// 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());
}
// single input that maps its input element-wise (e.g. Sigmoid)
void ComputationNodeBase::ValidateUnaryMap(bool isFinalValidationPass)
{
assert(m_inputs.size() == 1);
ComputationNodeBase::Validate(isFinalValidationPass);
InferMBLayoutFromInputsForStandardCase(isFinalValidationPass);
SetDims(Input(0));
}
/*virtual*/ void ReduceElementsNode<ElemType>::Validate(bool isFinalValidationPass) /*override*/
{
Base::Validate(isFinalValidationPass);
InferMBLayoutFromInputsForStandardCase(isFinalValidationPass);
// validate the opcode (in case we got instantiated empty and never updated)
ValidateOp();
let shape = Input(0)->GetSampleLayout();
auto dims = shape.GetDims();
size_t reducedDim = 0; // (init to keep compiler happy)
if (m_axis == 0)
{
reducedDim = shape.GetNumElements();
dims = { 1 }; // entire sample is reduced to a scalar
}
else if (m_axis - 1 >= 0 && m_axis - 1 < dims.size())
{
reducedDim = dims[m_axis - 1];
dims[m_axis - 1] = 1; // one axis is reduced to a scalar
}
else if (isFinalValidationPass)
InvalidArgument("The shape of %ls [%s] has no axis %d", NodeDescription().c_str(), string(shape).c_str(), m_axis);
// for "Mean", we must divide by #elements
if (isFinalValidationPass && m_operation == L"Mean")
m_scale = (ElemType)(1.0 / reducedDim);
else
m_scale = (ElemType)1;
SetDims(TensorShape(dims), Input(0)->HasMBLayout());
}
/*virtual*/ void ReduceElementsNode<ElemType>::Validate(bool isFinalValidationPass) /*override*/
{
Base::Validate(isFinalValidationPass);
InferMBLayoutFromInputsForStandardCase(isFinalValidationPass);
// validate the opcode (in case we got instantiated empty and never updated)
ValidateOp();
let shape = Input(0)->GetSampleLayout();
auto dims = shape.GetDims();
if (m_axis == 0)
dims = { 1 }; // entire sample is reduced to a scalar
else if (m_axis - 1 >= 0 && m_axis - 1 < dims.size())
dims[m_axis - 1] = 1; // one axis is reduced to a scalar
else if (isFinalValidationPass)
InvalidArgument("The shape of %ls [%s] has no axis %d", NodeDescription().c_str(), string(shape).c_str(), m_axis);
SetDims(TensorShape(dims), Input(0)->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());
}