Expose binmlf in python

Source/CNTKv2LibraryDll/API/CNTKLibrary.h

@@ -6069,6 +6069,12 @@ namespace CNTK
     ///
     CNTK_API  Deserializer HTKMLFDeserializer(const std::wstring& streamName, const std::wstring& labelMappingFile, size_t dimension, const std::vector<std::wstring>& mlfFiles, bool phoneBoundaries = false);
 
+    ///
+    /// Create an HTKMLFBinaryDeserializer with the specified options
+    ///
+    CNTK_API  Deserializer HTKMLFBinaryDeserializer(const std::wstring& streamName, const std::vector<std::wstring>& mlfFiles, size_t dimension);
+
+
     ///
     /// Create a LatticeDeserializer with the specified options
     ///

Source/CNTKv2LibraryDll/MinibatchSource.cpp

@@ -512,6 +512,24 @@ namespace CNTK
         return htk;
     }
 
+    Deserializer HTKMLFBinaryDeserializer(const std::wstring& streamName, const std::vector<std::wstring>& mlfFiles, size_t dimension)
+    {
+        Deserializer htk;
+        Dictionary stream;
+        Dictionary labels;
+        std::vector<DictionaryValue> actualFiles;
+        std::transform(mlfFiles.begin(), mlfFiles.end(), std::back_inserter(actualFiles), [](const std::wstring& s) {return static_cast<DictionaryValue>(s); });
+        if (actualFiles.size() > 1)
+            labels[L"mlfFileList"] = actualFiles;
+        else if (actualFiles.size() == 1)
+            labels[L"mlfFile"] = actualFiles[0];
+        else
+            LogicError("HTKMLFDeserializer: No mlf files were specified");
+        stream[streamName] = labels;
+        htk.Add(L"type", L"HTKMLFBinaryDeserializer", L"input", stream);
+        return htk;
+    }
+
     Deserializer LatticeDeserializer(const std::wstring& streamName, const std::wstring& latticeIndexFile)
     {
         Deserializer lattice;

Source/SequenceTrainingLib/gammacalculation.h

@@ -208,6 +208,9 @@ public:
                                                                    (const msra::math::ssematrixbase&) predstripe, (const msra::asr::simplesenonehmm&) m_hset,
                                                                    (msra::math::ssematrixbase&) dengammasstripe, (msra::math::ssematrixbase&) gammasbuffer /*empty, not used*/,
                                                                    lmf, wp, amf, boostmmifactor, seqsMBRmode, uidsstripe, boundariesstripe);
+            if (denavlogp < LOGZERO / 2)
+                return;
+
             objectValue += (ElemType)((numavlogp - denavlogp) * numframes);
 
             if (samplesInRecurrentStep == 1)

bindings/python/cntk/io/__init__.py

@@ -823,6 +823,26 @@ def HTKMLFDeserializer(label_mapping_file, streams, phoneBoundaries = False):
             master_label_files = [master_label_files]
         return cntk_py.htk_mlf_deserializer(stream_name, label_mapping_file, dimension, master_label_files, phoneBoundaries)
 
+def HTKMLFBinaryDeserializer(mlf_file, dimension, streams):
+    '''
+    Configures a binary HTK label reader that reads speech MLF (Master
+    Label File)
+
+    Args:
+        mlf_file (str): path to the label mapping file
+        dimension (int): dimension of the labels
+        streams: any dictionary-like object that contains a mapping from stream
+          names to :class:`StreamDef` objects. Each StreamDef object configures
+          a label stream.
+    '''
+    if len(streams) != 1:
+        raise ValueError("HTKMLFBinaryDeserializer only accepts a single stream")
+    for stream_name, stream in streams.items():
+        if stream.stream_alias is not None:
+            raise ValueError("HTKMLFBinaryDeserializer does not support stream names")
+        return cntk_py.htk_mlf_binary_deserializer(stream_name, mlf_file, dimension)
+
+
 def LatticeDeserializer(lattice_index_file, streams):
     '''
     Configures a lattice deserializer