Added simple deconv sample.

TonyChouZJU · Jun 14, 2016 · 75dbb2c · 75dbb2c
1 parent c5a6f11
commit 75dbb2c
Show file tree

Hide file tree

Showing 8 changed files with 185 additions and 12 deletions.
diff --git a/Examples/Image/MNIST/Config/04_DeConv.cntk b/Examples/Image/MNIST/Config/04_DeConv.cntk
@@ -0,0 +1,102 @@
+# Parameters can be overwritten on the command line
+# for example: cntk configFile=myConfigFile RootDir=../.. 
+# For running from Visual Studio add
+# currentDirectory=$(SolutionDir)/<path to corresponding data folder> 
+RootDir = ".."
+
+ConfigDir = "$RootDir$/Config"
+DataDir   = "$RootDir$/Data"
+OutputDir = "$RootDir$/Output"
+ModelDir  = "$OutputDir$/Models"
+
+deviceId = 0
+imageLayout = "cudnn"
+# Override the above as follows when running on CPU:
+# deviceId = -1
+# Note: Compared to GPU, this runs very slow.
+
+command = train:test
+
+precision = "float"
+modelPath = "$ModelDir$/04_DeConv"
+ndlMacros = "$ConfigDir$/Macros.ndl"
+
+# uncomment the following line to write logs to a file 
+# stderr = "$OutputDir$/04_DeConv_out"
+traceLevel=1
+numMBsToShowResult=500
+
+prefetch=true
+
+# If set to true, always initialize the network on CPU, making initialization consistent across CPU and GPU targets (for testing).
+initOnCPUOnly=true
+
+#######################################
+#  TRAINING CONFIG                    #
+#######################################
+
+train = [
+    action = "train"
+
+    NDLNetworkBuilder = [
+        networkDescription = "$ConfigDir$/04_DeConv.ndl"
+    ]
+
+    SGD = [
+        epochSize = 60000
+        minibatchSize = 32
+        learningRatesPerMB = 1*5:0.03
+        momentumPerMB = 0*10:0.7
+        maxEpochs = 15
+    ]
+
+    # Note: this reader crashes if randomization is turned on.
+    reader = [
+        readerType = "UCIFastReader"
+        # To get the data (Train-28x28.txt) please run `python mnist_convert.py` 
+        # from the 'AdditionalFiles' folder. See REAMDE.md for details.
+        file = "$DataDir$/Train-28x28.txt"
+
+        features = [
+            dim = 784
+            start = 1
+        ]
+
+        labels = [
+            dim = 1
+            start = 0
+            labelDim = 10
+            labelMappingFile = "$DataDir$/labelsmap.txt"
+        ]
+    ]    
+]
+
+#######################################
+#  TEST CONFIG                        #
+#######################################
+
+test = [
+    action = test
+    minibatchSize = 16
+
+    NDLNetworkBuilder = [
+        networkDescription = "$ConfigDir$/04_DeConv.ndl"
+    ]
+
+    reader = [
+        readerType = "UCIFastReader"
+        file = "$DataDir$/Test-28x28.txt"
+
+        features = [
+            dim = 784
+            start = 1
+        ]
+
+        labels = [
+            dim = 1
+            start = 0
+            labelDim = 10
+            labelMappingFile = "$DataDir$/labelsmap.txt"
+        ]
+    ]
+]
diff --git a/Examples/Image/MNIST/Config/04_DeConv.ndl b/Examples/Image/MNIST/Config/04_DeConv.ndl
@@ -0,0 +1,60 @@
+# macros to include
+load = ndlMnistMacros
+
+# the actual NDL that defines the network
+run = DNN
+
+ndlMnistMacros = [
+    imageW = 28
+    imageH = 28
+    imageC = 1
+    labelDim = 10
+
+    features = ImageInput(imageW, imageH, imageC, imageLayout=$imageLayout$)
+    featScale = Constant(0.00390625)
+    featScaled = Scale(featScale, features)
+    labels = InputValue(labelDim)
+]
+
+DNN=[
+    # conv1
+    kW1 = 5
+    kH1 = 5
+    cMap1 = 16
+    hStride1 = 2
+    vStride1 = 2
+    wScale1 = 10
+    bValue1 = 1
+    # weight[cMap1, kW1 * kH1 * inputChannels]
+    # Conv2DReLULayer is defined in Macros.ndl
+    conv1 = Conv2DReLULayer(featScaled, cMap1, 25, kW1, kH1, hStride1, vStride1, wScale1, bValue1)
+
+    # pool1
+    pool1W = 2
+    pool1H = 2
+    pool1hStride = 2
+    pool1vStride = 2
+    # MaxPooling is a standard NDL node.
+    pool1 = MaxPooling(conv1, pool1W, pool1H, pool1hStride, pool1vStride, imageLayout=$imageLayout$)
+
+    #unpool1
+    unpool1 = MaxNDUnpooling(pool1, conv1, pool1W, pool1H, pool1hStride, pool1vStride)
+
+    # deconv1
+    lpad1 = 2
+    upad1 = 1
+    # weight[cMap2, kW2 * kH2 * cMap1]
+    # DeconvReLULayer is defined in Macros.ndl
+    deconv1 = DeconvReLULayer(unpool1, kW1, kH1, imageC, 25, cMap1, hStride1, vStride1, lpad1, upad1, wScale1, bValue1)
+
+    mse = SquareError(featScaled, deconv1)
+    #err = ErrorPrediction(labels, ol)
+
+    # Special Nodes
+    FeatureNodes = (features)
+    #LabelNodes = (labels)
+    CriterionNodes = (mse)
+    #EvalNodes = (err)
+    OutputNodes = (deconv1)
+]
+
diff --git a/Examples/Image/MNIST/Config/Macros.ndl b/Examples/Image/MNIST/Config/Macros.ndl
@@ -48,8 +48,8 @@ ConvND(w, inp, kW, kH, inMap, outMap, hStride, vStride) = [
     c = Convolution(w, inp, {kW, kH, inMap}, mapCount=outMap, stride={hStride, vStride, inMap}, sharing={true, true, true}, autoPadding={true, true, false}, lowerPad=0, upperPad=0, imageLayout=$imageLayout$)
 ]
 
-DeConv(w, inp, k, inMap, outMap, hStride, vStride, lpad, upad) = [
-    c = Convolution(w, inp, {k, k, inMap}, mapCount=outMap, stride={hStride, vStride, inMap}, sharing={true, true, true}, autoPadding=false, lowerPad={lpad, lpad, 0}, upperPad={upad, upad, 0}, transpose=1, imageLayout=$imageLayout$)
+DeConv(w, inp, kW, kH, inMap, outMap, hStride, vStride, lpad, upad) = [
+    c = Convolution(w, inp, {kW, kH, inMap}, mapCount=outMap, stride={hStride, vStride, inMap}, sharing={true, true, true}, autoPadding=false, lowerPad={lpad, lpad, 0}, upperPad={upad, upad, 0}, transpose=1, imageLayout=$imageLayout$)
 ]
 
 Conv2DReLULayer(inp, outMap, inWCount, kW, kH, hStride, vStride, wScale, bValue) = [
@@ -88,11 +88,19 @@ ConvBNReLULayer(inp, outMap, inWCount, kW, kH, hStride, vStride, wScale, bValue,
     y = RectifiedLinear(c)
 ]
 
+DeconvReLULayer(inp, kW, kH, inMap, inWCount, outMap, hStride, vStride, lpad, upad, wScale, bValue) = [
+    w = ConvW(outMap, inWCount, wScale)
+    b = ConvB(inMap, bValue)
+    dc = DeConv(w, inp, kW, kH, inMap, outMap, hStride, vStride, lpad, upad)
+    cpb = Plus(dc, b);
+    out = RectifiedLinear(dc);
+]
+
 MaxNDPooling(inp, kW, kH, hStride, vStride) = [
-    mask = MaxPoolingMask(inp, {kW, kH, 1}, stride={hStride, vStride, 1}, autoPadding={true, true, false}, lowerPad=0, upperPad=0, imageLayout=$imageLayout$)
 		p = Pooling(inp, "max", {kW, kH, 1}, stride={hStride, vStride, 1}, autoPadding={true, true, false}, lowerPad=0, upperPad=0, imageLayout=$imageLayout$)
 ]
 
-MaxNDUnpooling(inp, mask, kW, kH, hStride, vStride) = [
+MaxNDUnpooling(inp, poolInp, kW, kH, hStride, vStride) = [
+    mask = MaxPoolingMask(poolInp, {kW, kH, 1}, stride={hStride, vStride, 1}, autoPadding={true, true, false}, lowerPad=0, upperPad=0, imageLayout=$imageLayout$)
 		up = MaxUnpooling(inp, mask, {kW, kH, 1}, stride={hStride, vStride, 1}, autoPadding={false, false, false}, lowerPad=0, upperPad=0, imageLayout=$imageLayout$)
 ]
diff --git a/Source/ComputationNetworkLib/ConvolutionalNodes.h b/Source/ComputationNetworkLib/ConvolutionalNodes.h
@@ -637,7 +637,10 @@ class MaxUnpoolingNode : public ConvolutionNodeBase<ElemType>, public NumInputs<
 
     void ForwardProp(const FrameRange& fr) override
     {
-        UNUSED(fr);
+        const Matrix<ElemType>& input = Input(0)->ValueFor(fr);
+        const Matrix<ElemType>& mask = Input(1)->ValueFor(fr);
+        Matrix<ElemType> sliceOutputValue = ValueFor(fr);
+        m_convEng->MaxUnpooling(input, mask, sliceOutputValue);
     }
 
     void Validate(bool isFinalValidationPass) override

diff --git a/Source/Math/CPUMatrix.cpp b/Source/Math/CPUMatrix.cpp
@@ -4358,7 +4358,7 @@ void CPUMatrix<ElemType>::MaxUnpooling(const CPUMatrix<int>& mpRowCol, const CPU
                                        const CPUMatrix<int>& indices, const CPUMatrix<ElemType>& mask,
                                        CPUMatrix<ElemType>& input) const
 {
-//#pragma omp parallel for
+#pragma omp parallel for
     for (int64_t sample = 0; sample < (int64_t)GetNumCols(); sample++)
     {
         for (size_t row = 0; row < GetNumRows(); row++)

diff --git a/Source/Math/ConvolutionEngine.cpp b/Source/Math/ConvolutionEngine.cpp
@@ -126,7 +126,7 @@ void ConvolutionEngine<ElemType>::MaxPoolingMask(const Mat& in, Mat& mask)
 }
 
 template <class ElemType>
-void ConvolutionEngine<ElemType>::MaxUnpooling(const Mat& out, Mat& mask, Mat& in)
+void ConvolutionEngine<ElemType>::MaxUnpooling(const Mat& out, const Mat& mask, Mat& in)
 {
     const auto& g = *m_geometry;
     assert(g.InputShape().GetNumElements() == in.GetNumRows());
@@ -253,7 +253,7 @@ class ReferenceConvolutionEngine : public ConvolutionEngine<ElemType>
         in.MaxPoolingMask(m_mpRowCol, *m_mpRowIndices, *m_indices, mask);
     }
 
-    void MaxUnpoolingCore(const Mat& out, Mat& mask, Mat& in) override
+    void MaxUnpoolingCore(const Mat& out, const Mat& mask, Mat& in) override
     {
         out.MaxUnpooling(m_mpRowCol, *m_mpRowIndices, *m_indices, mask, in);
     }
@@ -556,7 +556,7 @@ class LegacyConvolutionEngine : public ConvolutionEngine<ElemType>
         LogicError("MaxPoolingMask is not implemented for legacy engine.");
     }
 
-    void MaxUnpoolingCore(const Mat& out, Mat& mask, Mat& in) override
+    void MaxUnpoolingCore(const Mat& out, const Mat& mask, Mat& in) override
     {
         UNUSED(out);
         UNUSED(mask);

diff --git a/Source/Math/ConvolutionEngine.h b/Source/Math/ConvolutionEngine.h
@@ -57,7 +57,7 @@ class MATH_API ConvolutionEngine
 
     void MaxPoolingMask(const Mat& in, Mat& mask);
 
-    void MaxUnpooling(const Mat& out, Mat& mask, Mat& in);
+    void MaxUnpooling(const Mat& out, const Mat& mask, Mat& in);
 
     std::shared_ptr<const ConvolveGeometry> Geometry() const { return m_geometry; }
 
@@ -99,7 +99,7 @@ class MATH_API ConvolutionEngine
 
     virtual void MaxPoolingMaskCore(const Mat& in, Mat& mask) = 0;
 
-    virtual void MaxUnpoolingCore(const Mat& out, Mat& mask, Mat& in) = 0;
+    virtual void MaxUnpoolingCore(const Mat& out, const Mat& mask, Mat& in) = 0;
 
 protected:
     ConvolveGeometryPtr m_geometry;

diff --git a/Source/Math/CuDnnConvolutionEngine.cu b/Source/Math/CuDnnConvolutionEngine.cu
@@ -314,7 +314,7 @@ protected:
         LogicError("MaxPoolingMask is not implemented for cuDNN engine.");
     }
 
-    void MaxUnpoolingCore(const Mat& out, Mat& mask, Mat& in) override
+    void MaxUnpoolingCore(const Mat& out, const Mat& mask, Mat& in) override
     {
         UNUSED(out);
         UNUSED(mask);