Merge branch 'fseide/reshaping' of https://git01.codeplex.com/cntk into fseide/reshaping

DataReader/HTKMLFReader/HTKMLFReader.cpp

@@ -1010,6 +1010,8 @@ namespace Microsoft { namespace MSR { namespace CNTK {
                             {
                                 m_pMBLayout->SetAsNoInput(i, m_numValidFrames[i], m_mbNumTimeSteps);
                             }
+
+                            // TODO: Also blast the gaps in the features and labels matrices with NaNs to prevent them from being read
                         }
 
                         typename std::map<std::wstring, Matrix<ElemType>*>::iterator iter;
@@ -1192,14 +1194,20 @@ namespace Microsoft { namespace MSR { namespace CNTK {
                             }
                             m_processedFrame[i] += (endFr-startFr);
                             m_switchFrame[i] = actualmbsize[i];
-                            if (actualmbsize[i] < m_mbNumTimeSteps)
+                            if (actualmbsize[i] != 0)
-                                m_pMBLayout->Set(i, actualmbsize[i], MinibatchPackingFlags::SequenceStart); // NOTE: this ORs, while original code overwrote in matrix but ORed into vector
-                            if (actualmbsize[i] == m_mbNumTimeSteps)
                                 m_pMBLayout->Set(i, actualmbsize[i] - 1, MinibatchPackingFlags::SequenceEnd); // NOTE: this ORs, while original code overwrote in matrix but ORed into vector
+
+                            // TODO: We should fill in a loop until we fill the minibatch for the case where just one ReNew is not sufficient
+                            // to fill up the remaining slots in the minibatch
+                            bool reNewSucc = ReNewBufferForMultiIO(i);
+                            if (actualmbsize[i] < m_mbNumTimeSteps)
+                            {
+                                if (reNewSucc)
+                                {
+                                    m_pMBLayout->Set(i, actualmbsize[i], MinibatchPackingFlags::SequenceStart); // NOTE: this ORs, while original code overwrote in matrix but ORed into vector
                                     startFr = m_switchFrame[i];
                                     endFr = m_mbNumTimeSteps;
-                            bool reNewSucc = ReNewBufferForMultiIO(i);
+                                    for (iter = matrices.begin(); iter != matrices.end(); iter++)
-                            for (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
@@ -1211,7 +1219,7 @@ namespace Microsoft { namespace MSR { namespace CNTK {
                                             dim = m_featureNameToDimMap[iter->first];
                                             if (sizeof(ElemType) == sizeof(float))
                                             {
-                                        for (size_t j = startFr,k = 0; j < endFr; j++,k++) // column major, so iterate columns
+                                                for (size_t j = startFr, k = 0; j < endFr; j++, k++) // column major, so iterate columns
                                                 {
                                                     // copy over the entire column at once, need to do this because SSEMatrix may have gaps at the end of the columns
                                                     memcpy_s(&m_featuresBufferMultiIO[id].get()[(j * m_numSeqsPerMB + i) * dim], sizeof(ElemType) * dim, &m_featuresBufferMultiUtt[i].get()[k * dim + m_featuresStartIndexMultiUtt[id + i * numOfFea]], sizeof(ElemType) * dim);
@@ -1219,7 +1227,7 @@ namespace Microsoft { namespace MSR { namespace CNTK {
                                             }
                                             else
                                             {
-                                        for (size_t j=startFr,k=0; j < endFr; j++,k++) // column major, so iterate columns in outside loop
+                                                for (size_t j = startFr, k = 0; j < endFr; j++, k++) // column major, so iterate columns in outside loop
                                                 {
                                                     for (int d = 0; d < dim; d++)
                                                         m_featuresBufferMultiIO[id].get()[(j * m_numSeqsPerMB + i) * dim + d] = m_featuresBufferMultiUtt[i].get()[k * dim + d + m_featuresStartIndexMultiUtt[id + i * numOfFea]];
@@ -1230,7 +1238,7 @@ namespace Microsoft { namespace MSR { namespace CNTK {
                                         {
                                             id = m_labelNameToIdMap[iter->first];
                                             dim = m_labelNameToDimMap[iter->first];
-                                    for (size_t j = startFr,k=0; j < endFr; j++,k++)
+                                            for (size_t j = startFr, k = 0; j < endFr; j++, k++)
                                             {
                                                 for (int d = 0; d < dim; d++)
                                                     m_labelsBufferMultiIO[id].get()[(j * m_numSeqsPerMB + i) * dim + d] = m_labelsBufferMultiUtt[i].get()[k * dim + d + m_labelsStartIndexMultiUtt[id + i * numOfLabel]];
@@ -1238,8 +1246,16 @@ namespace Microsoft { namespace MSR { namespace CNTK {
                                         }
                                     }
 
-                            if (reNewSucc) m_processedFrame[i] += (endFr-startFr);
+                                    m_processedFrame[i] += (endFr - startFr);
+                                }
+                                else
+                                {
+                                    // Mark gaps with NoInput
+                                    m_pMBLayout->SetAsNoInput(i, actualmbsize[i], m_mbNumTimeSteps);
 
+                                    // TODO: Also blast the gaps in the features and labels matrices with NaNs to prevent them from being read
+                                }
+                            }
                         }
                     }
                     for (auto iter = matrices.begin();iter!=matrices.end(); iter++)

MachineLearning/CNTKSGDLib/SGD.cpp

@@ -1824,11 +1824,11 @@ namespace Microsoft { namespace MSR { namespace CNTK {
         if (useDistributedMBReading)
         {
             trainSetDataReader->StartDistributedMinibatchLoop(tunedMBSize, epochNumber, g_mpi->CurrentNodeRank(),
-                                                              g_mpi->NumNodesInUse(), m_epochSize);
+                                                              g_mpi->NumNodesInUse(), epochSize);
         }
         else
         {
-            trainSetDataReader->StartMinibatchLoop(tunedMBSize, epochNumber, m_epochSize);
+            trainSetDataReader->StartMinibatchLoop(tunedMBSize, epochNumber, epochSize);
         }
 
         net.StartEvaluateMinibatchLoop(evaluationNodes);
@@ -2160,9 +2160,6 @@ namespace Microsoft { namespace MSR { namespace CNTK {
             totalEpochSamples += aggregateNumSamplesWithLabel;
             totalSamplesSeen += aggregateNumSamplesWithLabel;
 
-            if (totalEpochSamples >= epochSize)
-                break;
-
             // call DataEnd function
             // This signals something from SGD to the reader.
             // DataEnd does reader specific process if sentence ending is reached