Merge branch 'master' of https://github.com/Microsoft/CNTK into amita…

…ga/parallelizationFix

Source/ComputationNetworkLib/PreComputeNodes.h

@@ -429,6 +429,47 @@ class PerDimMeanVarNormalizationNode : public ComputationNode<ElemType>, public
         Input(1)->ValidateInferInputDimsFrom(Input(0)->GetSampleLayout());
         Input(2)->ValidateInferInputDimsFrom(Input(0)->GetSampleLayout());
 
+
+#if 1
+        // support for legacy models when the mean and variance vectors were stored as column vectors (N,1)
+        // This code will copy the shape of Input(0) (source) to Input(1) and Input(2) (target) if:
+        //   1. The source is a 3-tensor with shape 1x1xM
+        //   2. The target is a vector (i.e., a 2-tensor with shape Nx1)
+        //   3. Both targets have the same number of elements
+        //   4. The number of elements in the target (N) is the same as the number of elements in the source (M)
+        // Note: This is somewhat ugly [Jasha Droppo].
+
+        auto dimsA = Input(0)->GetSampleLayout().GetDims();
+        auto dimsB = Input(1)->GetSampleLayout().GetDims();
+        auto dimsC = Input(2)->GetSampleLayout().GetDims();
+
+        if (
+            // Test condition 1.
+            (dimsA.size() == 3 && dimsA[0] == 1 && dimsA[1] == 1) &&
+            // Test condition 2.
+            (dimsB.size() == 2 && dimsB[1] == 1) &&
+            (dimsC.size() == 2 && dimsC[1] == 1) &&
+            // Test condition 3. and condition 4.
+            (dimsB[0] == dimsC[0] && dimsB[0] == dimsA[2])
+            )
+        {
+            // for error messages
+            string dimsBstring = string(Input(1)->GetSampleLayout());
+            string dimsCstring = string(Input(2)->GetSampleLayout());
+
+            // reshape Input(1)
+            Input(1)->SetDims(TensorShape(dimsA), false);
+            fprintf(stderr, "\n%ls %ls operation: For legacy compatibility, the sample layout of second input (%ls %ls operation) was patched to [%s] (from [%s])\n",
+                NodeName().c_str(), OperationName().c_str(), Input(1)->NodeName().c_str(), Input(1)->OperationName().c_str(), string(Input(1)->GetSampleLayout()).c_str(), dimsBstring.c_str());
+
+            // reshape Input(2)
+            Input(2)->SetDims(TensorShape(dimsA), false);
+            fprintf(stderr, "\n%ls %ls operation: For legacy compatibility, the sample layout of third input (%ls %ls operation) was patched to [%s] (from [%s])\n",
+                NodeName().c_str(), OperationName().c_str(), Input(2)->NodeName().c_str(), Input(2)->OperationName().c_str(), string(Input(2)->GetSampleLayout()).c_str(), dimsCstring.c_str());
+        }
+
+#endif
+
         if (isFinalValidationPass)
         {
             if (!Input(0)->GetSampleLayout().IsElementwiseCompatibleWith(Input(1)->GetSampleLayout()) || !Input(0)->GetSampleLayout().IsElementwiseCompatibleWith(Input(2)->GetSampleLayout()))

Source/ComputationNetworkLib/SpecialPurposeNodes.h

@@ -701,8 +701,8 @@ class DummyCriterionNode : public ComputationNodeNonLooping /*ComputationNode*/<
     virtual void /*ComputationNodeNonLooping::*/ ForwardPropNonLooping() override
     {
         Value().VerifySize(1, 1);
-        Input(0)->Value().VerifySize(1, 1);
-        Value().SetValue(Input(0)->Value());
+        assert(Input(0)->Value().GetNumRows() == 1);
+        Value().SetValue(Input(0)->Value().SumOfElements());
 #if NANCHECK
         Value().HasNan("DummyCriterionNode");
 #endif
@@ -715,15 +715,17 @@ class DummyCriterionNode : public ComputationNodeNonLooping /*ComputationNode*/<
 
         if (Input(0)->OperationName() != L"InputValue")
             LogicError("DummyCriterionNode criterion requires the first input to be computed objectives.");
-        if (Input(0)->OperationName() != L"InputValue")
-            LogicError("DummyCriterionNode criterion requires the first input to be computed derivatives.");
+        if (Input(1)->OperationName() != L"InputValue")
+            LogicError("DummyCriterionNode criterion requires the second input to be computed derivatives.");
         if (isFinalValidationPass)
         {
-            if (Input(0)->GetSampleMatrixNumRows() != 1)
-                LogicError("DummyCriterionNode criterion requires the first input to have dimension 1.");
-            if (Input(0)->GetSampleMatrixNumRows() == 0 || Input(1)->GetSampleMatrixNumRows() == 0 || Input(2)->GetSampleMatrixNumRows() == 0)
+            if (Input(0)->GetSampleMatrixNumRows() == 0
+                || Input(1)->GetSampleMatrixNumRows() == 0
+                || Input(2)->GetSampleMatrixNumRows() == 0)
                 LogicError("DummyCriterionNode operation: one of the operands has 0 elements.");
-            if (Input(1)->GetSampleMatrixNumRows() != Input(2)->GetSampleMatrixNumRows())
+            if (Input(1)->GetSampleMatrixNumRows() != Input(2)->GetSampleMatrixNumRows()
+                || Input(0)->GetSampleMatrixNumCols() != Input(2)->GetSampleMatrixNumCols()
+                || Input(1)->GetSampleMatrixNumCols() != Input(2)->GetSampleMatrixNumCols())
                 LogicError("The Matrix dimension in the DummyCriterionNode operation does not match.");
         }
 

Source/Math/GPUTensor.cu

@@ -655,9 +655,10 @@ void LaunchUnaryTensorOp(ElemType beta, const ElemType* pa, ElemType* pb, ElemTy
 #define CaseLaunchUnaryTensorOp(oper)                                                                                                        \
     case ElementWiseOperator::op##oper:                                                                                                      \
         if (beta == 0 && alpha == 1)                                                                                                         \
-            return _launchUnaryTensorOp<ElemType, Functor##oper><<<grid.m_blocksPerGrid, grid.m_threadsPerBlock, 0, t_stream>>>(pa, pb, NN); \
+            _launchUnaryTensorOp<ElemType, Functor##oper><<<grid.m_blocksPerGrid, grid.m_threadsPerBlock, 0, t_stream>>>(pa, pb, NN); \
         else                                                                                                                                 \
-            return _launchUnaryTensorOp<ElemType, Functor##oper><<<grid.m_blocksPerGrid, grid.m_threadsPerBlock, 0, t_stream>>>(beta, pa, pb, alpha, NN);
+            _launchUnaryTensorOp<ElemType, Functor##oper><<<grid.m_blocksPerGrid, grid.m_threadsPerBlock, 0, t_stream>>>(beta, pa, pb, alpha, NN);\
+        break;
 
     SyncGuard syncGuard;
     GridDim grid(NN);

Source/Readers/Kaldi2Reader/HTKMLFReader.cpp

@@ -1405,10 +1405,10 @@ void HTKMLFReader<ElemType>::CopyMinibatchFromBufferToMatrix(
             {
                 if (m_getMinibatchCopy)
                 {
-                    if (data.GetNumCols() != m_currentMBSize * m_numberOfuttsPerMinibatch)
+                    assert(m_currentMBSize * m_numberOfuttsPerMinibatch == m_pMBLayout->GetNumCols());
+                    if (data.GetNumCols() != m_pMBLayout->GetNumCols())
                     {
-                        matrices.GetInputMatrix<ElemType>(iter->first).Resize(data.GetNumRows(),
-                                                                    m_currentMBSize * m_numberOfuttsPerMinibatch);
+                        data.Resize(data.GetNumRows(), m_pMBLayout->GetNumCols());
                     }
                     matrices.GetInputMatrix<ElemType>(iter->first).SetValue(0);
                 }
@@ -1423,15 +1423,17 @@ void HTKMLFReader<ElemType>::CopyMinibatchFromBufferToMatrix(
             {
                 if (m_getMinibatchCopy)
                 {
-                    if (data.GetNumCols() != 1)
+                    assert(m_currentMBSize * m_numberOfuttsPerMinibatch == m_pMBLayout->GetNumCols());
+                    if (data.GetNumCols() != m_pMBLayout->GetNumCols())
                     {
-                        data.Resize(1, 1);
+                        data.Resize(1, m_pMBLayout->GetNumCols());
                     }
                     data.SetValue(0);
                 }
                 else
                 {
                     m_uttDerivBuffer->GetObjective(m_minibatchUttInfo,
+                                                   m_pMBLayout,
                                                    &matrices.GetInputMatrix<ElemType>(iter->first)); // TODO: use a reference instead of a ptr
                 }
             }
@@ -1473,18 +1475,19 @@ void HTKMLFReader<ElemType>::CopyMinibatchToMatrix(
         {
             if (m_nameToTypeMap.at(iter->first) == InputOutputTypes::readerDeriv)
             {
-                if (data.GetNumCols() != m_currentMBSize * m_numberOfuttsPerMinibatch)
+                assert(m_currentMBSize * m_numberOfuttsPerMinibatch == m_pMBLayout->GetNumCols());
+                if (data.GetNumCols() != m_pMBLayout->GetNumCols())
                 {
-                    data.Resize(data.GetNumRows(),
-                                m_currentMBSize * m_numberOfuttsPerMinibatch);
+                    data.Resize(data.GetNumRows(), m_pMBLayout->GetNumCols());
                 }
                 data.SetValue(0);
             }
             else if (m_nameToTypeMap.at(iter->first) == InputOutputTypes::readerObj)
             {
-                if (data.GetNumCols() != 1)
+                assert(m_currentMBSize * m_numberOfuttsPerMinibatch == m_pMBLayout->GetNumCols());
+                if (data.GetNumCols() != m_pMBLayout->GetNumCols())
                 {
-                    data.Resize(1, 1);
+                    data.Resize(1, m_pMBLayout->GetNumCols());
                 }
                 data.SetValue(0);
             }
@@ -1589,8 +1592,7 @@ bool HTKMLFReader<ElemType>::GetMinibatchToWrite(StreamMinibatchInputs& matrices
 
         // populate input matrices
         bool first = true;
-        typename StreamMinibatchInputs::iterator iter;
-        for (iter = matrices.begin(); iter != matrices.end(); iter++)
+        for (auto iter = matrices.begin(); iter != matrices.end(); iter++)
         {
             // dereference matrix that corresponds to key (input/output name) and
             // populate based on whether its a feature or a label
@@ -1907,7 +1909,7 @@ bool HTKMLFReader<ElemType>::SetNetOutput(
     const MatrixBase& outputsb,
     const MBLayoutPtr pMBLayout)
 {
-    const auto& outputs = dymamic_cast<Matrix<ElemType>>(outputsb); // TODO: a NULL check, to be sure
+    const auto& outputs = dynamic_cast<const Matrix<ElemType>&>(outputsb); // TODO: a NULL check, to be sure
     // Set the likelihoods for the utterance with which we can comput the
     // derivatives. Note that the minibatch may only contain partial output
     // for the utterance, <m_uttDerivBuffer> takes care of "gluing" them

Source/Readers/Kaldi2Reader/HTKMLFWriter.h

@@ -42,8 +42,6 @@ class HTKMLFWriter : public IDataWriter
     };
 
 public:
-    using LabelType = typename IDataWriter<ElemType>::LabelType;
-    using LabelIdType = typename IDataWriter<ElemType>::LabelIdType;
     template <class ConfigRecordType>
     void InitFromConfig(const ConfigRecordType& writerConfig);
     virtual void Init(const ConfigParameters& config)

Source/Readers/Kaldi2Reader/UtteranceDerivativeBuffer.cpp

@@ -91,7 +91,7 @@ bool UtteranceDerivativeBuffer<ElemType>::SetLikelihood(
     ProcessUttInfo(uttInfo, pMBLayout, &uttInfoInMinibatch);
 
     // Checks if we need to move data to CPU.
-    Matrix<ElemType> logLikelihood(logLikelihoodIn);
+    Matrix<ElemType> logLikelihood = logLikelihoodIn.DeepClone();
     if (logLikelihood.GetDeviceId() >= 0)
     {
         logLikelihood.TransferFromDeviceToDevice(
@@ -114,7 +114,7 @@ bool UtteranceDerivativeBuffer<ElemType>::SetLikelihood(
                 tmpUttUnit.streamID = i;
                 tmpUttUnit.logLikelihood.Resize(logLikelihood.GetNumRows(),
                                                 tmpUttUnit.uttLength);
-                m_uttPool[uttID] = tmpUttUnit;
+                m_uttPool[uttID] = std::move(tmpUttUnit);
             }
 
             // Sets the likelihood and computes derivatives.
@@ -244,6 +244,7 @@ bool UtteranceDerivativeBuffer<ElemType>::GetDerivative(
 template <class ElemType>
 bool UtteranceDerivativeBuffer<ElemType>::GetObjective(
     const std::vector<std::vector<std::pair<wstring, size_t>>>& uttInfo,
+    MBLayoutPtr pMBLayout,
     Matrix<ElemType>* objectivesIn)
 {
     assert(objectivesIn != NULL);
@@ -258,8 +259,9 @@ bool UtteranceDerivativeBuffer<ElemType>::GetObjective(
     }
 
     // Sets the objectives...
-    objectivesIn->Resize(1, 1);
-    objectivesIn->SetValue(m_currentObj);
+    objectivesIn->Resize(1, pMBLayout->GetNumCols());
+    objectivesIn->SetValue(
+        m_currentObj / static_cast<ElemType>(pMBLayout->GetNumCols()));
 
     return true;
 }

Source/Readers/Kaldi2Reader/UtteranceDerivativeBuffer.h

@@ -32,6 +32,7 @@ class UtteranceDerivativeBuffer
             progress = 0;
             streamID = 0;
         }
+
     };
 
     bool m_needLikelihood;
@@ -87,6 +88,7 @@ class UtteranceDerivativeBuffer
     // Gets the computed objectives for given utterance.
     bool GetObjective(
         const std::vector<std::vector<std::pair<wstring, size_t>>>& uttInfo,
+        const MBLayoutPtr pMBLayout,
         Matrix<ElemType>* objectivesIn);
 
     bool HasResourceForDerivative(const wstring& uttID) const;