Normalize files with mixed line endings (to DOS LE)

Common/Include/ScriptableObjects.h

@@ -43,7 +43,7 @@ namespace Microsoft { namespace MSR { namespace ScriptableObjects {
     //  - C++ primitives like 'double' -> wrap in a Wrapper first then in a BoxOf, e.g. Number = BoxOf<Wrapped<double>>
 
     struct Object { virtual ~Object() { } };
-
+
     // indicates that the object has a name should be set from the expression path
 
     struct HasName { virtual void SetName(const wstring & name) = 0; };
@@ -95,7 +95,7 @@ namespace Microsoft { namespace MSR { namespace ScriptableObjects {
     // TODO: unify with ComputationNodeBase
     // -----------------------------------------------------------------------
 
-    class ComputationNodeObject : public Object { };   // a base class for all nodes (that has no template parameter)
+    class ComputationNodeObject : public Object { };   // a base class for all nodes (that has no template parameter)
 
     // -----------------------------------------------------------------------
     // HasToString -- trait to indicate an object can print their content

Common/Include/Sequences.h

@@ -567,97 +567,97 @@ namespace Microsoft { namespace MSR { namespace CNTK {
     // -----------------------------------------------------------------------
 
     template<class ElemType>
-    static inline Matrix<ElemType> DataSliceWithMBLayout(Matrix<ElemType> & data,
-                                                         const FrameRange & frameRange/*select frame or entire batch*/,
-                                                         const MBLayoutPtr & pMBLayout/*the MB layout of 'data'*/)
-    {
-        // MBLayout of data and of FrameRange must be identical pointers,
-        // or in case of broadcasting, respective parent pointers.
-        // MBLayouts that are identical in content but not object identity (pointer) are not admissible.
-        // For those cases, use a ReconcileMBLayout node.
-        if (frameRange.m_pMBLayout != pMBLayout)
-        {
-            // if broadcast allowed then it is allowed to broadcast from an outer-loop value
-            // Currently, the only 'outer' loop we have is to have no layout.
-            if (frameRange.m_broadcastAllowed && !pMBLayout && data.GetNumCols() == 1)
-                return data.AsReference();
-            if (frameRange.m_pMBLayout && pMBLayout && *frameRange.m_pMBLayout == *pMBLayout)
-                LogicError("DataSlice: frameRange's MBLayout inconsistent with matrix. They are compatible though--are you missing a ReconcileMBLayout operation?");
-            else
-                LogicError("DataSlice: frameRange's MBLayout inconsistent with matrix");
-        }
-        // if FrameRange refers to whole minibatch (map mode)
-        // or if we don't even have a layout
-        // then return the whole matrix
-        // but as a reference (e.g. it cannot be resized)
-        if (!pMBLayout || frameRange.IsAllFrames())
-        {
-            if (frameRange.seqIndex == SIZE_MAX)
-                return data.AsReference();
-            else
-            {
-                if (!pMBLayout)
-                    LogicError("DataSlice: Attempting to retrieve a parallel sequence from data without layout.");
-#if 1
-                else
-                    LogicError("DataSlice: To retrieve a parallel sequence, implement Matrix::RowSlice() first!");
-#else
-                // get a reshaped view that stacks all sequences into T long vectors
-                auto mat = data.ColumnSlice(0, data.GetNumCols());
-                mat.Resize(data.GetNumRows() * pMBLayout->GetNumParallelSequences(), data.GetNumRows() / pMBLayout->GetNumParallelSequences());
-                return mat;   // .RowSlice(frameRange.seqIndex * data.GetNumRows());
-                // TODO: Why does RowSlice() not exist? Seems simple. Is there a hidden assumption of contiguous memory?#endif
-#endif
-            }
-        }
-        // FrameRange refers to a time slice -> return that
-        else
-        {
-            size_t numParallelSequences = pMBLayout->GetNumParallelSequences();
-            size_t startColumn = frameRange.t() * numParallelSequences;
-            if (frameRange.seqIndex == SIZE_MAX)
-                return data.ColumnSlice(startColumn, numParallelSequences);
-            else
-                return data.ColumnSlice(startColumn + frameRange.seqIndex, 1);
-        }
-    }
+    static inline Matrix<ElemType> DataSliceWithMBLayout(Matrix<ElemType> & data,
+                                                         const FrameRange & frameRange/*select frame or entire batch*/,
+                                                         const MBLayoutPtr & pMBLayout/*the MB layout of 'data'*/)
+    {
+        // MBLayout of data and of FrameRange must be identical pointers,
+        // or in case of broadcasting, respective parent pointers.
+        // MBLayouts that are identical in content but not object identity (pointer) are not admissible.
+        // For those cases, use a ReconcileMBLayout node.
+        if (frameRange.m_pMBLayout != pMBLayout)
+        {
+            // if broadcast allowed then it is allowed to broadcast from an outer-loop value
+            // Currently, the only 'outer' loop we have is to have no layout.
+            if (frameRange.m_broadcastAllowed && !pMBLayout && data.GetNumCols() == 1)
+                return data.AsReference();
+            if (frameRange.m_pMBLayout && pMBLayout && *frameRange.m_pMBLayout == *pMBLayout)
+                LogicError("DataSlice: frameRange's MBLayout inconsistent with matrix. They are compatible though--are you missing a ReconcileMBLayout operation?");
+            else
+                LogicError("DataSlice: frameRange's MBLayout inconsistent with matrix");
+        }
+        // if FrameRange refers to whole minibatch (map mode)
+        // or if we don't even have a layout
+        // then return the whole matrix
+        // but as a reference (e.g. it cannot be resized)
+        if (!pMBLayout || frameRange.IsAllFrames())
+        {
+            if (frameRange.seqIndex == SIZE_MAX)
+                return data.AsReference();
+            else
+            {
+                if (!pMBLayout)
+                    LogicError("DataSlice: Attempting to retrieve a parallel sequence from data without layout.");
+#if 1
+                else
+                    LogicError("DataSlice: To retrieve a parallel sequence, implement Matrix::RowSlice() first!");
+#else
+                // get a reshaped view that stacks all sequences into T long vectors
+                auto mat = data.ColumnSlice(0, data.GetNumCols());
+                mat.Resize(data.GetNumRows() * pMBLayout->GetNumParallelSequences(), data.GetNumRows() / pMBLayout->GetNumParallelSequences());
+                return mat;   // .RowSlice(frameRange.seqIndex * data.GetNumRows());
+                // TODO: Why does RowSlice() not exist? Seems simple. Is there a hidden assumption of contiguous memory?#endif
+#endif
+            }
+        }
+        // FrameRange refers to a time slice -> return that
+        else
+        {
+            size_t numParallelSequences = pMBLayout->GetNumParallelSequences();
+            size_t startColumn = frameRange.t() * numParallelSequences;
+            if (frameRange.seqIndex == SIZE_MAX)
+                return data.ColumnSlice(startColumn, numParallelSequences);
+            else
+                return data.ColumnSlice(startColumn + frameRange.seqIndex, 1);
+        }
+    }
 
     // -----------------------------------------------------------------------
     // MaskMissingColumnsTo() -- function to set gaps to zero or NaN
     // -----------------------------------------------------------------------
 
-    // This sets MB columns to 0 (or any 'val') that have the NoLabel or NoFeature flag set.
-    // Such situation happens when packing multiple sequences for parallel processing--there will be some gaps, which are flagged by these flags.
-    // Nodes that operate in 'map' style (input(j) -> output(j) independently) can ignore this; it will be garbage-in-garbage-out.
-    // However, nodes that 'reduce' minibatches (e.g. computing the sum of all frames across all sequences) must deal with the garbage.
-    // This function sets those to 0, assuming that now they can be reduced without affecting the result.
-    // This function can operate on the whole range or on a selected single frame and/or a single sequence.
-    // It is indirectly guarded by the m_maskMissingColumnsToZero flag, which, if false, will install a layout with IsAllNone() to be true. TODO: we better always install the same layout, and instead test m_maskMissingColumnsToZero here.
-    // Note that existing 'reduce' style operations--the criterion nodes and gradient computation--already call this.  --BUGBUG: They can't, wrong layout!
-    // Warning: The layout used here must match the matrix. E.g. don't pass a child's matrix from a criterion node (use Inputs(x)->MaskMissing{Values,Gradient}ColumnsToZero() instead.
-    template<class ElemType>
-    static inline bool MaskMissingColumnsTo(Matrix<ElemType>& matrixToBeMasked, const MBLayoutPtr & pMBLayout, const FrameRange & frameRange, ElemType val)
-    {
-        bool foundLabelOrFeatureMissing = false;    // return value: set to true if either nolabel or feature missing is processed
-
-        if (pMBLayout && !pMBLayout->IsAllNone())   // TODO: This should check whether there are any gaps.
-        {
-            size_t nT = pMBLayout->GetNumTimeSteps();
-            size_t nS = pMBLayout->GetNumParallelSequences();
-
-            if (matrixToBeMasked.GetNumCols() != nT * nS)
-                LogicError("MaskMissingColumnsToZero: pMBLayout->m_minibatchPackingFlags should have one element for each timestep of all streams. Check feature reader. ");
-
-            shared_ptr<Matrix<char>> columnsValidityMask = pMBLayout->GetColumnsValidityMask(frameRange, matrixToBeMasked.GetDeviceId());
-            if (columnsValidityMask != nullptr)
-            {
-                auto matrixSliceToMask = DataSliceWithMBLayout(matrixToBeMasked, frameRange, pMBLayout);
-                foundLabelOrFeatureMissing = true;
-                matrixSliceToMask.MaskColumnsValue(*columnsValidityMask, val);
-            }
-        }
-
-        return foundLabelOrFeatureMissing;
-    }
-
+    // This sets MB columns to 0 (or any 'val') that have the NoLabel or NoFeature flag set.
+    // Such situation happens when packing multiple sequences for parallel processing--there will be some gaps, which are flagged by these flags.
+    // Nodes that operate in 'map' style (input(j) -> output(j) independently) can ignore this; it will be garbage-in-garbage-out.
+    // However, nodes that 'reduce' minibatches (e.g. computing the sum of all frames across all sequences) must deal with the garbage.
+    // This function sets those to 0, assuming that now they can be reduced without affecting the result.
+    // This function can operate on the whole range or on a selected single frame and/or a single sequence.
+    // It is indirectly guarded by the m_maskMissingColumnsToZero flag, which, if false, will install a layout with IsAllNone() to be true. TODO: we better always install the same layout, and instead test m_maskMissingColumnsToZero here.
+    // Note that existing 'reduce' style operations--the criterion nodes and gradient computation--already call this.  --BUGBUG: They can't, wrong layout!
+    // Warning: The layout used here must match the matrix. E.g. don't pass a child's matrix from a criterion node (use Inputs(x)->MaskMissing{Values,Gradient}ColumnsToZero() instead.
+    template<class ElemType>
+    static inline bool MaskMissingColumnsTo(Matrix<ElemType>& matrixToBeMasked, const MBLayoutPtr & pMBLayout, const FrameRange & frameRange, ElemType val)
+    {
+        bool foundLabelOrFeatureMissing = false;    // return value: set to true if either nolabel or feature missing is processed
+
+        if (pMBLayout && !pMBLayout->IsAllNone())   // TODO: This should check whether there are any gaps.
+        {
+            size_t nT = pMBLayout->GetNumTimeSteps();
+            size_t nS = pMBLayout->GetNumParallelSequences();
+
+            if (matrixToBeMasked.GetNumCols() != nT * nS)
+                LogicError("MaskMissingColumnsToZero: pMBLayout->m_minibatchPackingFlags should have one element for each timestep of all streams. Check feature reader. ");
+
+            shared_ptr<Matrix<char>> columnsValidityMask = pMBLayout->GetColumnsValidityMask(frameRange, matrixToBeMasked.GetDeviceId());
+            if (columnsValidityMask != nullptr)
+            {
+                auto matrixSliceToMask = DataSliceWithMBLayout(matrixToBeMasked, frameRange, pMBLayout);
+                foundLabelOrFeatureMissing = true;
+                matrixSliceToMask.MaskColumnsValue(*columnsValidityMask, val);
+            }
+        }
+
+        return foundLabelOrFeatureMissing;
+    }
+
 }}}

DataReader/DSSMReader/DSSMReader.h

@@ -93,7 +93,7 @@ class DSSMReader : public IDataReader<ElemType>
     bool m_partialMinibatch;    // a partial minibatch is allowed
     LabelKind m_labelType;  // labels are categories, create mapping table
     msra::dbn::randomordering m_randomordering;   // randomizing class
-    MBLayoutPtr m_pMBLayout;
+    MBLayoutPtr m_pMBLayout;
 
     std::wstring m_labelsName;
     std::wstring m_featuresName;

DataReader/LibSVMBinaryReader/LibSVMBinaryReader.h

@@ -96,7 +96,7 @@ class LibSVMBinaryReader : public IDataReader<ElemType>
     bool m_partialMinibatch;    // a partial minibatch is allowed
     LabelKind m_labelType;  // labels are categories, create mapping table
     msra::dbn::randomordering m_randomordering;   // randomizing class
-    MBLayoutPtr m_pMBLayout;
+    MBLayoutPtr m_pMBLayout;
 
     std::wstring m_labelsName;
     std::wstring m_labelsCategoryName;

MachineLearning/CNTKComputationNetworkLib/InputAndParamNodes.h

@@ -70,7 +70,7 @@ namespace Microsoft { namespace MSR { namespace CNTK {
             fstream >> rows >> cols;
 
             SetDims(rows, cols);
-            LoadFunctionValues(fstream);
+            LoadFunctionValues(fstream);
 
             m_imageLayout = ImageLayoutWHC(1, rows, 1);
         }
@@ -350,8 +350,8 @@ namespace Microsoft { namespace MSR { namespace CNTK {
             Inputs(1)->FunctionValues().TransferFromDeviceToDevice(input0DeviceId, input1DeviceId);
         }
 
-        virtual void /*ComputationNode::*/ComputeInputPartial(const size_t inputIndex, const FrameRange & frameRange) override
-        {
+        virtual void /*ComputationNode::*/ComputeInputPartial(const size_t inputIndex, const FrameRange & frameRange) override
+        {
             if (frameRange.IsAllFrames()) { ComputeInputPartialMap(inputIndex); return; } // TODO: remove these one by one
             if (inputIndex > 1)
                 InvalidArgument("LookupTable operation only takes two inputs.");
@@ -573,8 +573,8 @@ namespace Microsoft { namespace MSR { namespace CNTK {
             Matrix<ElemType>::ScaleAndAdd(1.0, GradientValues(), Inputs(inputIndex)->GradientValues());
         }
 
-        virtual void /*ComputationNode::*/ComputeInputPartial(const size_t inputIndex, const FrameRange & frameRange) override
-        {
+        virtual void /*ComputationNode::*/ComputeInputPartial(const size_t inputIndex, const FrameRange & frameRange) override
+        {
             if (frameRange.IsAllFrames()) { ComputeInputPartialMap(inputIndex); return; } // TODO: remove these one by one
             assert(m_functionValues->GetNumRows() == GradientValues().GetNumRows()); // original used m_functionValues->GetNumRows() for loop dimension
             assert(m_pMBLayout);

MachineLearning/CNTKSGDLib/DataReaderHelpers.h

@@ -1,22 +1,22 @@
 // DataReaderHelper.h -- helper functions that understand both DataReader and ComputationNetwork
 
 #pragma once
-
-#include "Basics.h"
-#include "DataReader.h"
-#include "ComputationNetwork.h"
-#include "MPIWrapper.h"
-#include <string>
-#include <map>
+
+#include "Basics.h"
+#include "DataReader.h"
+#include "ComputationNetwork.h"
+#include "MPIWrapper.h"
+#include <string>
+#include <map>
 #include "TrainingCriterionNodes.h"
 
-namespace Microsoft { namespace MSR { namespace CNTK {
-
-    /*static*/ struct DataReaderHelpers
-    {
-        // -------------------------------------------------------------------
+namespace Microsoft { namespace MSR { namespace CNTK {
+
+    /*static*/ struct DataReaderHelpers
+    {
+        // -------------------------------------------------------------------
         // DecimateMinibatch - decimate minibatch for parallelization
-        // -------------------------------------------------------------------
+        // -------------------------------------------------------------------
 
         // We sub-sample the parallel utterances.
         template<class ElemType> 
@@ -96,15 +96,15 @@ namespace Microsoft { namespace MSR { namespace CNTK {
                 pNewMBLayout->Set(id, t, pMBLayout->Get(id + sent_start, t));
             pMBLayout->MoveFrom(pNewMBLayout);  // update layout in-place
         }
-
-        // -------------------------------------------------------------------
+
+        // -------------------------------------------------------------------
         // GetMinibatchIntoNetwork() -- get one minibatch from Reader (this->trainSetDataReader) into Network (this->net)
         // Returns false if end of epoch has been reached.
         // If not, then actualMBSize is set. Note that 0 is a valid value to be returned for actualMBSize, caller must handle that correctly.
-        // -------------------------------------------------------------------
+        // -------------------------------------------------------------------
 
         // Note: Later, a function like this will become part of the reader interface.
-        // TODO: callers of this often do ComputationNetwork::UpdateEvalTimeStamps(featureNodes) and also for labels; we should eliminate the need for this.
+        // TODO: callers of this often do ComputationNetwork::UpdateEvalTimeStamps(featureNodes) and also for labels; we should eliminate the need for this.
         template<class ElemType>
         static bool GetMinibatchIntoNetwork(IDataReader<ElemType>& trainSetDataReader,
                                             ComputationNetwork& net,
@@ -188,6 +188,6 @@ namespace Microsoft { namespace MSR { namespace CNTK {
 
             return true;
         }
-    };
-
-}}}
+    };
+
+}}}

MachineLearning/CNTKSequenceTrainingLib/parallelforwardbackward.cpp

@@ -2,8 +2,8 @@
 //
 // F. Seide, V-hansu
 
-#define _CRT_SECURE_NO_WARNINGS // "secure" CRT not available on all platforms  --add this at the top of all CPP files that give "function or variable may be unsafe" warnings
-
+#define _CRT_SECURE_NO_WARNINGS // "secure" CRT not available on all platforms  --add this at the top of all CPP files that give "function or variable may be unsafe" warnings
+
 #include "BestGpu.h"            // for CPUONLY
 #include "latticearchive.h"     // we implement parts of class lattice
 #include "simple_checked_arrays.h"
@@ -865,4 +865,4 @@ namespace msra { namespace lattices {
             emulatemmierrorsignal (thisedgealignments.getalignmentsbuffer(), thisedgealignments.getalignoffsets(), edges, nodes, logpps, errorsignal);
         }
     }
-};};
+};};

Math/Math/CPUMatrix.h

@@ -332,8 +332,8 @@ namespace Microsoft { namespace MSR { namespace CNTK {
         static void ElementWisePower (ElemType alpha, const CPUMatrix<ElemType>& a, CPUMatrix<ElemType>& c);
 
         static bool AreEqual(const CPUMatrix<ElemType>& a, const CPUMatrix<ElemType>& b, const ElemType threshold = 1e-8);
-
-        static void TensorShuffleScaleAndAdd(ElemType keepWeight, const CPUMatrix<ElemType>& a, size_t D, size_t S, size_t M, size_t K, size_t T, ElemType scaleFactor, const CPUMatrix<ElemType>& b, CPUMatrix<ElemType>& c);
+
+        static void TensorShuffleScaleAndAdd(ElemType keepWeight, const CPUMatrix<ElemType>& a, size_t D, size_t S, size_t M, size_t K, size_t T, ElemType scaleFactor, const CPUMatrix<ElemType>& b, CPUMatrix<ElemType>& c);
 
         static CPUMatrix<ElemType> Ones(const size_t rows, const size_t cols);
         static CPUMatrix<ElemType> Zeros(const size_t rows, const size_t cols);