Merge pull request #3534 from shrit/float

Float -- episode 1

src/mlpack/methods/ann/ffn_impl.hpp

@@ -468,7 +468,7 @@ typename MatType::elem_type FFN<
   res += EvaluateWithGradient(parameters, 0, gradient, 1);
   for (size_t i = 1; i < predictors.n_cols; ++i)
   {
-    arma::mat tmpGradient(gradient.n_rows, gradient.n_cols);
+    MatType tmpGradient(gradient.n_rows, gradient.n_cols);
     res += EvaluateWithGradient(parameters, i, tmpGradient, 1);
     gradient += tmpGradient;
   }

src/mlpack/methods/ann/layer/batch_norm_impl.hpp

@@ -277,7 +277,7 @@ void BatchNormType<MatType>::Backward(
     const MatType& gy,
     MatType& g)
 {
-  const arma::mat stdInv = 1.0 / arma::sqrt(variance + eps);
+  const MatType stdInv = 1.0 / arma::sqrt(variance + eps);
 
   const size_t batchSize = input.n_cols;
   const size_t inputSize = inputDimension;

src/mlpack/methods/ann/layer/c_relu_impl.hpp

@@ -68,13 +68,14 @@ template<typename MatType>
 void CReLUType<MatType>::Forward(
     const MatType& input, MatType& output)
 {
+  typename MatType::elem_type zero = 0.0;
   #pragma omp for
   for (size_t i = 0; i < (size_t) input.n_cols; ++i)
   {
     for (size_t j = 0; j < (size_t) input.n_rows; ++j)
     {
-      output(j, i) = std::max(input(j, i), 0.0);
-      output(j + input.n_rows, i) = std::max(-input(j, i), 0.0);
+      output(j, i) = std::max(input(j, i), zero);
+      output(j + input.n_rows, i) = std::max(-input(j, i), zero);
     }
   }
 }

src/mlpack/methods/ann/layer/convolution.hpp

@@ -374,10 +374,10 @@ class ConvolutionType : public Layer<MatType>
   arma::Cube<typename MatType::elem_type> gradientTemp;
 
   //! Locally-stored padding layer.
-  Padding padding;
+  PaddingType<MatType> padding;
 
   //! Locally-stored padding layer for backward pass.
-  Padding paddingBackward;
+  PaddingType<MatType> paddingBackward;
 
   //! Type of padding.
   std::string paddingType;

src/mlpack/methods/ann/layer/convolution_impl.hpp

@@ -573,7 +573,7 @@ void ConvolutionType<
     InitializeSamePadding();
   }
 
-  padding = Padding(padWLeft, padWRight, padHTop, padHBottom);
+  padding = PaddingType<MatType>(padWLeft, padWRight, padHTop, padHBottom);
   padding.InputDimensions() = this->inputDimensions;
   padding.ComputeOutputDimensions();
 
@@ -600,7 +600,7 @@ void ConvolutionType<
   apparentHeight = (this->outputDimensions[1] - 1) * strideHeight +
       kernelHeight;
 
-  paddingBackward = Padding(0, padding.OutputDimensions()[0] -
+  paddingBackward = PaddingType<MatType>(0, padding.OutputDimensions()[0] -
       apparentWidth, 0, padding.OutputDimensions()[1] - apparentHeight);
   paddingBackward.InputDimensions() = std::vector<size_t>({ apparentWidth,
       apparentHeight, inMaps * higherInDimensions });

src/mlpack/methods/ann/layer/grouped_convolution.hpp

@@ -390,10 +390,10 @@ class GroupedConvolutionType : public Layer<MatType>
   arma::Cube<typename MatType::elem_type> gradientTemp;
 
   //! Locally-stored padding layer.
-  Padding padding;
+  PaddingType<MatType> padding;
 
   //! Locally-stored padding layer for backward pass.
-  Padding paddingBackward;
+  PaddingType<MatType> paddingBackward;
 
   //! Type of padding.
   std::string paddingType;

src/mlpack/methods/ann/layer/grouped_convolution_impl.hpp

@@ -612,7 +612,7 @@ void GroupedConvolutionType<
     InitializeSamePadding();
   }
 
-  padding = Padding(padWLeft, padWRight, padHTop, padHBottom);
+  padding = PaddingType<MatType>(padWLeft, padWRight, padHTop, padHBottom);
   padding.InputDimensions() = this->inputDimensions;
   padding.ComputeOutputDimensions();
 
@@ -650,7 +650,7 @@ void GroupedConvolutionType<
   apparentHeight = (this->outputDimensions[1] - 1) * strideHeight + 
       kernelHeight;
 
-  paddingBackward = Padding(0, padding.OutputDimensions()[0] -
+  paddingBackward = PaddingType<MatType>(0, padding.OutputDimensions()[0] -
       apparentWidth, 0, padding.OutputDimensions()[1] - apparentHeight);
   paddingBackward.InputDimensions() = std::vector<size_t>({ apparentWidth,
       apparentHeight, inMaps * higherInDimensions });

src/mlpack/methods/ann/layer/leaky_relu.hpp

@@ -47,7 +47,7 @@ class LeakyReLUType : public Layer<MatType>
    *
    * @param alpha Non zero gradient.
    */
-  LeakyReLUType(const double alpha = 0.03);
+  LeakyReLUType(const typename MatType::elem_type alpha = 0.03);
 
   //! Clone the LeakyReLUType object. This handles polymorphism correctly.
   LeakyReLUType* Clone() const { return new LeakyReLUType(*this); }
@@ -85,9 +85,9 @@ class LeakyReLUType : public Layer<MatType>
   void Backward(const MatType& input, const MatType& gy, MatType& g);
 
   //! Get the non zero gradient.
-  double const& Alpha() const { return alpha; }
+  typename MatType::elem_type const& Alpha() const { return alpha; }
   //! Modify the non zero gradient.
-  double& Alpha() { return alpha; }
+  typename MatType::elem_type& Alpha() { return alpha; }
 
   //! Serialize the layer.
   template<typename Archive>
@@ -95,7 +95,7 @@ class LeakyReLUType : public Layer<MatType>
 
  private:
   //! Leakyness Parameter in the range 0 <alpha< 1
-  double alpha;
+  typename MatType::elem_type alpha;
 }; // class LeakyReLUType
 
 // Convenience typedefs.

src/mlpack/methods/ann/layer/leaky_relu_impl.hpp

@@ -20,7 +20,7 @@
 namespace mlpack {
 
 template<typename MatType>
-LeakyReLUType<MatType>::LeakyReLUType(const double alpha) :
+LeakyReLUType<MatType>::LeakyReLUType(const typename MatType::elem_type alpha) :
     Layer<MatType>(),
     alpha(alpha)
 {

src/mlpack/methods/ann/layer/mean_pooling_impl.hpp

@@ -376,7 +376,7 @@ void MeanPoolingType<MatType>::Unpooling(
           rowEnd = output.n_rows - 1;
         }
 
-        arma::mat OutputArea = output(arma::span(i, rowEnd), arma::span(j, colEnd));
+        MatType OutputArea = output(arma::span(i, rowEnd), arma::span(j, colEnd));
 
         unpooledError = arma::Mat<typename MatType::elem_type>(OutputArea.n_rows, OutputArea.n_cols);
         unpooledError.fill(error(rowidx, colidx) / OutputArea.n_elem);