(), OperationName().c_str());
// TODO: We also know that indexData and sourceData must have the same MBLayout. But that is checked at runtime.
// inherit tensor dimension from sourceData
SetDims(Input(SOURCEDATA)->GetSampleLayout(), HasMBLayout());
}
// binary zip operation, e.g. Plus
// If allowBroadcast then one can be a sub-dimension of the other (if layout then only for rows, otherwise for cols, too).
// This also helpfully resizes the children if not yet sized.
void ComputationNodeBase::ValidateBinaryZip(bool isFinalValidationPass, bool allowBroadcast)
{
assert(m_inputs.size() == 2);
ComputationNodeBase::Validate(isFinalValidationPass);
InferMBLayoutFromInputsForStandardCase(isFinalValidationPass);
ValidateInferBinaryInputDims();
if (isFinalValidationPass &&
Input(0)->GetMBLayout() != Input(1)->GetMBLayout() && Input(0)->HasMBLayout() && Input(1)->HasMBLayout())
{
LogicError("%ls: Minibatch layouts are not the same between arguments and might get out of sync during runtime. If this is by design, use ReconcileDynamicAxis() to forward layouts between nodes.", NodeDescription().c_str());
}
// result has tensor shape with dimensions being the max over both
let shape0 = GetInputSampleLayout(0);
let shape1 = GetInputSampleLayout(1);
SmallVector<size_t> dims = shape0.GetDims();
if (shape1.GetRank() > dims.size())
dims.resize(shape1.GetRank(), 1); // pad with ones
// If rank of [0] is higher than we only need to take max over rank [1].
// If rank of [1] is higher then we have padded to equal lentgh.
for (size_t k = 0; k < shape1.GetRank(); k++)
{
size_t dim1 = shape1[k];
// BUGBUG: We must consider the allowBroadcast flag here.
if (dims[k] == 1) // is [0] broadcasting?
dims[k] = dim1; // then use dimension we broadcast to
else if (dim1 == 1) // if [1] is broadcasting
; // dims is already correct
else if (isFinalValidationPass && dim1 != dims[k]) // no broadcasting: they must match
InvalidArgument("%ls: Input dimensions [%s] and [%s] are not compatible.",
NodeDescription().c_str(), string(shape0).c_str(), string(shape1).c_str());
}
SetDims(TensorShape(dims), HasMBLayout());
}
// N-nary zip operation, e.g. for TernaryZip for clip()
// If allowBroadcast then one can be a sub-dimension of the other (if layout then only for rows, otherwise for cols, too).
// This also helpfully resizes the children if not yet sized.
void ComputationNodeBase::ValidateNaryZip(bool isFinalValidationPass, bool allowBroadcast, size_t numInputs)
{
assert(m_inputs.size() == numInputs);
ComputationNodeBase::Validate(isFinalValidationPass);
InferMBLayoutFromInputsForStandardCase(isFinalValidationPass);
ValidateInferNaryInputDims(numInputs);
// check minibatch layout consistency for all possible pairs (n choose 2)
if (isFinalValidationPass)
for (size_t i = 0; i < numInputs; i++)
for (size_t j = i + 1; j < numInputs; j++)
ValidateMBLayout(Input(i), Input(j));
// result has tensor shape with dimensions being the max over all inputs
let shape0 = GetInputSampleLayout(0);
// dims is max over all inputs
size_t maxRank = shape0.GetRank();
for (size_t i = 1; i < numInputs; i++)
{
let shape = GetInputSampleLayout(i);
if (shape.GetRank() > maxRank)
maxRank = shape.GetRank();
}
SmallVector<size_t> dims = shape0.GetDims();
dims.resize(maxRank, 1); // pad with 1
// first check for invalid dimensions
for (size_t k = 0; k < maxRank; k++)
{
size_t maxDim = 0;
TensorShape maxShape = shape0; // arbitrary; this is just used for the error message
for (size_t i = 0; i < numInputs; i++)
{
let currentShape = GetInputSampleLayout(i);
size_t currentRank = currentShape.GetRank();
// make sure that the rank of this input is bigger than the current index (otherwise, these are implied singleton dimensions that do not need to be checked)
if (currentRank > k)
{
size_t currentDim = currentShape[k];
if (currentDim > 1 && maxDim != currentDim && maxDim > 1) // 1=broadcasting, 0=not known yet, meant to be inferred
{
InvalidArgument("%ls: Input dimensions [%s] and [%s] are not compatible.",
NodeDescription().c_str(), string(maxShape).c_str(), string(currentShape).c_str());
}
else if (currentDim > maxDim)
{
maxDim = currentDim;
maxShape = currentShape;
}
}
}
}
// now set up the right dims
for (size_t k = 0; k < maxRank; k++)
{
for (size_t i = 0; i < numInputs; i++)
{
let shape = GetInputSampleLayout(i);
if (shape.GetRank() > k)
{
size_t dim = shape[k];
if (dims[k] <= 1 && dim != 0)
dims[k] = dim;
}
}
}
SetDims(TensorShape(dims), HasMBLayout());
}
