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

renamed ComputationNode::m_children to m_inputs

This commit is contained in:
Frank Seide 2015-11-27 23:08:12 -08:00
Родитель f8ac1eae70
Коммит 9d0bc1b15e
13 изменённых файлов: 99 добавлений и 115 удалений
Показать статистику Скачать Patch файл Скачать Diff файл Показать все файлы Свернуть все файлы

28
BrainScript/BrainScriptEvaluator.cpp
Просмотреть файл

@ -139,7 +139,7 @@ namespace Microsoft { namespace MSR { namespace BS {
typedef shared_ptr<ComputationNode> ComputationNodePtr;
// inputs and output
vector<ComputationNodePtr> m_children; // these are the inputs
vector<ComputationNodePtr> m_inputs; // these are the inputs
MatrixPtr m_functionValue; // this is the result
// other
@ -183,29 +183,29 @@ namespace Microsoft { namespace MSR { namespace BS {
virtual void AttachInputs(ComputationNodePtr arg)
{
m_children.resize(1);
m_children[0] = arg;
m_inputs.resize(1);
m_inputs[0] = arg;
}
virtual void AttachInputs(ComputationNodePtr leftNode, ComputationNodePtr rightNode)
{
m_children.resize(2);
m_children[0] = leftNode;
m_children[1] = rightNode;
m_inputs.resize(2);
m_inputs[0] = leftNode;
m_inputs[1] = rightNode;
}
virtual void AttachInputs(ComputationNodePtr arg1, ComputationNodePtr arg2, ComputationNodePtr arg3)
{
m_children.resize(3);
m_children[0] = arg1;
m_children[1] = arg2;
m_children[2] = arg3;
m_inputs.resize(3);
m_inputs[0] = arg1;
m_inputs[1] = arg2;
m_inputs[2] = arg3;
}
void AttachInputs(vector<ComputationNodePtr> && inputs, size_t num = 0/*0 means all OK*/)
{
if (num != 0 && inputs.size() != num)
LogicError("AttachInputs: called with incorrect number of arguments");
m_children = inputs;
m_inputs = inputs;
}
const std::vector<ComputationNodePtr> & GetChildren() const { return m_children; }
const std::vector<ComputationNodePtr> & GetChildren() const { return m_inputs; }
/*HasToString::*/ wstring ToString() const
{
@ -213,12 +213,12 @@ namespace Microsoft { namespace MSR { namespace BS {
wstring result = TidyName(NodeName()) + L" : " + wstring(OperationName());
if (!m_tag.empty())
result += L" {tag: " + m_tag + L"}";
if (m_children.empty()) result.append(L"()");
if (m_inputs.empty()) result.append(L"()");
else
{
wstring args;
bool first = true;
for (auto & child : m_children)
for (auto & child : m_inputs)
{
if (first)
first = false;

8
Documentation/CNTK-TechReport/lyx/CNTKBook_CNTK_Programmer_Chapter.lyx
Просмотреть файл

@ -3032,17 +3032,17 @@ nNodePtr Value)
\begin_layout Plain Layout
m_children.resize(2);
m_inputs.resize(2);
\end_layout
\begin_layout Plain Layout
m_children[0] = scalarValue;
m_inputs[0] = scalarValue;
\end_layout
\begin_layout Plain Layout
m_children[1] = Value;
m_inputs[1] = Value;
\end_layout
\begin_layout Plain Layout
@ -3149,7 +3149,7 @@ virtual void Validate()
\begin_layout Plain Layout
if (m_children.size() != 2)
if (m_inputs.size() != 2)
\end_layout
\begin_layout Plain Layout

7
MachineLearning/CNTKComputationNetworkLib/CompositeComputationNodes.h
Просмотреть файл

@ -114,13 +114,6 @@ namespace Microsoft { namespace MSR { namespace CNTK {
InferImageDimsFromInput(0);
}
//virtual void AttachInputs(const ComputationNodePtr c1, const ComputationNodePtr c2)
//{
// m_children.resize(2);
// m_children[0] = c1;
// m_children[1] = c2;
//}
public:
virtual bool UnitTest() {
size_t nT = 3;

8
MachineLearning/CNTKComputationNetworkLib/ComputationNetwork.cpp
Просмотреть файл

@ -63,7 +63,7 @@ namespace Microsoft { namespace MSR { namespace CNTK {
m_cacheGradientCalcOrders.clear();
m_cachedOuterLoopNodes.clear();
m_inputs.clear();
m_inputValues.clear();
m_learnableParameters.clear();
m_nameToNodeMap.clear(); // will also deref and likely deallocate all nodes we hold in here
@ -551,12 +551,12 @@ namespace Microsoft { namespace MSR { namespace CNTK {
m_cachedOuterLoopNodes.clear();
}
// lazily reate the m_inputs[] and m_learnableParameters lists
// lazily reate the m_inputValues[] and m_learnableParameters lists
// The only other side effect is to call GetEvalOrder(), which will cache the evaluation order for the given root node.
void ComputationNetwork::CollectInputAndLearnableParameters(const ComputationNodeBasePtr& rootNode)
{
//not found
if (m_inputs.find(rootNode) == m_inputs.end())
if (m_inputValues.find(rootNode) == m_inputValues.end())
{
list<ComputationNodeBasePtr> inputs;
@ -570,7 +570,7 @@ namespace Microsoft { namespace MSR { namespace CNTK {
inputs.push_back(node);
}
}
m_inputs[rootNode] = inputs;
m_inputValues[rootNode] = inputs;
}
//not found

7
MachineLearning/CNTKComputationNetworkLib/ComputationNetwork.h
Просмотреть файл

@ -20,7 +20,6 @@
// EvaluateThisNode() -> ForwardProp() // the familiar names
// ComputeInputPartial() -> BackpropTo()
// OnEvaluateBeginIteration() -> BeginForwardProp() // and similar functions likewise
// m_children -> m_inputs // likewise related functions
// Inputs() -> Input() // or In()? or GetInput()?
// Children() -> Inputs()
// ChildrenSize() -> NumInputs()
@ -573,7 +572,7 @@ public:
void ClearCaches()
{
m_built.clear();
m_inputs.clear();
m_inputValues.clear();
m_learnableParameters.clear();
ClearCalcOrderCaches();
}
@ -594,7 +593,7 @@ public:
{
if (bNoBuild == false)
BuildAndValidateSubNetwork(rootNode);
return m_inputs[rootNode];
return m_inputValues[rootNode];
}
std::list<ComputationNodeBasePtr>& LearnableNodes(const ComputationNodeBasePtr& rootNode)
@ -990,7 +989,7 @@ private: // TODO: make all private that can be made private
std::map<const ComputationNodeBasePtr, std::list<ComputationNodeBasePtr>> m_cacheGradientCalcOrders;
std::map<const ComputationNodeBasePtr, ComputationNodeBasePtr> m_cachedOuterLoopNodes;
std::map<const ComputationNodeBasePtr, std::list<ComputationNodeBasePtr>> m_inputs; // [out node] -> all input nodes feeding into out node
std::map<const ComputationNodeBasePtr, std::list<ComputationNodeBasePtr>> m_inputValues; // [out node] -> all input nodes feeding into out node
std::map<const ComputationNodeBasePtr, std::list<ComputationNodeBasePtr>> m_learnableParameters; // [out node] -> all parameter nodes feeding into out node
// pool for matrices that can be shared across nodes

2
MachineLearning/CNTKComputationNetworkLib/ComputationNetworkEvaluation.cpp
Просмотреть файл

@ -581,7 +581,7 @@ namespace Microsoft { namespace MSR { namespace CNTK {
// TODO: not nice--why not always call this in ValidateSubNetwork() only?
FormRecurrentLoops(rootNode);
// for the m_inputs and m_learnableParameters sets for this rootNode
// for the m_inputValues and m_learnableParameters sets for this rootNode
CollectInputAndLearnableParameters(rootNode);
// validate the rootNode and all nodes it depends on, in evaluation order

12
MachineLearning/CNTKComputationNetworkLib/ComputationNode.cpp
Просмотреть файл

@ -29,7 +29,7 @@ namespace Microsoft { namespace MSR { namespace CNTK {
//wstring name = NodeName(); name;
//fprintf(stderr, "\nDetermining Layout --> %ls:", name.c_str());
MBLayoutPtr pMBLayout; // starts with NULL layout
for (auto child : m_children)
for (auto child : m_inputs)
{
//wstring cname = child->NodeName(); cname;
//fprintf(stderr, " %ls(%s)", cname.c_str(), child->m_pMBLayout ? "." : "NULL");
@ -54,10 +54,10 @@ namespace Microsoft { namespace MSR { namespace CNTK {
// single input that maps its input element-wise (e.g. Sigmoid)
void ComputationNodeBase::ValidateUnaryMap(bool isFinalValidationPass)
{
assert(m_children.size() == 1);
assert(m_inputs.size() == 1);
ComputationNodeBase::Validate(isFinalValidationPass);
InferMBLayoutFromInputsForStandardCase();
SetDims(m_children[0]->GetNumRows(), DetermineNumCols(m_children[0]));
SetDims(m_inputs[0]->GetNumRows(), DetermineNumCols(m_inputs[0]));
InferImageDimsFromInputs();
}
// binary zip operation, e.g. Plus
@ -65,7 +65,7 @@ namespace Microsoft { namespace MSR { namespace CNTK {
// This also helpfully resizes the children if not yet sized.
void ComputationNodeBase::ValidateBinaryZip(bool isFinalValidationPass, bool allowMultiples)
{
assert(m_children.size() == 2);
assert(m_inputs.size() == 2);
ComputationNodeBase::Validate(isFinalValidationPass);
InferMBLayoutFromInputsForStandardCase();
@ -89,7 +89,7 @@ namespace Microsoft { namespace MSR { namespace CNTK {
// unary reduce-to-(1,1) operation, e.g. MatrixL1RegNode
void ComputationNodeBase::ValidateUnaryReduce(bool isFinalValidationPass)
{
assert(m_children.size() == 1);
assert(m_inputs.size() == 1);
ComputationNodeBase::Validate(isFinalValidationPass);
m_pMBLayout = nullptr; // this node does not hold mini-batch data
SetDims(1, 1);
@ -121,7 +121,7 @@ namespace Microsoft { namespace MSR { namespace CNTK {
// if dimension not specified we assume two operands' dimensions should be the same
// NOTE: The assert is set to check if >= 2 since this is called from nodes which have more than two children.
// The number of children is formally verified elsewhere, so this will not break consistency.
assert(m_children.size() >= 2);
assert(m_inputs.size() >= 2);
for (size_t index = 0; index < 2; index++)
{
auto in = Inputs(index);

90
MachineLearning/CNTKComputationNetworkLib/ComputationNode.h
Просмотреть файл

@ -242,7 +242,7 @@ namespace Microsoft { namespace MSR { namespace CNTK {
RuntimeError("Cannot copy from one node type to another node type");
if (flags & CopyNodeFlags::copyNodeChildren)
{
node->m_children = m_children;
node->m_inputs = m_inputs;
}
if (flags & CopyNodeFlags::copyNodeValue)
{
@ -308,15 +308,15 @@ namespace Microsoft { namespace MSR { namespace CNTK {
virtual void Validate(bool isFinalValidationPass) // main base validation function
{
// check for NULL pointers
for (size_t i = 0; i < m_children.size(); i++)
for (size_t i = 0; i < m_inputs.size(); i++)
{
if (!m_children[i])
if (!m_inputs[i])
RuntimeError("Validate: Input [%d] of %ls node '%ls' is empty (NULL, not connected).", (int)i, OperationName().c_str(), NodeName().c_str());
}
// check for empty inputs
if (isFinalValidationPass)
{
for (const auto & child : m_children)
for (const auto & child : m_inputs)
{
if (child->GetNumRows() == 0 || (!child->HasMBLayout() && child->GetNumCols() == 0))
RuntimeError("%ls %ls operation: input %ls %ls has 0 elements.",
@ -356,7 +356,7 @@ namespace Microsoft { namespace MSR { namespace CNTK {
void AttachInputs(const ComputationNodeBasePtr& firstInput, const ComputationNodeBasePtr& secondInput, const ComputationNodeBasePtr &thirdInput, const ComputationNodeBasePtr& fourthInput, const ComputationNodeBasePtr& fifthInput) { AttachInputs(std::vector<ComputationNodeBasePtr> { firstInput, secondInput, thirdInput, fourthInput, fifthInput } ); }
void AttachInputs(const ComputationNodeBasePtr& firstInput, const ComputationNodeBasePtr& secondInput, const ComputationNodeBasePtr &thirdInput, const ComputationNodeBasePtr& fourthInput, const ComputationNodeBasePtr& fifthInput, const ComputationNodeBasePtr& sixthInput) { AttachInputs(std::vector<ComputationNodeBasePtr> { firstInput, secondInput, thirdInput, fourthInput, fifthInput, sixthInput } ); }
virtual void DetachInputs() { m_children.clear(); }
virtual void DetachInputs() { m_inputs.clear(); }
// helper for the factory function for ComputationNodes
static vector<ComputationNodeBasePtr> GetInputsFromConfig(const ScriptableObjects::IConfigRecordPtr configp)
@ -378,8 +378,8 @@ namespace Microsoft { namespace MSR { namespace CNTK {
return inputs;
}
const std::vector<ComputationNodeBasePtr> & GetChildren() const { return m_children; }
ComputationNodeBasePtr Inputs(size_t index) const { return m_children[index]; } // TODO: delete this; change to m_children
const std::vector<ComputationNodeBasePtr> & GetChildren() const { return m_inputs; }
ComputationNodeBasePtr Inputs(size_t index) const { return m_inputs[index]; } // TODO: delete this; change to m_inputs
//return true if the node's value should be computed before the normal training. e.g., mean and invStd of input features.
virtual bool /*IComputationNode::*/RequiresPreCompute() const { return false; }
@ -452,7 +452,7 @@ namespace Microsoft { namespace MSR { namespace CNTK {
fprintf(stderr, "(");
for (size_t i = 0; i<ChildrenSize(); i++)
{
const auto & child = m_children[i];
const auto & child = m_inputs[i];
if (i > 0)
fprintf(stderr, ", ");
@ -502,14 +502,14 @@ namespace Microsoft { namespace MSR { namespace CNTK {
bool IsChildAnImage(const size_t index) const
{
return m_children[index]->m_imageLayout.GetWidth() != 1 || m_children[index]->m_imageLayout.GetNumChannels() != 1;
return m_inputs[index]->m_imageLayout.GetWidth() != 1 || m_inputs[index]->m_imageLayout.GetNumChannels() != 1;
}
const ImageLayout & GetImageLayout() const { return m_imageLayout; }
pair<ImageLayout, ImageLayout> GetImageLayouts() const { return make_pair(m_inputImageLayout, m_imageLayout); } // helper for Validate()
const size_t ChildrenSize() const { return m_children.size(); } // TODO: rename to NumChildren() or NumInputs(); and inside here where we use m_children, use m_children.size() as well
const size_t ChildrenSize() const { return m_inputs.size(); } // TODO: rename to NumChildren() or NumInputs(); and inside here where we use m_inputs, use m_inputs.size() as well
virtual void SetInput(const size_t childIndex, const ComputationNodeBasePtr& node) = 0;
@ -555,7 +555,7 @@ namespace Microsoft { namespace MSR { namespace CNTK {
if (index >= ChildrenSize())
InvalidArgument("InferImageDimsFromInput: output index");
const auto & child = m_children[index];
const auto & child = m_inputs[index];
if (child != nullptr)
m_inputImageLayout = child->m_imageLayout;
if (outputSameAsInput)
@ -573,7 +573,7 @@ namespace Microsoft { namespace MSR { namespace CNTK {
bool IsEqualTo(const ComputationNodeBasePtr& other) const //this will be used to determine whehter two nodes are the same
{
if (OperationName() != other->OperationName() || m_children.size() != other->m_children.size())
if (OperationName() != other->OperationName() || m_inputs.size() != other->m_inputs.size())
return false;
if (NodeName() == other->NodeName()) //assume names are unique in the system
@ -582,8 +582,8 @@ namespace Microsoft { namespace MSR { namespace CNTK {
if (IsLeaf() && other->IsLeaf()) //since names are not equal otherwise will return above
return false;
for (size_t i=0; i<m_children.size(); i++)
if (!(m_children[i] == other->m_children[i]))
for (size_t i=0; i<m_inputs.size(); i++)
if (!(m_inputs[i] == other->m_inputs[i]))
return false;
return true;
@ -619,10 +619,10 @@ namespace Microsoft { namespace MSR { namespace CNTK {
// children first for function evaluation
if (OperationName() != L"PairNetwork" || !skipPairNetwork) // (don't step through network-pair boundary if called from FormRecurrentLoops())
{
for (int i = 0; i < m_children.size(); i++)
for (int i = 0; i < m_inputs.size(); i++)
{
if (m_children[i])
m_children[i]->EnumerateNodesRec(visited, result, skipPairNetwork);
if (m_inputs[i])
m_inputs[i]->EnumerateNodesRec(visited, result, skipPairNetwork);
}
}
@ -641,7 +641,7 @@ namespace Microsoft { namespace MSR { namespace CNTK {
// TODO: use range-based for
for (size_t i = 0; i < ChildrenSize(); i++)
{
if (IsOlderThan(*m_children[i]))
if (IsOlderThan(*m_inputs[i]))
return true;
}
@ -668,12 +668,12 @@ namespace Microsoft { namespace MSR { namespace CNTK {
if (visited.find(curNode) == visited.end())
{
for (size_t i = 0; i < curNode->m_children.size(); i++)
for (size_t i = 0; i < curNode->m_inputs.size(); i++)
{
arcs.push_back(ComputationArc(curNode, curNode->m_children[i]));
arcs.push_back(ComputationArc(curNode, curNode->m_inputs[i]));
if (visited.find(curNode->m_children[i]) == visited.end()) // this children has not been visited before
tovisit.push_front(curNode->m_children[i]); // going to visit each of the children
if (visited.find(curNode->m_inputs[i]) == visited.end()) // this children has not been visited before
tovisit.push_front(curNode->m_inputs[i]); // going to visit each of the children
}
visited.insert(curNode);
}
@ -715,7 +715,7 @@ namespace Microsoft { namespace MSR { namespace CNTK {
std::wstring m_nodeName;
// inputs
std::vector<ComputationNodeBasePtr> m_children;
std::vector<ComputationNodeBasePtr> m_inputs;
// dimensions and layout
// Data is stored as a matrix, but often it is interpreted as a more complex structure.
@ -829,12 +829,12 @@ namespace Microsoft { namespace MSR { namespace CNTK {
const auto * pNumInputs = dynamic_cast<INumInputs*>(this); // if this class also derives from NumInputs<N> then N is the expected number of inputs
if (pNumInputs && pNumInputs->GetExpectedNumInputs() != inputs.size())
RuntimeError("%ls operation '%ls' expects %d inputs (given: %d)", OperationName().c_str(), NodeName().c_str(), (int)pNumInputs->GetExpectedNumInputs(), (int)inputs.size());
m_children.resize(inputs.size());
for (size_t i = 0; i < m_children.size(); i++)
m_inputs.resize(inputs.size());
for (size_t i = 0; i < m_inputs.size(); i++)
if (inputs[i])
m_children[i] = UpCast(inputs[i]); // (UpCast() checks the type; the assignment then downcasts it again)
m_inputs[i] = UpCast(inputs[i]); // (UpCast() checks the type; the assignment then downcasts it again)
else
m_children[i] = nullptr; // during network creation, nullpts are possible
m_inputs[i] = nullptr; // during network creation, nullpts are possible
}
protected:
@ -878,10 +878,10 @@ namespace Microsoft { namespace MSR { namespace CNTK {
virtual void AllocateGradientMatricesForChildren(MatrixPool& matrixPool) override
{
for (int i = 0; i < m_children.size(); i++)
for (int i = 0; i < m_inputs.size(); i++)
{
if (m_children[i]->NeedGradient())
m_children[i]->RequestMatricesBeforeGradientComp(matrixPool);
if (m_inputs[i]->NeedGradient())
m_inputs[i]->RequestMatricesBeforeGradientComp(matrixPool);
}
}
@ -911,12 +911,12 @@ namespace Microsoft { namespace MSR { namespace CNTK {
// we format it like "name : type rows x cols ( args )"
wstring result = /*TidyName*/(NodeName()) + L" : " + OperationName();
result.append(msra::strfun::wstrprintf(L" %d x %d", (int)m_functionValues->GetNumRows(), (int)m_functionValues->GetNumCols()));
if (m_children.empty()) result.append(L" ()");
if (m_inputs.empty()) result.append(L" ()");
else
{
wstring args;
bool first = true;
for (auto & child : m_children)
for (auto & child : m_inputs)
{
if (first)
first = false;
@ -1006,7 +1006,7 @@ namespace Microsoft { namespace MSR { namespace CNTK {
{
if (i > 0)
fprintf(stderr, ", ");
fprintf(stderr, "%ls[%lu, %lu]", m_children[i] ? m_children[i]->NodeName().c_str():L"NULL", m_children[i]->GetNumRows(), m_children[i]->GetNumCols());
fprintf(stderr, "%ls[%lu, %lu]", m_inputs[i] ? m_inputs[i]->NodeName().c_str():L"NULL", m_inputs[i]->GetNumRows(), m_inputs[i]->GetNumCols());
}
fprintf(stderr, ")");
}
@ -1033,10 +1033,10 @@ namespace Microsoft { namespace MSR { namespace CNTK {
inline ComputationNodePtr Inputs(const size_t childIndex) const // TODO: rename to Input
{
#ifdef _DEBUG // profile shows this is range check very expensive in release mode, skip it
if (childIndex >= m_children.size())
if (childIndex >= m_inputs.size())
LogicError("Inputs: childIndex is out of range.");
#endif
return UpCast(m_children[childIndex]);
return UpCast(m_inputs[childIndex]);
}
void /*ComputationNodeBase::*/SetInput(const size_t childIndex, const ComputationNodeBasePtr& inode) override
@ -1044,15 +1044,15 @@ namespace Microsoft { namespace MSR { namespace CNTK {
const ComputationNodePtr node = UpCast(inode);
//require first nodes specified before the second to avoid null nodes condition.
if (childIndex > m_children.size())
if (childIndex > m_inputs.size())
InvalidArgument("SetInput: You must specify the input for children with index less than this one first.");
// expand the inputs to exist up to the desired index
while (childIndex >= m_children.size())
m_children.push_back(nullptr);
while (childIndex >= m_inputs.size())
m_inputs.push_back(nullptr);
// set the input value
m_children[childIndex] = node;
m_inputs[childIndex] = node;
}
const Matrix<ElemType>& FunctionValues() const { return *m_functionValues; }
@ -1154,10 +1154,10 @@ namespace Microsoft { namespace MSR { namespace CNTK {
MaskMissingGradientColumnsToZero(FrameRange(m_pMBLayout));
}
bool anyChildNeedsGradient = false;
for (size_t i = 0; i < m_children.size(); i++)
for (size_t i = 0; i < m_inputs.size(); i++)
anyChildNeedsGradient |= Inputs(i)->m_needsGradient;
if (anyChildNeedsGradient)
for (size_t i = 0; i < m_children.size(); i++)
for (size_t i = 0; i < m_inputs.size(); i++)
Inputs(i)->MaskMissingValuesColumnsToZero(FrameRange(Inputs(i)->GetMBLayout()));
#endif
}
@ -1167,7 +1167,7 @@ namespace Microsoft { namespace MSR { namespace CNTK {
{
Base::OnComputeGradientEndIteration();
#ifdef TRACK_GAP_NANS
for (size_t i = 0; i < m_children.size(); i++)
for (size_t i = 0; i < m_inputs.size(); i++)
{
ComputationNodePtr child = Inputs(i);
if (child->m_needsGradient)
@ -1188,7 +1188,7 @@ namespace Microsoft { namespace MSR { namespace CNTK {
if (frameRange.IsAllFrames() && IsPartOfLoop() && childrenInThisLoop)
LogicError("%ls %ls operation: ComputeGradientForChildren called with whole-batch FrameRange on node that participates in a loop", NodeName().c_str(), OperationName().c_str());
for (size_t i = 0; i < m_children.size(); i++)
for (size_t i = 0; i < m_inputs.size(); i++)
{
ComputationNodePtr child = Inputs(i);
if (child->m_needsGradient &&
@ -1234,7 +1234,7 @@ namespace Microsoft { namespace MSR { namespace CNTK {
void /*ComputationNodeBase::*/ClearGradientForChildren() override // TODO: bad naming--this just clears the lazy flags, whereas LazyZeroGradient() actually clears the values
{
for (size_t i = 0; i < m_children.size(); i++)
for (size_t i = 0; i < m_inputs.size(); i++)
Inputs(i)->m_gradientInitialized = false;
}
@ -1494,7 +1494,7 @@ protected: \
using Base::MaskMissingColumnsToZero; using Base::MaskMissingValuesColumnsToZero; using Base::MaskMissingGradientColumnsToZero; using Base::InvalidateMissingValuesColumns; using Base::InvalidateMissingGradientColumns; \
using Base::DataSlice; using Base::ValueSlice; using Base::GradientValues; using Base::GradientValuesPtr; using Base::GradientSlice; using Base::MaskedValueSlice; using Base::MaskedGradientSlice; \
using Base::EvaluateThisNode; using Base::ComputeInputPartial; \
using Base::m_children; using Base::m_deviceId; using Base::m_functionValues; using Base::m_gradientValues; \
using Base::m_inputs; using Base::m_deviceId; using Base::m_functionValues; using Base::m_gradientValues; \
using Base::m_inputImageLayout; using Base::m_imageLayout; \
using Base::m_parameterUpdateRequired; using Base::m_nodeName; \
using Base::CreateMatrixIfNull; using Base::RequestMatrixFromPool; using Base::ReleaseMatrixToPool; \

2
MachineLearning/CNTKComputationNetworkLib/ConvolutionalNodes.h
Просмотреть файл

@ -334,7 +334,7 @@ namespace Microsoft { namespace MSR { namespace CNTK {
ValidateInferChildDims(0, m_imageLayout.GetNumChannels(), weightCols);
if (isFinalValidationPass && (Inputs(0)->GetNumCols() != weightCols || Inputs(0)->GetNumRows() != m_imageLayout.GetNumChannels()))
LogicError("convolutionWeight matrix %ls should have dimension [%d, %d] which is [outputChannels, kernelWidth * kernelHeight * inputChannels]", m_children[0]->NodeName().c_str(), (int)m_imageLayout.GetNumChannels(), (int)weightCols);
LogicError("convolutionWeight matrix %ls should have dimension [%d, %d] which is [outputChannels, kernelWidth * kernelHeight * inputChannels]", m_inputs[0]->NodeName().c_str(), (int)m_imageLayout.GetNumChannels(), (int)weightCols);
size_t inputDim = m_inputImageLayout.GetWidth() * m_inputImageLayout.GetHeight() * m_inputImageLayout.GetNumChannels();
if (Inputs(1)->GetNumRows() == 0)

2
MachineLearning/CNTKComputationNetworkLib/EvaluationCriterionNodes.h
Просмотреть файл

@ -78,7 +78,7 @@ namespace Microsoft { namespace MSR { namespace CNTK {
m_topK = 1;
// TODO: Make topK a constructor parameter
if (m_children.size() == 3)
if (m_inputs.size() == 3)
{
if (Inputs(2)->GetNumRows() != 1 || Inputs(2)->GetNumCols() != 1)
throw std::logic_error("TopK in ErrorPredictionNode must be a scalar value.");

2
MachineLearning/CNTKComputationNetworkLib/LinearAlgebraNodes.h
Просмотреть файл

@ -484,7 +484,7 @@ namespace Microsoft { namespace MSR { namespace CNTK {
virtual void AllocateGradientMatricesForChildren(MatrixPool& matrixPool) override
{
//this is a special handling case. We need to allocate sparse matrix directly instead of from pool.
if (m_children[0]->NeedGradient() && Inputs(1)->FunctionValues().GetMatrixType() == SPARSE)
if (m_inputs[0]->NeedGradient() && Inputs(1)->FunctionValues().GetMatrixType() == SPARSE)
{
CreateMatrixIfNull(Inputs(0)->GradientValuesPtr());
Inputs(0)->GradientValues().SwitchToMatrixType(SPARSE, MatrixFormat::matrixFormatSparseBlockCol, false);

10
MachineLearning/CNTKComputationNetworkLib/RecurrentNodes.h
Просмотреть файл

@ -1393,16 +1393,6 @@ namespace Microsoft { namespace MSR { namespace CNTK {
InferImageDimsFromInput(1, false);
}
//virtual void AttachInputs(const ComputationNodePtr obs, const ComputationNodePtr inputGate, const ComputationNodePtr forgetGate, const ComputationNodePtr outputGate, const ComputationNodePtr memoryCellWgt)
//{
// m_children.resize(5);
// m_children[0] = obs;
// m_children[1] = inputGate;
// m_children[2] = forgetGate;
// m_children[3] = outputGate;
// m_children[4] = memoryCellWgt;
//}
virtual void DumpNodeInfo(const bool printValues, File& fstream) const override
{
Base::DumpNodeInfo(printValues, fstream);

36
MachineLearning/CNTKComputationNetworkLib/TrainingCriterionNodes.h
Просмотреть файл

@ -1528,7 +1528,7 @@ namespace Microsoft { namespace MSR { namespace CNTK {
m_temp->AssignDifferenceOf(Inputs(0)->ValueSlice(frameRange), *m_classZeroLabels); // TODO: need a slice for m_classZeroLabels?
// Multiply the vector by the Inputs(2)->FunctionValues()
if (m_children.size() == 3) // without weight
if (m_inputs.size() == 3) // without weight
m_temp->AssignElementProductOf(*m_temp, Inputs(2)->ValueSlice(frameRange)); // TODO: is Inputs(2) minibatch data? Confirm
// divide class by p (class 1) or (1-p) (class 0)
@ -1579,7 +1579,7 @@ namespace Microsoft { namespace MSR { namespace CNTK {
m_temp->AssignLogOf(*m_result);
// The error is the negative of the sum of the result
if (m_children.size() == 2)
if (m_inputs.size() == 2)
FunctionValues().AssignSumOfElements(*m_temp);
else
FunctionValues().AssignInnerProductOf(Inputs(2)->ValueSlice(frameRange), *m_temp, false);
@ -1588,27 +1588,29 @@ namespace Microsoft { namespace MSR { namespace CNTK {
virtual void /*ComputationNodeBase::*/Validate(bool isFinalValidationPass) override
{
if (m_children.size() != 2 && m_children.size() != 3)
InvalidArgument("%ls %ls operation requires two or three inputs.", NodeName().c_str(), OperationName().c_str());
if (m_inputs.size() != 2 && m_inputs.size() != 3)
InvalidArgument("%ls %ls operation requires two or three inputs.", NodeName().c_str(), OperationName().c_str());
ValidateBinaryReduce(isFinalValidationPass);
ValidateBinaryReduce(isFinalValidationPass);
/* Note that this is the same as ValidateInferBinaryChildrenDims, but done for the 3rd child if it exists */
if (m_children.size() == 3)
{
auto in = Inputs(2);
auto other = Inputs(1);
// borrow any unset dimension on one input from the other input
size_t rows = in->GetNumRows() == 0 ? other->GetNumRows()/*borrow from peer*/ : in->GetNumRows()/*keep as is*/;
size_t cols = (!in->HasMBLayout() && in->GetNumCols() == 0) ? other->GetNumCols()/*borrow from peer*/ : in->GetNumCols()/*keep as is*/;
/* Note that this is the same as ValidateInferBinaryChildrenDims, but done for the 3rd child if it exists */
if (m_inputs.size() == 3)
{
auto in = Inputs(2);
auto other = Inputs(1);
// borrow any unset dimension on one input from the other input
size_t rows = in->GetNumRows() == 0 ? other->GetNumRows()/*borrow from peer*/ : in->GetNumRows()/*keep as is*/;
size_t cols = (!in->HasMBLayout() && in->GetNumCols() == 0) ? other->GetNumCols()/*borrow from peer*/ : in->GetNumCols()/*keep as is*/;
ValidateInferChildDims(2, rows, cols);
ValidateInferChildDims(2, rows, cols);
if (isFinalValidationPass &&
if (isFinalValidationPass &&
!(Inputs(0)->GetNumRows() == Inputs(2)->GetNumRows() &&
(Inputs(0)->HasMBLayout() || (Inputs(0)->GetNumCols() == Inputs(2)->GetNumCols()))))
(Inputs(0)->HasMBLayout() || (Inputs(0)->GetNumCols() == Inputs(2)->GetNumCols()))))
{
LogicError("The Matrix dimensions of the second argument in the %ls %ls operation do not match.", NodeName().c_str(), OperationName().c_str());
}
}
}
}
//request matrices needed to do node function value evaluation