Add support for non square matrices

include/layers/avg_pooling.cuh

@@ -9,10 +9,10 @@ namespace CUDANet::Layers {
 class AvgPooling2D : public SequentialLayer {
   public:
     AvgPooling2D(
-        int            inputSize,
+        dim2d           inputSize,
         int            nChannels,
-        int            poolingSize,
-        int            stride,
+        dim2d           poolingSize,
+        dim2d           stride,
         ActivationType activationType
     );
     ~AvgPooling2D();
@@ -28,18 +28,18 @@ class AvgPooling2D : public SequentialLayer {
 
     /**
      * @brief Get input size
-     * 
+     *
      * @return int input size
      */
     int getInputSize();
 
   private:
-    int inputSize;
-    int nChannels;
-    int poolingSize;
-    int stride;
+    dim2d inputSize;
+    int  nChannels;
+    dim2d poolingSize;
+    dim2d stride;
 
-    int outputSize;
+    dim2d outputSize;
 
     float* d_output;
 

include/layers/batch_norm.cuh

@@ -10,7 +10,7 @@ namespace CUDANet::Layers {
 
 class BatchNorm2D : public WeightedLayer {
   public:
-    BatchNorm2D(int inputSize, int inputChannels, float epsilon, ActivationType activationType);
+    BatchNorm2D(dim2d inputSize, int inputChannels, float epsilon, ActivationType activationType);
 
     ~BatchNorm2D();
 
@@ -66,7 +66,7 @@ class BatchNorm2D : public WeightedLayer {
 
   private:
 
-    int inputSize;
+    dim2d inputSize;
     int inputChannels;
 
     int gridSize;

include/layers/conv2d.cuh

@@ -28,12 +28,12 @@ class Conv2d : public WeightedLayer {
      * 'SOFTMAX' or 'NONE')
      */
     Conv2d(
-        int            inputSize,
+        dim2d          inputSize,
         int            inputChannels,
-        int            kernelSize,
-        int            stride,
+        dim2d          kernelSize,
+        dim2d          stride,
         int            numFilters,
-        int            paddingSize,
+        dim2d          paddingSize,
         ActivationType activationType
     );
 
@@ -98,23 +98,23 @@ class Conv2d : public WeightedLayer {
      *
      * @return int
      */
-    int getPaddingSize() {
+    dim2d getPaddingSize() {
         return paddingSize;
     }
 
   private:
     // Inputs
-    int inputSize;
-    int inputChannels;
+    dim2d inputSize;
+    int   inputChannels;
 
     // Outputs
-    int outputSize;
+    dim2d outputSize;
 
     // Kernel
-    int kernelSize;
-    int stride;
-    int paddingSize;
-    int numFilters;
+    dim2d kernelSize;
+    dim2d stride;
+    dim2d paddingSize;
+    int   numFilters;
 
     // Kernels
     std::vector<float> weights;

include/layers/dense.cuh

@@ -81,8 +81,8 @@ class Dense : public WeightedLayer {
     int getInputSize();
 
   private:
-    unsigned int inputSize;
-    unsigned int outputSize;
+    int inputSize;
+    int outputSize;
 
     float* d_output;
 
@@ -95,8 +95,8 @@ class Dense : public WeightedLayer {
     Layers::Activation* activation;
 
     // Precompute kernel launch parameters
-    unsigned int forwardGridSize;
-    unsigned int biasGridSize;
+    int forwardGridSize;
+    int biasGridSize;
 
     /**
      * @brief Initialize the weights to zeros

include/layers/layer.cuh

@@ -7,6 +7,8 @@
 #define CUDANET_SAME_PADDING(inputSize, kernelSize, stride) \
     ((stride - 1) * inputSize - stride + kernelSize) / 2;
 
+typedef std::pair<int, int> dim2d;
+
 namespace CUDANet::Layers {
 
 /**

include/layers/max_pooling.cuh

@@ -9,10 +9,10 @@ namespace CUDANet::Layers {
 class MaxPooling2D : public SequentialLayer {
   public:
     MaxPooling2D(
-        int            inputSize,
+        dim2d           inputSize,
         int            nChannels,
-        int            poolingSize,
-        int            stride,
+        dim2d           poolingSize,
+        dim2d           stride,
         ActivationType activationType
     );
     ~MaxPooling2D();
@@ -28,18 +28,18 @@ class MaxPooling2D : public SequentialLayer {
 
     /**
      * @brief Get input size
-     * 
+     *
      * @return int input size
      */
     int getInputSize();
 
   private:
-    int inputSize;
-    int nChannels;
-    int poolingSize;
-    int stride;
+    dim2d inputSize;
+    int  nChannels;
+    dim2d poolingSize;
+    dim2d stride;
 
-    int outputSize;
+    dim2d outputSize;
 
     float* d_output;