mini checkin

BrunoBozza · Feb 28, 2016 · 570b6a8 · 570b6a8
1 parent d55374d
commit 570b6a8
Show file tree

Hide file tree

Showing 4 changed files with 13 additions and 11 deletions.
diff --git a/Examples/Text/PennTreebank/Config/rnn.cntk b/Examples/Text/PennTreebank/Config/rnn.cntk
@@ -28,7 +28,7 @@ confVocabSize = 10000
 confClassSize = 50
 
 trainFile = "ptb.train.txt"
-validFile = "ptb.valid.txt"
+validFile = "ptb.valid.txt-2"
 testFile  = "ptb.test.txt"
 
 #######################################
@@ -65,7 +65,7 @@ train = [
 
     SimpleNetworkBuilder = [
         trainingCriterion = "classCrossEntropyWithSoftmax"
-        evalCriterion = "classCrossEntropyWithSoftmax"
+        evalCriterion     = "classCrossEntropyWithSoftmax"
         nodeType = "sigmoid"
         initValueScale = 6.0
         layerSizes = "$confVocabSize$:150:200:10000"
@@ -85,8 +85,7 @@ train = [
         momentumPerMB = 0
         gradientClippingWithTruncation = true
         clippingThresholdPerSample = 15.0
-        maxEpochs = 16
-        unroll = false
+        maxEpochs = 16                      # BUGBUG: this seems to be ignored?
         numMBsToShowResult = 100
         gradUpdateType = "none"
         loadBestModel = true
@@ -111,6 +110,8 @@ train = [
         readerType = "LMSequenceReader"
         randomize = "none"              # BUGBUG: This is currently ignored
         nbruttsineachrecurrentiter = 10 # note: real-life configurations would use something much larger
+        #nbruttsineachrecurrentiter = 0  # means fill up the minibatch with as many parallel sequences as fit
+        #cacheBlockSize = 2000000        # just load it all
 
         # word class info
         wordclass = "$ModelDir$/vocab.txt"
@@ -218,6 +219,8 @@ train = [
         # reader to use
         readerType = "LMSequenceReader"
         randomize = "none"
+        nbruttsineachrecurrentiter = 0  # 0 means fill up the minibatch with as many parallel sequences as fit
+        cacheBlockSize = 2000000        # just load it all
 
         # word class info
         wordclass = "$ModelDir$/vocab.txt"
@@ -347,6 +350,7 @@ test = [
         # reader to use
         readerType = "LMSequenceReader"
         randomize = "none"
+        nbruttsineachrecurrentiter = 2 # means fill up the minibatch with as many parallel sequences as fit
 
         # word class info
         wordclass = "$ModelDir$/vocab.txt"

diff --git a/Source/ActionsLib/OtherActions.cpp b/Source/ActionsLib/OtherActions.cpp
@@ -274,7 +274,7 @@ void DoWriteWordAndClassInfo(const ConfigParameters& config)
 #if 1
     std::unordered_map<string, double> v_count;
 #else
-    // for unknown reasons, this gives a very different result (PPL of 500 instead of 190)
+    // TODO: For unknown reasons, this gives a very different result (PPL of 500 instead of 190). Should be tracked down.
     std::map<string, double> v_count;
     v_count[beginSequence] = 0;  // get these into the table upfront into position 0 (and 1 if different)
     v_count[endSequence]   = 0;

diff --git a/Source/Math/CPUMatrix.cpp b/Source/Math/CPUMatrix.cpp
@@ -3603,14 +3603,16 @@ struct PrintRange
             begin = 0;
             skipBegin = (size_t)(-last);
             skipEnd = (size_t)(total + first);
+            if (skipEnd <= skipBegin)
+                skipBegin = skipEnd = total;
             end = total;
         }
         else    // if other combinations are ever of interest then implement them here
             LogicError("Print: Bounds must be either both positive or both negative.");
     }
 };
 
-// use negative ranges to print corners, e.g. specify first=-3, last=-3 which will print the first 3 and last 3 rows/cols
+// use negative ranges to print corners, e.g. Print("name", -3, -3, -3, -3) will print the first 3 and last 3 rows/cols
 template <class ElemType>
 void CPUMatrix<ElemType>::Print(const char* matrixName, ptrdiff_t rowFirst, ptrdiff_t rowLast, ptrdiff_t colFirst, ptrdiff_t colLast) const
 {

diff --git a/Source/Readers/LMSequenceReader/SequenceReader.cpp b/Source/Readers/LMSequenceReader/SequenceReader.cpp
@@ -1673,14 +1673,12 @@ void BatchSequenceReader<ElemType>::StartMinibatchLoop(size_t mbSize, size_t epo
 template <class ElemType>
 size_t BatchSequenceReader<ElemType>::DetermineSequencesToProcess()
 {
-    if (mNumRead == 0) // no data loaded or input file empty (?)
+    if (mNumRead == 0) // no data loaded (either this is the first call, or the input file empty)
         return 0;
 
     // lazily create the mProcessed[] array
     if (mProcessed.size() == 0)
-    {
         mProcessed.resize(mNumRead, false);
-    }
 
     // lazily check if we are done
     // We are done if mToProcess[] contains a sequence for which mProcessed[] is set (...which likely applies to all?).
@@ -1701,9 +1699,7 @@ size_t BatchSequenceReader<ElemType>::DetermineSequencesToProcess()
             }
         }
         if (allDone) // if we are done
-        {
             mToProcess.clear();
-        }
     }
 
     // if we still have unfinished sequences then just return their length