Implement getOutputSize and getInputSize for seq layers

include/layers/avg_pooling.cuh

@@ -20,13 +20,18 @@ class AvgPooling2D : public SequentialLayer {
     float* forward(const float* d_input);
 
     /**
-     * @brief Get the output width (/ height) of the layer
+     * @brief Get output size
      *
-     * @return int
+     * @return int output size
      */
-    int getOutputSize() {
-        return outputSize;
-    }
+    int getOutputSize();
+
+    /**
+     * @brief Get input size
+     * 
+     * @return int input size
+     */
+    int getInputSize();
 
   private:
     int inputSize;

include/layers/conv2d.cuh

@@ -80,13 +80,18 @@ class Conv2d : public WeightedLayer {
     std::vector<float> getBiases();
 
     /**
-     * @brief Get the output width (/ height) of the layer
+     * @brief Get output size
      *
-     * @return int
+     * @return int output size
      */
-    int getOutputSize() {
-        return outputSize;
-    }
+    int getOutputSize();
+
+    /**
+     * @brief Get input size
+     *
+     * @return int input size
+     */
+    int getInputSize();
 
     /**
      * @brief Get the padding size of the layer

include/layers/dense.cuh

@@ -3,8 +3,8 @@
 
 #include <vector>
 
-#include "layer.cuh"
 #include "activation.cuh"
+#include "layer.cuh"
 
 namespace CUDANet::Layers {
 
@@ -19,7 +19,8 @@ class Dense : public WeightedLayer {
      *
      * @param inputSize Size of the input vector
      * @param outputSize Size of the output vector
-     * @param activationType Activation function type ('RELU', 'SIGMOID', 'SOFTMAX' or 'NONE')
+     * @param activationType Activation function type ('RELU', 'SIGMOID',
+     * 'SOFTMAX' or 'NONE')
      */
     Dense(int inputSize, int outputSize, Layers::ActivationType activationType);
 
@@ -65,6 +66,20 @@ class Dense : public WeightedLayer {
      */
     std::vector<float> getBiases();
 
+    /**
+     * @brief Get output size
+     *
+     * @return int output size
+     */
+    int getOutputSize();
+
+    /**
+     * @brief Get input size
+     *
+     * @return int input size
+     */
+    int getInputSize();
+
   private:
     unsigned int inputSize;
     unsigned int outputSize;

include/layers/input.cuh

@@ -25,13 +25,28 @@ class Input : public SequentialLayer {
     ~Input();
 
     /**
-     * @brief Forward pass of the input layer. Just copies the input to the device
+     * @brief Forward pass of the input layer. Just copies the input to the
+     * device
      *
      * @param input Host pointer to the input vector
      * @return Device pointer to the output vector
      */
     float* forward(const float* input);
 
+    /**
+     * @brief Get output size
+     *
+     * @return int output size
+     */
+    int getOutputSize();
+
+        /**
+     * @brief Get input size
+     * 
+     * @return int input size
+     */
+    int getInputSize();
+
   private:
     int    inputSize;
     float* d_output;

include/layers/layer.cuh

@@ -4,8 +4,8 @@
 
 #include <vector>
 
-#define CUDANET_SAME_PADDING(inputSize, kernelSize, stride) ((stride - 1) * inputSize - stride + kernelSize) / 2;
-
+#define CUDANET_SAME_PADDING(inputSize, kernelSize, stride) \
+    ((stride - 1) * inputSize - stride + kernelSize) / 2;
 
 namespace CUDANet::Layers {
 
@@ -28,6 +28,20 @@ class SequentialLayer {
      * @return float* Device pointer to the output
      */
     virtual float* forward(const float* input) = 0;
+
+    /**
+     * @brief Get output size
+     *
+     * @return int output size
+     */
+    virtual int getOutputSize() = 0;
+
+    /**
+     * @brief Get input size
+     * 
+     * @return int input size
+     */
+    virtual int getInputSize() = 0;
 };
 
 /**

include/layers/max_pooling.cuh

@@ -20,13 +20,18 @@ class MaxPooling2D : public SequentialLayer {
     float* forward(const float* d_input);
 
     /**
-     * @brief Get the output width (/ height) of the layer
+     * @brief Get output size
      *
-     * @return int
+     * @return int output size
      */
-    int getOutputSize() {
-        return outputSize;
-    }
+    int getOutputSize();
+
+    /**
+     * @brief Get input size
+     * 
+     * @return int input size
+     */
+    int getInputSize();
 
   private:
     int inputSize;

include/layers/output.cuh

@@ -6,34 +6,48 @@
 namespace CUDANet::Layers {
 
 class Output : public SequentialLayer {
-    public:
-        /**
-         * @brief Create a new Output layer
-         * 
-         * @param inputSize Size of the input vector
-         */
-        explicit Output(int inputSize);
+  public:
+    /**
+     * @brief Create a new Output layer
+     *
+     * @param inputSize Size of the input vector
+     */
+    explicit Output(int inputSize);
 
-        /**
-         * @brief Destroy the Output layer
-         * 
-         */
-        ~Output();
+    /**
+     * @brief Destroy the Output layer
+     *
+     */
+    ~Output();
 
-        /**
-         * @brief Forward pass of the output layer. Just copies the input from device to host
-         * 
-         * @param input Device pointer to the input vector
-         * @return Host pointer to the output vector
-         */
-        float* forward(const float* input);
+    /**
+     * @brief Forward pass of the output layer. Just copies the input from
+     * device to host
+     *
+     * @param input Device pointer to the input vector
+     * @return Host pointer to the output vector
+     */
+    float* forward(const float* input);
 
-    private:
-        int    inputSize;
-        float* h_output;
+    /**
+     * @brief Get output size
+     *
+     * @return int output size
+     */
+    int getOutputSize();
+
+    /**
+     * @brief Get input size
+     *
+     * @return int input size
+     */
+    int getInputSize();
+
+  private:
+    int    inputSize;
+    float* h_output;
 };
 
-
-} // namespace CUDANet::Layers
+}  // namespace CUDANet::Layers
 
 #endif  // CUDANET_OUTPUT_LAYER_H

src/layers/avg_pooling.cu

@@ -50,3 +50,11 @@ float* AvgPooling2D::forward(const float* d_input) {
 
     return d_output;
 }
+
+int AvgPooling2D::getOutputSize() {
+    return outputSize * outputSize * nChannels;
+}
+
+int AvgPooling2D::getInputSize() {
+    return inputSize * inputSize * nChannels;
+}

src/layers/conv2d.cu

@@ -130,3 +130,11 @@ float* Conv2d::forward(const float* d_input) {
 
     return d_output;
 }
+
+int Conv2d::getOutputSize() {
+    return outputSize * outputSize * numFilters;
+}
+
+int Conv2d::getInputSize() {
+    return inputSize * inputSize * inputChannels;
+}

src/layers/dense.cu

@@ -109,3 +109,11 @@ void Dense::setBiases(const float* biases_input) {
 std::vector<float> Dense::getBiases() {
     return biases;
 }
+
+int Dense::getOutputSize() {
+    return outputSize;
+}
+
+int Dense::getInputSize() {
+    return inputSize;
+}

src/layers/input.cu

@@ -20,3 +20,12 @@ float* Input::forward(const float* input) {
 
     return d_output;
 }
+
+int Input::getOutputSize() {
+    return inputSize;
+}
+
+
+int Input::getInputSize() {
+    return inputSize;
+}

src/layers/max_pooling.cu

@@ -53,3 +53,11 @@ float* MaxPooling2D::forward(const float* d_input) {
 
     return d_output;
 }
+
+int MaxPooling2D::getOutputSize() {
+    return outputSize * outputSize * nChannels;
+}
+
+int MaxPooling2D::getInputSize() {
+    return inputSize * inputSize * nChannels;
+}

src/layers/output.cu

@@ -21,3 +21,12 @@ float* Output::forward(const float* input) {
 
     return h_output;
 }
+
+int Output::getOutputSize() {
+    return inputSize;
+}
+
+
+int Output::getInputSize() {
+    return inputSize;
+}

src/model/model.cpp

@@ -43,7 +43,6 @@ float* Model::predict(const float* input) {
     float* d_input = inputLayer->forward(input);
 
     for (auto& layer : layers) {
-        std::cout << layer.first << std::endl;
         d_input = layer.second->forward(d_input);
     }
 

test/kernels/test_matmul.cu

@@ -187,6 +187,8 @@ TEST(MatMulTest, SumReduceTest) {
         cudaMemcpy(d_input, input.data(), sizeof(float) * n, cudaMemcpyHostToDevice);
     EXPECT_EQ(cudaStatus, cudaSuccess);
 
+    CUDANet::Utils::clear(d_sum, n);
+
     CUDANet::Kernels::sum_reduce<<<gridSize, BLOCK_SIZE>>>(
         d_input, d_sum, n
     );
@@ -209,6 +211,4 @@ TEST(MatMulTest, SumReduceTest) {
 
     cudaFree(d_input);
     cudaFree(d_sum);    
-
-    
 }

test/layers/test_avg_pooling.cu

@@ -51,10 +51,10 @@ TEST(AvgPoolingLayerTest, AvgPoolForwardTest) {
 
     int outputSize = avgPoolingLayer.getOutputSize();
 
-    std::vector<float> output(outputSize * outputSize * nChannels);
+    std::vector<float> output(outputSize);
     cudaStatus = cudaMemcpy(
         output.data(), d_output,
-        sizeof(float) * outputSize * outputSize * nChannels,
+        sizeof(float) * outputSize,
         cudaMemcpyDeviceToHost
     );
     EXPECT_EQ(cudaStatus, cudaSuccess);

test/layers/test_conv2d.cu

@@ -82,14 +82,15 @@ TEST_F(Conv2dTest, SimpleTest) {
         activationType, input, kernels.data(), d_input
     );
 
-    int outputSize = (inputSize - kernelSize) / stride + 1;
+    int outputWidth = (inputSize - kernelSize) / stride + 1;
+    int outputSize  = outputWidth * outputWidth * numFilters;
     EXPECT_EQ(outputSize, conv2d.getOutputSize());
 
     d_output = conv2d.forward(d_input);
 
     std::vector<float> expected = {44.0f,  54.0f,  64.0f,  84.0f, 94.0f,
                                    104.0f, 124.0f, 134.0f, 144.0f};
-    std::vector<float> output(outputSize * outputSize * numFilters);
+    std::vector<float> output(outputSize);
 
     cudaStatus = cudaMemcpy(
         output.data(), d_output, sizeof(float) * output.size(),
@@ -172,18 +173,16 @@ TEST_F(Conv2dTest, PaddedTest) {
         activationType, input, kernels.data(), d_input
     );
 
-    EXPECT_EQ(inputSize, conv2d.getOutputSize());
+    EXPECT_EQ(inputSize * inputSize * numFilters, conv2d.getOutputSize());
 
     d_output = conv2d.forward(d_input);
 
     std::vector<float> output(
-        conv2d.getOutputSize() * conv2d.getOutputSize() * numFilters
+        conv2d.getOutputSize()
     );
     cudaMemcpy(
         output.data(), d_output,
-        sizeof(float) * conv2d.getOutputSize() * conv2d.getOutputSize() *
-            numFilters,
-        cudaMemcpyDeviceToHost
+        sizeof(float) * conv2d.getOutputSize(), cudaMemcpyDeviceToHost
     );
 
     // Generated by tools/generate_conv2d_test.py
@@ -259,17 +258,16 @@ TEST_F(Conv2dTest, StridedPaddedConvolution) {
         activationType, input, kernels.data(), d_input
     );
 
-    EXPECT_EQ(inputSize, conv2d.getOutputSize());
+    EXPECT_EQ(inputSize * inputSize * numFilters, conv2d.getOutputSize());
 
     d_output = conv2d.forward(d_input);
 
     std::vector<float> output(
-        conv2d.getOutputSize() * conv2d.getOutputSize() * numFilters
+        conv2d.getOutputSize()
     );
     cudaMemcpy(
         output.data(), d_output,
-        sizeof(float) * conv2d.getOutputSize() * conv2d.getOutputSize() *
-            numFilters,
+        sizeof(float) * conv2d.getOutputSize(),
         cudaMemcpyDeviceToHost
     );
 

test/layers/test_max_pooling.cu

@@ -51,10 +51,10 @@ TEST(MaxPoolingLayerTest, MaxPoolForwardTest) {
 
     int outputSize = maxPoolingLayer.getOutputSize();
 
-    std::vector<float> output(outputSize * outputSize * nChannels);
+    std::vector<float> output(outputSize);
     cudaStatus = cudaMemcpy(
         output.data(), d_output,
-        sizeof(float) * outputSize * outputSize * nChannels,
+        sizeof(float) * outputSize,
         cudaMemcpyDeviceToHost
     );
     EXPECT_EQ(cudaStatus, cudaSuccess);