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

Add TopK error evaluation

This commit is contained in:
Alexey Kamenev 2015-09-10 15:24:26 -07:00
Родитель 35d4a54437
Коммит fa4383e7f5
7 изменённых файлов: 304 добавлений и 195 удалений
Показать статистику Скачать Patch файл Скачать Diff файл Показать все файлы Свернуть все файлы

104
DataReader/ImageReader/ImageReader.cpp
Просмотреть файл

@ -37,16 +37,31 @@ public:
class CropTransform : public ITransform
{
public:
CropTransform(unsigned int seed) : m_rng(seed), m_rndUniInt(0, INT_MAX)
CropTransform(unsigned int seed) : m_rng(seed)
{
}
void Init(const ConfigParameters& config)
{
m_cropType = ParseCropType(config("cropType", ""));
m_cropRatio = std::stof(config("cropRatio", "1"));
if (!(0 < m_cropRatio && m_cropRatio <= 1.0))
RuntimeError("Invalid cropRatio value: %f.", m_cropRatio);
std::stringstream ss{ config("cropRatio", "1") };
std::string token{ "" };
if (std::getline(ss, token, ':'))
{
m_cropRatioMin = std::stof(token);
m_cropRatioMax = std::getline(ss, token, ':') ? std::stof(token) : m_cropRatioMin;
}
if (!(0 < m_cropRatioMin && m_cropRatioMin <= 1.0) ||
!(0 < m_cropRatioMax && m_cropRatioMax <= 1.0) ||
m_cropRatioMin > m_cropRatioMax)
{
RuntimeError("Invalid cropRatio value, must be > 0 and <= 1. cropMin must <= cropMax");
}
m_jitterType = ParseJitterType(config("jitterType", ""));
if (!config.ExistsCurrent("hflip"))
m_hFlip = m_cropType == CropType::Random;
else
@ -55,21 +70,44 @@ public:
void Apply(cv::Mat& mat)
{
mat = mat(GetCropRect(m_cropType, mat.rows, mat.cols, m_cropRatio));
if (m_hFlip && (m_rndUniInt(m_rng) % 2) != 0)
double ratio = 1;
switch (m_jitterType)
{
case RatioJitterType::None:
ratio = m_cropRatioMin;
break;
case RatioJitterType::UniRatio:
ratio = UniRealT(m_cropRatioMin, m_cropRatioMax)(m_rng);
assert(m_cropRatioMin <= ratio && ratio < m_cropRatioMax);
break;
default:
RuntimeError("Jitter type currently not implemented.");
}
mat = mat(GetCropRect(m_cropType, mat.rows, mat.cols, ratio));
if (m_hFlip && std::bernoulli_distribution()(m_rng))
cv::flip(mat, mat, 1);
}
private:
using UniRealT = std::uniform_real_distribution<double>;
using UniIntT = std::uniform_int_distribution<int>;
enum class CropType { Center = 0, Random = 1 };
enum class RatioJitterType
{
None = 0,
UniRatio = 1,
UniLength = 2,
UniArea = 3
};
bool AreEqual(const std::string& s1, const std::string& s2)
{
return std::equal(s1.begin(), s1.end(), s2.begin(), [](const char& a, const char& b) { return std::tolower(a) == std::tolower(b); });
};
CropType ParseCropType(const std::string& src)
{
auto AreEqual = [](const std::string& s1, const std::string& s2) -> bool
{
return std::equal(s1.begin(), s1.end(), s2.begin(), [](const char& a, const char& b) { return std::tolower(a) == std::tolower(b); });
};
if (src.empty() || AreEqual(src, "center"))
return CropType::Center;
if (AreEqual(src, "random"))
@ -78,6 +116,20 @@ private:
RuntimeError("Invalid crop type: %s.", src.c_str());
}
RatioJitterType ParseJitterType(const std::string& src)
{
if (src.empty() || AreEqual(src, "none"))
return RatioJitterType::None;
if (AreEqual(src, "uniratio"))
return RatioJitterType::UniRatio;
if (AreEqual(src, "unilength"))
return RatioJitterType::UniLength;
if (AreEqual(src, "uniarea"))
return RatioJitterType::UniArea;
RuntimeError("Invalid jitter type: %s.", src.c_str());
}
cv::Rect GetCropRect(CropType type, int crow, int ccol, double cropRatio)
{
assert(crow > 0);
@ -87,7 +139,6 @@ private:
int cropSize = static_cast<int>(std::min(crow, ccol) * cropRatio);
int xOff = -1;
int yOff = -1;
switch (type)
{
case CropType::Center:
@ -95,8 +146,8 @@ private:
yOff = (crow - cropSize) / 2;
break;
case CropType::Random:
xOff = m_rndUniInt(m_rng) % std::max(ccol - cropSize, 1);
yOff = m_rndUniInt(m_rng) % std::max(crow - cropSize, 1);
xOff = UniIntT(0, ccol - cropSize)(m_rng);
yOff = UniIntT(0, crow - cropSize)(m_rng);
break;
default:
assert(false);
@ -108,18 +159,20 @@ private:
}
private:
std::default_random_engine m_rng;
std::uniform_int_distribution<int> m_rndUniInt;
// REVIEW alexeyk: currently not thread safe. Engines are expensive to create.
std::mt19937 m_rng;
CropType m_cropType;
double m_cropRatio;
double m_cropRatioMin;
double m_cropRatioMax;
RatioJitterType m_jitterType;
bool m_hFlip;
};
class ScaleTransform : public ITransform
{
public:
ScaleTransform(int dataType, unsigned int seed) : m_dataType(dataType), m_rng(seed), m_rndUniInt(0, INT_MAX)
ScaleTransform(int dataType, unsigned int seed) : m_dataType(dataType), m_rng(seed)
{
assert(m_dataType == CV_32F || m_dataType == CV_64F);
@ -160,12 +213,13 @@ public:
assert(m_interp.size() > 0);
cv::resize(mat, mat, cv::Size(static_cast<int>(m_imgWidth), static_cast<int>(m_imgHeight)), 0, 0,
m_interp[m_rndUniInt(m_rng) % m_interp.size()]);
m_interp[UniIntT(0, static_cast<int>(m_interp.size()) - 1)(m_rng)]);
}
private:
std::default_random_engine m_rng;
std::uniform_int_distribution<int> m_rndUniInt;
using UniIntT = std::uniform_int_distribution<int>;
// REVIEW alexeyk: currently not thread safe. Engines are expensive to create.
std::mt19937 m_rng;
int m_dataType;
@ -229,7 +283,7 @@ private:
// ImageReader
template<class ElemType>
ImageReader<ElemType>::ImageReader() : m_seed(0), m_rng(m_seed), m_rndUniInt(0, INT_MAX)
ImageReader<ElemType>::ImageReader() : m_seed(0), m_rng(m_seed)
{
m_transforms.push_back(std::make_unique<CropTransform>(m_seed));
m_transforms.push_back(std::make_unique<ScaleTransform>(sizeof(ElemType) == 4 ? CV_32F : CV_64F, m_seed));
@ -287,8 +341,6 @@ void ImageReader<ElemType>::Init(const ConfigParameters& config)
files.push_back({ imgPath, std::stoi(clsId) });
}
std::shuffle(files.begin(), files.end(), m_rng);
m_epochStart = 0;
m_mbStart = 0;
}
@ -304,6 +356,8 @@ void ImageReader<ElemType>::StartMinibatchLoop(size_t mbSize, size_t epoch, size
assert(mbSize > 0);
assert(requestedEpochSamples > 0);
std::shuffle(files.begin(), files.end(), m_rng);
m_epochSize = (requestedEpochSamples == requestDataSize ? files.size() : requestedEpochSamples);
m_mbSize = mbSize;
// REVIEW alexeyk: if user provides epoch size explicitly then we assume epoch size is a multiple of mbsize, is this ok?
@ -339,7 +393,7 @@ bool ImageReader<ElemType>::GetMinibatch(std::map<std::wstring, Matrix<ElemType>
std::fill(m_labBuf.begin(), m_labBuf.end(), static_cast<ElemType>(0));
#pragma omp parallel for ordered schedule(dynamic)
//#pragma omp parallel for ordered schedule(dynamic)
for (long long i = 0; i < static_cast<long long>(mbLim - m_mbStart); i++)
{
const auto& p = files[i + m_mbStart];

3
DataReader/ImageReader/ImageReader.h
Просмотреть файл

@ -39,8 +39,7 @@ public:
private:
unsigned int m_seed;
std::default_random_engine m_rng;
std::uniform_int_distribution<int> m_rndUniInt;
std::mt19937 m_rng;
std::vector<std::unique_ptr<ITransform>> m_transforms;

23
MachineLearning/CNTKComputationNetworkLib/EvaluationCriterionNodes.h
Просмотреть файл

@ -21,7 +21,7 @@ namespace Microsoft { namespace MSR { namespace CNTK {
// -----------------------------------------------------------------------
template<class ElemType>
class ErrorPredictionNode : public ComputationNodeNonLooping/*ComputationNode*/<ElemType>, public NumInputs<2>
class ErrorPredictionNode : public ComputationNodeNonLooping/*ComputationNode*/<ElemType>
{
typedef ComputationNodeNonLooping<ElemType> Base; UsingComputationNodeMembersBoilerplate;
static const std::wstring TypeName() { return L"ErrorPrediction"; }
@ -42,16 +42,16 @@ namespace Microsoft { namespace MSR { namespace CNTK {
virtual void /*ComputationNodeNonLooping::*/EvaluateThisNodeNonLooping() override
{
EvaluateThisNodeS(m_functionValues, Inputs(0)->FunctionValues(), Inputs(1)->FunctionValues(), m_maxIndexes0, m_maxIndexes1, m_maxValues, shared_from_this());
EvaluateThisNodeS(m_functionValues, Inputs(0)->FunctionValues(), Inputs(1)->FunctionValues(), m_maxIndexes0, m_maxIndexes1, m_maxValues, m_topK, shared_from_this());
}
void EvaluateThisNodeS(Matrix<ElemType>& functionValues, const Matrix<ElemType>& inputFunctionValues0, const Matrix<ElemType>& inputFunctionValues1, Matrix<ElemType>& maxIndexes0, Matrix<ElemType>& maxIndexes1, Matrix<ElemType>& maxValues, ComputationNodePtr curNode)
{
inputFunctionValues0.VectorMax(maxIndexes0, maxValues, true);
inputFunctionValues1.VectorMax(maxIndexes1, maxValues, true);
inputFunctionValues1.VectorMax(maxIndexes1, maxValues, true, topK);
curNode->MaskMissingColumnsToZero(maxIndexes0, Inputs(0)->GetMBLayout()); // we are fine since it will only be called with full minibatch
curNode->MaskMissingColumnsToZero(maxIndexes1, Inputs(1)->GetMBLayout());
functionValues.AssignNumOfDiff(maxIndexes0, maxIndexes1);
functionValues.AssignNumOfDiff(maxIndexes0, maxIndexes1, topK > 1);
#if NANCHECK
functionValues.HasNan("ErrorPrediction");
#endif
@ -81,6 +81,14 @@ namespace Microsoft { namespace MSR { namespace CNTK {
m_maxValues.Resize(1,cols);
}
m_topK = 1;
if (m_children.size() == 3)
{
if (Inputs(2)->FunctionValues().GetNumRows() != 1 || Inputs(2)->FunctionValues().GetNumCols() != 1)
throw std::logic_error("TopK in ErrorPredictionNode must be a scalar value.");
m_topK = static_cast<int>(Inputs(2)->FunctionValues().Get00Element());
}
//if (Inputs(0)->GetNumRows() == 0 || Inputs(1)->GetNumRows() == 0)
// LogicError("ErrorPrediction operation: one of the operands has 0 elements.");
@ -96,9 +104,9 @@ namespace Microsoft { namespace MSR { namespace CNTK {
// resize the temporaries to their proper size
size_t cols = Inputs(0)->GetNumCols();
m_maxIndexes0.Resize(1,cols);
m_maxIndexes1.Resize(1,cols);
m_maxValues.Resize(1,cols);
m_maxIndexes0.Resize(m_topK,cols);
m_maxIndexes1.Resize(m_topK,cols);
m_maxValues.Resize(m_topK,cols);
}
virtual void InferImageDimsFromInputs()
@ -133,6 +141,7 @@ protected:
private:
Matrix<ElemType> m_maxIndexes0, m_maxIndexes1;
Matrix<ElemType> m_maxValues;
int m_topK;
};
template class ErrorPredictionNode<float>;

87
Math/Math/CPUMatrix.cpp
Просмотреть файл

@ -20,6 +20,7 @@
#include <exception>
#include <thread>
#include<iostream>
#include <algorithm>
#ifdef _WIN32
#include <Windows.h>
#else
@ -3301,7 +3302,7 @@ namespace Microsoft { namespace MSR { namespace CNTK {
}
//I decided to use CPUMatrix<ElemType>& maxIndexes instead of integer vector because the result may be used to do additional calculation
template<class ElemType>
void CPUMatrix<ElemType>::VectorMax(CPUMatrix<ElemType>& maxIndexes, CPUMatrix<ElemType>& maxValues, const bool isColWise) const
void CPUMatrix<ElemType>::VectorMax(CPUMatrix<ElemType>& maxIndexes, CPUMatrix<ElemType>& maxValues, const bool isColWise, int topK) const
{
if (IsEmpty())
LogicError("VectorMax: Matrix is empty.");
@ -3309,33 +3310,63 @@ namespace Microsoft { namespace MSR { namespace CNTK {
auto& us=*this;
const int m = (int)GetNumRows();
const int n = (int)GetNumCols();
assert(topK <= m);
assert (m>0 && n>0); //converting from size_t to int may cause overflow
if (isColWise) //col-wise
{
maxValues.Resize(1, n);
maxIndexes.Resize(1, n);
maxValues.Resize(topK, n);
maxIndexes.Resize(topK, n);
#pragma omp parallel for
for (int j=0; j<n; j++)
if (topK == 1)
{
ElemType v = us(0, j);
size_t index = 0;
foreach_row(i,us)
#pragma omp parallel for
for (int j = 0; j < n; j++)
{
if (v < us(i,j))
ElemType v = us(0, j);
size_t index = 0;
foreach_row(i, us)
{
index = i;
v = us(i,j);
if (v < us(i, j))
{
index = i;
v = us(i, j);
}
}
maxValues(0, j) = v;
maxIndexes(0, j) = (ElemType)index;
}
}
else
{
std::vector<int> indices(m);
int i = 0;
std::generate(indices.begin(), indices.end(), [&i] { return i++; });
const ElemType* curVal = m_pArray;
ElemType* curIdx = maxIndexes.m_pArray;
ElemType* curMax = maxValues.m_pArray;
for (int icol = 0; icol < n; icol++, curVal += m, curIdx += topK, curMax += topK)
{
// Partial sort, descending order.
std::nth_element(indices.begin(), indices.begin() + topK, indices.end(),
[curVal](const int& a, const int& b) { return curVal[a] > curVal[b]; });
// REVIEW alexeyk: the following produces warning (see SCL_SECURE_NO_WARNINGS) so use loop instead.
//std::transform(indices.begin(), indices.begin() + topK, curIdx, [](const int& a) { return static_cast<ElemType>(a); });
for (int i = 0; i < topK; i++)
{
curIdx[i] = static_cast<ElemType>(indices[i]);
curMax[i] = curVal[indices[i]];
}
}
maxValues(0,j) = v;
maxIndexes(0,j) = (ElemType)index;
}
}
else
{
if (topK > 1)
RuntimeError("Row-wise TopK max is not supported.");
maxValues.Resize(m,1);
maxIndexes.Resize(m, 1);
@ -3420,19 +3451,35 @@ namespace Microsoft { namespace MSR { namespace CNTK {
}
template<class ElemType>
CPUMatrix<ElemType>& CPUMatrix<ElemType>::AssignNumOfDiff(const CPUMatrix<ElemType>& a, const CPUMatrix<ElemType>& b)
CPUMatrix<ElemType>& CPUMatrix<ElemType>::AssignNumOfDiff(const CPUMatrix<ElemType>& a, const CPUMatrix<ElemType>& b, bool searchInCol)
{
if (a.GetNumRows() != b.GetNumRows() || a.GetNumCols() != b.GetNumCols())
InvalidArgument("AssignNumOfDiff: a and b must have same dimension.");
if (a.GetNumCols() != b.GetNumCols())
throw std::invalid_argument("AssignNumOfDiff: a and b must have the same number of columns.");
if (!searchInCol && a.GetNumRows() != b.GetNumRows())
throw std::invalid_argument("AssignNumOfDiff: a and b must have the same number of rows.");
ElemType n = 0;
foreach_coord(i,j,a)
if (!searchInCol)
{
n += (a(i,j) != b(i,j));
foreach_coord(i, j, a)
{
n += (a(i, j) != b(i, j));
}
}
else
{
size_t crow = b.GetNumRows();
const ElemType* curCol = b.m_pArray;
for (size_t icol = 0; icol < a.GetNumCols(); icol++, curCol += crow)
{
auto res = std::find(curCol, curCol + crow, a(0, icol));
if (res == curCol + crow)
n++;
}
}
Resize(1,1); //result should be one element
(*this)(0,0) = n;
Resize(1, 1); //result should be one element
(*this)(0, 0) = n;
return *this;
}

4
Math/Math/CPUMatrix.h
Просмотреть файл

@ -252,10 +252,10 @@ namespace Microsoft { namespace MSR { namespace CNTK {
CPUMatrix<ElemType>& AssignPositiveAndShiftedNegSample(const CPUMatrix<ElemType>& a, const size_t posNumber, const size_t negNumber, const size_t shiftNumber);
CPUMatrix<ElemType>& AddFoldedPositiveAndShiftedNegSample(const CPUMatrix<ElemType>& a, const size_t posNumber, const size_t negNumber, const size_t shiftNumber);
void VectorMax(CPUMatrix<ElemType>& maxIndexes, CPUMatrix<ElemType>& maxValues, const bool isColWise) const;
void VectorMax(CPUMatrix<ElemType>& maxIndexes, CPUMatrix<ElemType>& maxValues, const bool isColWise, int topK = 1) const;
void VectorMin(CPUMatrix<ElemType>& mainndexes, CPUMatrix<ElemType>& minValues, const bool isColWise) const;
CPUMatrix<ElemType>& AssignNumOfDiff(const CPUMatrix<ElemType>& a, const CPUMatrix<ElemType>& b);
CPUMatrix<ElemType>& AssignNumOfDiff(const CPUMatrix<ElemType>& a, const CPUMatrix<ElemType>& b, bool searchInCol = false);
void Print(const char* matrixName, size_t rowStart, size_t rowEnd, size_t colStart, size_t colEnd) const;
void Print(const char* matrixName = nullptr) const; //print whole matrix. can be expensive

8
Math/Math/Matrix.cpp
Просмотреть файл

@ -2457,7 +2457,7 @@ namespace Microsoft { namespace MSR { namespace CNTK {
}
template<class ElemType>
Matrix<ElemType>& Matrix<ElemType>::AssignNumOfDiff (const Matrix<ElemType>& a, const Matrix<ElemType>& b)
Matrix<ElemType>& Matrix<ElemType>::AssignNumOfDiff (const Matrix<ElemType>& a, const Matrix<ElemType>& b, bool searchInCol)
{
DecideAndMoveToRightDevice(a, b, *this);
//WARNING: a and b must have same type
@ -2468,7 +2468,7 @@ namespace Microsoft { namespace MSR { namespace CNTK {
DISPATCH_MATRIX_ON_FLAG(this,
this,
this->m_CPUMatrix->AssignNumOfDiff(*a.m_CPUMatrix, *b.m_CPUMatrix),
this->m_CPUMatrix->AssignNumOfDiff(*a.m_CPUMatrix, *b.m_CPUMatrix, searchInCol),
this->m_GPUMatrix->AssignNumOfDiff(*a.m_GPUMatrix, *b.m_GPUMatrix),
NOT_IMPLEMENTED,
NOT_IMPLEMENTED
@ -3380,7 +3380,7 @@ namespace Microsoft { namespace MSR { namespace CNTK {
//I decided to use Matrix<ElemType>& maxIndexes instead of integer vector because the result may be used to do additional calculation
template<class ElemType>
void Matrix<ElemType>::VectorMax(Matrix<ElemType>& maxIndexes, Matrix<ElemType>& maxValues, const bool isColWise) const
void Matrix<ElemType>::VectorMax(Matrix<ElemType>& maxIndexes, Matrix<ElemType>& maxValues, const bool isColWise, int topK) const
{
if (IsEmpty())
LogicError("VectorMax: Matrix is empty.");
@ -3391,7 +3391,7 @@ namespace Microsoft { namespace MSR { namespace CNTK {
DISPATCH_MATRIX_ON_FLAG(this,
&maxValues,
this->m_CPUMatrix->VectorMax(*maxIndexes.m_CPUMatrix,*maxValues.m_CPUMatrix,isColWise); maxIndexes.SetDataLocation(CPU, DENSE),
this->m_CPUMatrix->VectorMax(*maxIndexes.m_CPUMatrix,*maxValues.m_CPUMatrix,isColWise,topK); maxIndexes.SetDataLocation(CPU, DENSE),
this->m_GPUMatrix->VectorMax(*maxIndexes.m_GPUMatrix,*maxValues.m_GPUMatrix,isColWise); maxIndexes.SetDataLocation(GPU, DENSE),
NOT_IMPLEMENTED,
NOT_IMPLEMENTED

4
Math/Math/Matrix.h
Просмотреть файл

@ -367,10 +367,10 @@ namespace Microsoft { namespace MSR { namespace CNTK {
ElemType MatrixNorm0() const; //number of non-zero elemets
Matrix<ElemType>& AssignSignOf(const Matrix<ElemType>& a);
Matrix<ElemType>& AddSignOf(const Matrix<ElemType>& a);
void VectorMax(Matrix<ElemType>& maxIndexes, Matrix<ElemType>& maxValues, const bool isColWise) const;
void VectorMax(Matrix<ElemType>& maxIndexes, Matrix<ElemType>& maxValues, const bool isColWise, int topK = 1) const;
void VectorMin(Matrix<ElemType>& mainndexes, Matrix<ElemType>& minValues, const bool isColWise) const;
Matrix<ElemType>& AssignNumOfDiff(const Matrix<ElemType>& a, const Matrix<ElemType>& b);
Matrix<ElemType>& AssignNumOfDiff(const Matrix<ElemType>& a, const Matrix<ElemType>& b, bool searchInCol = false);
Matrix<ElemType>& AssignInnerProductOfMatrices(const Matrix<ElemType>& a, const Matrix<ElemType>& b); //this method will resize(1,1) first