/*virtual*/ string ComputationNodeBase::FormatOperationPrototype(const string& extraArgs) const
{
string prototype;
prototype += msra::strfun::strprintf("%ls = %ls", NodeName().c_str(), OperationName().c_str());
// arguments of operation
if (IsLeaf())
prototype += "()";
else
{
prototype += " (";
for (size_t i = 0; i < GetNumInputs(); i++)
{
const auto& child = m_inputs[i];
if (i > 0)
prototype += ", ";
if (child)
prototype += msra::strfun::strprintf("%ls", child->NodeName().c_str());
else
prototype += "NULL";
}
prototype += extraArgs;
prototype += ")";
}
// type (tensor dimensions) of operation
prototype += " : ";
if (!IsLeaf())
{
//prototype += "(";
for (size_t i = 0; i < GetNumInputs(); i++)
{
const auto& child = m_inputs[i];
if (i > 0)
prototype += ", ";
if (child == nullptr)
{
prototype += "NULL";
continue;
}
prototype += child->ShapeDescription().c_str();
}
prototype += extraArgs;
//prototype += ")";
}
prototype += msra::strfun::strprintf(" -> %s", ShapeDescription().c_str());
return prototype;
}
/*virtual*/ void ComputationNode<ElemType>::DumpNodeInfo(const bool /*printValues*/, const bool printMetadata, File& fstream) const
{
if (printMetadata)
{
fstream << L"\n" + NodeName() + L"=" + OperationName();
if (!IsLeaf())
{
fstream << wstring(L"(");
for (size_t i = 0; i < GetNumInputs(); i++)
{
if (i > 0)
fstream << wstring(L",");
fstream << (Input(i) ? Input(i)->NodeName() : L"NULL");
}
fstream << wstring(L")");
}
}
}
xml_schema::Qname SubscribeTraits::ResponseType() {
return xml_schema::Qname(WS_NS_EVENTING, OperationName() + "Response");
}
xml_schema::Qname SubscribeTraits::RequestType() {
return xml_schema::Qname(WS_NS_EVENTING, OperationName());
}
xml_schema::Qname GetMetadataTraits::RequestType() {
return xml_schema::Qname(WS_NS_METADATA_EXCHANGE, OperationName());
}
/*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());
}
/*virtual*/ void ScatterPackedNode<ElemType>::ForwardPropNonLooping() /*override*/
{
if (*Input(INDEXDATA)->GetMBLayout() != *Input(SOURCEDATA)->GetMBLayout())
InvalidArgument("%ls %ls operation requires the minibatch layout of index and source data to be the same.", NodeName().c_str(), OperationName().c_str());
Input(INDEXDATA)->MaskMissingValueColumnsTo(FrameRange(Input(INDEXDATA)->GetMBLayout()), -1); // indicates an invalid column to Gather/Scatter
let& index = Input(INDEXDATA)->Value(); // column indices to copy from
let& source = Input(SOURCEDATA)->Value(); // source data to copy
auto& output = Value(); // output goes here
output.DoScatterColumnsOf(/*beta=*/0, index, source, /*alpha=*/1);
}
/*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 WhereNode<ElemType>::Validate(bool isFinalValidationPass) /*override*/
{
ComputationNodeBase::Validate(isFinalValidationPass);
// we generate its own MBLayout
if (isFinalValidationPass && !Input(0)->HasMBLayout())
InvalidArgument("%ls %ls operation can only operate on minibatch data (which have a layout).", NodeName().c_str(), OperationName().c_str());
if (!m_pMBLayout)
{
m_pMBLayout = make_shared<MBLayout>(); // this generates a new layout
m_pMBLayout->SetUniqueAxisName(L"WhereNodeAxis");
}
// we map scalars to scalars
if (isFinalValidationPass && Input(0)->GetSampleLayout().GetNumElements() != 1)
InvalidArgument("%ls %ls operation can only operate on scalar input.", NodeName().c_str(), OperationName().c_str());
SetDims(TensorShape(1), true);
}
/*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());
}
void LearnableParameter<ElemType>::ReviseFromFile(const std::wstring& reviseFromFilePath)
{
try
{
InitFromFile(reviseFromFilePath);
}
catch (const std::exception & e)
{
RuntimeError("ReviseFromFile: Failed to reload %ls %ls operation from file %ls: %s", NodeName().c_str(), OperationName().c_str(), reviseFromFilePath.c_str(), e.what());
}
}
void LearnableParameter<ElemType>::InitFromArray(const std::vector<ElemType>& array, size_t numRows, size_t numCols)
{
// infer tensor dimensions from input file if not set
// Note: The mapping of dimensions of the input matrix to tensor dimensions are somewhat confusing.
// The file contains a 2D matrix (one row per text line) that is saved into our column-major representation.
// That representation is then reshaped into a column-major tensor.
if (GetSampleLayout().GetNumElements() == 0) // at least one dimension is 0
{
auto dims = GetSampleLayout().GetDims();
// infer rank
if (dims.size() == 0)
dims.push_back(0);
if (dims.size() == 1 && numCols != 1)
dims.push_back(0);
// infer #rows
if (dims[0] == 0) // infer row dimension as input matrix row dimension
dims[0] = numRows; // (if already set, then mismatch will be caught in VerifyDataSize() below)
// infer #cols: product of all dimensions but the first must match matrix #cols; if there is a single 0 position, we infer it
size_t zeroDim = 0; // 0 means not found
size_t prod = 1;
for (size_t k = 1; k < dims.size(); k++)
{
auto dim = dims[k];
if (dim != 0)
prod *= dim;
else if (zeroDim == 0)
zeroDim = k;
else
InvalidArgument("%ls %ls operation's specified shape [%s] cannot be inferred: Too many unknown dimensions.", NodeName().c_str(), OperationName().c_str(), string(GetSampleLayout()).c_str());
}
if (zeroDim != 0) // we found a zero
{
dims[zeroDim] = numCols / prod;
if (prod * dims[zeroDim] != numCols)
InvalidArgument("%ls %ls operation's specified shape [%s] cannot be inferred: Tensor shape cannot hold a [%d x %d] matrix.", NodeName().c_str(), OperationName().c_str(), string(GetSampleLayout()).c_str(), (int)numRows, (int)numCols);
}
SetDims(TensorShape(dims), false);
}
// BUGBUG: We should allow to read an arbitrary tensor from a single-column file.
// Currently, this would cause a matrix/tensor dimension mismatch. --TODO: Is this comment up-to-date?
Value().SetValue(numRows, numCols, m_deviceId, const_cast<ElemType*>(array.data()), matrixFlagNormal);
// TODO: Get rid of that const_cast, as soon as after Ryan's Matrix-lib refactoring separated out SetValue() from external vs. from deep copy
VerifyDataSize(Value()); // sanity check
}
