https://github.com/Microsoft/CNTK
Tip revision: 3bef0f4214cb0b2aa4e646cba731076483036eb4 authored by liqun fu on 13 February 2019, 19:08:06 UTC
protobuf 3.6.0 to 3.6.1
protobuf 3.6.0 to 3.6.1
Tip revision: 3bef0f4
LinearAlgebraNodes.h
//
// Copyright (c) Microsoft. All rights reserved.
// Licensed under the MIT license. See LICENSE.md file in the project root for full license information.
//
#pragma once
#include "Basics.h"
#include "ComputationNode.h"
#include "Constants.h"
#include "Matrix.h"
#include "TensorView.h"
#include <unordered_set>
#include <map>
#include <string>
#include <vector>
#include <stdexcept>
#include <list>
#include <memory>
#include <algorithm>
#include <utility>
#include <assert.h>
#include <set>
#include "Quantizers.h"
#include "InputAndParamNodes.h"
namespace Microsoft { namespace MSR { namespace CNTK {
// -----------------------------------------------------------------------
// PlusNode (summand1, summand2)
// -----------------------------------------------------------------------
template <class ElemType>
class PlusNode : public BinaryElementWiseNode<ElemType>
{
typedef BinaryElementWiseNode<ElemType> Base; UsingBinaryElementwiseNodeBaseMembers;
static const std::wstring TypeName() { return L"Plus"; }
public:
DeclareConstructorFromConfigWithNumInputs(PlusNode);
PlusNode(DEVICEID_TYPE deviceId, const wstring& name)
: Base(deviceId, name)
{
}
virtual void /*ComputationNode::*/ ForwardProp(const FrameRange& fr) override
{
size_t rank = DetermineElementwiseTensorRank();
auto result = ValueTensorFor(rank, fr);
auto input0 = InputRef(0).ValueTensorFor(rank, fr.AllowBroadcast());
auto input1 = InputRef(1).ValueTensorFor(rank, fr.AllowBroadcast());
result.AssignSumOf(input0, input1);
}
virtual void /*ComputationNode::*/ BackpropTo(const size_t inputIndex, const FrameRange& fr) override
{
size_t rank = DetermineElementwiseTensorRank();
auto gradient = GradientTensorFor(rank, fr);
auto inputGradient = Input(inputIndex)->GradientTensorFor(rank, fr.AllowBroadcast());
// if reduction then mask the respective input(s) (zero out the gaps)
if (Input(inputIndex)->ReducesInTimeWrt(shared_from_this()))
MaskMissingGradientColumnsToZero(fr);
if (Input(inputIndex)->IsGradientOptimized(this))
{
if (Input(inputIndex)->ParentGradientReused())
{
if (inputGradient.GetSOBPtr() != gradient.GetSOBPtr())
LogicError("Gradients should be reused.");
}
else
inputGradient.AssignCopyOf(gradient);
}
else
inputGradient.AddCopyOf(gradient);
}
virtual ParentGradientOptimization ImplementsGradientOptimization(const ComputationNodeBase* input) const override
{
size_t i;
for (i = 0; i < GetNumInputs(); i++)
{
if (Input(i).get() == input) break;
}
if (i == GetNumInputs())
LogicError("Cannot find input.");
return this->InputMatchesOutput(i) ? ParentGradientOptimization::Reuse : ParentGradientOptimization::Overwrite;
}
};
template class PlusNode<float>;
template class PlusNode<double>;
template class PlusNode<half>;
// -----------------------------------------------------------------------
// LogPlusNode (summand1, summand2)
// Computes ln(exp(summand1) + exp(summand2)) in an overflow safe way.
// Useful e.g. for computing softmax over sequence.
// -----------------------------------------------------------------------
template <class ElemType>
class LogPlusNode : public BinaryElementWiseNode<ElemType>
{
typedef BinaryElementWiseNode<ElemType> Base; UsingBinaryElementwiseNodeBaseMembers;
static const std::wstring TypeName() { return L"LogPlus"; }
public:
DeclareConstructorFromConfigWithNumInputs(LogPlusNode);
LogPlusNode(DEVICEID_TYPE deviceId, const wstring& name)
: Base(deviceId, name)
{
}
virtual void /*ComputationNode::*/ ForwardProp(const FrameRange& fr) override
{
size_t rank = DetermineElementwiseTensorRank();
auto result = ValueTensorFor(rank, fr);
auto input0 = InputRef(0).ValueTensorFor(rank, fr.AllowBroadcast());
auto input1 = InputRef(1).ValueTensorFor(rank, fr.AllowBroadcast());
result.AssignLogSumOf(input0, input1);
}
virtual void /*ComputationNode::*/ BackpropTo(const size_t inputIndex, const FrameRange& fr) override
{
size_t rank = DetermineElementwiseTensorRank();
auto gradient = GradientTensorFor(rank, fr);
auto inputGradient = InputRef(inputIndex).GradientTensorFor(rank, fr.AllowBroadcast());
auto input0 = InputRef(0).ValueTensorFor(rank, fr.AllowBroadcast());
auto input1 = InputRef(1).ValueTensorFor(rank, fr.AllowBroadcast());
// if reduction then mask the respective input(s) (zero out the gaps)
if (Input(inputIndex)->ReducesInTimeWrt(shared_from_this()))
MaskMissingGradientColumnsToZero(fr);
if (Input(inputIndex)->ReducesInTimeWrt(Input(1 - inputIndex)))
Input(1 - inputIndex)->MaskMissingValueColumnsToZero(fr);
if (inputIndex == 0)
{
// d/dx (ln( exp(x) + (exp(y)) = exp(x) / (exp(x) + exp(y)) = 1 / (1 + exp(y-x)) = sigmoid(x-y)
if (Input(inputIndex)->IsGradientInitializedBy(this))
inputGradient.AssignElementwiseProductWithLogSumDerivativeOf(gradient, input1, input0);
else
inputGradient.AddElementwiseProductWithLogSumDerivativeOf(gradient, input1, input0);
}
else
{
// d/dy (ln( exp(x) + (exp(y)) = exp(y) / (exp(x) + exp(y)) = 1 / (1 + exp(x-y)) = sigmoid(y-x)
if (Input(inputIndex)->IsGradientInitializedBy(this))
inputGradient.AssignElementwiseProductWithLogSumDerivativeOf(gradient, input0, input1);
else
inputGradient.AddElementwiseProductWithLogSumDerivativeOf(gradient, input0, input1);
}
}
virtual bool InputUsedInComputingInputNodesGradients(size_t /*childIndex*/) const override
{
return true;
}
virtual ParentGradientOptimization ImplementsGradientOptimization(const ComputationNodeBase*) const override
{
return ParentGradientOptimization::Overwrite;
}
};
template class LogPlusNode<float>;
template class LogPlusNode<double>;
template class LogPlusNode<half>;
// -----------------------------------------------------------------------
// PowNode (base, exponent)
// Computes base ** exponent.
// -----------------------------------------------------------------------
template <class ElemType>
class PowNode : public BinaryElementWiseNode<ElemType>
{
typedef BinaryElementWiseNode<ElemType> Base; UsingBinaryElementwiseNodeBaseMembers;
static const std::wstring TypeName() { return L"Pow"; }
public:
DeclareConstructorFromConfigWithNumInputs(PowNode);
PowNode(DEVICEID_TYPE deviceId, const wstring& name)
: Base(deviceId, name)
{
}
virtual void /*ComputationNode::*/ ForwardProp(const FrameRange& fr) override
{
size_t rank = DetermineElementwiseTensorRank();
auto result = ValueTensorFor(rank, fr);
auto base = InputRef(0).ValueTensorFor(rank, fr.AllowBroadcast());
auto expo = InputRef(1).ValueTensorFor(rank, fr.AllowBroadcast());
result.AssignPowOf(base, expo);
}
virtual void /*ComputationNode::*/ BackpropTo(const size_t inputIndex, const FrameRange& fr) override
{
size_t rank = DetermineElementwiseTensorRank();
auto gradient = GradientTensorFor(rank, fr);
auto inputGradient = InputRef(inputIndex).GradientTensorFor(rank, fr.AllowBroadcast());
auto base = InputRef(0).ValueTensorFor(rank, fr.AllowBroadcast());
// if reduction then mask the respective input(s) (zero out the gaps)
if (Input(inputIndex)->ReducesInTimeWrt(shared_from_this()))
MaskMissingGradientColumnsToZero(fr);
if (Input(inputIndex)->ReducesInTimeWrt(Input(1 - inputIndex)))
Input(1 - inputIndex)->MaskMissingValueColumnsToZero(fr);
if (inputIndex == 0)
{
auto exponent = InputRef(1).ValueTensorFor(rank, fr.AllowBroadcast());
// d/dx x**y = y * x**(y-1)
inputGradient.AddElementwiseProductWithPowBaseDerivativeOf(gradient, base, exponent);
}
else
{
auto result = ValueTensorFor(rank, fr);
// d/dy x**y = ln(x) * x**y
inputGradient.AddElementwiseProductWithPowExponentDerivativeOf(gradient, result, base);
}
}
};
template class PowNode<float>;
template class PowNode<double>;
template class PowNode<half>;
// -----------------------------------------------------------------------
// MinusNode (minuend, subtrahend)
// -----------------------------------------------------------------------
template <class ElemType>
class MinusNode : public BinaryElementWiseNode<ElemType>
{
typedef BinaryElementWiseNode<ElemType> Base; UsingBinaryElementwiseNodeBaseMembers;
static const std::wstring TypeName() { return L"Minus"; }
public:
DeclareConstructorFromConfigWithNumInputs(MinusNode);
MinusNode(DEVICEID_TYPE deviceId, const wstring& name)
: Base(deviceId, name)
{
}
virtual void /*ComputationNode::*/ ForwardProp(const FrameRange& fr) override
{
size_t rank = DetermineElementwiseTensorRank();
auto result = ValueTensorFor(rank, fr);
auto input0 = InputRef(0).ValueTensorFor(rank, fr.AllowBroadcast());
auto input1 = InputRef(1).ValueTensorFor(rank, fr.AllowBroadcast());
result.AssignDifferenceOf(input0, input1);
}
virtual void /*ComputationNode::*/ BackpropTo(const size_t inputIndex, const FrameRange& fr) override
{
size_t rank = DetermineElementwiseTensorRank();
auto gradient = GradientTensorFor(rank, fr);
auto inputGradient = Input(inputIndex)->GradientTensorFor(rank, fr.AllowBroadcast());
// if reduction then mask the respective input(s) (zero out the gaps)
if (Input(inputIndex)->ReducesInTimeWrt(shared_from_this()))
MaskMissingGradientColumnsToZero(fr);
ElemType sign = inputIndex == 0 ? 1.0f : -1.0f;
if (Input(inputIndex)->IsGradientOptimized(this))
{
if (Input(inputIndex)->ParentGradientReused())
{
if (inputGradient.GetSOBPtr() != gradient.GetSOBPtr())
LogicError("Gradients should be reused.");
}
else
inputGradient.AssignCopyOf(gradient, sign);
}
else
inputGradient.AddCopyOf(gradient, sign);
}
virtual ParentGradientOptimization ImplementsGradientOptimization(const ComputationNodeBase* input) const override
{
// only left operand can use gradient overwrite optimization
return (Input(0).get() == input && this->InputMatchesOutput(0)) ? ParentGradientOptimization::Reuse : ParentGradientOptimization::Overwrite;
}
};
template class MinusNode<float>;
template class MinusNode<double>;
template class MinusNode<half>;
// -----------------------------------------------------------------------
// ElementTimesNode (factor1, factor2)
// This allows broadcasting, and can thus also scale with a row, a column, or a scalar,
// as well as multiplying with a diagonal matrix (if represented as a column vector).
// -----------------------------------------------------------------------
template <class ElemType>
class ElementTimesNode : public BinaryElementWiseNode<ElemType>
{
typedef BinaryElementWiseNode<ElemType> Base;
UsingBinaryElementwiseNodeBaseMembers;
static const std::wstring TypeName()
{
return L"ElementTimes";
}
public:
DeclareConstructorFromConfigWithNumInputs(ElementTimesNode);
ElementTimesNode(DEVICEID_TYPE deviceId, const wstring& name)
: Base(deviceId, name)
{
}
virtual void /*ComputationNode::*/ ForwardProp(const FrameRange& fr) override
{
ForwardPropImpl(*this, fr, true/*allowBroadcast*/);
}
virtual void /*ComputationNode::*/ BackpropTo(const size_t inputIndex, const FrameRange& fr) override
{
BackpropToImpl(*this, inputIndex, fr, true/*allowBroadcast*/);
}
virtual bool InputUsedInComputingInputNodesGradients(size_t /*childIndex*/) const override { return true; }
virtual ParentGradientOptimization ImplementsGradientOptimization(const ComputationNodeBase*) const override
{
return ParentGradientOptimization::Overwrite;
}
template <typename classType>
static void ForwardPropImpl(classType& c, const FrameRange& fr, bool allowBroadcast)
{
size_t rank = c.DetermineElementwiseTensorRank();
auto result = c.ValueTensorFor(rank, fr);
auto input0 = c.InputRef(0).ValueTensorFor(rank, allowBroadcast ? fr.AllowBroadcast() : fr);
auto input1 = c.InputRef(1).ValueTensorFor(rank, allowBroadcast ? fr.AllowBroadcast() : fr);
result.AssignElementwiseProductOf(input0, input1);
}
template <typename classType>
static void BackpropToImpl(classType& c, const size_t inputIndex, const FrameRange& fr, bool allowBroadcast)
{
size_t rank = c.DetermineElementwiseTensorRank();
auto gradient = c.GradientTensorFor(rank, fr);
auto inputGradient = c.Input( inputIndex)->GradientTensorFor(rank, allowBroadcast ? fr.AllowBroadcast() : fr);
auto otherInputValue = c.Input(1 - inputIndex)->ValueTensorFor (rank, allowBroadcast ? fr.AllowBroadcast() : fr);
// if reduction then mask the respective input(s) (zero out the gaps)
if (c.Input(inputIndex)->ReducesInTimeWrt(c.shared_from_this()))
c.MaskMissingGradientColumnsToZero(fr);
if (c.Input(inputIndex)->ReducesInTimeWrt(c.Input(1 - inputIndex)))
c.Input(1 - inputIndex)->MaskMissingValueColumnsToZero(fr);
if (c.Input(inputIndex)->IsGradientInitializedBy(&c))
inputGradient.AssignElementwiseProductOf(gradient, otherInputValue);
else
inputGradient.AddElementwiseProductOf(gradient, otherInputValue);
}
};
template class ElementTimesNode<float>;
template class ElementTimesNode<double>;
template class ElementTimesNode<half>;
// -----------------------------------------------------------------------
// TimesNodeBase (A, B, outputRank=1)
// shared code of TimesNode and TransposeTimesNode (which transposes A)
// The common case, W * v with weights W and minibatch data v is efficiently
// implemented as a per-minibatch BLAS GEMM call.
// If A is minibatch data, then this operation is currently not efficient.
// TODO: Implement this with TensorView::DoElementwiseProductOf() and stride magic
// TODO: Transpose flags for all matrices, inputs and outputs?
// TODO: allow outputRank < 0 meaning to denote "all but", from right
// -----------------------------------------------------------------------
template <class ElemType, bool m_transpose>
class TimesNodeBase : public ComputationNode<ElemType>, public NumInputs<2>
{
friend class ElementTimesNode<ElemType>;
typedef ComputationNode<ElemType> Base; UsingComputationNodeMembers; using Base::OperationName; \
public:
enum : int
{
NoInferredInputRank = -1, // the default, do not infer left operand input rank from right operand
ReduceSequenceAxisWithoutInferredInputRank = -2, // reduce sequence axis. Currently only support cases like (m x k) x (k) -> (m) for sequences
};
public:
TimesNodeBase(DEVICEID_TYPE deviceId, const wstring& name, size_t outputRank = 1, int inferInputRankToMap = NoInferredInputRank)
: Base(deviceId, name), m_outputRank(outputRank), m_inferInputRankToMap(inferInputRankToMap), m_beingUnrolled(false)
{
}
virtual void 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<TimesNodeBase<ElemType, m_transpose>>(nodeP);
node->m_outputRank = m_outputRank;
node->m_inferInputRankToMap = m_inferInputRankToMap;
}
}
void Save(File& fstream) const
{
Base::Save(fstream);
fstream << m_outputRank;
fstream << m_inferInputRankToMap;
}
virtual void Load(File& fstream, size_t modelVersion) override
{
Base::Load(fstream, modelVersion);
if (modelVersion >= CNTK_MODEL_VERSION_3)
fstream >> m_outputRank;
else
m_outputRank = 1;
if (modelVersion >= CNTK_MODEL_VERSION_12)
fstream >> m_inferInputRankToMap;
else
m_inferInputRankToMap = NoInferredInputRank;
}
protected:
// if the left argument of the matrix product (A) has a time axis, it can only be applied sample by sample
// where each sample is treated as a separate matrix object (as a consequence, it then also applies to B and the result as well)
TensorView<ElemType> OneSampleTensorFor(int inputIndex/*-1 for output*/, bool gradient/*instead of value*/, const FrameRange& fr)
{
auto input = inputIndex < 0 ? this : Input(inputIndex).get();
auto data = gradient ? input->GradientPtr() : input->ValuePtr();
size_t rank = input->GetSampleLayout().GetRank();
if (inputIndex == 0 && m_transpose && rank == 1) // transposing a 1D tensor implies it is really a 2D tensor. Note that m_transpose applies to left operand only.
rank = 2;
if (!InputRef(0).HasMBLayout()) // left input is no MB data: run normally
return input->DataTensorFor(data, rank, fr);
auto tensorShape = input->GetOneSampleTensorSliceFor(rank, fr);
return TensorView<ElemType>(data, tensorShape);
}
static std::pair<size_t, size_t> CalcOutputMatrixSize(const size_t leftRank , const size_t rightRank, const TensorShape& outShape)
{
size_t outRank = outShape.GetRank();
size_t m = 1;
size_t n = 1;
size_t firstReducedDim = leftRank - (leftRank + rightRank - outRank) / 2;
for (size_t i = 0; i < firstReducedDim; i++)
m *= outShape.GetDim(i);
for (size_t i = firstReducedDim; i < outRank; i++)
n *= outShape.GetDim(i);
return std::make_pair(m, n);
}
private:
// Check if TimesNodeBase could be simplified to ElementTimes to avoid unroll when:
// 1. input0: is rank-1 and transposed, or is rank-2 with Dim(0)==1
// 2. input1: is rank-1
// 3. output: rank-1 and reduced to a scalar (Dim(0)==1)
// 4. m_transpose (becomes Matrix::InnerProduct), or both input being dense
bool IsReduceableDotProduct(const FrameRange& fr, bool& hasSparse)
{
const auto& shape0 = InputRef(0).GetSampleLayout();
const auto& shape1 = InputRef(1).GetSampleLayout();
const auto& shapeOut = GetSampleLayout();
bool input0Sparse = (InputRef(0).Value().GetMatrixType() != DENSE);
bool input1Sparse = (InputRef(1).Value().GetMatrixType() != DENSE);
bool input0_ok =
(shape0.GetRank() == 1 && m_transpose) ||
(shape0.GetRank() == 2 && shape0.GetDim(0) == 1);
bool input1_ok =
(shape1.GetRank() == 1);
bool outputScalar =
(shapeOut.GetRank() == 1) &&
(shapeOut.GetDim(0) == 1);
bool notBothSparse = !(input0Sparse && input1Sparse);
hasSparse = (input0Sparse || input1Sparse);
return input0_ok && input1_ok && outputScalar && notBothSparse && (m_transpose || !hasSparse);
}
void RequestReduceSequenceAxisMatricesIfNeeded(MatrixPool& matrixPool)
{
if (!ReduceSequenceAxis()) return;
for (int i = 0; i < NumInputs; i++)
{
RequestMatrixFromPool(m_tempScatterIndices[i], matrixPool, InputRef(i).GetMBLayout()->GetNumCols(), true);
const auto& packedData = InputRef(i).Value();
if (packedData.GetMatrixType() == DENSE)
RequestMatrixFromPool(m_tempUnpackedValue[i], matrixPool, InputRef(i).GetSampleLayout().GetNumElements(), true);
else
m_tempUnpackedValue[i] = std::make_shared<Matrix<ElemType>>(packedData.GetNumRows(), packedData.GetNumCols(), packedData.GetDeviceId(), packedData.GetMatrixType(), packedData.GetFormat());
}
}
void ReleaseReduceSequenceAxisMatricesIfNeeded(MatrixPool& matrixPool)
{
if (!ReduceSequenceAxis()) return;
for (int i = 0; i < NumInputs; i++)
{
ReleaseMatrixToPool(m_tempScatterIndices[i], matrixPool);
if (InputRef(i).Value().GetMatrixType() == DENSE)
ReleaseMatrixToPool(m_tempUnpackedValue[i], matrixPool);
else
m_tempUnpackedValue[i].reset();
}
}
void ForwardProp_ReduceSequenceAxis()
{
// Input are stored as (m * k) x b*(batch axis) x s*(sequence axis) and k x b* x s*
// We unpack them to (m * k) x s* x b* and k x s* x b*
// Then perform b* matrix multiplies to get m x b* with both k and s* being resolved
auto inputMBLayout = InputRef(0).GetMBLayout();
auto numSequences = inputMBLayout->GetNumSequences(); // b*
auto maxNumTimeSteps = inputMBLayout->GetNumTimeSteps(); // s*
size_t m = InputRef(0).GetSampleLayout()[0];
size_t k = InputRef(1).GetSampleLayout()[0];
if (InputRef(1).Value().GetMatrixType() == SPARSE)
LogicError("Right operand cannot be sparse in times reduce sequence axis");
GetMBLayout()->InitAsFrameMode(numSequences);
UpdateFunctionValuesSize();
TensorView<ElemType> unpackedInput[NumInputs];
for (int i = 0; i < NumInputs; i++)
{
ElemType gapPadValue = 0;
unpackedInput[i] = ComputationNode<ElemType>::Unpack(
InputRef(i).GetSampleLayout(),
InputRef(i).Value(),
inputMBLayout,
m_tempUnpackedValue[i],
m_tempScatterIndices[i],
std::shared_ptr<Matrix<char>>(nullptr),
/*batchMajor=*/ false,
&gapPadValue);
}
// note the unpacked input is not the normal MBLayout (batchMajor) so do ColumnSlice directly
const Matrix<ElemType>& mat0 = unpackedInput[0].GetSOB();
const Matrix<ElemType>& mat1 = unpackedInput[1].GetSOB();
// unroll in the batch axis, we may use batched GEMM in future
for (int s = 0; s < numSequences; s++)
{
Matrix<ElemType> mat0Slice = mat0.ColumnSlice(s * maxNumTimeSteps, maxNumTimeSteps); // (m * k) x s*
mat0Slice.Reshape(m, k * maxNumTimeSteps); // m x (k * s*)
Matrix<ElemType> mat1Slice = mat1.ColumnSlice(s * maxNumTimeSteps, maxNumTimeSteps); // k x s*
mat1Slice.Reshape(k * maxNumTimeSteps, 1); // (k * s*) x 1
Matrix<ElemType> value = Value().ColumnSlice(s, 1);
Matrix<ElemType>::Multiply(mat0Slice, false, mat1Slice, false, value);
}
}
void BackpropTo_ReduceSequenceAxis(size_t inputIndex)
{
auto input0MBLayout = InputRef(0).GetMBLayout();
auto numSequences = input0MBLayout->GetNumSequences(); // b*
auto maxNumTimeSteps = input0MBLayout->GetNumTimeSteps(); // s*
size_t m = InputRef(0).GetSampleLayout()[0];
size_t k = InputRef(1).GetSampleLayout()[0];
TensorView<ElemType> unpackedInput[NumInputs];
bool unpacked[NumInputs];
for (int i = 0; i < NumInputs; i++)
{
ElemType gapPadValue = 0;
unpackedInput[i] = ComputationNode<ElemType>::Unpack(
InputRef(i).GetSampleLayout(),
InputRef(i).Value(),
input0MBLayout, // the same for both operands
m_tempUnpackedValue[i],
m_tempScatterIndices[i],
std::shared_ptr<Matrix<char>>(nullptr),
/*batchMajor=*/ false,
&gapPadValue);
unpacked[i] = ((input0MBLayout->GetNumTimeSteps() > 1) && (input0MBLayout->GetNumSequences() > 1));
}
const auto& unpackedInputValue = unpackedInput[1 - inputIndex].GetSOB();
ElemType beta = InputRef(inputIndex).IsGradientInitializedBy(this) ? (ElemType)0 : (ElemType)1;
// note the unpacked input is not the normal MBLayout (batchMajor), so do ColumnSlice directly
if (inputIndex == 0)
{
Matrix<ElemType> tempGradientUnpacked(m * k, maxNumTimeSteps * numSequences, InputRef(inputIndex).GetDeviceId());
Matrix<ElemType>& inputGradientUnpacked = unpacked[inputIndex] ? tempGradientUnpacked : InputRef(inputIndex).Gradient();
for (int s = 0; s < numSequences; s++)
{
Matrix<ElemType> inputGradientSlice = inputGradientUnpacked.ColumnSlice(s * maxNumTimeSteps, maxNumTimeSteps); // (m * k) x s*
inputGradientSlice.Reshape(m, k * maxNumTimeSteps); // m x (k * s*)
Matrix<ElemType> inputValueSlice = unpackedInputValue.ColumnSlice(s * maxNumTimeSteps, maxNumTimeSteps); // k x s*
inputValueSlice.Reshape(k * maxNumTimeSteps, 1); // (k * s*) x 1
Matrix<ElemType> gradientSlice = Gradient().ColumnSlice(s, 1); // m x 1
Matrix<ElemType>::MultiplyAndWeightedAdd(1, gradientSlice, false, inputValueSlice, true, unpacked[inputIndex] ? (ElemType)0 : beta, inputGradientSlice);
}
if (unpacked[inputIndex])
InputRef(inputIndex).Gradient().DoGatherColumnsOf(beta, *m_tempScatterIndices[inputIndex], inputGradientUnpacked, (ElemType)1);
}
else
{
Matrix<ElemType> tempGradientUnpacked(k, maxNumTimeSteps * numSequences, InputRef(inputIndex).GetDeviceId());
Matrix<ElemType>& inputGradientUnpacked = unpacked[inputIndex] ? tempGradientUnpacked : InputRef(inputIndex).Gradient();
for (int s = 0; s < numSequences; s++)
{
Matrix<ElemType> inputGradientSlice = inputGradientUnpacked.ColumnSlice(s * maxNumTimeSteps, maxNumTimeSteps); // k x s*
inputGradientSlice.Reshape(k * maxNumTimeSteps, 1); // (k * s*) x 1
Matrix<ElemType> inputValueSlice = unpackedInputValue.ColumnSlice(s * maxNumTimeSteps, maxNumTimeSteps); // (m * k) x s*
inputValueSlice.Reshape(m, k * maxNumTimeSteps); // m x (k * s*)
Matrix<ElemType> gradientSlice = Gradient().ColumnSlice(s, 1); // m x 1
Matrix<ElemType>::MultiplyAndWeightedAdd(1, inputValueSlice, true, gradientSlice, false, unpacked[inputIndex] ? (ElemType)0 : beta, inputGradientSlice);
}
if (unpacked[inputIndex])
InputRef(inputIndex).Gradient().DoGatherColumnsOf(beta, *m_tempScatterIndices[inputIndex], inputGradientUnpacked, (ElemType)1);
}
}
public:
virtual void /*ComputationNode::*/ ForwardProp(const FrameRange& fr) override
{
// If argument A is minibatch data, then this must be performed frame-by-frame, sequence-by-sequence, one GEMM call each.
// This will be inefficient. We hope this will be the baseline of a future, more efficient TensorView-based implementation.
auto inputMBLayout = InputRef(0).GetMBLayout();
if (!fr.IsOneColumnWrt(inputMBLayout))
{
if (ReduceSequenceAxis())
{
// only works in PAR mode
if (!fr.IsAllFrames())
RuntimeError("%ls %ls operation can perform sequence axis reduction only for all frames.", NodeName().c_str(), OperationName().c_str());
if (inputMBLayout->HasSequenceBeyondBegin() || inputMBLayout->HasSequenceBeyondEnd())
RuntimeError("%ls %ls operation cannot perform sequence axis reduction for truncated sequence.", NodeName().c_str(), OperationName().c_str());
ForwardProp_ReduceSequenceAxis();
return;
}
// speed up using ElementTimes or InnerProduct to avoid unroll if possible
bool hasSparse;
if (IsReduceableDotProduct(fr, hasSparse))
{
// for sparse transposed, use InnerProduct
if (hasSparse)
{
Matrix<ElemType> value = ValueFor(fr);
Matrix<ElemType> input0 = InputRef(0).ValueFor(fr);
Matrix<ElemType> input1 = InputRef(1).ValueFor(fr);
if (input0.GetMatrixType() == SPARSE)
Matrix<ElemType>::InnerProduct(input0, input1, value, true/*isColWise*/);
else
Matrix<ElemType>::InnerProduct(input1, input0, value, true/*isColWise*/);
// TODO: better move this special-casing into TensorView::AssignElementwiseProductOf()
}
else
{
ElementTimesNode<ElemType>::ForwardPropImpl(*this, fr, true/*allowBroadcast*/);
}
return;
}
if (fr.IsBatchMatmul(inputMBLayout) && fr.IsBatchMatmul(InputRef(1).GetMBLayout()) && !hasSparse)
{
auto mn = CalcOutputMatrixSize(InputRef(0).GetSampleLayout().GetRank(), InputRef(1).GetSampleLayout().GetRank(), GetSampleLayout());
Matrix<ElemType> value = ValueFor(fr);
Matrix<ElemType> input0 = InputRef(0).ValueFor(fr);
Matrix<ElemType> input1 = InputRef(1).ValueFor(fr);
Matrix<ElemType>::BatchMatMul(ElemType(0.0), input0, m_transpose, mn.first, input1, false, mn.second, value, true);
return;
}
// recursively call ourselves for each individual time and sequence
// note this is not performant, warn user about the slow path being used
if (Base::HasEnvironmentPtr() && Base::Environment().traceLevel > 0)
std::call_once(m_unrollWarningOnceFlag, [this]{ fprintf(stderr, "WARNING: %ls %ls operation: being unrolled, execution may be slow\n", NodeName().c_str(), OperationName().c_str()); });
auto timeRange = fr.GetTimeRange();
auto sequenceRange = fr.GetSequenceRange();
m_beingUnrolled = true;
for (auto t = timeRange.first; t < timeRange.second; t++)
for (auto s = sequenceRange.first; s < sequenceRange.second; s++)
ForwardProp(fr.WithTimeStep(t).Sequence(s));
m_beingUnrolled = false;
return;
}
// TensorView::DoMatrixProductOf() will reduce each tensor object into a 2D tensor (or fail if it cannot)
// and recreate actual Matrix objects (in case of sparse, they must be identical to the original tensor storage object).
// Transposition is applied after flattening into 2D, but only allowed if the input sample is 2D anyway.
auto input0 = OneSampleTensorFor(0, /*gradient=*/false, fr.AllowBroadcast());
auto input1 = OneSampleTensorFor(1, /*gradient=*/false, fr.AllowBroadcast());
auto output = OneSampleTensorFor(-1, /*gradient=*/false, fr);
output.AssignMatrixProductOf(false/*transC*/, input0, m_transpose/*transA*/, input1, false/*transB*/, 1.0f, this->m_pQuantizedMultiplier);
}
virtual void /*ComputationNode::*/ BackpropTo(const size_t inputIndex, const FrameRange& fr) override
{
// special treatment if A is minibatch data; see Forward() for comment
if (!fr.IsOneColumnWrt(InputRef(0).GetMBLayout()))
{
if (ReduceSequenceAxis())
{
// only works in PAR mode
if (!fr.IsAllFrames())
RuntimeError("%ls %ls operation can perform sequence axis reduction only for all frames.", NodeName().c_str(), OperationName().c_str());
BackpropTo_ReduceSequenceAxis(inputIndex);
return;
}
// speed up using ElementTimes or InnerProduct to avoid unroll if possible
bool hasSparse;
if (IsReduceableDotProduct(fr, hasSparse))
{
if (hasSparse)
{
Matrix<ElemType> gradient = GradientFor(fr);
Matrix<ElemType> inputValue = InputRef(1 - inputIndex).ValueFor(fr);
Matrix<ElemType> inputGradient = InputRef(inputIndex).GradientFor(fr);
Matrix<ElemType>::ColumnwiseScaleAndWeightedAdd(
(ElemType)1.0, inputValue, gradient,
Input(inputIndex)->IsGradientInitializedBy(this) ? (ElemType)0.0 : (ElemType)1.0,
inputGradient);
// TODO: better move this special-casing into TensorView::AssignElementwiseProductOf()
// Note: We do not need to mask gaps here, since this code branch operates sample by sample (no reduction over samples).
}
else
{
ElementTimesNode<ElemType>::BackpropToImpl(*this, inputIndex, fr, true/*allowBroadcast*/);
}
return;
}
ElemType beta = Input(inputIndex)->IsGradientInitializedBy(this) ? (ElemType)0.0 : (ElemType)1.0;
if (inputIndex == 0)
{
if (fr.IsBatchMatmul(InputRef(0).GetMBLayout()) && fr.IsBatchMatmul(InputRef(1).GetMBLayout()) && !hasSparse)
{
Matrix<ElemType> outputGradient = GradientFor(fr);
Matrix<ElemType> input1 = InputRef(1).ValueFor(fr);
Matrix<ElemType> input0Gradient = InputRef(0).GradientFor(fr);
if (!m_transpose)
{
auto mn = CalcOutputMatrixSize(GetSampleLayout().GetRank(), InputRef(1).GetSampleLayout().GetRank(), InputRef(0).GetSampleLayout());
Matrix<ElemType>::BatchMatMul(beta, outputGradient, false, mn.first, input1, true, mn.second, input0Gradient, true);
}
else
{
auto mn = CalcOutputMatrixSize(InputRef(1).GetSampleLayout().GetRank(), GetSampleLayout().GetRank(), InputRef(0).GetSampleLayout());
Matrix<ElemType>::BatchMatMul(beta, input1, false, mn.first, outputGradient, true, mn.second, input0Gradient, true);
}
return;
}
}
else if (inputIndex == 1)
{
if (fr.IsBatchMatmul(InputRef(0).GetMBLayout()) && fr.IsBatchMatmul(InputRef(1).GetMBLayout()) && !hasSparse)
{
auto mn = CalcOutputMatrixSize(InputRef(0).GetSampleLayout().GetRank(), GetSampleLayout().GetRank(), InputRef(1).GetSampleLayout());
Matrix<ElemType> input0 = InputRef(0).ValueFor(fr);
Matrix<ElemType> input1Gradient = InputRef(1).GradientFor(fr);
Matrix<ElemType> outputGradient = GradientFor(fr);
Matrix<ElemType>::BatchMatMul(beta, input0, !m_transpose, mn.first, outputGradient, false, mn.second, input1Gradient, true);
return;
}
}
// note this is not performant, warn user about the slow path being used
if (Base::HasEnvironmentPtr() && Base::Environment().traceLevel > 0)
std::call_once(m_unrollWarningOnceFlag, [this] { fprintf(stderr, "WARNING: %ls %ls operation: being unrolled in backprop, execution may be slow\n", NodeName().c_str(), OperationName().c_str()); });
auto timeRange = fr.GetTimeRange();
auto sequenceRange = fr.GetSequenceRange();
// when unroll, parent overwrite gradient should be ignored
m_beingUnrolled = true;
if (Input(inputIndex)->IsGradientInitializedBy(this))
{
Input(inputIndex)->Gradient().SetValue(0);
}
for (auto t = timeRange.first; t < timeRange.second; t++) // step left to right to allow to build a sparse matrix
for (auto s = sequenceRange.first; s < sequenceRange.second; s++)
BackpropTo(inputIndex, fr.WithTimeStep(t).Sequence(s));
m_beingUnrolled = false;
return;
}
// this potentially computes inner products over time, so we must mask gaps to 0
if (Input(inputIndex)->ReducesInTimeWrt(shared_from_this()))
MaskMissingGradientColumnsToZero(fr);
if (Input(inputIndex)->ReducesInTimeWrt(Input(1 - inputIndex)))
Input(1 - inputIndex)->MaskMissingValueColumnsToZero(fr);
bool overwriteInputGradient = (Input(inputIndex)->IsGradientInitializedBy(this) && !m_beingUnrolled);
if (inputIndex == 0) // left derivative
{
if (InputRef(1).Value().GetMatrixType() == SPARSE &&
InputRef(0).GetPreferredGradientMatrixType() == UNDETERMINED &&
Gradient().GetMatrixType() == DENSE)
{
// Special case for DENSE * SPARSE -> DENSE, which leads to a SPARSE gradient for input0 (common for embedding)
// We need a sparse matrix for the gradient. We allocate a new one instead of switching the type in place
// since switching in place may affect other nodes who share this matrix due to memory sharing
//
// Note this normally only happen once (per minibatch) for embedding cases where all calculations are dense * sparse
// This is needed in addition to AllocateGradientMatricesForInputs as some sparse inputs that are not input_variable
// cannot be determined as sparse when AllocateGradientMatricesForInputs happens.
auto& currentInput0GradientMatrixRef = InputRef(0).Gradient();
auto newInput0SparseGradientMatrix =
std::make_shared<Matrix<ElemType>>(
currentInput0GradientMatrixRef.GetNumRows(),
currentInput0GradientMatrixRef.GetNumCols(),
currentInput0GradientMatrixRef.GetPreferredDeviceId(),
SPARSE,
MatrixFormat::matrixFormatSparseBlockCol);
// no need not set sparse with value from dense as its just cleared to zero when PreferredGradientMatrixType == UNDETERMINED
InputRef(0).GradientPtrRef() = newInput0SparseGradientMatrix;
InputRef(0).SetPreferredGradientMatrixType(SPARSE);
}
else if (InputRef(1).Value().GetMatrixType() == DENSE &&
InputRef(0).GetPreferredGradientMatrixType() != DENSE)
{
// for dense * dense, if we found previously gradient accumulated with sparse for input0, switch to dense
// this is a rare case and the performance is not optimized
if (InputRef(0).GetPreferredGradientMatrixType() == SPARSE)
InputRef(0).Gradient().SwitchToMatrixType(DENSE, matrixFormatDense, !overwriteInputGradient);
InputRef(0).SetPreferredGradientMatrixType(DENSE);
}
auto input0Gradient = OneSampleTensorFor(0, /*gradient=*/true, fr.AllowBroadcast());
auto input1 = OneSampleTensorFor(1, /*gradient=*/false, fr.AllowBroadcast());
auto outputGradient = OneSampleTensorFor(-1, /*gradient=*/true, fr);
if (overwriteInputGradient)
input0Gradient.AssignMatrixProductOf(m_transpose/*transC*/, outputGradient, false/*transA*/, input1, true/*transB*/);
else
input0Gradient.AddMatrixProductOf(m_transpose/*transC*/, outputGradient, false/*transA*/, input1, true/*transB*/);
}
else if (inputIndex == 1) // right derivative
{
auto input0 = OneSampleTensorFor(0, /*gradient=*/false, fr.AllowBroadcast());
auto input1Gradient = OneSampleTensorFor(1, /*gradient=*/true, fr.AllowBroadcast());
auto outputGradient = OneSampleTensorFor(-1, /*gradient=*/true, fr);
if (InputRef(1).Gradient().GetMatrixType() == SPARSE)
{
// we only support dense * sparse to have sparse gradient for input0, so if input1 has sparse gradient, switch to dense
// this is a rare case and the performance is not optimized
InputRef(1).Gradient().SwitchToMatrixType(DENSE, matrixFormatDense, !overwriteInputGradient);
}
InputRef(1).SetPreferredGradientMatrixType(DENSE);
if (overwriteInputGradient)
input1Gradient.AssignMatrixProductOf(false/*transC*/, input0, !m_transpose/*transA*/, outputGradient, false/*transB*/);
else
input1Gradient.AddMatrixProductOf(false/*transC*/, input0, !m_transpose/*transA*/, outputGradient, false/*transB*/);
}
}
virtual bool OutputUsedInComputingInputNodesGradients() const override { return false; }
// but both *inputs* are used, so we don't overload the InputUsed-() function which defaults to 'true'
virtual ParentGradientOptimization ImplementsGradientOptimization(const ComputationNodeBase*) const override { return ParentGradientOptimization::Overwrite; }
virtual void /*ComputationNodeBase::*/ Validate(bool isFinalValidationPass) override
{
Base::Validate(isFinalValidationPass);
if (ReduceSequenceAxis())
{
// generate MBLayout without sequence axis
if (!Input(0)->HasMBLayout() || !Input(1)->HasMBLayout() || *(Input(0)->GetMBLayout()) != *(Input(1)->GetMBLayout()))
InvalidArgument("%ls %ls operation can perform sequence axis reduction only on matching dynamic axes for both operands (which have the same layouts).", NodeName().c_str(), OperationName().c_str());
const auto& input0SampleLayout = InputRef(0).GetSampleLayout();
const auto& input1SampleLayout = InputRef(1).GetSampleLayout();
if ((input0SampleLayout.GetRank() > 1 && input0SampleLayout[1] != input1SampleLayout[0]) ||
(input1SampleLayout.GetRank() > 1 && input1SampleLayout[1] != 1) ||
(input0SampleLayout.GetRank() <= 1 && input1SampleLayout[0] != 1) ||
m_transpose)
InvalidArgument("%ls %ls operation can perform sequence axis reduction only in forms of (m x k) * (k x 1), without transpose.", NodeName().c_str(), OperationName().c_str());
if (m_pMBLayout == nullptr)
{
m_pMBLayout = make_shared<MBLayout>(); // this generates a new layout
m_pMBLayout->SetUniqueAxisName(ComputationNodeBase::DefaultNoSequenceAxisName);
}
}
else
{
InferMBLayoutFromInputsForStandardCase(isFinalValidationPass);
}
bool transpose = m_transpose; // (assigning to a non-const variable avoids a compiler warning C4127: conditional expression is constant)
// get tensor shapes
auto dimsA = Input(0)->GetSampleLayout().GetDims();
auto dimsB = Input(1)->GetSampleLayout().GetDims();
string dimsAstring = string(Input(0)->GetSampleLayout()); // for error messages
string dimsBstring = string(Input(1)->GetSampleLayout());
// validate and infer
if (isFinalValidationPass || (dimsA.size() > 0 && dimsB.size() > 0)) // only if we got at least some input dimensions to work with or need to wrap up
{
// if transposing then only support actual matrices or column vectors
if (transpose)
{
if (dimsA.size() == 1) // column vector transposed becomes a 2D tensor
dimsA.push_back(1);
else if (dimsA.size() != 2)
InvalidArgument("%ls %ls operation: Transposition requires a 2D tensor (matrix) or a 1D tensor (column vector), instead of a [%s].", NodeName().c_str(), OperationName().c_str(), dimsAstring.c_str());
else if (m_outputRank != 1)
InvalidArgument("%ls %ls operation: The outputRank (%d) must be 1 when transposing.", NodeName().c_str(), OperationName().c_str(), (int)m_outputRank);
// swap them temporarily, to get transposition out of the way for validation
std::swap(dimsA[0], dimsA[1]);
}
if (m_outputRank > dimsA.size()) // note: it may be equal in case of dyadic product uv'
InvalidArgument("%ls %ls operation: outputRank %d exceeds left argument's shape [%s].", NodeName().c_str(), OperationName().c_str(), (int)m_outputRank, dimsAstring.c_str());
auto numReductionDims = dimsA.size() - m_outputRank; // we reduce over the remaining dims; this is their number. Can be 0 in case of dyadic product uv'
if (numReductionDims > dimsB.size())
InvalidArgument("%ls %ls operation: right argument shape [%s] has too few dimensions for outputRank %d.", NodeName().c_str(), OperationName().c_str(), dimsBstring.c_str(), (int)m_outputRank);
#if 1 // support for legacy models when only the matrix dimensions had to match
// Note: This is non-ambiguous w.r.t. valid new configurations because this condition would otherwise just be considered an error.
// But it will fail to discover trailing reduction dimensions that are 1. We assume that no such legacy models exist.
// Note: This is very ugly [Wayne Xiong]. I agree [fseide].
if (dimsA.size() == 2 && !transpose && m_outputRank == 1 && dimsA[1] != dimsB[0] && dimsB[0] != 0)
{
// search whether we can interpret dimsA[1] as the flattening of the first dimensions
size_t dim = 1;
for (size_t k = 0; k < dimsB.size(); k++)
{
dim *= dimsB[k];
if (dim == dimsA[1])
{
// OK, we have an explanation: Patch dimsA and back-patch the sample layout of Input(0)
numReductionDims = k + 1;
dimsA.resize(m_outputRank + numReductionDims);
for (size_t kk = 0; kk < numReductionDims; kk++)
dimsA[m_outputRank + kk] = dimsB[kk];
Input(0)->SetDims(TensorShape(dimsA), false);
fprintf(stderr, "\n%ls %ls operation: For legacy compatibility, the sample layout of left input (%ls %ls operation) was patched to [%s] (from [%s])\n",
NodeName().c_str(), OperationName().c_str(), Input(0)->NodeName().c_str(), Input(0)->OperationName().c_str(), string(Input(0)->GetSampleLayout()).c_str(), dimsAstring.c_str());
dimsAstring = string(Input(0)->GetSampleLayout()); // for error messages
break; // we will continue with this patched up model from here on
}
}
}
#endif
// validate or automatically infer dimension inference for learnable parameters
for (size_t k = 0; k < m_outputRank; k++) // outputRank dimensions cannot be inferred
if (dimsA[k] == 0)
InvalidArgument("%ls %ls operation: The outputRank (%d) dimensions in left argument's shape [%s] must not be 0.", NodeName().c_str(), OperationName().c_str(), (int)m_outputRank, dimsAstring.c_str());
// infer rank of dimsA
// For purpose of dimension inference, Times() accepts an optional parameter inferInputRankToMap (default -1=unspecified).
// The last 'inferInputRankToMap' axes are considered those that the matrix product should keep (Times()
// is applied one by one, like a "map" operation) rather than reducing over.
// Specifically, inferInputRankToMap=0 means to reduce over all input axes, e.g. for an image input that
// should be flattened.
// Examples:
// [I x Inferred] * [J x K], inferInputRankToMap=n/a --> Inferred := J, result is [I x K]
// [I x Inferred] * [W x H x C], inferInputRankToMap=n/a --> Inferred := W, result is [I x H x C] (not desired)
// [I x Inferred x Inferred] * [W x H x C], inferInputRankToMap=n/a --> Inf x Inf := [W x H], result is [I x C]
// [I x Inferred] * [W x H x C], inferInputRankToMap=0 --> Inferred := W x H x C, result is [I] (desired)
// [I x Inferred] * [W x H x C x R], inferInputRankToMap=1 --> Inferred := W x H x C, result is [I x R] (desired)
// If W's shape is too short, it will be padded with 0 (i.e. inferred in a subsequent step).
// (the second check below (dimsA.back() == 0) is required to infer dimensions correctly for fixed input tensors where a new dimension is added,
// e.g. when adding an ROI dimension to a pretrained weights tensor of a dense layer after ROI pooling)
if (m_inferInputRankToMap >= 0 && dimsA.back() == 0) // if given, we pad if needed
{
if ((size_t)m_inferInputRankToMap >= dimsB.size() && isFinalValidationPass) // at least one axis must be left to reduce over
InvalidArgument("%ls %ls operation: 'inferInputRankToMap' argument %d must be less than rank of second operand [%s].", NodeName().c_str(), OperationName().c_str(), m_inferInputRankToMap, dimsBstring.c_str());
assert(dimsA.size() == m_outputRank + numReductionDims);
while (numReductionDims + (size_t)m_inferInputRankToMap < dimsB.size())
{
dimsA.push_back(0);
numReductionDims++;
}
}
// fill in the missing ones
// We fill in dimensions given as 0. The tensor rank is not inferred here (that is done above).
for (size_t k = m_outputRank; k < dimsA.size(); k++)
{
auto& dimA = dimsA[k];
auto& dimB = dimsB[k - m_outputRank];
if (isFinalValidationPass && dimB == 0)
InvalidArgument("%ls %ls operation: Right [%s] operand must not have zero dimensions.", NodeName().c_str(), OperationName().c_str(), dimsBstring.c_str());
else if (dimA == 0)
dimA = dimB; // infer dimension
else if (isFinalValidationPass && dimA != dimB)
InvalidArgument("%ls %ls operation: Left [%s] and right [%s] operands' shapes are not compatible.", NodeName().c_str(), OperationName().c_str(), dimsAstring.c_str(), dimsBstring.c_str());
}
// now determine result dimensions
auto dimsC = dimsA;
dimsC.resize(m_outputRank); // output dims
for (size_t k = numReductionDims; k < dimsB.size(); k++)
dimsC.push_back(dimsB[k]); // input dims
SetDims(TensorShape(dimsC), HasMBLayout());
// update dimensions of A
if (transpose)
std::swap(dimsA[0], dimsA[1]);
// update if LearnableParameter
Input(0)->ValidateInferInputDimsFrom(TensorShape(dimsA));
}
}
void RequestMatricesBeforeForwardProp(MatrixPool& matrixPool) override
{
Base::RequestMatricesBeforeForwardProp(matrixPool);
RequestReduceSequenceAxisMatricesIfNeeded(matrixPool);
}
void ReleaseMatricesAfterForwardProp(MatrixPool& matrixPool) override
{
Base::ReleaseMatricesAfterForwardProp(matrixPool);
ReleaseReduceSequenceAxisMatricesIfNeeded(matrixPool);
}
void RequestMatricesBeforeBackprop(MatrixPool& matrixPool) override
{
Base::RequestMatricesBeforeBackprop(matrixPool);
RequestReduceSequenceAxisMatricesIfNeeded(matrixPool);
}
void ReleaseMatricesAfterBackprop(MatrixPool& matrixPool) override
{
Base::ReleaseMatricesAfterBackprop(matrixPool);
ReleaseReduceSequenceAxisMatricesIfNeeded(matrixPool);
}
size_t OutputRank() const { return m_outputRank; }
int InferInputRankToMap() const { return m_inferInputRankToMap; }
protected:
shared_ptr<QuantizedMultiplier<ElemType>> m_pQuantizedMultiplier;
private:
size_t m_outputRank;
int m_inferInputRankToMap; // -1 (not specified) or says how to expand shape of W, to keep this many mapping dims
bool m_beingUnrolled;
std::once_flag m_unrollWarningOnceFlag;
bool ReduceSequenceAxis() const { return m_inferInputRankToMap == ReduceSequenceAxisWithoutInferredInputRank; }
static const int NumInputs = 2;
shared_ptr<Matrix<ElemType>> m_tempScatterIndices[NumInputs];
shared_ptr<Matrix<ElemType>> m_tempUnpackedValue[NumInputs];
};
// -----------------------------------------------------------------------
// TimesNode (A, B, outputRank=1) -- matrix product
// Right operand and output can have MB layout, while left operand cannot.
// This is generalized to tensors, in that B's leading dimension(s) must match
// the trailing dimension(s) of A, and those dimensions get flattened into a single
// dimension, after which the regular matrix product is applied.
// Leading dimension(s) of A and trailing ones of B remain unchanged.
// For example:
// [I x J x K] * [J x K x L x *] = [I x L x *]
// How many dimensions must match is controlled by an optional parameter
// 'outputRank', where all but the first 'outputRank' dimensions of A must match.
// 'outputRank' defaults to 1, which is used in the above example.
// Example for outputRank = 2:
// [I x J x K] * [K x L x M x *] = [I x J x L x M x *]
// -----------------------------------------------------------------------
template <class ElemType>
class TimesNode : public TimesNodeBase<ElemType, false>
{
typedef TimesNodeBase<ElemType, false> Base;
UsingComputationNodeMembersBoilerplate;
static const std::wstring TypeName() { return L"Times"; }
public:
TimesNode(DEVICEID_TYPE deviceId, const wstring& name, size_t outputRank = 1, int inferInputRankToMap = Base::NoInferredInputRank)
: Base(deviceId, name, outputRank, inferInputRankToMap)
{
}
TimesNode(const ScriptableObjects::IConfigRecordPtr configp)
: TimesNode(configp->Get(L"deviceId"), L"<placeholder>", configp->Get(L"outputRank"), configp->Get(L"inferInputRankToMap"))
{
AttachInputsFromConfig(configp, this->GetExpectedNumInputs());
}
};
template class TimesNode<float>;
template class TimesNode<double>;
template class TimesNode<half>;
// -----------------------------------------------------------------------
// TransposeTimesNode (A', B)
// Right operand and output can have MB layout, while left operand cannot.
// This differs from TimesNode in that A is transposed, where A must be a
// rank-1 or rank-2 tensor.
// A common use of transposition is trace(X'X) where X is a matrix of samples.
// This can be more efficiently implemented as ReduceSum (ElementTimes (X, X))
// -----------------------------------------------------------------------
template <class ElemType>
class TransposeTimesNode : public TimesNodeBase<ElemType, true>
{
typedef TimesNodeBase<ElemType, true> Base;
UsingComputationNodeMembersBoilerplate;
static const std::wstring TypeName() { return L"TransposeTimes"; }
public:
DeclareConstructorFromConfigWithNumInputs(TransposeTimesNode);
TransposeTimesNode(DEVICEID_TYPE deviceId, const wstring& name, size_t outputRank = 1)
: Base(deviceId, name, outputRank, Base::NoInferredInputRank)
{
if (outputRank != 1)
LogicError("TransposeTimes does not yet support outputRank other than 1");
}
};
template class TransposeTimesNode<float>;
template class TransposeTimesNode<double>;
template class TransposeTimesNode<half>;
// Fixed-point matrix product. This scales inputs to 16bit signed integers by Symmetric quantizers, performs
// integer multiplication using SSE/AVX2, and transforms the results back.
// Only dense untransposed matrix multiplication will be quantized. If at least one matrix is sparse then it will fall back to un-quantized default evaluation
// Currently it works for CPU only. On GPU logicError will be thrown.
// One way to include this node to the network is with the Edit command:
// ...
// node => if node.name == 'LSTMoutput1.output' then QuantizedTimes(node.inputs[0], node.inputs[1], bitShiftA=1, bitShiftB=2) else node,
// ...
// bitShift(A|B) - bit shift parameters of quantizers for matrices A and B, see the quantizers for more details. Decreases the maximum range of quantziation by 2^bitShift to prevent integer overflow during BLAS routines.
// bitShift=0 doesn't change the range; higher bitShift will decrease precision of quantization, but will make BLAS routines less prone to overflow.
// Other parameters - refer to the base multiplication class
template <class ElemType>
class QuantizedTimesNode : public TimesNodeBase<ElemType, false>
{
typedef TimesNodeBase<ElemType, false> Base;
UsingComputationNodeMembersBoilerplate;
static const std::wstring TypeName()
{
return L"QuantizedTimes";
}
private:
// Quantizer bit shift for matrices A and B
size_t m_bitShiftA;
size_t m_bitShiftB;
public:
QuantizedTimesNode(DEVICEID_TYPE deviceId, const wstring& name, size_t bitShiftA = 1, size_t bitShiftB = 1, size_t outputRank = 1, int inferInputRankToMap = Base::NoInferredInputRank)
: Base(deviceId, name, outputRank, inferInputRankToMap), m_bitShiftA(bitShiftA), m_bitShiftB(bitShiftB)
{
// TODO support multiplication on GPUs as well.
if (deviceId != CPUDEVICE)
LogicError("Quantized operation is supposed to be used on CPU device only.");
shared_ptr<SymmetricQuantizer<ElemType, short>> pQA(new SymmetricQuantizer<ElemType, short>(m_bitShiftA));
shared_ptr<SymmetricQuantizer<ElemType, short>> qQB(new SymmetricQuantizer<ElemType, short>(m_bitShiftB));
this->m_pQuantizedMultiplier = shared_ptr<QuantizedMultiplier<ElemType>>(new QuantizedMultiplier<ElemType>(pQA, qQB));
}
QuantizedTimesNode(const ScriptableObjects::IConfigRecordPtr configp)
: QuantizedTimesNode(configp->Get(L"deviceId"), L"<placeholder>", configp->Get(L"bitShiftA"), configp->Get(L"bitShiftB"), configp->Get(L"outputRank"), configp->Get(L"inferInputRankToMap"))
{
AttachInputsFromConfig(configp, this->GetExpectedNumInputs());
}
virtual void 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<QuantizedTimesNode<ElemType>>(nodeP);
node->m_bitShiftA = m_bitShiftA;
node->m_bitShiftB = m_bitShiftB;
}
}
void Save(File& fstream) const
{
Base::Save(fstream);
fstream << m_bitShiftA;
fstream << m_bitShiftB;
}
virtual void Load(File& fstream, size_t modelVersion) override
{
Base::Load(fstream, modelVersion);
fstream >> m_bitShiftA;
fstream >> m_bitShiftB;
}
virtual void /*ComputationNode::*/ ForwardProp(const FrameRange& fr) override
{
if (dynamic_pointer_cast<LearnableParameter<ElemType>>(Input(0)))
this->m_pQuantizedMultiplier->SetIsAConstant(true);
if (dynamic_pointer_cast<LearnableParameter<ElemType>>(Input(1)))
this->m_pQuantizedMultiplier->SetIsBConstant(true);
Base::ForwardProp(fr);
}
virtual void /*ComputationNode::*/ BackpropTo(const size_t /*inputIndex*/, const FrameRange& /*fr*/) override
{
// This operation is intended only for inference
NOT_IMPLEMENTED;
}
};
template class QuantizedTimesNode<float>;
template class QuantizedTimesNode<double>;
template class QuantizedTimesNode<half>;
// -----------------------------------------------------------------------
// SumElementsNode (input)
// Sums up all elements in the input across all samples into a single scalar.
// When applied to minibatch data, this will sum across all sequences in the
// minibatch, like a training-criterion node. This is one of the few operations
// that cross the boundary between input sequences.
// Note that SGD itself aggregates over samples in a criterion node.
// So the only proper use of this node is for multi-task learning, where
// different nodes have different numbers of samples (sequence lenth).
// -----------------------------------------------------------------------
template <class ElemType>
class SumElementsNode : public ComputationNodeNonLooping /*ComputationNode*/<ElemType>, public NumInputs<1>
{
typedef ComputationNodeNonLooping<ElemType> Base;
UsingComputationNodeMembersBoilerplate;
static const std::wstring TypeName()
{
return L"SumElements";
}
public:
DeclareConstructorFromConfigWithNumInputs(SumElementsNode);
SumElementsNode(DEVICEID_TYPE deviceId, const wstring& name)
: Base(deviceId, name)
{
}
virtual void /*ComputationNodeNonLooping::*/ ForwardPropNonLooping() override
{
FrameRange fr(InputRef(0).GetMBLayout());
// TODO: change to TensorView and AssignCopyOf() with reduction
Value().AssignSumOfElements(InputRef(0).MaskedValueFor(fr)); // since we are reducing over frames, we must first mask gaps in input to zero
}
virtual void /*ComputationNodeNonLooping::*/ BackpropToNonLooping(size_t inputIndex) override
{
FrameRange fr(InputRef(0).GetMBLayout());
InputRef(0).GradientFor(fr) += Gradient(); // here the assumption is that gradientValues are 1x1 matrix
}
virtual bool OutputUsedInComputingInputNodesGradients() const override { return false; }
virtual bool InputUsedInComputingInputNodesGradients(size_t /*childIndex*/) const override { return false; }
virtual void /*ComputationNodeBase::*/ Validate(bool isFinalValidationPass) override
{
ValidateUnaryReduce(isFinalValidationPass);
}
};
template class SumElementsNode<float>;
template class SumElementsNode<double>;
// -----------------------------------------------------------------------
// TransposeDimensions (input, axis1, axis2)
// - swaps index dimensions axis1 and axis2. The values are 1-based; 1 stands for the leading dimension.
// - new dimensions can be created; e.g. a column vector can be transposed into a row vector, which is a [1 x N] tensor
// - transposing into the time dimension is currently not supported
// - internally implemented with tensor lib by shuffling dimensions with their strides
// - input may be minibatch data or not
// Transpose (input) = TransposeDimensions (input, 1, 2)
// -----------------------------------------------------------------------
template <class ElemType>
class TransposeDimensionsNode : public ComputationNode /*ComputationNode*/<ElemType>, public NumInputs<1>
{
typedef ComputationNode<ElemType> Base; UsingComputationNodeMembersBoilerplate;
static const std::wstring TypeName() { return L"TransposeDimensions"; }
public:
TransposeDimensionsNode(DEVICEID_TYPE deviceId, const wstring& name, int axis1 = 1, int axis2 = 2)
: Base(deviceId, name), m_axis1(axis1), m_axis2(axis2), m_perm({})
{
}
TransposeDimensionsNode(DEVICEID_TYPE deviceId, const wstring& name, const std::vector<int>& perm)
: Base(deviceId, name), m_axis1(0), m_axis2(0), m_perm({})
{
if (!std::all_of(perm.begin(), perm.end(), [](int p) {return p >= 1; }))
InvalidArgument("%ls %ls operation: _internal_ indices for axes must be >= 1.", NodeName().c_str(), OperationName().c_str());
// undo the annoying +1
std::transform(perm.begin(), perm.end(), std::back_inserter(m_perm), [](const int& p) { return (size_t)(p - 1); });
}
TransposeDimensionsNode(const ScriptableObjects::IConfigRecordPtr configp)
: TransposeDimensionsNode(configp->Get(L"deviceId"), L"<placeholder>", configp->Get(L"axis1"), configp->Get(L"axis2"))
{
AttachInputsFromConfig(configp, this->GetExpectedNumInputs());
}
void Save(File& fstream) const
{
Base::Save(fstream);
fstream << m_axis1 << m_axis2;
if (m_axis1 == 0 && m_axis2 == 0)
{
fstream << m_perm.size();
for (const auto p : m_perm)
fstream << p;
}
}
virtual void Load(File& fstream, size_t modelVersion) override
{
Base::Load(fstream, modelVersion);
if (modelVersion >= CNTK_MODEL_VERSION_3)
{
fstream >> m_axis1 >> m_axis2;
if (modelVersion >= CNTK_MODEL_VERSION_25 && m_axis1 == 0 && m_axis2 == 0)
{
size_t size = 0;
fstream >> size;
m_perm.resize(size);
for (size_t i = 0; i < size; ++i)
fstream >> m_perm[i];
}
}
else
m_axis1 = 1, m_axis2 = 2; // default
}
int Axis1() const { return m_axis1; }
int Axis2() const { return m_axis2; }
private:
// compute the transposed tensor shape (in-place)
void TransposeShape(TensorShape& shape) const
{
if (m_axis1 > 0 && m_axis2 > 0)
{
size_t i = m_axis1 - 1;
size_t j = m_axis2 - 1;
shape.SwapDimsInPlace(i, j);
}
else /* if (m_axis1 == 0 && m_axis2 == 0) */
{
shape.PermuteDimsInPlace(m_perm);
}
}
// get the transposed input shape
// Using this shape yields the same tensor but index positions are swapped.
TensorShape GetTransposedTensorSliceFor(size_t rank, const FrameRange& fr)
{
auto shape = InputRef(0).GetTensorSliceFor(rank, fr);
TransposeShape(shape);
return shape;
}
// verify that the argument is a valid permutation
// We pass by value because the function mutates perm
static bool IsPermutation(std::vector<size_t> perm)
{
std::sort(perm.begin(), perm.end());
for (auto i = 0; i < perm.size(); ++i)
if (i != perm[i])
return false;
return true;
}
public:
virtual void /*ComputationNode::*/ ForwardProp(const FrameRange& fr) override
{
size_t rank = DetermineElementwiseTensorRank();
auto output = ValueTensorFor( rank, fr);
auto input = TensorView<ElemType>(InputRef(0).ValuePtr(), GetTransposedTensorSliceFor(rank, fr));
output.AssignCopyOf(input);
}
virtual void /*ComputationNode::*/ BackpropTo(const size_t inputIndex, const FrameRange& fr) override
{
size_t rank = DetermineElementwiseTensorRank();
auto outputGradient = GradientTensorFor( rank, fr);
auto inputGradient = TensorView<ElemType>(InputRef(0).GradientPtr(), GetTransposedTensorSliceFor(rank, fr));
inputGradient.AddCopyOf(outputGradient);
}
virtual bool OutputUsedInComputingInputNodesGradients() const override { return false; }
virtual bool InputUsedInComputingInputNodesGradients(size_t /*childIndex*/) const override { return false; }
virtual void /*ComputationNodeBase::*/ Validate(bool isFinalValidationPass) override
{
assert(m_inputs.size() == 1);
ComputationNodeBase::Validate(isFinalValidationPass);
InferMBLayoutFromInputsForStandardCase(isFinalValidationPass);
// input shape
auto shape = Input(0)->GetSampleLayout();
if (m_perm.size() == 0 && shape.GetRank() > 0)
{
// we are swapping two axes
// validate indices
if (m_axis1 < 1 || m_axis2 < 1)
InvalidArgument("%ls %ls operation: Indices to transpose must be >= 1.", NodeName().c_str(), OperationName().c_str());
size_t i = m_axis1 - 1;
size_t j = m_axis2 - 1;
if (i >= shape.GetRank() && j >= shape.GetRank())
InvalidArgument("%ls %ls operation: At least one index must refer to an existing index.", NodeName().c_str(), OperationName().c_str());
// pad
// Permutation is allowed to create new dimensions, specifically to be able to transpose a [N] column vector into a [1 x N] row vector.
// One can also use SplitDimensions() for this, but this seems a natural thing to do.
size_t maxij = std::max(i, j);
if (maxij >= shape.GetRank())
shape.PadRankInPlace(maxij + 1);
}
else
{
//we are transposing a tensor using a permutation
if (m_axis1 != 0 || m_axis2 != 0)
InvalidArgument("%ls %ls operation: Cannot transpose via permutation and axes indices != 0", NodeName().c_str(), OperationName().c_str());
//verify shapes are matching
if (m_perm.size() != shape.GetRank())
InvalidArgument("%ls %ls operation: Specified permutation doesn't match operand", NodeName().c_str(), OperationName().c_str());
//verify it is a permutation
if (!TransposeDimensionsNode::IsPermutation(m_perm))
InvalidArgument("%ls %ls operation: Specified permutation is not a permutation of (0, 1, ..., rank - 1)", NodeName().c_str(), OperationName().c_str());
}
// apply the permutation
TransposeShape(shape);
// drop the strides, since output is dense (swapped strides will be used with the input in ForwardProp())
SetDims(TensorShape(shape.GetDims()), HasMBLayout());
}
private:
int m_axis1, m_axis2; // the two dimensions (axes, 1-based) to swap
std::vector<size_t> m_perm; // permutation to use for transposition
};
template class TransposeDimensionsNode<float>;
template class TransposeDimensionsNode<double>;
template class TransposeDimensionsNode<half>;
// -----------------------------------------------------------------------
// CosDistanceNode (left, right)
// column-wise cos distance
// TODO: Would it be useful to allow one of the two to be a single column?
// TODO: Allow to reduce only over a single dimension, or a subset.
// -----------------------------------------------------------------------
template <class ElemType>
class CosDistanceNode : public ComputationNode<ElemType>, public NumInputs<2>
{
typedef ComputationNode<ElemType> Base;
UsingComputationNodeMembersBoilerplate;
static const std::wstring TypeName()
{
return L"CosDistance";
}
public:
DeclareConstructorFromConfigWithNumInputs(CosDistanceNode);
CosDistanceNode(DEVICEID_TYPE deviceId, const wstring& name)
: Base(deviceId, name)
{
}
virtual void /*ComputationNode::*/ BackpropTo(const size_t inputIndex, const FrameRange& fr) override
{
// functionValues, invNorm0, invNorm1 - output from the EvaluateNode() method
// temp, rightTerm, leftTerm - temporary matrices
if (inputIndex == 0) // left derivative
m_temp->AssignElementProductOf(*m_invNorm0, *m_invNorm0);
else // right derivative
m_temp->AssignElementProductOf(*m_invNorm1, *m_invNorm1);
int rank = DetermineElementwiseTensorRank();
auto tempView = TensorView<ElemType>(m_temp, ValueTensorFor(rank, fr).GetShape());
tempView.AssignElementwiseProductOf(ValueTensorFor(rank, fr), tempView);
tempView.AssignElementwiseProductOf(GradientTensorFor(rank, fr), tempView);
auto gradientView = Input(inputIndex)->GradientTensorFor(rank, fr);
gradientView.AddElementwiseProductOf(tempView, Input(inputIndex)->ValueTensorFor(rank, fr), -1);
m_temp->AssignElementProductOf(*m_invNorm0, *m_invNorm1);
tempView.AssignElementwiseProductOf(GradientTensorFor(rank, fr), tempView);
gradientView.AddElementwiseProductOf(tempView, Input(1 - inputIndex)->ValueTensorFor(rank, fr));
}
virtual void /*ComputationNode::*/ ForwardProp(const FrameRange& fr) override
{
Matrix<ElemType> sliceInput0Value = InputRef(0).ValueFor(fr);
Matrix<ElemType> sliceInput1Value = InputRef(1).ValueFor(fr);
Matrix<ElemType> sliceOutputValue = ValueFor(fr);
m_invNorm0->AssignVectorNorm2Of(sliceInput0Value, true);
m_invNorm0->AssignElementInverseOf(*m_invNorm0);
m_invNorm1->AssignVectorNorm2Of(sliceInput1Value, true);
m_invNorm1->AssignElementInverseOf(*m_invNorm1);
sliceOutputValue.AssignInnerProductOf(sliceInput0Value, sliceInput1Value, true);
sliceOutputValue.ElementMultiplyWith(*m_invNorm0);
sliceOutputValue.ElementMultiplyWith(*m_invNorm1);
}
virtual void /*ComputationNodeBase::*/ Validate(bool isFinalValidationPass) override
{
Base::Validate(isFinalValidationPass);
InferMBLayoutFromInputsForStandardCase(isFinalValidationPass);
ValidateInferBinaryInputDims();
// TODO: We could do something more interesting with tensors.
// E.g. apply a cos distance of a whole set of data with a single reference.
SetDims(TensorShape::Scalar(Environment().IsV2Library()), Input(1)->HasMBLayout());
}
virtual void 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<CosDistanceNode<ElemType>>(nodeP);
node->m_invNorm0->SetValue(*m_invNorm0);
node->m_invNorm1->SetValue(*m_invNorm1);
node->m_temp->SetValue(*m_temp);
}
}
// request matrices needed to do node function value evaluation
virtual void RequestMatricesBeforeForwardProp(MatrixPool& matrixPool)
{
Base::RequestMatricesBeforeForwardProp(matrixPool);
RequestMatrixFromPool(m_invNorm0, matrixPool, 1, true, true);
RequestMatrixFromPool(m_invNorm1, matrixPool, 1, true, true);
}
// request matrices that are needed for gradient computation
virtual void RequestMatricesBeforeBackprop(MatrixPool& matrixPool)
{
Base::RequestMatricesBeforeBackprop(matrixPool);
RequestMatrixFromPool(m_temp, matrixPool, 1, true, true);
}
// release gradient and temp matrices that no longer needed after all the children's gradients are computed.
virtual void ReleaseMatricesAfterBackprop(MatrixPool& matrixPool)
{
Base::ReleaseMatricesAfterBackprop(matrixPool);
ReleaseMatrixToPool(m_invNorm0, matrixPool);
ReleaseMatrixToPool(m_invNorm1, matrixPool);
ReleaseMatrixToPool(m_temp, matrixPool);
}
private:
// invNorm nodes tranfer data between ForwardProp and BackpropTo
shared_ptr<Matrix<ElemType>> m_invNorm0;
shared_ptr<Matrix<ElemType>> m_invNorm1;
// the rest are temporaries, values don't need to be maintained
shared_ptr<Matrix<ElemType>> m_temp;
};
template class CosDistanceNode<float>;
template class CosDistanceNode<double>;
template class CosDistanceNode<half>;
// -----------------------------------------------------------------------
// KhatriRaoProductNode (left, right)
// Compute an outer product of column vectors (for each sample).
// TODO: This is a special kind of tensor product, and calls for a tensor representation.
// -----------------------------------------------------------------------
template <class ElemType>
class KhatriRaoProductNode : public ComputationNode<ElemType>, public NumInputs<2>
{
typedef ComputationNode<ElemType> Base;
UsingComputationNodeMembersBoilerplate;
static const std::wstring TypeName()
{
return L"KhatriRaoProduct";
}
public:
DeclareConstructorFromConfigWithNumInputs(KhatriRaoProductNode);
KhatriRaoProductNode(DEVICEID_TYPE deviceId, const wstring& name)
: Base(deviceId, name)
{
}
virtual void /*ComputationNode::*/ BackpropTo(const size_t inputIndex, const FrameRange& fr) override
{
Matrix<ElemType> sliceOutputGrad = GradientFor(fr);
if (inputIndex == 0) // left derivative
{
Matrix<ElemType> sliceInput0Grad = InputRef(0).GradientFor(fr);
Matrix<ElemType> sliceInput1Value = InputRef(1).ValueFor(fr);
sliceInput0Grad.AddColumnReshapeProductOf(sliceOutputGrad, sliceInput1Value, false);
}
else // right derivative
{
Matrix<ElemType> sliceInput0Value = InputRef(0).ValueFor(fr);
Matrix<ElemType> sliceInput1Grad = InputRef(1).GradientFor(fr);
sliceInput1Grad.AddColumnReshapeProductOf(sliceOutputGrad, sliceInput0Value, true);
}
}
virtual bool OutputUsedInComputingInputNodesGradients() const override
{
// The KhatriRaoProductNode does not require its output value for computing
// the gradients of its input nodes
return false;
}
virtual void /*ComputationNode::*/ ForwardProp(const FrameRange& fr) override
{
ValueFor(fr).AssignKhatriRaoProductOf(InputRef(0).ValueFor(fr), InputRef(1).ValueFor(fr));
}
virtual void /*ComputationNodeBase::*/ Validate(bool isFinalValidationPass) override
{
Base::Validate(isFinalValidationPass);
InferMBLayoutFromInputsForStandardCase(isFinalValidationPass);
size_t rows0 = Input(0)->GetSampleMatrixNumRows();
size_t rows1 = Input(1)->GetSampleMatrixNumRows();
// after KhatriRaoProduct the structure is lost
// TODO: ^^ Is that correct? Should we use a tensor here, TensorShape(rows0, rows1)?
SetDims(TensorShape(rows0 * rows1), HasMBLayout());
}
};
template class KhatriRaoProductNode<float>;
template class KhatriRaoProductNode<double>;
// -----------------------------------------------------------------------
// CosDistanceWithNegativeSamplesNode (left, right, shift, neg)
//
// Left and right forms pairs of positive samples. They are symmetric but usually
// the search key is used as the left. For example, Left is search query and right is document.
// The negative samples are formed on the fly by shifting the right side.
// The 'shift' indicates how many samples in the right node you should shift to form each negative sample pair.
// It is often choose to be one. 'Neg' indicates how many negative samples you want to generate.
// -----------------------------------------------------------------------
template <class ElemType>
class CosDistanceWithNegativeSamplesNode : public ComputationNode<ElemType>, public NumInputs<4>
{
typedef ComputationNode<ElemType> Base;
UsingComputationNodeMembersBoilerplate;
static const std::wstring TypeName()
{
return L"CosDistanceWithNegativeSamples";
}
public:
DeclareConstructorFromConfigWithNumInputs(CosDistanceWithNegativeSamplesNode);
CosDistanceWithNegativeSamplesNode(DEVICEID_TYPE deviceId, const wstring& name)
: Base(deviceId, name)
{
}
virtual void /*ComputationNode::*/ BackpropTo(const size_t inputIndex, const FrameRange& fr) override
{
Matrix<ElemType> sliceInput0Value = InputRef(0).ValueFor(fr);
Matrix<ElemType> sliceInput1Value = InputRef(1).ValueFor(fr);
Matrix<ElemType> sliceOutputValue = ValueFor(fr);
Matrix<ElemType> sliceInputGrad = Input(inputIndex)->GradientFor(fr);
Matrix<ElemType> sliceThisGrad = GradientFor(fr);
BackpropToS(inputIndex, *m_invNorm0, *m_invNorm1, sliceOutputValue, *m_temp, *m_rightTerm, *m_leftTerm, *m_invNormSquare, sliceInput0Value, sliceInput1Value, InputRef(2).Value(), InputRef(3).Value(), sliceInputGrad, sliceThisGrad);
}
// functionValues, invNorm0, invNorm1 - output from the EvaluateNode() method
// temp, rightTerm, leftTerm - temporary matrices
// in0, in1, in2, in3 - input functionValues from other nodes
// inputGradientValues(x) - gradients to update, where x matches inputIndex
/*TODO: merge with call site*/ void BackpropToS(const size_t inputIndex, const Matrix<ElemType>& invNorm0, const Matrix<ElemType>& invNorm1, const Matrix<ElemType>& functionValues,
Matrix<ElemType>& temp, Matrix<ElemType>& rightTerm, Matrix<ElemType>& leftTerm, Matrix<ElemType>& invNormSquare, // the temporary variables
const Matrix<ElemType>& in0, const Matrix<ElemType>& in1, const Matrix<ElemType>& in2, const Matrix<ElemType>& in3,
Matrix<ElemType>& inputGradientValues, Matrix<ElemType>& thisGradientValues)
{
size_t shift = (size_t) in2.Get00Element();
size_t negNumber = (size_t) in3.Get00Element();
size_t numCols = in0.GetNumCols(); // used in computing right child's graident
if (inputIndex == 0) // left derivative
{
invNormSquare.AssignElementProductOf(invNorm0, invNorm0);
for (long m = 0; m < negNumber + 1; m++)
{
temp.GetARowByIndex(functionValues, m); // set this matrx to be the m-th row in functionValues
temp.ElementMultiplyWith(invNormSquare);
Matrix<ElemType>::ConductRowElementMultiplyWithShift(temp, in0, rightTerm, 0, true);
if (m == 0)
{
temp.AssignElementProductOf(invNorm0, invNorm1);
Matrix<ElemType>::ConductRowElementMultiplyWithShift(temp, in1, leftTerm, 0, true);
}
else
{
size_t currshift = m + shift - 1; // for current line, how much should we shift
temp.AssignElementProductOfWithShift(invNorm0, invNorm1, currshift); // this is a row vector
Matrix<ElemType>::ConductRowElementMultiplyWithShift(temp, in1, leftTerm, currshift, true);
}
leftTerm -= rightTerm;
temp.GetARowByIndex(thisGradientValues, m);
Matrix<ElemType>::ConductRowElementMultiplyWithShift(temp, leftTerm, rightTerm, 0, true);
inputGradientValues += rightTerm;
}
}
else // right part
{
invNormSquare.AssignElementProductOf(invNorm1, invNorm1); // this matrix should be save and unchanged. It should not be changed
for (long m = 0; m < negNumber + 1; m++)
{
temp.GetARowByIndex(functionValues, m); // set this matrx to be the m-th row in functionValues
if (m == 0) // this is the first line. computation should be symmetric
{
// the following is for the right part
temp.ElementMultiplyWith(invNormSquare);
Matrix<ElemType>::ConductRowElementMultiplyWithShift(temp, in1, rightTerm, 0, true);
// the following is for the left part
temp.AssignElementProductOf(invNorm0, invNorm1);
Matrix<ElemType>::ConductRowElementMultiplyWithShift(temp, in0, leftTerm, 0, true);
leftTerm -= rightTerm;
temp.GetARowByIndex(thisGradientValues, m);
Matrix<ElemType>::ConductRowElementMultiplyWithShift(temp, leftTerm, rightTerm, 0, true);
inputGradientValues += rightTerm;
}
else // this requires shift
{
size_t currshift = (m + shift - 1) % numCols;
size_t reverseshift = numCols - currshift;
leftTerm.AssignElementProductOfWithShift(invNormSquare, temp, reverseshift); // use leftTerm as a temp variable here
Matrix<ElemType>::ConductRowElementMultiplyWithShift(leftTerm, in1, rightTerm, 0, true);
temp.AssignElementProductOfWithShift(invNorm1, invNorm0, reverseshift);
Matrix<ElemType>::ConductRowElementMultiplyWithShift(temp, in0, leftTerm, reverseshift, true);
leftTerm -= rightTerm;
temp.GetARowByIndex(thisGradientValues, m);
Matrix<ElemType>::ConductRowElementMultiplyWithShift(temp, leftTerm, rightTerm, reverseshift, false);
inputGradientValues += rightTerm;
}
}
}
}
virtual void /*ComputationNode::*/ ForwardProp(const FrameRange& fr) override
{
Matrix<ElemType> sliceInput0Value = InputRef(0).ValueFor(fr);
Matrix<ElemType> sliceInput1Value = InputRef(1).ValueFor(fr);
Matrix<ElemType> sliceOutputValue = ValueFor(fr);
ForwardPropS(*m_invNorm0, *m_invNorm1, sliceOutputValue, sliceInput0Value, sliceInput1Value, InputRef(2).Value(), InputRef(3).Value(), *m_leftTerm, *m_rightTerm);
}
/*TODO: merge with call site*/ void ForwardPropS(Matrix<ElemType>& invNorm0, Matrix<ElemType>& invNorm1, Matrix<ElemType>& functionValues, Matrix<ElemType>& in0, Matrix<ElemType>& in1, Matrix<ElemType>& in2, Matrix<ElemType>& in3, Matrix<ElemType>& leftTermTemp, Matrix<ElemType>& rightTermTemp)
{
invNorm0.AssignVectorNorm2Of(in0, true); // seems to modify input (in0)
invNorm0.AssignElementInverseOf(invNorm0);
invNorm1.AssignVectorNorm2Of(in1, true); // seems to modify the input (in1)
invNorm1.AssignElementInverseOf(invNorm1);
size_t shift = (size_t) in2.Get00Element();
size_t negNumber = (size_t) in3.Get00Element();
// mutiply invNorm0 and invNorm1 with shift and neg.
// The result is a matrix of (numberneg+1, invNorm0.Cols)
leftTermTemp.AssignElementProductOfWithShiftNeg(invNorm0, invNorm1, shift, negNumber);
// compute the right values
// Again, the output is a matrix of (negNumber+1, invNorm0.cols)
rightTermTemp.AssignInnerProductOfWithShiftNeg(in0, in1, true, shift, negNumber);
// compute the evaluation result matrix by multiply these two matrices, element by element
// we get a (negNumber+1, n) matrix
functionValues.AssignElementProductOf(leftTermTemp, rightTermTemp);
}
virtual void /*ComputationNodeBase::*/ Validate(bool isFinalValidationPass) override
{
Base::Validate(isFinalValidationPass);
InferMBLayoutFromInputsForStandardCase(isFinalValidationPass);
ValidateInferBinaryInputDims();
if (isFinalValidationPass &&
(Input(0)->GetSampleMatrixNumRows() != Input(1)->GetSampleMatrixNumRows() || !Input(0)->IsMBLayoutCompatibleWith(Input(1))))
{
LogicError("The tensor dimension in the %ls %ls operation does not match.", NodeName().c_str(), OperationName().c_str());
}
auto input3AsLearnableParameterNode = Input(3)->template As<LearnableParameter<ElemType>>();
if (isFinalValidationPass && (!input3AsLearnableParameterNode || input3AsLearnableParameterNode->GetLearningRateMultiplier() != 0) || (Input(3)->GetSampleLayout().GetNumElements() != 1))
LogicError("%ls %ls operation expects a constant scalar for Input(3) which corresponds to number of negative samples.", NodeName().c_str(), OperationName().c_str());
// input(2) is shift, input(3) is the #neg
size_t negNumber = (size_t) Input(3)->Get00Element();
// TODO: This calls for a tensor representation!
SetDims(TensorShape(negNumber + 1), HasMBLayout());
}
virtual void 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<CosDistanceWithNegativeSamplesNode<ElemType>>(nodeP);
node->m_invNorm0->SetValue(*m_invNorm0);
node->m_invNorm1->SetValue(*m_invNorm1);
node->m_invNormSquare->SetValue(*m_invNormSquare);
node->m_leftTerm->SetValue(*m_leftTerm);
node->m_rightTerm->SetValue(*m_rightTerm);
node->m_temp->SetValue(*m_temp);
}
}
// request matrices needed to do node function value evaluation
virtual void RequestMatricesBeforeForwardProp(MatrixPool& matrixPool)
{
Base::RequestMatricesBeforeForwardProp(matrixPool);
RequestMatrixFromPool(m_invNorm0, matrixPool);
RequestMatrixFromPool(m_invNorm1, matrixPool);
RequestMatrixFromPool(m_leftTerm, matrixPool);
RequestMatrixFromPool(m_rightTerm, matrixPool);
}
// request matrices that are needed for gradient computation
virtual void RequestMatricesBeforeBackprop(MatrixPool& matrixPool)
{
Base::RequestMatricesBeforeBackprop(matrixPool);
RequestMatrixFromPool(m_invNormSquare, matrixPool);
RequestMatrixFromPool(m_temp, matrixPool);
}
// release gradient and temp matrices that no longer needed after all the children's gradients are computed.
virtual void ReleaseMatricesAfterBackprop(MatrixPool& matrixPool)
{
Base::ReleaseMatricesAfterBackprop(matrixPool);
ReleaseMatrixToPool(m_invNorm0, matrixPool);
ReleaseMatrixToPool(m_invNorm1, matrixPool);
ReleaseMatrixToPool(m_leftTerm, matrixPool);
ReleaseMatrixToPool(m_rightTerm, matrixPool);
ReleaseMatrixToPool(m_invNormSquare, matrixPool);
ReleaseMatrixToPool(m_temp, matrixPool);
}
private:
// invNorm nodes tranfer data between ForwardProp and BackpropTo
shared_ptr<Matrix<ElemType>> m_invNorm0;
shared_ptr<Matrix<ElemType>> m_invNorm1;
shared_ptr<Matrix<ElemType>> m_leftTerm;
shared_ptr<Matrix<ElemType>> m_rightTerm;
// the rest are temporaries, values don't need to be maintained
shared_ptr<Matrix<ElemType>> m_invNormSquare;
shared_ptr<Matrix<ElemType>> m_temp;
};
template class CosDistanceWithNegativeSamplesNode<float>;
template class CosDistanceWithNegativeSamplesNode<double>;
template class CosDistanceWithNegativeSamplesNode<half>;
template <class ElemType>
void UpdateRunningAverage(ComputationNode<ElemType>& newInput, TensorView<ElemType>& runningAverage,
size_t& runningCount);
class MPIWrapper;
struct DistGradHeader;
template <typename ElemType>
void AggregateAccumulatorValuesAndUpdateEvaluation(
shared_ptr<ComputationNetwork> net,
set<shared_ptr<ComputationNodeBase>> evalNodesWhichAccumulateResult,
shared_ptr<DistGradHeader> gradHeader,
shared_ptr<MPIWrapper> mpi,
size_t packThresholdSizeInBytes = (size_t)DEFAULT_PACK_THRESHOLD_SIZE_IN_BYTES);
// -----------------------------------------------------------------------
// EpochAccumulatorNode calculates mean values of all samples used in forward pass.
// During training, mean sample value is calculated in each epoch. Value of the node will contain mean sample value of
// its input node values since the beginning of epoch.
// This node is useful for creating "per class" metrics like average class recall or mean intersection over union (mean
// IOU) which is standard metric in semantic labeling.
// For mean IOU, we calculate ratio true_positives / (true_positives + false_negatives + false_positives) for all target
// classes and then get mean of those values. true_positives, false_negatives, false_positives should be calculated over
// the whole data set. Here we cannot calculate mean IOU per sample and then average the result. Instead, we use
// EpochAccumulatorNode to store those values over the whole data set.
// -----------------------------------------------------------------------
template <class ElemType>
class EpochAccumulatorNode : public ComputationNodeNonLooping<ElemType>, public NumInputs<1>
{
typedef ComputationNodeNonLooping<ElemType> Base;
UsingComputationNodeMembersBoilerplate;
static const std::wstring TypeName() { return L"EpochAccumulator"; }
public:
EpochAccumulatorNode(DEVICEID_TYPE deviceId, const wstring& name);
EpochAccumulatorNode(const ScriptableObjects::IConfigRecordPtr configp);
virtual void BackpropToNonLooping(size_t inputIndex) override;
virtual bool OutputUsedInComputingInputNodesGradients() const override { return false; }
virtual bool InputUsedInComputingInputNodesGradients(size_t /*childIndex*/) const override { return false; }
virtual void OnEpochStart() override;
virtual void /*ComputationNodeNonLooping::*/ ForwardPropNonLooping() override;
virtual void CopyTo(ComputationNodeBasePtr nodeP, const std::wstring& newName,
const CopyNodeFlags flags) const override;
virtual void Validate(bool isFinalValidationPass);
// Returns tensor view for the accumulator matrix. If accumulator matrix memory is not allocated
// accumulator matrix will be resized (memory will be allocated).
TensorView<ElemType> EnsureAccumlator();
protected:
friend void AggregateAccumulatorValuesAndUpdateEvaluation<ElemType>(
shared_ptr<ComputationNetwork> net,
set<shared_ptr<ComputationNodeBase>> evalNodesWhichAccumulateResult,
shared_ptr<DistGradHeader> gradHeader,
shared_ptr<MPIWrapper> mpi,
size_t packThresholdSize);
void Reset();
size_t GetNumberOfSamples() const { return m_numSamples; }
void SetNumberOfSamples(size_t samples) { m_numSamples = samples; }
shared_ptr<Matrix<ElemType>> GetAccumulator() { return m_accumulator; }
// Copies internal accumulator to the output.
void CopyAccumulatorToValue();
shared_ptr<Matrix<ElemType>> m_accumulator;
size_t m_numSamples;
};
// -----------------------------------------------------------------------
// CastNode converts data types from InputType to ElemType
// -----------------------------------------------------------------------
template <class ElemType, class InputType>
class CastNode : public UnaryElementWiseNode<ElemType>
{
typedef UnaryElementWiseNode<ElemType> Base; UsingUnaryElementwiseNodeBaseMembers;
static const std::wstring TypeName() { return L"Cast"; }
public:
CastNode(DEVICEID_TYPE deviceId, const wstring& name)
: Base(deviceId, name)
{
}
virtual void /*ComputationNode::*/ ForwardProp(const FrameRange& fr) override
{
auto result = ValueFor(fr);
auto input = static_cast<ComputationNode<InputType>&>(*m_inputs[0].get()).ValueFor(fr);
result.CastAssignValuesOf(input);
}
virtual void /*ComputationNode::*/ BackpropTo(const size_t /*inputIndex*/, const FrameRange& fr) override
{
auto grad = GradientFor(fr);
auto inputGrad = static_cast<ComputationNode<InputType>&>(*m_inputs[0].get()).GradientFor(fr);
inputGrad.CastAssignValuesOf(grad);
}
virtual void /*ComputationNodeBase::*/ Validate(bool isFinalValidationPass) override
{
ValidateUnaryMap(isFinalValidationPass);
}
virtual bool OutputUsedInComputingInputNodesGradients() const override { return false; }
virtual bool InputUsedInComputingInputNodesGradients(size_t /*childIndex*/) const override { return false; }
};
template class CastNode<half, float>;
template class CastNode<half, double>;
template class CastNode<float, half>;
template class CastNode<float, double>;
template class CastNode<double, half>;
template class CastNode<double, float>;
}}}
