updated beam decoder to new structure

CNTK.sln

@@ -666,6 +666,7 @@ Project("{2150E333-8FDC-42A3-9474-1A3956D46DE8}") = "Config", "Config", "{850008
 	ProjectSection(SolutionItems) = preProject
 		Examples\Text\PennTreebank\Config\rnn.cntk = Examples\Text\PennTreebank\Config\rnn.cntk
 		Examples\Text\PennTreebank\Config\S2SAutoEncoder.cntk = Examples\Text\PennTreebank\Config\S2SAutoEncoder.cntk
+		Examples\Text\PennTreebank\Config\S2SLib.bs = Examples\Text\PennTreebank\Config\S2SLib.bs
 	EndProjectSection
 EndProject
 Project("{2150E333-8FDC-42A3-9474-1A3956D46DE8}") = "SLU", "SLU", "{E6DC3B7D-303D-4A54-B040-D8DCF8C56E17}"

Examples/Text/PennTreebank/Config/S2SAutoEncoder.cntk

@@ -18,7 +18,8 @@ ExpRootDir = "$RunRootDir$"
 #ExpId = _run
 
 deviceId = 0
-ExpId = 68-$deviceId$-s2sae-bigmodel
+#ExpId = 68-$deviceId$-s2sae-bigmodel
+ExpId = 67-1-s2sae-bigmodel # for decoding a different model
 
 # directories
 ExpDir    = "$ExpRootDir$/$ExpId$"
@@ -42,7 +43,7 @@ command = writeWordAndClassInfo:train:test:write
 precision  = "float"
 traceLevel = 1
 modelPath  = "$ModelDir$/S2SAutoEncoder.dnn"
-decodeModelPath = "$modelPath$.13" # epoch to decode. Has best CV WER
+decodeModelPath = "$modelPath$" # epoch to decode. Has best CV WER
 
 confVocabSize = 10000
 confClassSize = 50
@@ -55,6 +56,7 @@ validFile = "ptb.valid.txt"
 #validFile = "ptb.small.valid.txt"
 testFile  = "ptb.test.txt"
 #testFile  = "ptb.test.txt-econ1"
+#testFile = "ptb.small.train.txt" # test on train, to see whether model makes sense at all
 
 #######################################
 #  network definition                 #
@@ -63,43 +65,10 @@ testFile  = "ptb.test.txt"
 BrainScriptNetworkBuilder = (new ComputationNetwork [
 
 # TODO: move this somewhere shared
-    Trace (node, say='', logFrequency=traceFrequency, logFirst=10, logGradientToo=false, onlyUpToRow=100000000, onlyUpToT=100000000, format=[], tag='') = new ComputationNode [
-        operation = 'Trace' ; inputs = node
-    ]
-
-    formatDense = [
-        type = "real"
-        transpose = false
-        precisionFormat = ".4"
-    ]
-    formatOneHot = [
-        type = "category"
-        transpose = false
-        labelMappingFile = "$ModelDir$/vocab.wl"
-    ]
-    formatSparse = [
-        type = "sparse"
-        transpose = false
-        labelMappingFile = "$ModelDir$/vocab.wl"
-    ]
     enableTracing = true
     traceFrequency = 1000
-    TraceState (h, what) =
-        if enableTracing
-        then Transpose (Trace (Transpose (h), say=what, logFirst=10, logFrequency=traceFrequency, logGradientToo=false, onlyUpToRow=beamDepth*beamDepth, onlyUpToT=3, format=formatDense))
-        else h
-    TraceDense (h, what) =
-        if enableTracing
-        then Trace (h, say=what, logFirst=10, logFrequency=traceFrequency, logGradientToo=false, onlyUpToRow=beamDepth*beamDepth, onlyUpToT=25, format=formatDense)
-        else h
-    TraceOneHot (h, what) =
-        if enableTracing
-        then Trace (h, say=what, logFirst=10, logFrequency=traceFrequency, logGradientToo=false, /*onlyUpToRow=beamDepth*beamDepth, onlyUpToT=15,*/ format=formatOneHot)
-        else h
-    TraceSparse (h, what) =
-        if enableTracing
-        then Trace (h, say=what, logFirst=10, logFrequency=traceFrequency, logGradientToo=false, /*onlyUpToRow=beamDepth*beamDepth, onlyUpToT=3,*/ format=formatSparse)
-        else h
+    tracingLabelMappingFile = "$ModelDir$/vocab.wl"
+    include "S2SLib.bs"
     beamDepth=3 // for above Trace macros only
 
     # import general config options from outside config values
@@ -537,6 +506,11 @@ write = [
     # We need to make a change:
     BrainScriptNetworkBuilder = ([
 
+        enableTracing = true
+        traceFrequency = 1000
+        tracingLabelMappingFile = "$ModelDir$/vocab.wl"
+        include "S2SLib.bs"
+
         beamDepth = 3 // 0=predict; 1=greedy; >1=beam
 
         # import some names
@@ -575,8 +549,7 @@ write = [
             logP = LogSoftmax (model.z)
 
             offset = Constant (10000)
-            top1a = Hardmax (logP)  .* (logP + offset)/*for tracing*/
-            top1b = top1a
+            top1b = Hardmax (logP)  .* (logP + offset)/*for tracing*/
             top1 = TraceSparse (top1b, 'logP') # TODO: get the accumulated logP out, it's a little more involved
 
             topN = 10
@@ -591,10 +564,21 @@ write = [
         ]
 
         # replace old decoderFeedback node by newDecoderFeedback
+        EmbedLabels (x) = TransposeTimes (modelAsTrained.labelsEmbedded.TransposeTimesArgs[0], x)
+        decoderFeedback = EmbedLabels (Hardmax (modelAsTrained.z))  # in training, this is decoderFeedback = labelsEmbedded
 
-        decoderFeedback        = modelAsTrained.decoderOutputEmbedded  # in training, this is decoderFeedback = labelsEmbedded
-        sentenceStartEmbedded  = Boolean.If (Loop.IsFirst (decoderFeedback), modelAsTrained.inputEmbedded, Previous (sentenceStartEmbedded)) # enforces no leaking of labels
-        delayedDecoderFeedback = Boolean.If (Loop.IsFirst (decoderFeedback), sentenceStartEmbedded, Loop.Previous (decoderFeedback)) # same expression as in training
+        # TODO: fold this in
+        PreviousOrDefault1 (x, defaultValue=Constant (0)) =   # a delay node with initial value  --TODO: merge the two, then do in C++
+        [
+            flags = BS.Loop.IsFirst (defaultValue/*x*/)
+            out = BS.Boolean.If (flags,
+                        /*then*/ defaultValue,
+                        /*else*/ Previous (x))
+        ].out
+
+        labelSentenceStartEmbeddedScattered = TraceDense (BS.Sequences.Scatter (BS.Loop.IsFirst (modelAsTrained.labelsEmbedded), modelAsTrained.labelSentenceStartEmbedded), 'sest')
+
+        delayedDecoderFeedback = TraceDense (/*Loop.*/PreviousOrDefault1 (defaultValue=labelSentenceStartEmbeddedScattered, TraceDense (decoderFeedback, 'lemb'))   , 'prev lemb')
 
         greedyDecodingModel = BS.Network.Edit (modelAsTrained,
                                                BS.Network.Editing.ReplaceLinksToNode (modelAsTrained.delayedDecoderFeedback, delayedDecoderFeedback),
@@ -630,17 +614,19 @@ write = [
             #  - traceback is a right-to-left recurrence
             #     - output best hypo conditioned on the path (it is already known)
 
-            propagationEdits[i:0..2] = // TODO: implement and use { } syntax
+            propagationEdits[i:0..6] = // TODO: implement and use { } syntax
                 if      i == 0 then (node => if node.name == 'decoder[0].prevState.h' then TraceState (Previous (PropagateTopN (node.PastValueArgs[0])), 'propagated') else node) # inject reshuffling of hypotheses
                 else if i == 1 then (node => if node.name == 'decoder[0].prevState.c' then TraceState (Previous (PropagateTopN (node.PastValueArgs[0])), 'propagated') else node)
+                else if i == 2 then (node => if node.name == 'decoder[1].prevState.h' then TraceState (Previous (PropagateTopN (node.PastValueArgs[0])), 'propagated') else node) # inject reshuffling of hypotheses
+                else if i == 3 then (node => if node.name == 'decoder[1].prevState.c' then TraceState (Previous (PropagateTopN (node.PastValueArgs[0])), 'propagated') else node)
+                else if i == 4 then (node => if node.name == 'decoder[2].prevState.h' then TraceState (Previous (PropagateTopN (node.PastValueArgs[0])), 'propagated') else node) # inject reshuffling of hypotheses
+                else if i == 5 then (node => if node.name == 'decoder[2].prevState.c' then TraceState (Previous (PropagateTopN (node.PastValueArgs[0])), 'propagated') else node)
                 else                BS.Network.Editing.ReplaceLinksToNode (modelAsTrained.delayedDecoderFeedback, delayedDecoderFeedback)
 
             # decoderFeedback must be updated to take actual decoder output
 
-            Elabel = modelAsTrained.decoderOutputEmbedded.TransposeTimesArgs[0]
-            decoderFeedback = TraceState (TransposeTimes (Elabel, TraceSparse (topWords, 'topWords')), 'feedback')
-
-            delayedDecoderFeedback = Boolean.If (Loop.IsFirst (decoderFeedback), sentenceStartEmbedded, Loop.Previous (decoderFeedback))
+            decoderFeedback = TraceState (EmbedLabels (TraceSparse (topWords, 'topWords')), 'feedback')
+            delayedDecoderFeedback = Boolean.If (Loop.IsFirst (labelSentenceStartEmbeddedScattered), labelSentenceStartEmbeddedScattered, Loop.Previous (decoderFeedback))
 
             m2 = BS.Network.Edit (modelAsTrained,
                                   propagationEdits,
@@ -679,7 +665,7 @@ write = [
             LOGZERO = -1e30
             initialPathScores = FirstAndOther (0, LOGZERO, beamDepth, axis = 2)  # row vector: [ 0, -INF, -INF, -INF, ... ]
 
-            expandedPathScores = logLLs + PreviousOrDefault (PropagateTopN (pathScores), initialPathScores)        # [V x Dprev] un-normalized log (P(w|hist) * P(hist)) for all top D hypotheses
+            expandedPathScores = logLLs + Boolean.If (Loop.IsFirst (pathScores), initialPathScores, Previous (PropagateTopN (pathScores)))        # [V x Dprev] un-normalized log (P(w|hist) * P(hist)) for all top D hypotheses
             # ^^ path expansion, [V x 1] + [1 x D] -> [V x D]
 
             tokenSet = TraceSparse (GetTopNTensor (beamDepth, expandedPathScores), 'tokenSet') # [V x Dprev] -> [V x Dprev x Dnew]
@@ -746,7 +732,8 @@ write = [
             # previous states: multiply wth respective backPointers matrix
             # -> hyp index for every time step
             # then finally use that to select the actual output   TODO: That's a sample-wise matrix product between two sequences!!!
-            traceback = TraceDense (NextOrDefault (backPointers * traceback, finalHyp), 'traceback')    # [D] one-hot, multiplying backPointers from the left will select another one-hot row of backPointers
+            # TODO: condition must be 1-dim, not 2-dim tensor, so we use labelSentenceStartEmbeddedScattered instead of backPointers
+            traceback = TraceDense (Boolean.If (Loop.IsLast (labelSentenceStartEmbeddedScattered/*backPointers*/), finalHyp, Loop.Next (backPointers * traceback)), 'traceback')    # [D] one-hot, multiplying backPointers from the left will select another one-hot row of backPointers
             # +-+
             # |0|
             # |1|  means at this time step, hyp[1] was the best globally
@@ -758,7 +745,7 @@ write = [
             # This is the one to output (top sentence-level hypothesis after traceback).
             decode = [
                 hyp = Times (tokenSet, traceback, outputRank = 2)   # [V x Dprev] 2D one-hot
-                out = TraceOneHot (hyp * ConstantTensor (1, beamDepth), 'out')           # reduces over Dprev -> 1D one-hot
+                out = TraceOneHot (hyp * ConstantTensor (1, (beamDepth)), 'out')           # reduces over Dprev -> 1D one-hot
             ].out
             # traceback : [Dnew]
             # tokenSet : [V x Dprev x Dnew]

Examples/Text/PennTreebank/Config/S2SLib.bs

@@ -0,0 +1,39 @@
+# TODO: must sort this out. For now, this is just shared stuff between training and decoding.
+
+    # these depend on beamDepth parameter for now, fix this
+    TraceState (h, what) =
+        if enableTracing
+        then Transpose (Trace (Transpose (h), say=what, logFirst=10, logFrequency=traceFrequency, logGradientToo=false, onlyUpToRow=beamDepth*beamDepth, onlyUpToT=3, format=formatDense))
+        else h
+    TraceDense (h, what) =
+        if enableTracing
+        then Trace (h, say=what, logFirst=10, logFrequency=traceFrequency, logGradientToo=false, onlyUpToRow=beamDepth*beamDepth, onlyUpToT=25, format=formatDense)
+        else h
+    TraceOneHot (h, what) =
+        if enableTracing
+        then Trace (h, say=what, logFirst=10, logFrequency=traceFrequency, logGradientToo=false, /*onlyUpToRow=beamDepth*beamDepth, onlyUpToT=15,*/ format=formatOneHot)
+        else h
+    TraceSparse (h, what) =
+        if enableTracing
+        then Trace (h, say=what, logFirst=10, logFrequency=traceFrequency, logGradientToo=false, /*onlyUpToRow=beamDepth*beamDepth, onlyUpToT=3,*/ format=formatSparse)
+        else h
+
+    Trace (node, say='', logFrequency=traceFrequency, logFirst=10, logGradientToo=false, onlyUpToRow=100000000, onlyUpToT=100000000, format=[], tag='') = new ComputationNode [
+        operation = 'Trace' ; inputs = node
+    ]
+
+    formatDense = [
+        type = "real"
+        transpose = false
+        precisionFormat = ".4"
+    ]
+    formatOneHot = [
+        type = "category"
+        transpose = false
+        labelMappingFile = tracingLabelMappingFile
+    ]
+    formatSparse = [
+        type = "sparse"
+        transpose = false
+        labelMappingFile = tracingLabelMappingFile
+    ]

Source/ComputationNetworkLib/ComputationNetworkEvaluation.cpp

@@ -192,8 +192,9 @@ ComputationNetwork::PARTraversalFlowControlNode::PARTraversalFlowControlNode(con
     for (auto& node : m_nestedNodes)
     {
         if (node->GetMBLayout() != GetMBLayout())
-            LogicError("Evaluate: all nodes inside a recurrent loop must have a layout that is identical; mismatch found for nodes '%ls' vs. '%ls'",
-                       node->NodeName().c_str(), m_nestedNodes[0]->NodeName().c_str());
+            LogicError("Evaluate: All nodes inside a recurrent loop must have a layout that is identical; mismatch found for nodes '%ls' (%ls) vs. '%ls' (%ls)",
+                       node            ->NodeName().c_str(), node            ->GetMBLayoutAxisString().c_str(),
+                       m_nestedNodes[0]->NodeName().c_str(), m_nestedNodes[0]->GetMBLayoutAxisString().c_str());
     }
 
     // tell all that loop is about to commence
@@ -672,7 +673,7 @@ size_t ComputationNetwork::ValidateNodes(list<ComputationNodeBasePtr> nodes, boo
             {
                 unchanged = !ValidateNode(node, isFinalValidationPass);
                 string updatedPrototype = node->FormatOperationPrototype("");
-#if 1           // print prototype in final validation pass
+#if 0           // print prototype in final validation pass. Problematic for tracking down validation errors in loops.
                 unchanged;
                 if (isFinalValidationPass)
 #else           // print prototype upon every change (useful for debugging)