marked CRFNode as #ifdef COMING_SOON

Source/CNTK/NetworkDescriptionLanguage.cpp

@@ -155,7 +155,9 @@ bool CheckFunction(std::string& p_nodeType, bool* allowUndeterminedVariable)
     bool ret = false;
     if (EqualInsensitive(nodeType, OperationNameOf(AveragePoolingNode))) ret = true;
     else if (EqualInsensitive(nodeType, OperationNameOf(BatchNormalizationNode))) ret = true;
+#ifdef COMING_SOON
     else if (EqualInsensitive(nodeType, OperationNameOf(CRFNode), L"CRF")) ret = true;
+#endif
     else if (EqualInsensitive(nodeType, OperationNameOf(ClassBasedCrossEntropyWithSoftmaxNode), L"CBCEWithSM")) ret = true;
     else if (EqualInsensitive(nodeType, OperationNameOf(ConvolutionNode), L"Convolve")) ret = true;
     else if (EqualInsensitive(nodeType, OperationNameOf(CosDistanceNode), L"CosDist")) ret = true;

Source/CNTK/SimpleNetworkBuilder.cpp

@@ -1688,10 +1688,12 @@ shared_ptr<ComputationNode<ElemType>> SimpleNetworkBuilder<ElemType>::AddTrainAn
             tinput = builder.Times(matrix, input);
         output = builder.Logistic(label, tinput, (trainNodeName == L"") ? L"Logistic" : trainNodeName);
         break;
+#ifdef COMING_SOON
     case TrainingCriterion::CRF:
         assert(trans != nullptr);
         output = builder.CRF(label, input, trans, (trainNodeName == L"") ? L"CRF" : trainNodeName);
         break;
+#endif
     case TrainingCriterion::ClassCrossEntropyWithSoftmax:
         output = builder.ClassCrossEntropyWithSoftmax(label, input, matrix, clspostprob, (trainNodeName == L"") ? L"ClassCrossEntropyWithSoftmax" : trainNodeName);
         break;
@@ -1743,12 +1745,14 @@ shared_ptr<ComputationNode<ElemType>> SimpleNetworkBuilder<ElemType>::AddTrainAn
                 tinput = builder.Times(matrix, input);
             output = builder.ErrorPrediction(label, tinput, (evalNodeName == L"") ? L"EvalErrorPrediction" : evalNodeName);
             break;
+#ifdef COMING_SOON
         case EvalCriterion::CRF:
             assert(trans != nullptr);
             if (matrix != nullptr && tinput == input)
                 tinput = builder.Times(matrix, input);
             output = builder.CRF(label, tinput, trans, (evalNodeName == L"") ? L"EvalCRF" : evalNodeName);
             break;
+#endif
         default:
             LogicError("Unsupported training criterion.");
         }
@@ -1769,14 +1773,10 @@ template class SimpleNetworkBuilder<double>;
 
 TrainingCriterion ParseTrainingCriterionString(wstring s)
 {
-    if (!_wcsicmp(s.c_str(), L"crossEntropyWithSoftmax"))
-        return TrainingCriterion::CrossEntropyWithSoftmax;
-    if (!_wcsicmp(s.c_str(), L"sequenceWithSoftmax"))
-        return TrainingCriterion::SequenceWithSoftmax;
-    else if (!_wcsicmp(s.c_str(), L"squareError"))
-        return TrainingCriterion::SquareError;
-    else if (!_wcsicmp(s.c_str(), L"logistic"))
-        return TrainingCriterion::Logistic;
+         if (!_wcsicmp(s.c_str(), L"crossEntropyWithSoftmax"))  return TrainingCriterion::CrossEntropyWithSoftmax;
+    else if (!_wcsicmp(s.c_str(), L"sequenceWithSoftmax"))      return TrainingCriterion::SequenceWithSoftmax;
+    else if (!_wcsicmp(s.c_str(), L"squareError"))              return TrainingCriterion::SquareError;
+    else if (!_wcsicmp(s.c_str(), L"logistic"))                 return TrainingCriterion::Logistic;
     else if (!_wcsicmp(s.c_str(), L"noiseContrastiveEstimation") || !_wcsicmp(s.c_str(), L"noiseContrastiveEstimationNode" /*spelling error, deprecated*/))
         return TrainingCriterion::NCECrossEntropyWithSoftmax;
     else if (!!_wcsicmp(s.c_str(), L"classCrossEntropyWithSoftmax")) // (twisted logic to keep compiler happy w.r.t. not returning from LogicError)
@@ -1786,20 +1786,16 @@ TrainingCriterion ParseTrainingCriterionString(wstring s)
 
 EvalCriterion ParseEvalCriterionString(wstring s)
 {
-    if (!_wcsicmp(s.c_str(), L"errorPrediction"))
-        return EvalCriterion::ErrorPrediction;
-    else if (!_wcsicmp(s.c_str(), L"crossEntropyWithSoftmax"))
-        return EvalCriterion::CrossEntropyWithSoftmax;
-    else if (!_wcsicmp(s.c_str(), L"sequenceWithSoftmax"))
-        return EvalCriterion::SequenceWithSoftmax;
-    else if (!_wcsicmp(s.c_str(), L"classCrossEntropyWithSoftmax"))
-        return EvalCriterion::ClassCrossEntropyWithSoftmax;
+         if (!_wcsicmp(s.c_str(), L"errorPrediction"))              return EvalCriterion::ErrorPrediction;
+    else if (!_wcsicmp(s.c_str(), L"crossEntropyWithSoftmax"))      return EvalCriterion::CrossEntropyWithSoftmax;
+    else if (!_wcsicmp(s.c_str(), L"sequenceWithSoftmax"))          return EvalCriterion::SequenceWithSoftmax; 
+    else if (!_wcsicmp(s.c_str(), L"classCrossEntropyWithSoftmax")) return EvalCriterion::ClassCrossEntropyWithSoftmax;
+    else if (!_wcsicmp(s.c_str(), L"logistic"))                     return EvalCriterion::Logistic;
     else if (!_wcsicmp(s.c_str(), L"noiseContrastiveEstimation") || !_wcsicmp(s.c_str(), L"noiseContrastiveEstimationNode" /*spelling error, deprecated*/))
         return EvalCriterion::NCECrossEntropyWithSoftmax;
-    else if (!_wcsicmp(s.c_str(), L"logistic"))
-        return EvalCriterion::Logistic;
     else if (!!_wcsicmp(s.c_str(), L"squareError"))
         LogicError("evalCriterion: Invalid trainingCriterion value. Valid values are (errorPrediction | crossEntropyWithSoftmax | squareError | logistic | sequenceWithSoftmax)");
     return EvalCriterion::SquareError;
 }
+
 } } }

Source/Common/Include/Sequences.h

@@ -464,10 +464,6 @@ typedef MBLayout::MBLayoutPtr MBLayoutPtr;
 // TODO: This will in the future be able to hold sub-ranges for nested loops as well.
 // -----------------------------------------------------------------------
 
-// TODO: We should also have a FrameRange that selects all frames of a single sequence. Currently now possible since that would require Matrix::RowSlice()
-//  - likewise, LSTMNode does its own iteration, hence needs access to GetNumParallelSequences() or NumCols() in the whole-batch iterator
-// BUGBUG: These nodes are currently broken and will need to be fixed:
-//  - CRFNode does not support > 1 parallel sequence
 class FrameRange
 {
 public:                       // TODO: make private (currently used from masking and DataFor) ; TODO: rename all members with m_ prefix

Source/ComputationNetworkLib/ComputationNetwork.cpp

@@ -418,7 +418,9 @@ bool ComputationNetwork::IsTypicalCriterionNode(ComputationNodeBasePtr nodePtr)
         nodePtr->OperationName() == OperationNameOf(CrossEntropyNode) ||
         nodePtr->OperationName() == OperationNameOf(ClassBasedCrossEntropyWithSoftmaxNode) ||
         nodePtr->OperationName() == OperationNameOf(ErrorPredictionNode) ||
+#ifdef COMING_SOON
         nodePtr->OperationName() == OperationNameOf(CRFNode) ||
+#endif
         nodePtr->OperationName() == OperationNameOf(DummyCriterionNode))
         return true;
 

Source/ComputationNetworkLib/ComputationNetwork.h

@@ -981,10 +981,5 @@ typedef ComputationNetwork::ComputationNetworkPtr ComputationNetworkPtr;
 //  - code prettification:
 //     - sort all node implementations' methods into the same order; esp, ForwardProp() comes before partial
 //     - sort important nodes first; move unused/experimental nodes into source files named accordingly
-//  - finish the job:
-//     - everywhere complete folding ForwardPropS() into ForwardProp(FrameRange()), same for partial
-//     - revise node constructors, merge by means of default parameters
-//  - known issues that need actual test cases to be fixed:
-//     - CRFNode::BackpropTo() fails for >1 parallel sequence due to DataFor() not being able to return whole sequences
-//     - implement reading of MB Layout in Binary, DSSM, and LivbSVM readers    --is DSSM already done?
+
 } } }

Source/ComputationNetworkLib/ComputationNetworkBuilder.cpp

@@ -33,8 +33,11 @@ template <class ElemType, class... _Types>
 static shared_ptr<ComputationNode<ElemType>> CreateStandardNode(const std::wstring& nodeType, _Types&&... _Args)
 {
     // please keep this table sorted
-    if      (nodeType == OperationNameOf(CRFNode))                              return New<CRFNode<ElemType>>(forward<_Types>(_Args)...);
-    else if (nodeType == OperationNameOf(ClassBasedCrossEntropyWithSoftmaxNode))return New<ClassBasedCrossEntropyWithSoftmaxNode<ElemType>>(forward<_Types>(_Args)...);
+#ifdef COMING_SOON
+         if (nodeType == OperationNameOf(CRFNode))                              return New<CRFNode<ElemType>>(forward<_Types>(_Args)...);
+    else
+#endif
+         if (nodeType == OperationNameOf(ClassBasedCrossEntropyWithSoftmaxNode))return New<ClassBasedCrossEntropyWithSoftmaxNode<ElemType>>(forward<_Types>(_Args)...);
     else if (nodeType == OperationNameOf(CosDistanceNode))                      return New<CosDistanceNode<ElemType>>(forward<_Types>(_Args)...);
     else if (nodeType == OperationNameOf(CosDistanceWithNegativeSamplesNode))   return New<CosDistanceWithNegativeSamplesNode<ElemType>>(forward<_Types>(_Args)...);
     else if (nodeType == OperationNameOf(CosineNode))                           return New<CosineNode<ElemType>>(forward<_Types>(_Args)...);
@@ -346,6 +349,7 @@ shared_ptr<ComputationNode<ElemType>> ComputationNetworkBuilder<ElemType>::Class
     return net.AddNodeToNetAndAttachInputs(New<ClassBasedCrossEntropyWithSoftmaxNode<ElemType>>(net.GetDeviceId(), nodeName), label, prediction, input_weight, cls_log_post_prob);
 }
 
+#ifdef COMING_SOON
 template <class ElemType>
 shared_ptr<ComputationNode<ElemType>> ComputationNetworkBuilder<ElemType>::CRF(const ComputationNodePtr label,
                                                                                const ComputationNodePtr postDepScore,
@@ -354,6 +358,7 @@ shared_ptr<ComputationNode<ElemType>> ComputationNetworkBuilder<ElemType>::CRF(c
 {
     return net.AddNodeToNetAndAttachInputs(New<CRFNode<ElemType>>(net.GetDeviceId(), nodeName), label, postDepScore, transition_score);
 }
+#endif
 
 template <class ElemType>
 shared_ptr<ComputationNode<ElemType>> ComputationNetworkBuilder<ElemType>::DummyCriterion(const ComputationNodePtr objectives, const ComputationNodePtr derivatives, const ComputationNodePtr prediction, const std::wstring nodeName)
@@ -605,4 +610,5 @@ shared_ptr<ComputationNode<ElemType>> ComputationNetworkBuilder<ElemType>::Batch
 
 template class ComputationNetworkBuilder<float>;
 template class ComputationNetworkBuilder<double>;
+
 } } }

Source/ComputationNetworkLib/ComputationNetworkBuilder.h

@@ -73,7 +73,9 @@ class ComputationNetworkBuilder
     ComputationNodePtr AveragePooling(const ComputationNodePtr inputValues,
                                       const size_t windowWidth, const size_t windowHeight, const size_t horizontalSubsample, const size_t verticalSubsample, ImageLayoutKind imageLayoutKind,
                                       const std::wstring nodeName = L"");
+#ifdef COMING_SOON
     ComputationNodePtr CRF(const ComputationNodePtr label, const ComputationNodePtr postDepScore, const ComputationNodePtr transition_score, const std::wstring nodeName = L"");
+#endif
     ComputationNodePtr ClassCrossEntropyWithSoftmax(const ComputationNodePtr label, const ComputationNodePtr prediction, const ComputationNodePtr input_weight, const ComputationNodePtr cls_log_post_prob, const std::wstring nodeName = L"");
     ComputationNodePtr Cos(const ComputationNodePtr a, const std::wstring nodeName = L"");
     ComputationNodePtr CosDistance(const ComputationNodePtr a, const ComputationNodePtr b, const std::wstring nodeName = L"");

Source/ComputationNetworkLib/ComputationNode.h

@@ -1599,7 +1599,7 @@ inline shared_ptr<C> New(_Types&&... _Args)
 
 // =======================================================================
 // ComputationNodeNonLooping -- abstract base class for computation nodes that do not implement eval/partial for individual frames
-// Such as CRFNode, LSTMNode, ParallelNode, SequenceDecoderNode, TimeReverseNode (BatchModeNode), and TransposeNode.
+// Such as CRFNode, SequenceDecoderNode, and training criteria.
 // =======================================================================
 
 // This will provide default implementations for those two functions that will fail at runtime with a meaningful error.

Source/ComputationNetworkLib/TrainingNodes.h

@@ -1016,6 +1016,8 @@ class ClassBasedCrossEntropyWithSoftmaxNode : public ComputationNodeNonLooping /
 template class ClassBasedCrossEntropyWithSoftmaxNode<float>;
 template class ClassBasedCrossEntropyWithSoftmaxNode<double>;
 
+#ifdef COMING_SOON
+
 // -----------------------------------------------------------------------
 // CRFNode (labels, position_dependent_scores, transition_scores)
 //  - labels: output label vector of [0:T-1]
@@ -1315,6 +1317,8 @@ class CRFNode : public ComputationNodeNonLooping /*ComputationNode*/<ElemType>,
     int mEndLbl;
 };
 
+#endif
+
 // -----------------------------------------------------------------------
 // LogisticNode (labels, prediction, weight)
 // calculates: -sum(left * log(right) + (1-left)*log(1-right)) (optionally * weight)