https://github.com/Microsoft/CNTK
Revision 9fc6fa4b112f3225b0035b3e65fae8b88e041465 authored by Zhou Wang on 19 May 2017, 06:30:25 UTC, committed by Zhou Wang on 19 May 2017, 06:30:25 UTC
1 parent 8b5035c
Tip revision: 9fc6fa4b112f3225b0035b3e65fae8b88e041465 authored by Zhou Wang on 19 May 2017, 06:30:25 UTC
make FindAllWithName private for all
make FindAllWithName private for all
Tip revision: 9fc6fa4
ReshapingNodes.cpp
//
// Copyright (c) Microsoft. All rights reserved.
// Licensed under the MIT license. See LICENSE.md file in the project root for full license information.
//
// ReshapingNodes.cpp -- collection of nodes that reshape or sub-sample matrices leading to layout changes
//
#include "Basics.h"
#include "ReshapingNodes.h"
#include "Matrix.h"
#include "ComputationNode.h"
#include "Sequences.h"
#include <unordered_set>
#include <map>
#include <string>
#include <vector>
#include <stdexcept>
#include <list>
#include <memory>
#include <algorithm>
#include <assert.h>
#include <stack>
#include <unordered_map>
namespace Microsoft { namespace MSR { namespace CNTK {
// -----------------------------------------------------------------------
// ReduceElements (op, axis=, input)
// -----------------------------------------------------------------------
template <class ElemType>
/*virtual*/ void ReduceElementsNode<ElemType>::CopyTo(ComputationNodeBasePtr nodeP, const std::wstring& newName, const CopyNodeFlags flags) const /*override*/
{
Base::CopyTo(nodeP, newName, flags);
if (flags & CopyNodeFlags::copyNodeValue)
{
auto node = dynamic_pointer_cast<ReduceElementsNode<ElemType>>(nodeP);
node->m_axis = m_axis;
node->m_operation = m_operation;
node->m_reductionOp = m_reductionOp;
node->m_scale = m_scale;
node->m_keepDimensions = m_keepDimensions;
}
}
template <class ElemType>
/*virtual*/ void ReduceElementsNode<ElemType>::Load(File& fstream, size_t modelVersion) /*override*/
{
Base::Load(fstream, modelVersion);
fstream >> m_axis >> m_operation;
if (modelVersion >= CNTK_MODEL_VERSION_24)
fstream >> m_keepDimensions;
else
m_keepDimensions = DefaultKeepDimensionsSetting(m_axis);
ValidateOp();
}
template <class ElemType>
/*virtual*/ void ReduceElementsNode<ElemType>::Save(File& fstream) const /*override*/
{
Base::Save(fstream);
fstream << m_axis << m_operation; // note: we serialize the string and not the opcode, since opcodes may change
fstream << m_keepDimensions;
}
template <class ElemType>
/*virtual*/ void ReduceElementsNode<ElemType>::ForwardProp(const FrameRange& fr) /*override*/
{
// We are mixing two kinds of operations here; elementwise and whole-batch or sequence reduction (ReduceAllAxes()).
// In the latter case, we must mimic the behaviour of ComputationNodeNonLooping.
if ((ReduceAllAxes() || ReduceSequenceAxis() || ReduceBatchAxis()) && !fr.IsAllFrames())
LogicError("%ls: %s node should never be in a loop when reducing over all static and dynamic axes or just the sequence axis.", Base::NodeDescription().c_str(), typeid(*this).name());
const auto frInput = (ReduceAllAxes() || ReduceBatchAxis()) ? FrameRange(InputRef(0).GetMBLayout()) : fr; // can't use 'fr' for ReduceAllAxes() and ReduceBatchAxis() as it refers to the result (same as for training criteria)
// when reducing all, we must mask gaps
if (ReduceAllAxes() || ReduceBatchAxis())
{
InputRef(0).MaskMissingValueColumnsTo(frInput, NeutralValue(m_reductionOp));
if (IsMean())
{
// In mean reduction for all axes or batch axis, we need to carefully compute the scaling factor
auto actual_samples = InputRef(0).HasMBLayout() ? InputRef(0).GetMBLayout()->GetActualNumSamples() : 1;
m_scale = ElemType(1.0 / actual_samples);
if (ReduceAllAxes())
m_scale /= ElemType(GetInputSampleLayout(0).GetNumElements());
}
}
// Create a new layout if we are reducing the sequence axis
if (ReduceSequenceAxis())
{
auto inputMBLayout = InputRef(0).GetMBLayout();
if (inputMBLayout->HasSequenceBeyondBegin() || inputMBLayout->HasSequenceBeyondEnd())
LogicError("%ls: %s node cannot perform sequence axis reduction for truncated sequence.", Base::NodeDescription().c_str(), typeid(*this).name());
GetMBLayout()->InitAsFrameMode(inputMBLayout->GetNumSequences());
UpdateFunctionValuesSize();
}
// get the args
size_t rank = DetermineElementwiseTensorRank();
TensorView<ElemType> input;
if (ReduceSequenceAxis())
{
ElemType gapPadValue = NeutralValue(m_reductionOp);
input = ComputationNode<ElemType>::Unpack(GetSampleLayout(), InputRef(0).Value(), InputRef(0).GetMBLayout(), m_tempUnpackedData, m_tempScatterIndices, m_tempMask, /*batchMajor=*/ true, &gapPadValue);
}
else
input = InputRef(0).ValueTensorFor(rank, frInput);
auto result = ReduceAllAxes() ? TensorView<ElemType>(ValuePtr(), GetSampleLayout()) : ValueTensorFor(rank, fr);
switch (m_reductionOp)
{
case ElementWiseOperator::opArgmin:
case ElementWiseOperator::opArgmax:
result.DoArgReductionOpOf(input, m_reductionOp);
break;
default:
// the actual operation is a Copy with reduction, where the magic is in the reduction op
// For "Mean", m_scale is 1/#elements, and 1 otherwise.
result.DoUnaryOpOf(0, input, m_scale, ElementWiseOperator::opCopy, m_reductionOp);
}
}
template <class ElemType>
/*virtual*/ void ReduceElementsNode<ElemType>::BackpropTo(const size_t inputIndex, const FrameRange& fr) /*override*/
{
assert(inputIndex == 0), inputIndex;
if (ReduceSequenceAxis())
{
// Broadcast along the sequence
auto result = ValueFor(fr);
ComputationNode<ElemType>::BroadcastToPacked(Gradient(), GetMBLayout(), /*beta =*/ 1, InputRef(0).Gradient(), FrameRange(InputRef(0).GetMBLayout()), m_tempGatherIndices);
}
else
{
const auto frInput = (ReduceAllAxes() || ReduceBatchAxis()) ? FrameRange(InputRef(0).GetMBLayout()) : fr; // can't use 'fr' for ReduceAllAxes() as it refers to the result (same as for training criteria)
// get the args
size_t rank = DetermineElementwiseTensorRank();
auto sliceOutputGrad = ReduceAllAxes() ? TensorView<ElemType>(GradientPtr(), GetSampleLayout()) : GradientTensorFor(rank, fr); // propagate from this one...
auto sliceInputGrad = InputRef(0).GradientTensorFor(rank, frInput); // ...to this one
// gradients are not as simple as passing an op-code, unfortunately
switch (m_reductionOp)
{
case ElementWiseOperator::opSum:
// "Sum": broadcast the gradient
// "Mean": same as "Sum" with scaling by 1/#dims
sliceInputGrad.AddCopyOf(sliceOutputGrad, m_scale);
break;
case ElementWiseOperator::opLogSum:
{
auto input = InputRef(inputIndex).ValueTensorFor(rank, frInput);
auto output = ValueTensorFor(rank, fr.AllowBroadcast());
// Let: f(x, y, z) = log(exp x + exp y + exp z)
// For the derivative we get:
// df / dx = exp(x)/exp(f)
// = exp(x - f)
sliceInputGrad.AddElementwiseProductWithExpOfDiffOf(sliceOutputGrad, input, output);
}
break;
case ElementWiseOperator::opMin:
case ElementWiseOperator::opMax:
{
auto input = InputRef(inputIndex).ValueTensorFor(rank, frInput);
auto output = ValueTensorFor(rank, fr.AllowBroadcast());
// POTENTIAL PROBLEM:
// For ReduceMin/Max there are combinations of input values where the gradient is not defined because the function has an edge at these points.
// E.g. for ReduceMin this is the case when the minimum input value is attained by several inputs at the same time.
// In these cases there is no correct gradient.The question is if this could lead to any problems.
// Let's look at two scenarios where this might happen:
//
// * Scenario 1: The input comes from a layer of nodes like e.g. ReLU and some of them might operate in the regime where they clip to a constant value.
// In this case it's not a problem that the input gradient is kind of bad as the derivative of the concerning input nodes will be zero anyway.
//
// * Scenario 2: The input data is directly coming from training data. Here bad gradients don't matter as we wouldn't wan't to propagate gradients to the training data.
//
// So as we don't have a better solution yet and it probably doesn't have impact let's stay with the current solution.
// Also note that for Clip , Min, Max and ReLU we have the same kind of problem.
sliceInputGrad.AddCopyIfEqualOf(input, output, sliceOutputGrad);
}
break;
case ElementWiseOperator::opElementwiseProduct:
{
auto input = InputRef(inputIndex).ValueTensorFor(rank, frInput);
auto output = ValueTensorFor(rank, fr.AllowBroadcast());
sliceInputGrad.AddElementwiseProductWithQuotientOf(sliceOutputGrad, output, input);
break;
}
case ElementWiseOperator::opArgmin:
case ElementWiseOperator::opArgmax:
break;
// more coming
}
}
}
template <class ElemType>
/*virtual*/ bool ReduceElementsNode<ElemType>::OutputUsedInComputingInputNodesGradients() const /*override*/
{
switch (m_reductionOp)
{
case ElementWiseOperator::opSum: return false;
case ElementWiseOperator::opLogSum: return true;
case ElementWiseOperator::opMin: return true;
case ElementWiseOperator::opMax: return true;
case ElementWiseOperator::opElementwiseProduct: return true;
case ElementWiseOperator::opArgmin: return false;
case ElementWiseOperator::opArgmax: return false;
}
LogicError("Should not get here.");
}
template <class ElemType>
/*virtual*/ bool ReduceElementsNode<ElemType>::InputUsedInComputingInputNodesGradients(size_t inputIndex) const /*override*/
{
switch (m_reductionOp)
{
case ElementWiseOperator::opSum: return false;
case ElementWiseOperator::opLogSum: return true;
case ElementWiseOperator::opMin: return true;
case ElementWiseOperator::opMax: return true;
case ElementWiseOperator::opElementwiseProduct: return true;
case ElementWiseOperator::opArgmin: return false;
case ElementWiseOperator::opArgmax: return false;
}
LogicError("Should not get here.");
}
// map the operation specified as a string to an ElementWiseOperator value.
template <class ElemType>
void ReduceElementsNode<ElemType>::ValidateOp()
{
m_reductionOp = ReductionOpEnumValue(m_operation);
}
template <class ElemType>
/*virtual*/ void ReduceElementsNode<ElemType>::Validate(bool isFinalValidationPass) /*override*/
{
// validate the opcode (in case we got instantiated empty and never updated)
ValidateOp();
m_scale = (ElemType)1;
if (ReduceAllAxes())
Base::ValidateUnaryReduce(isFinalValidationPass, m_keepDimensions);
else if (ReduceSequenceAxis())
{
Base::Validate(isFinalValidationPass);
// we generate its own MBLayout
if (isFinalValidationPass && !Input(0)->HasMBLayout())
InvalidArgument("%ls %ls operation can perform sequence axis reduction only on minibatch data (which have a layout).", NodeName().c_str(), OperationName().c_str());
if ((m_operation != L"Sum") && (m_operation != L"Plus"))
InvalidArgument("%ls %ls operation can perform sequence axis reduction only for the 'sum' reduction operation, specified operation %ls.", NodeName().c_str(), OperationName().c_str(), m_operation.c_str());
if (!m_pMBLayout)
{
m_pMBLayout = make_shared<MBLayout>(); // this generates a new layout
m_pMBLayout->SetUniqueAxisName(ComputationNodeBase::DefaultNoSequenceAxisName);
}
SetDims(Input(0)->GetSampleLayout(), HasMBLayout());
}
else if (ReduceBatchAxis())
{
Base::Validate(isFinalValidationPass);
if (isFinalValidationPass && !Input(0)->HasMBLayout())
InvalidArgument("%ls %ls operation can perform batch axis reduction only on minibatch data (which have a layout).", NodeName().c_str(), OperationName().c_str());
SetDims(Input(0)->GetSampleLayout(), false);
}
else
{
Base::Validate(isFinalValidationPass);
InferMBLayoutFromInputsForStandardCase(isFinalValidationPass);
let shape = Input(0)->GetSampleLayout();
auto dims = shape.GetDims();
size_t reducedDim = 0; // (init to keep compiler happy)
if (ReduceAllStaticAxes())
{
reducedDim = shape.GetNumElements();
dims = m_keepDimensions ? SmallVector<size_t>(shape.GetRank(), 1) : (Environment().IsV2Library() ? SmallVector<size_t>({}) : SmallVector<size_t>({ 1 })); // entire sample is reduced to a scalar
}
else if (m_axis - 1 >= 0 && m_axis - 1 < dims.size())
{
reducedDim = dims[m_axis - 1];
// one axis is reduced to a scalar
if (m_keepDimensions)
dims[m_axis - 1] = 1;
else
{
SmallVector<size_t> reducedDims(dims.size() - 1);
for (size_t i = 0, j = 0; i < dims.size(); ++i)
{
if (i == (m_axis - 1))
continue;
reducedDims[j] = dims[i];
j++;
}
dims = reducedDims;
}
}
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 && IsMean())
m_scale = (ElemType)(1.0 / reducedDim);
SetDims(TensorShape(dims), Input(0)->HasMBLayout());
}
}
template class ReduceElementsNode<float>;
template class ReduceElementsNode<double>;
// -----------------------------------------------------------------------
// Where(bitVector) -- extract indices of non-0 values in a sequence
// -----------------------------------------------------------------------
// wrapper class to pass MBLayout sequence vector to PackSequences()
struct SequenceLengthVector
{
typedef vector<vector<size_t>> SequenceVector;
typedef MBLayout::SequenceInfo SequenceInfo;
const SequenceVector& m_sequenceVector; // vector of sequences (to get sequence length)
const vector<SequenceInfo>& m_sequenceInfo; // original sequence info (for seqId)
SequenceLengthVector(const vector<SequenceInfo>& sequenceInfo, const SequenceVector& sequenceVector) : m_sequenceInfo(sequenceInfo), m_sequenceVector(sequenceVector) { }
size_t size() const { return m_sequenceInfo.size(); }
MBLayout::SequenceInfo operator[](size_t i) const // return a descriptor of the new sequence
{
SequenceInfo seq;
seq.seqId = m_sequenceInfo[i].seqId;
seq.s = i;
seq.tBegin = 0;
seq.tEnd = m_sequenceVector[i].size();
return seq;
}
void operator=(const SequenceLengthVector&) = delete;
};
// TODO: Where should the MBLayout be created--in BeginForwardProp() or ForwardProp()?
// BeginForwardProp() should generally have no access to the actual values,
// while ForwardProp() might be too late. We may have to define the semantics here.
// BUGBUG: This is the first node with value-dependent MBLayout. It resizes Value(), which we otherwise always do before.
template <class ElemType>
/*virtual*/ void WhereNode<ElemType>::ForwardPropNonLooping() /*override*/
{
// gather all sequences
let& inMBLayout = InputRef(0).GetMBLayout();
let& input = InputRef(0).Value();
let& sequences = inMBLayout->GetAllSequences();
auto& indexSequences = m_indexSequenceBuffer;
if (indexSequences.size() < sequences.size())
indexSequences.resize(sequences.size());
for (size_t i = 0; i < sequences.size(); i++)
{
let& seq = sequences[i];
if (seq.seqId == GAP_SEQUENCE_ID)
continue;
auto& indexSequence = indexSequences[i];
// create index map for one sequence
// this is the condition check that this node performs; the meat
indexSequence.clear();
double desiredCount = 0.0;
for (size_t t = 0; t < seq.GetNumTimeSteps(); t++)
{
double delta = input(0, inMBLayout->GetColumnIndex(seq, t)); // how many frames the current time step should expand into
desiredCount += delta; // this is now how many frames we should have
// use a margin against round-off errors, so that we get non-binary ratios like 1/3 and 1/5 right
// This really means generate a frame if too few, unless we are within machine accuracy of the target.
// The assumption is that the delta has this error, while accumulation (in double) has no error.
ElemType relativeMargin = 1 - std::numeric_limits<ElemType>::epsilon();
while ((indexSequence.empty() && desiredCount > 0) // no margin for the first frame (always include unless flag is 0)
|| indexSequence.size() < desiredCount * relativeMargin)
indexSequence.push_back(t);
}
// Note: The above accesses m_value directly on the CPU, putting it into BOTH state, possibly for other consumers as well.
}
input.CollapseDataLocation(); // BUGBUG: Move back, since BOTH state is broken at present.
// create a new MBLayout
let& outMBLayout = GetMBLayout();
outMBLayout->InitAsPackedSequences(SequenceLengthVector(sequences, indexSequences), /*temp*/m_placementBuffer, /*temp*/m_rowAllocationsBuffer);
// copy to output
vector<ElemType> buf(outMBLayout->GetNumCols(), numeric_limits<ElemType>::quiet_NaN()); // STL cannot easily avoid initializing, so we might as well init with NaN for gaps
let size = min(sequences.size(), outMBLayout->GetAllSequences().size()); // no non-gap sequence has an index beyond this
for (size_t i = 0; i < size; i++)
{
let& seq = outMBLayout->GetAllSequences()[i];
if (seq.seqId == GAP_SEQUENCE_ID) // gaps will keep the NaN
continue;
let& indexSequence = indexSequences[i];
for (size_t t = 0; t < seq.GetNumTimeSteps(); t++)
buf[outMBLayout->GetColumnIndex(seq, t)] = (ElemType)indexSequence[t];
}
// there may be dangling gaps at the end. Take the opportunity to verify this.
for (size_t i = size; i < sequences.size(); i++)
assert(sequences[i].seqId == GAP_SEQUENCE_ID);
for (size_t i = size; i < outMBLayout->GetAllSequences().size(); i++)
assert(outMBLayout->GetAllSequences()[i].seqId == GAP_SEQUENCE_ID);
// the result will be kept in CPUDEVICE, since most likely we will access it again in PackedIndexNode
Value().TransferToDeviceIfNotThere(CPUDEVICE, /*isBeingMoved=*/ true, /*emptyTransfer=*/ true, /*updatePreferredDevice=*/ true);
Value().SetValue(1, outMBLayout->GetNumCols(), CPUDEVICE, buf.data(), MatrixFormat::matrixFormatColMajor);
}
template <class ElemType>
/*virtual*/ void WhereNode<ElemType>::BackpropToNonLooping(size_t /*inputIndex*/) /*override*/
{
// we cannot backprop through a condition
return;
}
template <class ElemType>
/*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(m_dynamicAxisName);
}
// 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::Scalar(Environment().IsV2Library()), true);
}
template class WhereNode<float>;
template class WhereNode<double>;
// -----------------------------------------------------------------------
// PackedIndexNode(targetObject, indexSequence) -- map sequence
// -----------------------------------------------------------------------
template <class ElemType>
/*virtual*/ void PackedIndexNode<ElemType>::ForwardPropNonLooping() /*override*/
{
let& sourceMBLayout = InputRef(SOURCEDATA).GetMBLayout(); // only used for index conversion
let& indexMBLayout = InputRef(INDEXDATA).GetMBLayout();
let& index = InputRef(INDEXDATA).Value(); // per-seq index values that are to be mapped
auto& result = Value(); // packed index values as mapped to sourceData's layout
// loop over sourceSequences
// Input matrix contains time indices for each sequence that refer to frames inside that sequence.
// We replace every per-sequence index by the resolved column index w.r.t. the same MBLayout.
let& sourceSequences = sourceMBLayout->GetAllSequences();
for (size_t i = 0; i < sourceSequences.size(); i++)
{
let& sourceSeq = sourceSequences[i];
if (sourceSeq.seqId == GAP_SEQUENCE_ID)
continue;
let& indexSeq = indexMBLayout->FindMatchingSequence(sourceSequences, i); // find corresponding entry in indexMBLayout
for (size_t tIndex = 0; tIndex < indexSeq.GetNumTimeSteps(); tIndex++) // map all index values in index sequence
{
let jIndex = indexMBLayout->GetColumnIndex(indexSeq, tIndex); // map time index to actual location in the matrix storage object
let tSource = (size_t)index(0, jIndex); // the new time location (relative to source sequence)
let jSource = sourceMBLayout->GetColumnIndex(sourceSeq, tSource); // map new time index as well. This performs a range check.
result(0, jIndex) = (ElemType)jSource;
}
}
// Note: maybe this is no longer needed, now that we do the same inside UpdateFunctionValueSize() for all nodes.
result.CollapseDataLocation(); // BUGBUG: Move back, since BOTH state is broken at present.
}
template <class ElemType>
/*virtual*/ void PackedIndexNode<ElemType>::BackpropToNonLooping(size_t /*inputIndex*/) /*override*/
{
// we cannot backprop through a condition
// Can we?
return;
}
template <class ElemType>
/*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());
}
template class PackedIndexNode<float>;
template class PackedIndexNode<double>;
// -----------------------------------------------------------------------
// GatherPackedNode(packedIndex, sourceData) -- gather operation
// -----------------------------------------------------------------------
template <class ElemType>
/*virtual*/ void GatherPackedNode<ElemType>::ForwardPropNonLooping() /*override*/
{
InputRef(INDEXDATA).MaskMissingValueColumnsTo(FrameRange(InputRef(INDEXDATA).GetMBLayout()), -1); // indicates an invalid column to Gather/Scatter
let& index = InputRef(INDEXDATA) .Value(); // column indices to copy from
let& source = InputRef(SOURCEDATA).Value(); // source data to copy
#ifdef _MSC_VER
auto& outputValuePtrRef = ValuePtrRef();
#else
auto& outputValuePtrRef = this->template ValuePtrRef();
#endif
if ((source.GetMatrixType() == SPARSE) && (outputValuePtrRef->GetMatrixType() != SPARSE))
outputValuePtrRef = std::make_shared<Matrix<ElemType>>(outputValuePtrRef->GetNumRows(),
outputValuePtrRef->GetNumCols(),
outputValuePtrRef->GetPreferredDeviceId(),
source.GetMatrixType(),
source.GetFormat());
auto& output = Value(); // output goes here
output.DoGatherColumnsOf(/*beta=*/0, index, source, /*alpha=*/1);
}
template <class ElemType>
/*virtual*/ void GatherPackedNode<ElemType>::BackpropToNonLooping(size_t inputIndex) /*override*/
{
if (inputIndex == SOURCEDATA)
{
let& index = InputRef(INDEXDATA) .Value(); // column indices to copy from
auto& sourceGradient = InputRef(SOURCEDATA).Gradient(); // source to propagate the gradient intpu
auto& outputGradient = Gradient(); // output gradient to propagate
sourceGradient.DoScatterColumnsOf(/*beta=*/1, index, outputGradient, /*alpha=*/1);
}
}
template <class ElemType>
/*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());
}
}
template class GatherPackedNode<float>;
template class GatherPackedNode<double>;
// -----------------------------------------------------------------------
// ScatterPackedNode(layoutData, packedIndex, sourceData) -- scatter operation
// -----------------------------------------------------------------------
template <class ElemType>
/*virtual*/ void ScatterPackedNode<ElemType>::ForwardPropNonLooping() /*override*/
{
if (*InputRef(INDEXDATA).GetMBLayout() != *InputRef(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());
InputRef(INDEXDATA).MaskMissingValueColumnsTo(FrameRange(InputRef(INDEXDATA).GetMBLayout()), -1); // indicates an invalid column to Gather/Scatter
let& index = InputRef(INDEXDATA) .Value(); // column indices to copy from
let& source = InputRef(SOURCEDATA).Value(); // source data to copy
#ifdef _MSC_VER
auto& outputValuePtrRef = ValuePtrRef();
#else
auto& outputValuePtrRef = this->template ValuePtrRef();
#endif
if ((source.GetMatrixType() == SPARSE) && (outputValuePtrRef->GetMatrixType() != SPARSE))
outputValuePtrRef = std::make_shared<Matrix<ElemType>>(outputValuePtrRef->GetNumRows(),
outputValuePtrRef->GetNumCols(),
outputValuePtrRef->GetPreferredDeviceId(),
source.GetMatrixType(),
source.GetFormat());
auto& output = Value(); // output goes here
output.DoScatterColumnsOf(/*beta=*/0, index, source, /*alpha=*/1);
}
template <class ElemType>
/*virtual*/ void ScatterPackedNode<ElemType>::BackpropToNonLooping(size_t inputIndex) /*override*/
{
if (inputIndex == SOURCEDATA)
{
let& index = InputRef(INDEXDATA).Value(); // column indices to copy from
auto& sourceGradient = Input(SOURCEDATA)->Gradient(); // source to propagate the gradient input
auto& outputGradient = Gradient(); // output gradient to propagate
sourceGradient.DoGatherColumnsOf(/*beta=*/1, index, outputGradient, /*alpha=*/1);
}
}
template <class ElemType>
/*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());
}
template class ScatterPackedNode<float>;
template class ScatterPackedNode<double>;
// -----------------------------------------------------------------------
// CropNode -- crop operation, crops first input according to shape of second
// input at offsets which are directly given or automatically calculated.
// -----------------------------------------------------------------------
template <class ElemType>
CropNode<ElemType>::CropNode(DEVICEID_TYPE deviceId, const wstring& name)
: Base(deviceId, name), m_xOffset(numeric_limits<double>::max()), m_yOffset(numeric_limits<double>::max())
{
}
template <class ElemType>
CropNode<ElemType>::CropNode(size_t offsetX, size_t offsetY, DEVICEID_TYPE deviceId, const wstring& name)
: CropNode(deviceId, name)
{
m_xOffset = (double)(offsetX);
m_yOffset = (double)(offsetY);
}
template <class ElemType>
CropNode<ElemType>::CropNode(const ScriptableObjects::IConfigRecordPtr configp)
: CropNode(configp->Get(L"deviceId"), L"<placeholder>")
{
// We may have 2 or 4 node inputs, check that and attach them.
const auto inputs = GetInputsFromConfig(configp);
if (inputs.size() != 2 && inputs.size() != 4)
LogicError("Crop node must have 2 or 4 node inputs.");
AttachInputs(inputs);
// Here we have 3 possibilities:
// 1. 2 input nodes -> auto crop calculation without equivalence nodes
// 2. 2 input nodes + 2 parameters -> manual crop with given offsets
// 3. 4 inputs -> auto crop calculation with equivalence nodes
if (inputs.size() == 2)
{
// We have 2 input nodes related to cropping (no equivalence node inputs given). Check if we have offsets
// directly given.
if (configp->Exists(L"yOffset") && configp->Exists(L"xOffset"))
{
// We have manual crop with given offsets (option 2. above). Save given offsets.
m_xOffset = configp->Get(L"xOffset");
m_yOffset = configp->Get(L"yOffset");
}
// else: Offsets not given (option 1. above), we have automatic crop calculation without equivalence nodes.
}
// else: We have 4 node inputs (option 3. above), we have automatic crop calculation with equivalence nodes.
}
template <class ElemType>
void CropNode<ElemType>::Validate(bool isFinalValidationPass)
{
Base::Validate(isFinalValidationPass);
InferMBLayoutFromInputsForStandardCase(isFinalValidationPass);
// Here we need to determine output dimensions which are same as dimensions of second input.
TensorShape inputShape0 = Input(0)->GetSampleLayout();
TensorShape inputShape1 = Input(1)->GetSampleLayout();
SmallVector<size_t> inDims = inputShape0.GetDims();
SmallVector<size_t> inDimsCropped = inputShape1.GetDims();
// We assume we have at least two dimensions (first two are to be cropped).
if (inDims.size() < 2 || inDimsCropped.size() < 2)
RuntimeError("Crop input samples must have at least two dimensions.");
// Output dimensions are equal to input dimensions with first two axis copied from cropped dimensions.
SmallVector<size_t> outDims = inDims;
outDims[0] = inDimsCropped[0];
outDims[1] = inDimsCropped[1];
// Set output dimensions.
SetDims(TensorShape(outDims), HasMBLayout());
if (isFinalValidationPass)
{
// In final validation pass we compute crop offsets if needed.
ComputeCropOffsets();
// Cropped input must be large enough to allow cropping at given offset.
if (inDims[0] < outDims[0] + m_xOffset)
RuntimeError("Input is small to be cropped along x dimension in crop node.");
if (inDims[1] < outDims[1] + m_yOffset)
RuntimeError("Input is small to be cropped along y dimension in crop node.");
}
}
template <class ElemType>
void CropNode<ElemType>::ForwardProp(const FrameRange& /*fr*/)
{
// Our offsets must be initialized here.
if (m_xOffset == numeric_limits<double>::max() || m_yOffset == numeric_limits<double>::max())
LogicError("Crop offsets not initialized in ForwardProp.");
// Retrieve input and output views for the values. Input and output views are tensor views
// that define parts of first input and output that we operate on (we copy input from input view
// to output).
CroppedIOViews ioViews = CreateIOViews(&ComputationNode<ElemType>::ValuePtr);
// Copy values from cropped input to output.
ioViews.outputView.AssignCopyOf(ioViews.inputViewCropped);
}
template <class ElemType>
void CropNode<ElemType>::BackpropTo(const size_t inputIndex, const FrameRange& /*fr*/)
{
// We propagate gradients just to the cropped input.
if (inputIndex == 0)
{
// Reset input gradients to ensure that non-cropped parts do not affect backprop.
Input(0)->Gradient().SetValue(0);
// Retrieve input and output views for the gradients. Input and output views are tensor views
// that define parts of first input and output that we operate on (we copy gradients from output view
// to input view).
CroppedIOViews ioViews = CreateIOViews(&ComputationNode<ElemType>::GradientPtr);
// Copy gradients from output to cropped input.
ioViews.inputViewCropped.AddCopyOf(ioViews.outputView);
}
}
template <class ElemType>
void CropNode<ElemType>::Save(File& fstream) const
{
Base::Save(fstream);
fstream << m_xOffset;
fstream << m_yOffset;
}
template <class ElemType>
void CropNode<ElemType>::Load(File& fstream, size_t modelVersion)
{
Base::Load(fstream, modelVersion);
fstream >> m_xOffset;
fstream >> m_yOffset;
}
template <class ElemType>
void CropNode<ElemType>::CopyTo(ComputationNodeBasePtr nodeP, const wstring& newName, const CopyNodeFlags flags) const
{
Base::CopyTo(nodeP, newName, flags);
if (flags & CopyNodeFlags::copyNodeValue)
{
auto node = dynamic_pointer_cast<CropNode<ElemType>>(nodeP);
node->m_xOffset = m_xOffset;
node->m_yOffset = m_yOffset;
}
}
template <class ElemType>
typename CropNode<ElemType>::CroppedIOViews CropNode<ElemType>::CreateIOViews(MatrixGetter matrixGetter)
{
// Get the shapes of the inputs.
TensorShape inputShape0 = Input(0)->GetTensorShape(Input(0)->GetSampleLayout().GetRank());
TensorShape inputShape1 = Input(1)->GetTensorShape(Input(1)->GetSampleLayout().GetRank());
// Calculate cropped shape of the input.
TensorShape inputShapeCropped = inputShape0;
inputShapeCropped.NarrowTo(0, (size_t)(m_xOffset), (size_t)(m_xOffset) + inputShape1.GetDim(0));
inputShapeCropped.NarrowTo(1, (size_t)(m_yOffset), (size_t)(m_yOffset) + inputShape1.GetDim(1));
// Get output shape.
TensorShape outputShape = GetTensorShape(GetSampleLayout().GetRank());
// Cropped input and output dimensions must be same.
if (inputShapeCropped.GetDims() != outputShape.GetDims())
LogicError("Cropped input and output must have same rank.");
// Create proper views using calculated shapes.
return CroppedIOViews(this, matrixGetter, inputShapeCropped, outputShape);
}
// ComputeCropOffsets computes offsets to be used for cropping if manual offsets are absent. The offsets are computed
// by traversing the network graph and finding common ancestor of crop node inputs. Once ancestor is found affine transform
// is computed along the paths from first and second input to common ancestor. Complete transform from one input to other it
// finally calculated composing these two transforms. Translate components of final transform define crop offsets.
template <class ElemType>
void CropNode<ElemType>::ComputeCropOffsets()
{
// Helper method for traversing the tree and calculating node transforms.
// For currNode, calculates coordinate maps of its inputs based on known coordinate maps of its outputs.
// nodeToTransformMap contains coordinate maps for all nodes traversed so far, and is updated by this function.
// Traversal stack contains all nodes traversed so far. Inputs of currNode are pushed to traversal stack so that their
// inputs can be processed later on.
auto ProcessInputs = [](ComputationNodeBase* currNode, stack<ComputationNodeBase*>& traversalStack, unordered_map<ComputationNodeBase*, SpaceTransform>& nodeToTransformMap)
{
if (!currNode->Is<TransformerNode>())
RuntimeError("Node does not support affine transform for cropping.");
auto transformerNode = currNode->As<TransformerNode>();
// Go over the nodes inputs.
for (size_t i = 0; i < currNode->GetNumInputs(); i++)
{
// Check if input-output transform is supported on the node.
if (transformerNode->SupportsTransformOnInput(i))
{
// Transform is supported, take the input.
ComputationNodeBase* currInput = currNode->GetInputs()[i].get();
// Take node transform from input to output.
const SpaceTransform& nodeTransform = transformerNode->GetTransformForInput(i);
// Calculate composite transform from node input to crop node.
SpaceTransform nodeToCropTransform = nodeToTransformMap.find(currNode)->second.Compose(nodeTransform);
// Check if we already visited this input node.
auto it = nodeToTransformMap.find(currInput);
if (it == nodeToTransformMap.end())
{
// We have not visited this node before. Add it to the transform map and to traversal stack to continue
// traversing its children.
nodeToTransformMap.insert(make_pair(currInput, nodeToCropTransform));
traversalStack.push(currInput);
}
else
{
// We have been here before, check that transforms along two different paths are same.
if (it->second != nodeToCropTransform)
{
// Different transforms along two different paths, should never happen.
RuntimeError("Different transforms along different paths in Crop node.");
}
}
}
}
};
if (m_xOffset != numeric_limits<double>::max() && m_yOffset != numeric_limits<double>::max())
{
// Offsets are already available, skip compute.
return;
}
// Used to keep nodes while traversing the network graph.
stack<ComputationNodeBase*> traversalStack;
// Maps node to transform between its output and crop node.
unordered_map<ComputationNodeBase*, SpaceTransform> nodeToCropInput0TransformMap;
unordered_map<ComputationNodeBase*, SpaceTransform> nodeToCropInput1TransformMap;
// Take equivalence nodes if provided.
ComputationNodeBase* equivalenceNode1 = nullptr;
ComputationNodeBase* equivalenceNode2 = nullptr;
if (GetInputs().size() == 4)
{
equivalenceNode1 = GetInputs()[2].get();
equivalenceNode2 = GetInputs()[3].get();
}
// Push first input to traversal stack to start exploring paths starting from there.
traversalStack.push(GetInputs()[0].get());
// Push first input transform as identity to enable composing transforms.
nodeToCropInput0TransformMap.insert(make_pair(GetInputs()[0].get(), SpaceTransform::Identity(2)));
// Start traversing graph starting from the first input.
while (!traversalStack.empty())
{
ComputationNodeBase* currNode = traversalStack.top();
traversalStack.pop();
ProcessInputs(currNode, traversalStack, nodeToCropInput0TransformMap);
}
// Now traverse from second input.
traversalStack.push(GetInputs()[1].get());
// Push second input transform as identity to enable composing transforms.
nodeToCropInput1TransformMap.insert(make_pair(GetInputs()[1].get(), SpaceTransform::Identity(2)));
// Once we meet node that is in nodeToCropInput0TransformMap or equivalence node we will compute offsets.
double xOffset = numeric_limits<double>::max();
double yOffset = numeric_limits<double>::max();
while (!traversalStack.empty())
{
ComputationNodeBase* currNode = traversalStack.top();
traversalStack.pop();
// Check if node is in the map corresponding to the first input (path connected over common ancestor).
auto it = nodeToCropInput0TransformMap.find(currNode);
const SpaceTransform* firstInputTransform = nullptr;
if (it != nodeToCropInput0TransformMap.end())
{
// We have closed the path between nodes, save the first input transform.
firstInputTransform = &it->second;
}
// Check if node is equivalent to one from the first subtree (path connected over equivalence nodes).
else if (currNode == equivalenceNode2)
{
// We have closed the path between nodes using equivalence nodes, save the first equivalence node transform.
firstInputTransform = &nodeToCropInput0TransformMap.find(equivalenceNode1)->second;
}
if (firstInputTransform)
{
// Calculate final transform.
SpaceTransform finalTransform = nodeToCropInput1TransformMap.find(currNode)->second.Compose(firstInputTransform->Inverse());
for (size_t ia = 0; ia < finalTransform.m_axisTransforms.size(); ia++)
{
// In crop node we expect no scaling.
if (finalTransform.m_axisTransforms[ia].scale != 1.0f)
RuntimeError("Composite transform has non 1 scale in crop node.");
if (finalTransform.m_axisTransforms[ia].translate > 0)
RuntimeError("Composite transform has positive translate (negative offset) in crop node.");
}
// Crop offsets are defined with transform translations.
xOffset = -finalTransform.m_axisTransforms[0].translate;
yOffset = -finalTransform.m_axisTransforms[1].translate;
// Finished.
break;
}
// No connected path, keep searching.
ProcessInputs(currNode, traversalStack, nodeToCropInput1TransformMap);
}
if (xOffset == numeric_limits<double>::max() || yOffset == numeric_limits<double>::max())
LogicError("Connected path between crop inputs not found. Unable to compute crop offsets.");
// Save computed offsets.
m_xOffset = xOffset;
m_yOffset = yOffset;
}
template <class ElemType>
void CropNode<ElemType>::ComputeTransforms()
{
if (m_transforms[0].m_axisTransforms.empty())
{
m_transforms[0].m_axisTransforms.resize(2);
m_transforms[0].m_axisTransforms[0].scale = 1;
m_transforms[0].m_axisTransforms[0].translate = -m_xOffset;
m_transforms[0].m_axisTransforms[1].scale = 1;
m_transforms[0].m_axisTransforms[1].translate = -m_yOffset;
}
// else: already computed.
}
template <class ElemType>
bool CropNode<ElemType>::SupportsTransformOnInput(size_t inputIndex)
{
// We support transform on cropped input.
return (inputIndex == 0);
}
template class CropNode<float>;
template class CropNode<double>;
}}}
Computing file changes ...