addressed CR feedback; modifed an e2e test to cover inferInputRankToMap

Source/CNTK/BrainScript/CNTKCoreLib/CNTK.core.bs

@@ -32,13 +32,19 @@
 # Note: outDim may describe a tensor as well.
 LinearLayer {outDim, bias = true, init='heNormal', initValueScale=1, inputRank=None, mapRank=None} =
 {
-    inputShape = if BS.Constants.IsNone (inputRank) then Inferred else Repeat (inputRank, Inferred)
+    # inputRank given: number of zeroes to add to W (mapRank must not be given)
+    # mapRank   given: expand W to leave exactly mapRank axes (inputRank must not be given)
+    # none      given: expand W to all (same as mapRank=0)
+    inputShape =
+        if       BS.Constants.IsNone (inputRank) then Inferred  # not given: one Inferred, which will get expanded
+        else if !BS.Constants.IsNone (mapRank)   then Fail ("'inputRank' and 'mapRank' cannot be specified at the same time.")
+        else Repeat (inputRank, Inferred)
     W = ParameterTensor {_ConcatArrays (outDim, inputShape), init=init, initValueScale=initValueScale}
     b = ParameterTensor {outDim, initValue=0}
     outputRank = Length (_AsArray (outDim)) # support outputs with tensor layouts
     inferInputRankToMap =
-        if       BS.Constants.IsNone (mapRank) then -1  # means not specified
-        else if !BS.Constants.IsNone (inputRank) Fail ("'inputRank' and 'mapRank' cannot be specified at the same time.")
+        if      !BS.Constants.IsNone (inputRank) then -1  # means not specified
+        else if  BS.Constants.IsNone (mapRank)   then 0   # default to 'use all input dims'
         else mapRank
     apply (x) =
         if bias

Source/ComputationNetworkLib/LinearAlgebraNodes.h

@@ -238,7 +238,7 @@ class TimesNodeBase : public ComputationNode<ElemType>, public NumInputs<2>
     typedef ComputationNode<ElemType> Base; UsingComputationNodeMembers; using Base::OperationName;                                                                                                                           \
 
 public:
-    TimesNodeBase(DEVICEID_TYPE deviceId, const wstring& name, size_t outputRank = 1, int inferInputRankToMap = 1)
+    TimesNodeBase(DEVICEID_TYPE deviceId, const wstring& name, size_t outputRank = 1, int inferInputRankToMap = -1)
         : Base(deviceId, name), m_outputRank(outputRank), m_inferInputRankToMap(inferInputRankToMap)
     {
     }
@@ -249,8 +249,8 @@ public:
         if (flags & CopyNodeFlags::copyNodeValue)
         {
             auto node = dynamic_pointer_cast<TimesNodeBase<ElemType, m_transpose>>(nodeP);
-            node->m_outputRank      = m_outputRank;
-            node->m_inferInputRankToMap  = m_inferInputRankToMap;
+            node->m_outputRank          = m_outputRank;
+            node->m_inferInputRankToMap = m_inferInputRankToMap;
         }
     }
 
@@ -271,7 +271,7 @@ public:
         if (modelVersion >= CNTK_MODEL_VERSION_11)
             fstream >> m_inferInputRankToMap;
         else
-            m_inferInputRankToMap = 1;
+            m_inferInputRankToMap = -1;
     }
 
 private:
@@ -525,7 +525,7 @@ class TimesNode : public TimesNodeBase<ElemType, false>
     static const std::wstring TypeName() { return L"Times"; }
 
 public:
-    TimesNode(DEVICEID_TYPE deviceId, const wstring& name, size_t outputRank = 1, int inferInputRankToMap = 1)
+    TimesNode(DEVICEID_TYPE deviceId, const wstring& name, size_t outputRank = 1, int inferInputRankToMap = -1)
         : Base(deviceId, name, outputRank, inferInputRankToMap)
     {
     }
@@ -558,7 +558,7 @@ class TransposeTimesNode : public TimesNodeBase<ElemType, true>
 public:
     DeclareConstructorFromConfigWithNumInputs(TransposeTimesNode);
     TransposeTimesNode(DEVICEID_TYPE deviceId, const wstring& name, size_t outputRank = 1)
-        : Base(deviceId, name, outputRank, /*inferInputRankToMap=*/1)
+        : Base(deviceId, name, outputRank, /*inferInputRankToMap=*/-1)
     {
     }
 };

Tests/EndToEndTests/Speech/LSTM/cntk.cntk

@@ -90,8 +90,8 @@ speechTrain = [
                     // parameter macros--these carry their own weight matrices
                     B() = BS.Parameters.BiasParam (cellDim)
 
-                    W(v) = BS.Parameters.WeightParam (cellDim, inputDim)  * Stabilize (v, enabled=enableSelfStabilization) // input-to-hidden
-                    H(h) = BS.Parameters.WeightParam (cellDim, outputDim) * Stabilize (h, enabled=enableSelfStabilization) // hidden-to-hidden
+                    W(v) = BS.Parameters.WeightParam (cellDim, Inferred/*inputDim*/)  * Stabilize (v, enabled=enableSelfStabilization) // input-to-hidden
+                    H(h) = BS.Parameters.WeightParam (cellDim, Inferred/*outputDim*/) * Stabilize (h, enabled=enableSelfStabilization) // hidden-to-hidden
                     C(c) = BS.Parameters.DiagWeightParam (cellDim)       .* Stabilize (c, enabled=enableSelfStabilization) // cell-to-hiddden (note: applied elementwise)
 
                     // note: the W(x) here are all different, they all come with their own set of weights; same for H(dh), C(dc), and B()
@@ -133,10 +133,11 @@ speechTrain = [
                 else /*BS.RNNs.*/ RecurrentLSTMP (hiddenDim, cellDim=innerCellDim, /*S*/ (LSTMoutput[k-1]), inputDim=hiddenDim,   enableSelfStabilization=useSelfStabilization).h
 
             // and add a softmax layer on top
-            W = BS.Parameters.WeightParam (labelDim, hiddenDim)
+            W = BS.Parameters.WeightParam (labelDim, Inferred)
             B = BS.Parameters.BiasParam   (labelDim)
 
-            z = W * S(LSTMoutput[numLSTMLayers]) + B; // top-level input to Softmax
+            // (unnecessarily using explicit Times with inferInputRankToMap in order to have a test for inferInputRankToMap parameter)
+            z = Times (W, S(LSTMoutput[numLSTMLayers]), inferInputRankToMap=0) + B; // top-level input to Softmax
         ].z
 
         // features