changed sample 02_Convolution to use imageLayout to enable cudnn use;

and fixed a refactoring bug in ImageParameter NDL definition

Examples/Image/MNIST/Config/02_Convolution.config

@@ -5,11 +5,15 @@
 RootDir = ".."
 
 ConfigDir = "$RootDir$/Config"
-DataDir = "$RootDir$/Data"
+DataDir   = "$RootDir$/Data"
 OutputDir = "$RootDir$/Output"
-ModelDir = "$OutputDir$/Models"
+ModelDir  = "$OutputDir$/Models"
 
-deviceId = "auto"
+deviceId = 0
+imageLayout = "cudnn"
+# override the above as follows when running from CPU:
+#deviceId = -1
+#imageLayout = "legacy"
 
 command = train:test
 

Examples/Image/MNIST/Config/02_Convolution.ndl

@@ -9,7 +9,7 @@ ndlMnistMacros = [
     imageH = 28
     labelDim = 10
 
-    features = ImageInput(imageW, imageH, 1)
+    features = ImageInput(imageW, imageH, 1, imageLayout=$imageLayout$)
     featScale = Const(0.00390625)
     featScaled = Scale(featScale, features)
     labels = Input(labelDim)
@@ -31,7 +31,7 @@ DNN=[
     pool1H = 2
     pool1hStride = 2
     pool1vStride = 2
-    pool1 = MaxPooling(conv1_act, pool1W, pool1H, pool1hStride, pool1vStride)
+    pool1 = MaxPooling(conv1_act, pool1W, pool1H, pool1hStride, pool1vStride, imageLayout=$imageLayout$)
 
     # conv2
     kW2 = 5
@@ -48,7 +48,7 @@ DNN=[
     pool2H = 2
     pool2hStride = 2
     pool2vStride = 2
-    pool2 = MaxPooling(conv2_act, pool2W, pool2H, pool2hStride, pool2vStride)
+    pool2 = MaxPooling(conv2_act, pool2W, pool2H, pool2hStride, pool2vStride, imageLayout=$imageLayout$)
 
     h1Dim = 128
     # DNNSigmoidLayer and DNNLayer are defined in Macros.ndl

Examples/Image/MNIST/Config/Macros.ndl

@@ -15,17 +15,8 @@ DNNLayer(inDim, outDim, x, parmScale) = [
 
 ConvReLULayer(inp, outMap, inWCount, kW, kH, hStride, vStride, wScale, bValue) = [
     convW = Parameter(outMap, inWCount, init="uniform", initValueScale=wScale)
-    convB = Parameter(outMap, 1,        init="fixedValue", value=bValue)
-    conv = Convolution(convW, inp, kW, kH, outMap, hStride, vStride, zeroPadding=false)
+    convB = ImageParameter(1, 1, outMap,     init="fixedValue", value=bValue, imageLayout=$imageLayout$)
+    conv = Convolution(convW, inp, kW, kH, outMap, hStride, vStride, zeroPadding=false, imageLayout=$imageLayout$)
     convPlusB = Plus(conv, convB);
     act = RectifiedLinear(convPlusB);
 ]
-
-BatchNorm(dim, x, scaleInit, biasInit) = [
-    m = Mean(x)
-    isd = InvStdDev(x)
-    norm = ColumnElementTimes(Minus(x, m), isd)
-    sc = Parameter(dim, 1, init="uniform", initValueScale=scaleInit)
-    b  = Parameter(dim, 1, init="uniform", initValueScale=biasInit)
-    bn_norm = Plus(ColumnElementTimes(norm, sc), b)
-]

Source/CNTK/SynchronousExecutionEngine.cpp

@@ -92,8 +92,8 @@ void SynchronousNodeEvaluator<ElemType>::Evaluate(NDLNode<ElemType>* node, const
         {
             // evaluate only scalar parameters
             vector<void*> params = EvaluateParameters(node, baseName, 0, parameter.size(), pass);
-            size_t imageWidth = ((NDLNode<ElemType>*) params[0])->GetScalar();
-            size_t imageHeight = ((NDLNode<ElemType>*) params[1])->GetScalar();
+            size_t imageWidth    = ((NDLNode<ElemType>*) params[0])->GetScalar();
+            size_t imageHeight   = ((NDLNode<ElemType>*) params[1])->GetScalar();
             size_t imageChannels = ((NDLNode<ElemType>*) params[2])->GetScalar();
             ImageLayoutKind imageLayoutKind = ImageLayoutKindFrom(node->GetOptionalParameter("imageLayout", "HWC"));
 
@@ -123,8 +123,6 @@ void SynchronousNodeEvaluator<ElemType>::Evaluate(NDLNode<ElemType>* node, const
             vector<void*> params = EvaluateParameters(node, baseName, 0, parameter.size(), pass);
             size_t i = 0;
             auto tensorShape = ProcessTensorShapeParameters(node, params, i, isImage, cnNodeType);
-            if (isImage)
-                tensorShape.AppendInPlace(3, 1); // this goes into the column dimension
             bool needGradient = node->GetOptionalParameter("needGradient", "true");
 
             nodePtr = builder.CreateLearnableParameter(name, tensorShape);

Source/Common/Include/Basics.h

@@ -58,7 +58,7 @@ __declspec_noreturn static inline void ThrowFormatted(const char* format, ...)
     va_start(args, format);
     vsprintf(buffer, format, args);
 #ifdef _DEBUG // print this to log before throwing, so we can see what the error is
-    fprintf(stderr, "About to throw exception '%s'\n", buffer);
+    fprintf(stderr, "\nAbout to throw exception '%s'\n", buffer);
 #endif
     Microsoft::MSR::CNTK::DebugUtil::PrintCallStack();
     throw E(buffer);

Source/Common/Include/TensorShape.h

@@ -754,7 +754,7 @@ static inline ImageLayoutKind ImageLayoutKindFrom(const wstring& s)
 }
 
 // interpret TensorShape as an image descriptor
-// considering that we support two ways of storingimages
+// considering that we support two ways of storing images
 struct ImageDimensions
 {
     size_t m_width, m_height, m_numChannels;

Source/ComputationNetworkLib/ConvolutionalNodes.h

@@ -198,20 +198,6 @@ class ConvolutionNode : public ComputationNode<ElemType>, public NumInputs<2>
 #endif
     }
 
-    // BUGBUG: Should not be here. Use PlusNode and m_sampleLayout.  TODO: Bad naming:'output' is actually an 'input'
-    void AddBias(const Matrix<ElemType>& output, const Matrix<ElemType>& bias, Matrix<ElemType>& dst)
-    {
-        assert(m_convEng != nullptr);
-        m_convEng->AddBias(*m_outT, output, *m_biasT, bias, dst);
-    }
-
-    void BackwardBias(const Matrix<ElemType>& srcGrad, Matrix<ElemType>& biasGrad)
-    {
-        assert(m_convEng != nullptr);
-        m_convEng->BackwardBias(*m_outT, srcGrad, *m_biasT, biasGrad);
-    }
-
-    // note: this also infers dimensions from chilren
     void /*ComputationNodeBase::*/ Validate(bool isFinalValidationPass) override
     {
         Base::Validate(isFinalValidationPass);
@@ -224,11 +210,11 @@ class ConvolutionNode : public ComputationNode<ElemType>, public NumInputs<2>
             InvalidArgument("%ls %ls operation requires that input width be >= kernelWidth and input height >= kernelHeight.", NodeName().c_str(), OperationName().c_str());
 
         // determine output tensor shape
-        const int kernelWidthCenter = m_zeroPadding ? m_kernelWidth % 2 : m_kernelWidth;
+        const int kernelWidthCenter  = m_zeroPadding ? m_kernelWidth  % 2 : m_kernelWidth;
         const int kernelHeightCenter = m_zeroPadding ? m_kernelHeight % 2 : m_kernelHeight;
         auto outDims = ImageDimensions(
-            (inDims.m_width - kernelWidthCenter) / m_horizontalSubsample + 1,
-            (inDims.m_height - kernelHeightCenter) / m_verticalSubsample + 1,
+            (inDims.m_width  - kernelWidthCenter)  / m_horizontalSubsample + 1,
+            (inDims.m_height - kernelHeightCenter) / m_verticalSubsample   + 1,
             m_outputChannels);
 
         size_t weightCols = m_kernelWidth * m_kernelHeight * inDims.m_numChannels;

Tests/EndToEndTests/Examples/Image/MNIST/02_Convolution/run-test

@@ -10,5 +10,10 @@ else
 fi
 
 # cntkrun <CNTK config file name> <additional CNTK args>
-cntkrun 02_Convolution.config "train=[reader=[file=$CleanDataDir/Train.txt]] test=[reader=[file=$CleanDataDir/Test.txt]] train=[SGD=[maxEpochs=1]] train=[SGD=[epochSize=128]]  train=[reader=[randomize=none]]" || exit $?
+if [ "$TEST_DEVICE" == "cpu" ]; then
+  imageLayout='"legacy"'
+else
+  imageLayout='"cudnn"'
+fi
 
+cntkrun  02_Convolution.config  "train=[reader=[file=$CleanDataDir/Train.txt]]  test=[reader=[file=$CleanDataDir/Test.txt]]  train=[SGD=[maxEpochs=1]]  train=[SGD=[epochSize=128]]  train=[reader=[randomize=none]]  imageLayout=$imageLayout" || exit $?