#include "Arduino.h"
#include "embedia.h"
#include "embedia_model.h"
#include "example_file.h"

data3d_t input = { INPUT_CHANNELS, INPUT_WIDTH, INPUT_HEIGHT,  NULL};
data1d_t results;

void setup(){
  // Serial inicialization
  Serial.begin(115200);

  // Model initialization
  model_init();
}

void loop(){
  // sample intitialization
  input.data = sample_data;

  // model initialization
  model_init();

  // make model prediction
  int prediction = model_predict_class(input, &results);

  // print predicted class id
  Serial.print("Prediction class id: ");
  Serial.println(prediction);

  Serial.print("   Example class id: ");
  Serial.println(sample_data_id);

  delay(5000);
}

sketch.ino

diagram.json

/*
 * EmbedIA
 * C LIBRARY FOR THE IMPLEMENTATION OF NEURAL NETWORKS ON MICROCONTROLLERS
 */

#include "embedia.h"


typedef struct{
    size_t  size;
    void  * data;
} raw_buffer;


static raw_buffer buffer1 = {0, NULL};
static raw_buffer buffer2 = {0, NULL};
static raw_buffer * last_buff = &buffer2;

void prepare_buffers(){
    last_buff = &buffer2;
}

void * swap_alloc(size_t s){

    last_buff = (last_buff==&buffer1) ? &buffer2 : &buffer1;
    
    if (last_buff->size < s){
        last_buff->data = realloc(last_buff->data, s);
        last_buff->size = s;
    }
    return last_buff->data;
}

/*
 * conv2d()
 *  Function that performs convolution between a filter and a data set.
 * Parameters:
 *  filter => filter structure with weights loaded.
 *  input => input data of type data3d_t
 *  *output => pointer to the data3d_t structure where the result will be stored
 *  delta => positioning of feature_map inside output.data
 */

static void conv2d(filter_t filter, data3d_t input, data3d_t * output, uint32_t delta){
    uint32_t i,j,k,l,c;
    float suma;

    for(i=0; i<output->height; i++){
        for(j=0; j<output->width; j++){
            suma = 0;
            for(c=0; c<filter.channels; c++){
                for(k=0; k<filter.kernel_size; k++){
                    for(l=0; l<filter.kernel_size; l++){
                        suma += (filter.weights[(c*filter.kernel_size*filter.kernel_size)+k*filter.kernel_size+l] * input.data[(c*input.height*input.width)+(i+k)*input.width+(j+l)]);
                    }
                }
            }
            output->data[delta + i*output->width + j] = suma + filter.bias;
        }
    }

}

/*
 * conv2d_layer()
 *  Function in charge of applying the convolution of a filter layer (conv_layer_t) on a given input data set.
 * Parameters:
 *  layer => convolutional layer with loaded filters.
 *  input => input data of type data3d_t
 *  *output => pointer to the data3d_t structure where the result will be saved.
 */

void conv2d_layer(conv2d_layer_t layer, data3d_t input, data3d_t * output){
    uint32_t delta, i;

    output->channels = layer.n_filters; //cantidad de filtros
    output->height   = input.height - layer.filters[0].kernel_size + 1;
    output->width    = input.width - layer.filters[0].kernel_size + 1;
    output->data     = (float*)swap_alloc( sizeof(float)*output->channels*output->height*output->width );

    for(i=0; i<layer.n_filters; i++){
        delta = i*(output->height)*(output->width);
        conv2d(layer.filters[i],input,output,delta);
    }

}

static void depthwise(filter_t filter, data3d_t input, data3d_t * output){
    uint32_t i,j,k,l,c;
    float suma;

    for(i=0; i<output->height; i++){
        for(j=0; j<output->width; j++){
            for(c=0; c<filter.channels; c++){
                suma=0;
                for(k=0; k<filter.kernel_size; k++){
                    for(l=0; l<filter.kernel_size; l++){
                        suma += (filter.weights[(c*filter.kernel_size*filter.kernel_size)+k*filter.kernel_size+l] * input.data[(c*input.height*input.width)+(i+k)*input.width+(j+l)]);
                    }
                }
                output->data[c*output->width*output->height + i*output->width + j] = suma;
            }
        }
    }
}

static void pointwise(filter_t filter, data3d_t input, data3d_t * output, uint32_t delta){
    uint32_t i,j,c;
    float suma;

    for(i=0; i<output->height; i++){
        for(j=0; j<output->width; j++){
            suma = 0;
            for(c=0; c<filter.channels; c++){
                suma += (filter.weights[c] * input.data[(c*input.height*input.width)+i*input.width+j]);
            }
            output->data[delta + i*output->width + j] = suma + filter.bias;
        }
    }
}

/* 
 * separable_conv2d_layer()
 *  Function in charge of applying the convolution of a filter layer (conv_layer_t) on a given input data set.
 * Parameters:
 *  layer => convolutional layer with loaded filters.
 *  input => input data of type data3d_t
 *  *output => pointer to the data3d_t structure where the result will be saved.
 */

void separable_conv2d_layer(separable_conv2d_layer_t layer, data3d_t input, data3d_t * output){
    uint32_t delta, i;
    data3d_t depth_output;

    depth_output.channels = input.channels; //cantidad de canales
    depth_output.height   = input.height - layer.depth_filter.kernel_size + 1;
    depth_output.width    = input.width - layer.depth_filter.kernel_size + 1;
    depth_output.data     = (float*)swap_alloc( sizeof(float)*depth_output.channels*depth_output.height*depth_output.width );

    depthwise(layer.depth_filter,input,&depth_output);

    output->channels = layer.n_filters; //cantidad de filtros
    output->height   = depth_output.height;
    output->width    = depth_output.width;
    output->data     = (float*)swap_alloc( sizeof(float)*output->channels*output->height*output->width );

    for(i=0; i<layer.n_filters; i++){
        delta = i*(output->height)*(output->width);
        pointwise(layer.point_filters[i],depth_output,output,delta);
    }
}

/*
 * neuron_forward()
 *  Function that performs the forward of a neuron in front of a given set of input data.
 * Parameters:
 *  neuron_t neuron => neuron with its weights and bias loaded.
 *  flatten_data_t input => input data in vector form (flatten_data_t).
 * Returns:
 *  float => result of the operation         
 */

static float neuron_forward(neuron_t neuron, data1d_t input){
    uint32_t i;
    float result = 0;

    for(i=0;i<input.length;i++){
        result += input.data[i]*neuron.weights[i];
    }

    return result + neuron.bias;
}

/*
 * dense_layer()
 *  Performs feed forward of a dense layer (dense_layer_t) on a given input data set.
 * Parameters:
 *  dense_layer => structure with the weights of the neurons of the dense layer.
 *  input       => structure data1d_t with the input data to process.
 *  *output     => structure data1d_t to store the output result.
 */
void dense_layer(dense_layer_t dense_layer, data1d_t input, data1d_t * output){
    uint32_t i;

    output->length = dense_layer.n_neurons;
    output->data = (float*)swap_alloc(sizeof(float)*dense_layer.n_neurons);

    for(i=0;i<dense_layer.n_neurons;i++){
        output->data[i] = neuron_forward(dense_layer.neurons[i],input);
    }
}


/*
 * max_pooling2d_layer()
 *  Maxpooling layer, for now supports square size and stride. No support for padding
 * Parameters:
 *  pool_size => size for pooling
 *  stride    => stride for pooling
 *  input     => input data
 *  output    => output data
 */

void max_pooling2d_layer(pooling2d_layer_t pool, data3d_t input, data3d_t* output){
    uint32_t c,i,j,aux1,aux2;
    float max = -INFINITY;
    float num;

    // output->height = (input.height)/pool_size ;
    // output->width =  (input.width )/pool_size ;
    output->height = ((uint16_t) ((input.height - pool.size)/pool.strides)) + 1;
    output->width  = ((uint16_t) ((input.width - pool.size)/pool.strides)) + 1;
    output->channels = input.channels;
    output->data = (float*)swap_alloc(sizeof(float)*(output->channels)*(output->height)*(output->width));

    for(c=0; c<output->channels; c++){
        for(i=0; i<output->height; i++){
            for(j=0; j<output->width; j++){

                max = -INFINITY;

                for(aux1=0; aux1<pool.size; aux1++){
                        for(aux2=0; aux2<pool.size; aux2++){
                        
                        num = input.data[c*input.width*input.height + (i*pool.strides + aux1)*input.width + j*pool.strides + aux2];
                        
                        if(num>max){
                            max = num;
                        }
                    }
                }

                output->data[c*output->width*output->height + i*output->width + j] = max;
            }
        }    
    }
}


/*
 * avg_pooling_2d()
 *  Function that applies an average pooling to an input with a window size of received
 *  by parameter (uint16_t strides)
 * Parameters:
 *  input => input data of type data3d_t.
 *  *output => pointer to the data3d_t structure where the result will be stored.
 */

void avg_pooling2d_layer(pooling2d_layer_t pool, data3d_t input, data3d_t* output){
    uint32_t c,i,j,aux1,aux2;
    uint32_t cant = pool.size*pool.size;
    float avg = 0;
    float num;

    // output->height = (input.height)/strides ;
    // output->width =  (input.width )/strides ;
    output->height = ((uint32_t) ((input.height - pool.size)/pool.strides)) + 1;
    output->width  = ((uint32_t) ((input.width - pool.size)/pool.strides)) + 1;
    output->channels = input.channels;
    output->data = (float*)swap_alloc(sizeof(float)*(output->channels)*(output->height)*(output->width));

    for(c=0; c<output->channels; c++){
        for(i=0; i<output->height; i++){
            for(j=0; j<output->width; j++){

                avg = 0;

                for(aux1=0; aux1<pool.size; aux1++){
                    for(aux2=0; aux2<pool.size; aux2++){
                        num = input.data[c*input.width*input.height + (i*pool.strides + aux1)*input.width + j*pool.strides + aux2];
                        avg += num;
                    }                  
                }

                output->data[c*output->width*output->height + i*output->width + j] = avg/cant;
            }
        }    
    }
}

/*
 * softmax activation function
 * Parameters:
 *  *data  => array of values to update
 *  length => numbers of values to update
 */
void softmax_activation(float *data, uint32_t length){
    uint32_t i;
    float m = -INFINITY;
    for(i = 0; i < length; i++) {
        if (data[i] > m) {
            m = data[i];
        }
    }

    float sum = (0.0);
    for(i = 0; i < length; i++) {
        sum += exp(data[i] - m);
    }

    float offset = m + log(sum);
    for(i = 0; i < length; i++) {
        data[i] = exp(data[i] - offset);
    }
}


/*
 * relu activation function
 * Parameters:
 *  *data  => array of values to update
 *  length => numbers of values to update
 */
void relu_activation(float *data, uint32_t length){
    uint32_t i;

    for(i=0;i<(length);i++){
        data[i] = data[i] < 0 ? 0 : data[i];
    }
}

/*
 * leaky relu activation function
 * Parameters:
 *  alfa   => coeficient to multiply negative values
 *  *data  => array of values to update
 *  length => numbers of values to update
 */
void leakyrelu_activation(float *data, uint32_t length, float alpha){
    uint32_t i;

    for(i=0;i<(length);i++){
        data[i] = data[i] < 0 ? alpha*data[i] : data[i];
    }
}


/*
 * tanh activation function: (2 / (1+e^(-2x)) -1
 * Parameters:
 *  *data  => array of values to update
 *  length => numbers of values to update
 */
void tanh_activation(float *data, uint32_t length){
    uint32_t i;

    for(i=0;i<length;i++){
        // data.data[i] = tanh(data.data[i]);
        data[i] = 2/(1+exp(-2*data[i])) - 1;
    }
}

/*
 * sigmoid activation function: 1 / (1 + exp(-x))
 * Parameters:
 *  *data  => array of values to update
 *  length => numbers of values to update
 */
void sigmoid_activation(float *data, uint32_t length){
    uint32_t i;

    for(i=0;i<length;i++){
        data[i] = 1 / (1 + exp(-data[i]));
    }
}

/*
 * softsign activation function: x / (|x| + 1)
 * Parameters:
 *  *data  => array of values to update
 *  length => numbers of values to update
 */
void softsign_activation(float *data, uint32_t length){
    uint32_t i;

    for(i=0;i<length;i++){
        data[i] = data[i] / (abs(data[i])+1);
    }
}

/*
 * softplus activation function: log(e^x + 1)
 * Parameters:
 *  *data  => array of values to update
 *  length => numbers of values to update
 */
void softplus_activation(float *data, uint32_t length){
    uint32_t i;

    for(i=0;i<length;i++){
        data[i] = log( exp(data[i])+1 );
    }
}


/*
 * flatten3d_layer()
 *  Performs a variable shape change.
 *  Converts the data format from data3d_t array format to data1d_t vector.
 *  (prepares data for input into a layer of type dense_layer_t).
 * Parameters:
 *  input => input data of type data3d_t.
 *  *output => pointer to the data1d_t structure where the result will be stored.
 */
 void flatten3d_layer(data3d_t input, data1d_t * output){
    uint32_t c,i,j;
    uint32_t cantidad = 0;

    output->length = input.channels * input.height * input.width;
    output->data = (float*)swap_alloc(sizeof(float)*output->length);

    for(i=0;i<input.height;i++){
        for(j=0;j<input.width;j++){
            for(c=0;c<input.channels;c++){
                output->data[cantidad++] = input.data[(c*input.width*input.height)+(i*input.width)+j];
            }
        }
    }
}


/*
 * argmax()
 *  Finds the index of the largest value within a vector of data (data1d_t)
 * Parameters:
 *  data => data of type data1d_t to search for max.
 * Returns:
 *  search result - index of the maximum value
 */

uint32_t argmax(data1d_t data){
    float max = data.data[0];
    uint32_t i, pos = 0;

    for(i=1;i<data.length;i++){
        if(data.data[i]>max){
            max = data.data[i];
            pos = i;
        }
    }

    return pos;
}

/*********************************************************************************************************************************/
/* Normalization layers */


void normalization1(normalization_layer_t n, data1d_t input, data1d_t * output){

    uint32_t i;

    output->length = input.length;
    output->data = (float*)swap_alloc(sizeof(float)*output->length);

    for(i=0; i<input.length; i++){
        output->data[i] = (input.data[i]-n.sub_val[i])*n.inv_div_val[i];
    }
}

void normalization2(normalization_layer_t n, data1d_t input, data1d_t * output){

    uint32_t i;

    output->length = input.length;
    output->data = (float*)swap_alloc(sizeof(float)*output->length);

    for(i=0; i<input.length; i++){
        output->data[i] = input.data[i]*n.inv_div_val[i];
    }
}

/*
 * batch_normalization{X}d_layer()
 * Keras Batch Normalization
 * Parameters:
 *      batch_normlization_t norm =>  structure with batch normalization layer parameters
 *      *data  =>   pointer to data{X}d_t
 */

void batch_normalization1d_layer(batch_normalization_layer_t layer, data1d_t *data) {
    uint32_t i;

    for(i = 0; i < data->length; i++) {
        data->data[i] = data->data[i] * layer.moving_inv_std_dev[i] + layer.std_beta[i];
    }
}


void batch_normalization3d_layer(batch_normalization_layer_t layer, data3d_t *data) {
    uint32_t i, j, ilen = 0;
    uint32_t length = data->height * data->width;

    for(i = 0; i < data->channels; i++, ilen += length) {
        for(j = 0; j < length; j++) {
            data->data[ilen+j] = data->data[ilen+j] * layer.moving_inv_std_dev[i] + layer.std_beta[i];
        }
    }
}


/* image_adapt_layer()
 *  Converts Tensorflow/Keras Image (Height, Width, Channel) to Embedia format (Channel, Height, Width).
 *  Usually required for first convolutional layer
 * Parameters:
 *  input   => input data of type data3d_t.
 *  *output => pointer to the data3d_t structure where the result will be stored.
 */
void image_adapt_layer(data3d_t input, data3d_t * output){

    uint32_t i, j, c, l;

    output->channels = input.channels;
    output->height   = input.height;
    output->width    = input.width;
    output->data     = (float*)swap_alloc( sizeof(float)*output->channels*output->height*output->width );

    for(c=0, l=0; c < input.channels; c++){
        for(i=0; i < input.height; i++) {
            for(j=0; j < input.width; j++, l++ ){
                output->data[l] = input.data[i*input.channels*input.width+input.channels*j+c];
            }
        }
    }
 }

embedia.cpp

/* 
 * EmbedIA 
 * C LIBRARY FOR THE IMPLEMENTATION OF NEURAL NETWORKS ON MICROCONTROLLERS
 */
#ifndef _EMBEDIA_H
#define _EMBEDIA_H

#include <stdint.h>
#include <math.h>
#include "Arduino.h"



/* STRUCTURE DEFINITION */

/*
 * Structure that stores an array of float data (float * data) in vector form.
 * Specifies the number of channels, the width and the height of the array.
 */

typedef struct{
    uint16_t channels;
    uint16_t width;
    uint16_t height;
    float  * data;
}data3d_t;

typedef struct{
    uint16_t width;
    uint16_t height;
    float  * data;
}data2d_t;

typedef struct{
    uint32_t length;
    float  * data;
}data1d_t;


/*
 * Structure that stores the weights of a filter.
 Specifies the number of channels (uint16_t channels), their size (uint16_t kernel_size), the weights (float * weights) and the bias (float bias),
 * the weights (float * weights) and the bias (float bias).
 */
typedef struct{
    uint16_t channels;
    uint16_t kernel_size;
    const float  * weights;
    float  bias; 
}filter_t;

/*
 * Structure that models a convolutional layer.
 * Specifies the number of filters (uint16_t n_filters) and a vector of filters (filter_t * filters). 
 */
typedef struct{
    uint16_t n_filters;
    filter_t * filters; 
}conv2d_layer_t;

/*
 * Structure that models a separable layer...
 * Specifies the number of filters (uint16_t n_filters,
 * a filter of the specified size (filter_t depth_filter) 
 * a vector of 1x1 filters (filter_t * point_filters) 
 */
typedef struct{
    uint16_t n_filters;
    filter_t depth_filter;
    filter_t * point_filters; 
}separable_conv2d_layer_t;

/*
 * Structure that models a neuron.
 * Specifies the weights of the neuron as a vector (fixed * weights) and the bias (fixed bias).
 */
typedef struct{
    const float  * weights;
    float  bias;
}neuron_t;

/*
 * Structure that models a dense layer.
 * Specifies the number of neurons (uint16_t n_neurons) and a vector of neurons (neuron_t * neurons). 
 */
typedef struct{
    uint16_t n_neurons;
    neuron_t * neurons;
}dense_layer_t;


/*
 * Pooling Structure
 */
 
typedef struct{
    uint16_t size;
    uint16_t strides;
} pooling2d_layer_t;

/*********************************************************************************************************************************/

/* structure for normalization type (x_i-s_i)/d_i and (x_i)/d_i
 *  standard normalization  : (x_i-mean_i)/std_dev_i
 *  min_max normalization   : (x_i-min_i) / (max_i-min_i)
 *  robust normalization    : (x_i-q2_i)  / (q3_i-q1_i)
 *  abs_max_normalization   : (x_i)/(abs_max_xi)
 */

typedef struct{
    const float *sub_val;
    const float *inv_div_val;
} normalization_layer_t;


/* 
 * Structure for BatchNormalization layer.
 * Contains vectors for the two parameters used for normalization.
 * The number of each of the parameters is determined by the number of channels of the previous layer.
 */
typedef struct {
    uint32_t length;
    const float *moving_inv_std_dev; // = gamma / sqrt(moving_variance + epsilon)
    const float *std_beta;           // = beta - moving_mean * moving_inv_std_dev 
} batch_normalization_layer_t;



/* LIBRARY FUNCTIONS PROTOTYPES */


/*
 * prepare_buffers()
 *  This function should be invoked only at the beginning of the predict function of the model file.
 * Its purpose is to align the exchange buffers used by the different functions of the model. Due to
 * the allocation strategy that never frees the memory, it happens that if the swap_alloc function
 * is invoked an odd number of times in the 2nd invocation the predict reserves more memory than
 * necessary  (something that usually happens with convolutional layers)
 */
void prepare_buffers();


/*
 * conv2d_layer()
 *  Function in charge of applying the convolution of a filter layer (conv_layer_t) on a given input data set.
 * Parameters:
 *  layer => convolutional layer with loaded filters.
 *  input => input data of type data3d_t
 *  *output => pointer to the data3d_t structure where the result will be saved.
 */
void conv2d_layer(conv2d_layer_t layer, data3d_t input, data3d_t * output);

/* 
 * separable_conv2d_layer()
 *  Function in charge of applying the convolution of a filter layer (conv_layer_t) on a given input data set.
 * 
 * Parameters:
 *   layer => convolutional layer with loaded filters.
 *   input => input data of type data3d_t
 *   *output => pointer to the data3d_t structure where the result will be saved.
 */
void separable_conv2d_layer(separable_conv2d_layer_t layer, data3d_t input, data3d_t * output);


/* 
 * dense_layer()
 * Performs feed forward of a dense layer (dense_layer_t) on a given input data set.
 * Parameters:
 *   dense_layer => structure with the weights of the neurons of the dense layer.  
 *   input       => structure data1d_t with the input data to process. 
 *   *output     => structure data1d_t to store the output result.
 */
void dense_layer(dense_layer_t dense_layer, data1d_t input, data1d_t * output);

/* 
 * max_pooling2d_layer()
 * Maxpooling layer, for now supports square size and stride. No support for padding 
 * Parameters:
 *   pool_size => size for pooling
 *   stride    => stride for pooling
 *   input     => input data 
 *   output    => output data
 */
void max_pooling2d_layer(pooling2d_layer_t pool, data3d_t input, data3d_t* output);

/* 
 * avg_pooling_2d()
 * Function that applies an average pooling to an input with a window size of received 
 * by parameter (uint16_t strides)
 *
 * Parameters:
 *   input => input data of type data3d_t.
 *   *output => pointer to the data3d_t structure where the result will be stored.
 */
void avg_pooling2d_layer(pooling2d_layer_t pool, data3d_t input, data3d_t* output);


/* 
 * flatten3d_layer()
 * Performs a variable shape change. 
 * Converts the data format from data3d_t array format to data1d_t vector.
 * (prepares data for input into a layer of type dense_layer_t).
 * Parameters:
 *    input => input data of type data3d_t.
 *    *output => pointer to the data1d_t structure where the result will be stored.
 */
 void flatten3d_layer(data3d_t input, data1d_t * output);
     
/* 
 * argmax()
 * Finds the index of the largest value within a vector of data (data1d_t)
 * Parameters:
 *   data => data of type data1d_t to search for max.
 *
 * Returns:
 *  search result - index of the maximum value
 */
uint32_t argmax(data1d_t data);


/***************************************************************************************************************************/
/* Activation functions/layers */

void softmax_activation(float *data, uint32_t length);

void relu_activation(float *data, uint32_t length);

void leakyrelu_activation(float *data, uint32_t length, float alpha);

void tanh_activation(float *data, uint32_t length);

void sigmoid_activation(float *data, uint32_t length);

void softsign_activation(float *data, uint32_t length);



/***************************************************************************************************************************/
/* Normalization layers */

/* Normalization function for:
 *  standard normalization  : (x_i-mean_i)/std_dev_i
 *  min_max normalization   : (x_i-min_i) / (max_i-min_i)
 *  robust normalization    : (x_i-q2_i)  / (q3_i-q1_i)
 */
void normalization1(normalization_layer_t s, data1d_t input, data1d_t * output);

#define standard_norm_layer(norm, input, output) normalization1(norm, input, output)

#define min_max_norm_layer(norm, input, output) normalization1(norm, input, output)

#define robust_norm_layer(norm, input, output) normalization1(norm, input, output)


/* Normalization function for:
 *  abs_max_normalization   : (x_i)/(abs_max_xi)
 */
void normalization2(normalization_layer_t s, data1d_t input, data1d_t * output);

#define max_abs_norm_layer(norm, input, output) normalization2(norm, input, output)


void batch_normalization_layer(batch_normalization_layer_t norm, uint32_t length, float *data);


void batch_normalization3d_layer(batch_normalization_layer_t layer, data3d_t *data);

void batch_normalization1d_layer(batch_normalization_layer_t layer, data1d_t *data);


/* Tranformation Layers
 *
 */

/* Converts Tensorflow/Keras Image (Height, Width, Channel) to Embedia format (Channel, Height, Width).
   Usually required for first convolutional layer
*/
void image_adapt_layer(data3d_t input, data3d_t * output);


#endif

embedia.h

#include "embedia_debug.h"
#include "embedia_debug_def.h"


const char* STR_CONTENT = "  Content:";
const char* SEP_LINE = "";
const char* STR_CHN = "MCh: ";
const char* STR_MHE = " - MHe: ";
const char* STR_MWI = " - MWi: ";
const char* STR_LEN = "  Length : ";
const char* STR_FSZ = " - F.K_size: ";
const char* STR_FBS = " - F.bias: ";



void print_datad2_t(const char *head_text, data2d_t data){

    #if EMBEDIA_DEBUG > 0

    PRINT_TXT_LN(head_text);

    PRINT_INT(STR_MHE, data.height);
    PRINT_INT_LN(STR_MWI,data.width);

    #if EMBEDIA_DEBUG > 1

    PRINT_TXT_LN(STR_CONTENT);

    uint16_t h,w;
    for(h=0;h<data.height;h++){
        for(w=0;w<data.width;w++){
            PRINT_FL("", data.data[data.width+h*data.width+w] );
        }
        PRINT_TXT_LN("");
    }
    #endif // EMBEDIA_DEBUG 1

    PRINT_TXT_LN(SEP_LINE);
    
    #endif // EMBEDIA_DEBUG 0


}

void print_data3d_t(const char *head_text, data3d_t data){


    #if EMBEDIA_DEBUG > 0

    PRINT_TXT_LN(head_text);

    PRINT_INT(STR_CHN, data.channels);
    PRINT_INT(STR_MHE, data.height);
    PRINT_INT_LN(STR_MWI,data.width);

    #if EMBEDIA_DEBUG > 1

    PRINT_TXT_LN(STR_CONTENT);

    uint16_t c,h,w;
    for(c=0;c<data.channels;c++){
        for(h=0;h<data.height;h++){
            for(w=0;w<data.width;w++){
                PRINT_FL("", data.data[c*data.height*data.width+h*data.width+w] );
            }
            PRINT_TXT_LN("");
            }
        PRINT_TXT_LN("");
    }
    #endif // EMBEDIA_DEBUG 1

    PRINT_TXT_LN(SEP_LINE);
    
    #endif // EMBEDIA_DEBUG 0

}

/*
 * print_flatten_data_t()
 * Imprime los valores presentes en un vector de datos y su largo
 * Parámetros:
 *            flatten_data_t data => vector de datos a imprimir
 */
void print_data1d_t(const char *head_text, data1d_t data){


    #if EMBEDIA_DEBUG > 0

    PRINT_TXT_LN(head_text);

    PRINT_INT_LN(STR_LEN, data.length);

    #if EMBEDIA_DEBUG > 1

    PRINT_TXT_LN(STR_CONTENT);

    uint16_t i;
    for(i=0;i<data.length;i++){
        PRINT_FL("", data.data[i]);
    }
    PRINT_TXT_LN("");

    #endif // EMBEDIA_DEBUG 1

    PRINT_TXT_LN(SEP_LINE);

    #endif // EMBEDIA_DEBUG 0

}


/*
 * print_filter_t()
 * Imprime los valores de los pesos del filter y sus dimensiones
 * Parámetros:
 *                filter_t filter => filter a imprimir
 */
void print_filter_t(const char *head_text, filter_t filter){

    #if EMBEDIA_DEBUG > 0
    
    PRINT_TXT_LN(head_text);

    PRINT_INT_LN(STR_CHN, filter.channels);
    PRINT_INT_LN(STR_FSZ, filter.kernel_size);
    PRINT_FL_LN(STR_FBS, filter.bias);

    #if EMBEDIA_DEBUG > 1

    printf(STR_CONTENT);

    uint16_t c,h,w;
    for(c=0;c<filter.channels;c++){
        for(h=0;h<filter.kernel_size;h++){
            for(w=0;w<filter.kernel_size;w++){
                PRINT_FL("", DBG_FL(filter.weights[h*filter.kernel_size+w]));
            }
            PRINT_TXT_LN("");
        }
        PRINT_TXT_LN("");
    }
    #endif // EMBEDIA_DEBUG 1

    PRINT_TXT_LN(SEP_LINE);
    
    #endif // EMBEDIA_DEBUG 0
}



embedia_debug.cpp

#ifndef EMBEDIA_DEBUG_H_INCLUDED
#define EMBEDIA_DEBUG_H_INCLUDED

#include "embedia.h"

// exporter must define EMBEDIA_DEBUG macro
// 0 => NO DEBUG, 1 => DATA HEADER, 2 => DATA CONTENT
#define EMBEDIA_DEBUG 2





void print_data1d_t(const char *head_text, data1d_t data);

void print_data2d_t(const char *head_text, data2d_t data);

void print_data3d_t(const char *head_text, data3d_t data);

/*
 * print_filter_t()
 * Imprime los valores de los pesos del filtro y sus dimensiones
 * Parámetros:
 *                filter_t filtro => filtro a imprimir
 */

void print_filter_t(const char *head_text, filter_t filter);


#endif // EMBEDIA_DEBUG_H_INCLUDED


embedia_debug.h

#include <Arduino.h>

#ifdef FIX_SIZE
#define DBG_FL(x) FX2FL(x)
#else
#define DBG_FL(x) x
#endif // FIX_SIZE

#define DBG_INT_PART 4
#define DBG_FRC_PART 6
#define DBG_FLOAT_SIZE (DBG_INT_PART+DBG_FRC_PART+1)

#if EMBEDIA_DEBUG > 0
static char temp_str[DBG_FLOAT_SIZE+1];
#endif

#if defined (__AVR__) || (__avr__) || (ARDUINO_ARCH_AVR)
    // sprintf of many AVR doesn't support float numbers
    #define PRINT_FL(txt, fl)                                     \
        Serial.print(txt);                                        \
        dtostrf(DBG_FL(fl),DBG_FLOAT_SIZE,DBG_FRC_PART,temp_str); \
        Serial.print(temp_str)

    #define PRINT_FL_LN(txt, fl)                                  \
        Serial.print(txt);                                        \
        dtostrf(DBG_FL(fl),DBG_FLOAT_SIZE,DBG_FRC_PART,temp_str); \
        Serial.println(temp_str)
#else

   #define DBG_FLOAT_FMT "%11.6f"

   #define PRINT_FL(txt, fl)                        \
      Serial.print(txt);                            \
      sprintf(temp_str, DBG_FLOAT_FMT, DBG_FL(fl)); \
      Serial.print(temp_str)

   #define PRINT_FL_LN(txt, fl)                     \
      Serial.print(txt);                            \
      sprintf(temp_str, DBG_FLOAT_FMT, DBG_FL(fl)); \
      Serial.println(temp_str)

#endif // AVR


#define PRINT_INT(txt, n)           \
    Serial.print(txt);              \
    Serial.print(n)

#define PRINT_INT_LN(txt, n)        \
    Serial.print(txt);              \
    Serial.println(n)

#define PRINT_TXT_LN(txt)           \
    Serial.println(txt)


embedia_debug_def.h

embedia_model.cpp

/* EmbedIA model definition file*/
#ifndef EMBEDIA_MODEL_H
#define EMBEDIA_MODEL_H

/*

+-------------------+---------------+------------+------------+------+------------+
| Layer(activation) | Name          | #Param(NT) |   Shape    | MACs | Size (KiB) |
+-------------------+---------------+------------+------------+------+------------+
| Conv2D(relu)      | conv2d        |         80 | (6, 6, 8)  | 2592 |     0.375  |
| MaxPooling2D      | max_pooling2d |          0 | (3, 3, 8)  |    0 |     0.000  |
| Conv2D(relu)      | conv2d_1      |        528 | (2, 2, 16) | 2048 |     2.188  |
| Flatten           | flatten       |          0 |   (64,)    |    0 |     0.000  |
| Dense(relu)       | dense         |       1040 |   (16,)    | 1024 |     4.125  |
| Dense(softmax)    | dense_1       |        170 |   (10,)    |  160 |     0.703  |
+-------------------+---------------+------------+------------+------+------------+
Total params (NT)....: 1818
Total size in KiB....: 7.391
Total MACs operations: 5824

*/

#include "embedia.h"

#define INPUT_CHANNELS 1
#define INPUT_WIDTH 8
#define INPUT_HEIGHT 8

#define INPUT_SIZE 64


void model_init();

void model_predict(data3d_t input, data1d_t * output);

int model_predict_class(data3d_t input, data1d_t * results);

#endif


embedia_model.h

#ifndef _EXAMPLE_FILE_H
#define _EXAMPLE_FILE_H

#include "embedia.h"



#define MAX_SAMPLE 9

#define SELECT_SAMPLE 1

#if SELECT_SAMPLE == 0
uint16_t sample_data_id = 8;

static float sample_data[]= {
  -0.8136885985884057, -0.8136885985884057, 0.18298813446391424, 1.511890445200341, 1.8441160228844478, 0.016875345621860906, 
  -0.8136885985884057, -0.8136885985884057, -0.8136885985884057, -0.4814630209042991, 1.8441160228844478, 1.8441160228844478, 
  1.8441160228844478, 0.3491009233059676, -0.8136885985884057, -0.8136885985884057, -0.8136885985884057, -0.4814630209042991, 
  1.6780032340423943, 1.8441160228844478, 1.6780032340423943, -0.4814630209042991, -0.8136885985884057, -0.8136885985884057, 
  -0.8136885985884057, -0.8136885985884057, 0.18298813446391424, 1.8441160228844478, 1.6780032340423943, 0.3491009233059676, 
  -0.8136885985884057, -0.8136885985884057, -0.8136885985884057, -0.8136885985884057, 1.511890445200341, 0.8474392898321276, 
  0.18298813446391424, 1.8441160228844478, -0.31535023206224577, -0.8136885985884057, -0.8136885985884057, -0.6475758097463524, 
  1.8441160228844478, -0.31535023206224577, -0.8136885985884057, 1.8441160228844478, 0.3491009233059676, -0.8136885985884057, 
  -0.8136885985884057, -0.8136885985884057, 0.8474392898321276, 1.013552078674181, 1.013552078674181, 1.6780032340423943, 
  -0.31535023206224577, -0.8136885985884057, -0.8136885985884057, -0.8136885985884057, -0.31535023206224577, 1.511890445200341, 
  1.8441160228844478, 0.18298813446391424, -0.8136885985884057, -0.8136885985884057
};

#endif
#if SELECT_SAMPLE == 1
uint16_t sample_data_id = 5;

static float sample_data[]= {
  -0.8136885985884057, -0.8136885985884057, 1.3457776563582877, 1.8441160228844478, 1.8441160228844478, 1.8441160228844478, 
  1.3457776563582877, -0.8136885985884057, -0.8136885985884057, -0.14923744322019242, 1.8441160228844478, 0.6813265009900742, 
  0.5152137121480209, 0.016875345621860906, -0.14923744322019242, -0.8136885985884057, -0.8136885985884057, 0.6813265009900742, 
  1.6780032340423943, 0.3491009233059676, 0.5152137121480209, -0.4814630209042991, -0.8136885985884057, -0.8136885985884057, 
  -0.8136885985884057, 1.013552078674181, 1.8441160228844478, 1.8441160228844478, 1.511890445200341, 1.6780032340423943, 
  -0.6475758097463524, -0.8136885985884057, -0.8136885985884057, -0.6475758097463524, -0.31535023206224577, -0.8136885985884057, 
  -0.14923744322019242, 1.8441160228844478, -0.14923744322019242, -0.8136885985884057, -0.8136885985884057, -0.8136885985884057, 
  -0.8136885985884057, -0.8136885985884057, 0.5152137121480209, 1.511890445200341, -0.8136885985884057, -0.8136885985884057, 
  -0.8136885985884057, -0.8136885985884057, 1.013552078674181, 0.5152137121480209, 1.8441160228844478, 0.18298813446391424, 
  -0.8136885985884057, -0.8136885985884057, -0.8136885985884057, -0.6475758097463524, 1.6780032340423943, 1.8441160228844478, 
  0.8474392898321276, -0.8136885985884057, -0.8136885985884057, -0.8136885985884057
};

#endif
#if SELECT_SAMPLE == 2
uint16_t sample_data_id = 0;

static float sample_data[]= {
  -0.8136885985884057, -0.8136885985884057, -0.4814630209042991, 1.6780032340423943, 0.016875345621860906, -0.8136885985884057, 
  -0.8136885985884057, -0.8136885985884057, -0.8136885985884057, -0.8136885985884057, 1.013552078674181, 1.8441160228844478, 
  1.8441160228844478, 0.18298813446391424, -0.8136885985884057, -0.8136885985884057, -0.8136885985884057, -0.31535023206224577, 
  1.511890445200341, -0.31535023206224577, 0.3491009233059676, 1.8441160228844478, -0.31535023206224577, -0.8136885985884057, 
  -0.8136885985884057, 0.3491009233059676, 0.5152137121480209, -0.8136885985884057, -0.8136885985884057, 0.5152137121480209, 
  0.5152137121480209, -0.8136885985884057, -0.8136885985884057, 0.016875345621860906, 0.5152137121480209, -0.8136885985884057, 
  -0.8136885985884057, -0.14923744322019242, 0.5152137121480209, -0.8136885985884057, -0.8136885985884057, -0.14923744322019242, 
  1.1796648675162342, -0.8136885985884057, -0.8136885985884057, 0.5152137121480209, 0.5152137121480209, -0.8136885985884057, 
  -0.8136885985884057, -0.8136885985884057, 1.511890445200341, 0.6813265009900742, 0.5152137121480209, 1.8441160228844478, 
  -0.4814630209042991, -0.8136885985884057, -0.8136885985884057, -0.8136885985884057, -0.31535023206224577, 1.511890445200341, 
  1.6780032340423943, 0.18298813446391424, -0.8136885985884057, -0.8136885985884057
};

#endif
#if SELECT_SAMPLE == 3
uint16_t sample_data_id = 2;

static float sample_data[]= {
  -0.8136885985884057, -0.4814630209042991, 1.6780032340423943, 1.3457776563582877, -0.4814630209042991, -0.8136885985884057, 
  -0.8136885985884057, -0.8136885985884057, -0.8136885985884057, 0.3491009233059676, 1.8441160228844478, 1.3457776563582877, 
  1.6780032340423943, -0.8136885985884057, -0.8136885985884057, -0.8136885985884057, -0.8136885985884057, 1.013552078674181, 
  1.013552078674181, 0.016875345621860906, 1.8441160228844478, -0.14923744322019242, -0.8136885985884057, -0.8136885985884057, 
  -0.8136885985884057, -0.6475758097463524, -0.6475758097463524, 0.3491009233059676, 1.8441160228844478, -0.6475758097463524, 
  -0.8136885985884057, -0.8136885985884057, -0.8136885985884057, -0.8136885985884057, -0.8136885985884057, 1.1796648675162342, 
  1.013552078674181, -0.8136885985884057, -0.8136885985884057, -0.8136885985884057, -0.8136885985884057, -0.8136885985884057, 
  -0.14923744322019242, 1.8441160228844478, 0.5152137121480209, 0.5152137121480209, -0.4814630209042991, -0.8136885985884057, 
  -0.8136885985884057, -0.8136885985884057, 1.1796648675162342, 1.8441160228844478, 1.8441160228844478, 1.8441160228844478, 
  1.013552078674181, -0.8136885985884057, -0.8136885985884057, -0.4814630209042991, 1.6780032340423943, 1.3457776563582877, 
  0.18298813446391424, -0.14923744322019242, -0.6475758097463524, -0.8136885985884057
};

#endif
#if SELECT_SAMPLE == 4
uint16_t sample_data_id = 1;

static float sample_data[]= {
  -0.8136885985884057, -0.8136885985884057, -0.8136885985884057, -0.8136885985884057, 0.016875345621860906, 1.6780032340423943, 
  -0.31535023206224577, -0.8136885985884057, -0.8136885985884057, -0.8136885985884057, -0.8136885985884057, -0.31535023206224577, 
  1.6780032340423943, 1.8441160228844478, -0.14923744322019242, -0.8136885985884057, -0.8136885985884057, -0.31535023206224577, 
  1.3457776563582877, 1.8441160228844478, 1.511890445200341, 1.8441160228844478, -0.6475758097463524, -0.8136885985884057, 
  -0.8136885985884057, -0.4814630209042991, 0.3491009233059676, -0.14923744322019242, 0.5152137121480209, 1.8441160228844478, 
  -0.8136885985884057, -0.8136885985884057, -0.8136885985884057, -0.8136885985884057, -0.8136885985884057, -0.8136885985884057, 
  0.5152137121480209, 1.511890445200341, -0.8136885985884057, -0.8136885985884057, -0.8136885985884057, -0.8136885985884057, 
  -0.8136885985884057, -0.8136885985884057, 0.3491009233059676, 1.8441160228844478, -0.6475758097463524, -0.8136885985884057, 
  -0.8136885985884057, -0.8136885985884057, -0.8136885985884057, -0.8136885985884057, 0.6813265009900742, 1.8441160228844478, 
  0.18298813446391424, -0.8136885985884057, -0.8136885985884057, -0.8136885985884057, -0.8136885985884057, -0.8136885985884057, 
  0.5152137121480209, 1.6780032340423943, -0.4814630209042991, -0.8136885985884057
};

#endif
#if SELECT_SAMPLE == 5
uint16_t sample_data_id = 5;

static float sample_data[]= {
  -0.8136885985884057, -0.8136885985884057, 1.013552078674181, 1.1796648675162342, 1.3457776563582877, 1.511890445200341, 
  -0.14923744322019242, -0.8136885985884057, -0.8136885985884057, -0.8136885985884057, 1.3457776563582877, 0.5152137121480209, 
  -0.14923744322019242, -0.14923744322019242, -0.4814630209042991, -0.8136885985884057, -0.8136885985884057, -0.8136885985884057, 
  1.013552078674181, -0.8136885985884057, -0.8136885985884057, -0.8136885985884057, -0.8136885985884057, -0.8136885985884057, 
  -0.8136885985884057, 0.016875345621860906, 1.1796648675162342, 0.18298813446391424, -0.31535023206224577, -0.8136885985884057, 
  -0.8136885985884057, -0.8136885985884057, -0.8136885985884057, 0.016875345621860906, 1.1796648675162342, 1.1796648675162342, 
  1.3457776563582877, 1.013552078674181, -0.8136885985884057, -0.8136885985884057, -0.8136885985884057, -0.8136885985884057, 
  -0.4814630209042991, -0.8136885985884057, -0.8136885985884057, 1.3457776563582877, -0.8136885985884057, -0.8136885985884057, 
  -0.8136885985884057, -0.14923744322019242, 1.013552078674181, -0.8136885985884057, -0.31535023206224577, 1.6780032340423943, 
  -0.8136885985884057, -0.8136885985884057, -0.8136885985884057, -0.8136885985884057, 0.6813265009900742, 1.8441160228844478, 
  1.8441160228844478, 0.18298813446391424, -0.8136885985884057, -0.8136885985884057
};

#endif
#if SELECT_SAMPLE == 6
uint16_t sample_data_id = 2;

static float sample_data[]= {
  -0.8136885985884057, -0.8136885985884057, -0.6475758097463524, 1.3457776563582877, 1.8441160228844478, 0.8474392898321276, 
  -0.8136885985884057, -0.8136885985884057, -0.8136885985884057, -0.6475758097463524, 1.3457776563582877, 1.6780032340423943, 
  0.5152137121480209, 1.8441160228844478, -0.31535023206224577, -0.8136885985884057, -0.8136885985884057, 0.5152137121480209, 
  1.6780032340423943, -0.31535023206224577, -0.14923744322019242, 1.6780032340423943, -0.8136885985884057, -0.8136885985884057, 
  -0.8136885985884057, -0.6475758097463524, -0.31535023206224577, -0.8136885985884057, 1.1796648675162342, 0.5152137121480209, 
  -0.8136885985884057, -0.8136885985884057, -0.8136885985884057, -0.8136885985884057, -0.8136885985884057, -0.14923744322019242, 
  1.511890445200341, -0.6475758097463524, -0.8136885985884057, -0.8136885985884057, -0.8136885985884057, -0.8136885985884057, 
  -0.8136885985884057, 1.013552078674181, 0.5152137121480209, -0.8136885985884057, -0.14923744322019242, -0.8136885985884057, 
  -0.8136885985884057, -0.8136885985884057, -0.6475758097463524, 1.8441160228844478, 0.5152137121480209, 1.3457776563582877, 
  0.6813265009900742, -0.8136885985884057, -0.8136885985884057, -0.8136885985884057, -0.8136885985884057, 1.511890445200341, 
  1.8441160228844478, 1.013552078674181, -0.8136885985884057, -0.8136885985884057
};

#endif
#if SELECT_SAMPLE == 7
uint16_t sample_data_id = 3;

static float sample_data[]= {
  -0.8136885985884057, -0.8136885985884057, 0.3491009233059676, 1.6780032340423943, 1.3457776563582877, -0.6475758097463524, 
  -0.8136885985884057, -0.8136885985884057, -0.8136885985884057, 0.5152137121480209, 1.3457776563582877, 0.18298813446391424, 
  1.6780032340423943, -0.14923744322019242, -0.8136885985884057, -0.8136885985884057, -0.8136885985884057, -0.4814630209042991, 
  -0.6475758097463524, 1.3457776563582877, 1.3457776563582877, -0.8136885985884057, -0.8136885985884057, -0.8136885985884057, 
  -0.8136885985884057, -0.8136885985884057, -0.4814630209042991, 1.6780032340423943, 1.013552078674181, -0.6475758097463524, 
  -0.8136885985884057, -0.8136885985884057, -0.8136885985884057, -0.8136885985884057, -0.8136885985884057, -0.6475758097463524, 
  1.1796648675162342, 1.1796648675162342, -0.6475758097463524, -0.8136885985884057, -0.8136885985884057, -0.8136885985884057, 
  -0.8136885985884057, -0.8136885985884057, -0.6475758097463524, 0.8474392898321276, 0.5152137121480209, -0.8136885985884057, 
  -0.8136885985884057, -0.8136885985884057, 0.5152137121480209, -0.14923744322019242, 0.016875345621860906, 1.511890445200341, 
  0.6813265009900742, -0.8136885985884057, -0.8136885985884057, -0.8136885985884057, 0.3491009233059676, 1.3457776563582877, 
  1.3457776563582877, 0.6813265009900742, -0.8136885985884057, -0.8136885985884057
};

#endif
#if SELECT_SAMPLE == 8
uint16_t sample_data_id = 0;

static float sample_data[]= {
  -0.8136885985884057, -0.8136885985884057, -0.8136885985884057, 0.6813265009900742, 0.5152137121480209, -0.6475758097463524, 
  -0.8136885985884057, -0.8136885985884057, -0.8136885985884057, -0.8136885985884057, 0.016875345621860906, 1.8441160228844478, 
  1.8441160228844478, 0.5152137121480209, -0.8136885985884057, -0.8136885985884057, -0.8136885985884057, -0.14923744322019242, 
  1.8441160228844478, 1.6780032340423943, -0.6475758097463524, 1.6780032340423943, -0.8136885985884057, -0.8136885985884057, 
  -0.8136885985884057, 0.18298813446391424, 1.8441160228844478, 1.1796648675162342, -0.8136885985884057, 1.1796648675162342, 
  -0.6475758097463524, -0.8136885985884057, -0.8136885985884057, 0.016875345621860906, 1.8441160228844478, 1.013552078674181, 
  -0.8136885985884057, 1.013552078674181, 0.18298813446391424, -0.8136885985884057, -0.8136885985884057, -0.6475758097463524, 
  1.6780032340423943, 0.5152137121480209, -0.14923744322019242, 1.6780032340423943, 0.18298813446391424, -0.8136885985884057, 
  -0.8136885985884057, -0.8136885985884057, 0.016875345621860906, 1.8441160228844478, 1.8441160228844478, 1.6780032340423943, 
  -0.8136885985884057, -0.8136885985884057, -0.8136885985884057, -0.8136885985884057, -0.8136885985884057, 0.3491009233059676, 
  1.511890445200341, 0.6813265009900742, -0.8136885985884057, -0.8136885985884057
};

#endif
#if SELECT_SAMPLE == 9
uint16_t sample_data_id = 7;

static float sample_data[]= {
  -0.8136885985884057, -0.8136885985884057, 0.8474392898321276, 1.8441160228844478, 1.6780032340423943, -0.8136885985884057, 
  -0.8136885985884057, -0.8136885985884057, -0.8136885985884057, -0.14923744322019242, 1.511890445200341, 0.5152137121480209, 
  1.8441160228844478, -0.6475758097463524, -0.8136885985884057, -0.8136885985884057, -0.8136885985884057, -0.8136885985884057, 
  -0.6475758097463524, -0.14923744322019242, 1.8441160228844478, -0.8136885985884057, -0.4814630209042991, -0.8136885985884057, 
  -0.8136885985884057, -0.8136885985884057, -0.31535023206224577, 1.013552078674181, 1.8441160228844478, 1.8441160228844478, 
  1.3457776563582877, -0.8136885985884057, -0.8136885985884057, -0.8136885985884057, 1.1796648675162342, 1.8441160228844478, 
  1.013552078674181, 0.3491009233059676, -0.4814630209042991, -0.8136885985884057, -0.8136885985884057, -0.8136885985884057, 
  0.18298813446391424, 1.8441160228844478, -0.8136885985884057, -0.8136885985884057, -0.8136885985884057, -0.8136885985884057, 
  -0.8136885985884057, -0.8136885985884057, 0.3491009233059676, 1.6780032340423943, -0.8136885985884057, -0.8136885985884057, 
  -0.8136885985884057, -0.8136885985884057, -0.8136885985884057, -0.8136885985884057, 1.1796648675162342, 1.013552078674181, 
  -0.8136885985884057, -0.8136885985884057, -0.8136885985884057, -0.8136885985884057
};

#endif


#endif