File content as of revision 6:3a0fd4eef371:

#ifndef LOWPOWERNEURALNETWORK_DEEPNEURALNETWORK_H
#define LOWPOWERNEURALNETWORK_DEEPNEURALNETWORK_H

/**
 * Default Hyperparameters for a test set accuracy equals to 95.61%
 * */
#define INPUT_IMAGE_SIZE 28
#define CONV2D_SIZE 7
//#define CONV2D_SIZE 2
#define MAX_POOL_SIZE 2
#define DROPOUT_RATE 0.25
#define N_OUT_CLASSES 36

/**
 * Layers inputs-outputs dimensions (each dimension refers to a single dimension -> matrix are squares)
 */
#define CONV2D_INPUT_SIZE INPUT_IMAGE_SIZE
//#define CONV2D_INPUT_SIZE 3
//#define CONV2D_OUTPUT_SIZE 22
#define CONV2D_OUTPUT_SIZE CONV2D_INPUT_SIZE-CONV2D_SIZE+1

#define MAX_POOL_INPUT_SIZE CONV2D_OUTPUT_SIZE
//#define MAX_POOL_INPUT_SIZE 4
#define MAX_POOL_OUTPUT_SIZE 11

#define FLATTEN_INPUT_SIZE MAX_POOL_OUTPUT_SIZE
#define FLATTEN_OUTPUT_SIZE FLATTEN_INPUT_SIZE * FLATTEN_INPUT_SIZE

#define DENSE_INPUT_SIZE FLATTEN_OUTPUT_SIZE
//#define DENSE_INPUT_SIZE 5

#define DENSE_OUTPUT_SIZE N_OUT_CLASSES
//#define DENSE_OUTPUT_SIZE 3

#define DENSE_WEIGHTS_LEN DENSE_INPUT_SIZE * DENSE_OUTPUT_SIZE


#include <iostream>
#include <algorithm>
#include <limits>
#include "mbed.h"

/**
 * Static dimensioned deep neural network without dynamic allocation
 */
class DeepNeuralNetwork {
//private:
public:
    float conv2d_output[CONV2D_OUTPUT_SIZE][CONV2D_OUTPUT_SIZE];
    float maxpool_output[MAX_POOL_OUTPUT_SIZE][MAX_POOL_OUTPUT_SIZE];
    float flatten_output[FLATTEN_OUTPUT_SIZE];
    float dense_output[DENSE_OUTPUT_SIZE];

    float leaky_relu(float x);

    void conv2d(const float input[][CONV2D_INPUT_SIZE], const float kernel[][CONV2D_SIZE], float output[][CONV2D_OUTPUT_SIZE]);
    void maxpool2d(const float input[][MAX_POOL_INPUT_SIZE], float output[][MAX_POOL_OUTPUT_SIZE]);
    void flatten(const float input[][FLATTEN_INPUT_SIZE], float *output);
    void dense(const float *input, float *output);

    static char label_to_char(int label);
//public:
    DeepNeuralNetwork();

    int inference(const int image[][INPUT_IMAGE_SIZE]);

};


#endif //LOWPOWERNEURALNETWORK_DEEPNEURALNETWORK_H