[GRASSLIST:1902] RE: Neural Net

[Martin Wass]

Dear List !

How do I make a snns pat file ?

I have 8 layer in a raster group and want to use this layers ass
input units.

I believe the way should look simular to this nonsense

1 r.mask (some good training area)
2 r.stat -xxxxxx
3 awk {%%% } printf(/nt  #)
4 repet fore the other layers and other train area


No It looks more like this at lesst for an LNKnet pat file

1 r.mask area
2 r.stat -l input=lay1,lay2...... | awk '{printf("class_int %d %d %d %d %d 
%d %d...more %d\n",$1,$2,$3,$3,$4,$5,$6,$7,$8.....more layer)}'
3 repete for other areas

But the rel problem is how to bulid in the the C soure that LNKnet
produse.


Please GRASS C writes. LNKnet produse C soures classyfy funktion
How make this funktion classyfy Grass raster data ????????

AN MLP 2_input,25_hidden,10_output  example

/* X2mlp.param */
/* norm: SIMPLE */
/* 2 (all) features */

/* cost_func: SQ_ERR */
/* sigm: SIGM */
/* net: 2,25,10 */

#include <stdio.h>
#include <math.h>


/* function declarations */
extern int classify_X2mlp(/* float *,float * */);
extern int normalize_X2mlp(/* float * */);


#define NRAW_X2mlp 2
#define NINPUTS_X2mlp 2
#define NH1_X2mlp 25
#define NCLASSES_X2mlp 10

int classify_X2mlp (inputs, outputs)
   float *inputs,*outputs;
{
  int n, best, j, k;
  float h1[NH1_X2mlp];
  float y[NCLASSES_X2mlp];

static float x_h1[NH1_X2mlp+1][NINPUTS_X2mlp+1] = {
        {  1.048780e+00,  -2.359010e+00,  -1.970170e+00, },
.......... more floats .................
        {  4.419050e-02,  -4.314130e-02,   4.770690e-02, },
};

static float h1_y[NCLASSES_X2mlp+1][NH1_X2mlp+1] = {
        { -1.312270e+00,  -3.287920e-01,  -1.603640e-01,  -3.694100e-01,
   ...... simular flots are boring and take lots of space .......
   6.598780e-02,  -2.112230e-02, },
};
  float x[NRAW_X2mlp];
  int i;

  /* load inputs */
  for(i = 0; i < NRAW_X2mlp; i++)
    x[i] = inputs[i];

  /* normalize loaded data */
  normalize_X2mlp(x);

  /* calculate hidden layer outputs */
  for(j = 0; j < NH1_X2mlp; j++){
    h1[j] = x_h1[j][NINPUTS_X2mlp];
    for(k = 0; k < NINPUTS_X2mlp; k++){
      h1[j] += x[k] * x_h1[j][k];
    }
  }
  /* take sigmoids */
  for(j = 0; j < NH1_X2mlp; j++){
    h1[j] = 1. / (1. + exp(- 1.000000 * h1[j]));
  }

  /* calculate outputs */
  for(j = 0; j < NCLASSES_X2mlp; j++){
    y[j] = h1_y[j][NH1_X2mlp];
    for(k = 0; k < NH1_X2mlp; k++){
      y[j] += h1[k] * h1_y[j][k];
    }
  }
  /* take output sigmoid */
  for(j = 0; j < NCLASSES_X2mlp; j++){
    y[j] = 1. / (1. + exp(- 1.000000 * y[j]));
  }

  /* copy outputs */
  for(n = 0; n < NCLASSES_X2mlp; n++)
    outputs[n] = y[n];

  /* find highest output */
  for(best = n = 0; n < NCLASSES_X2mlp; n++)
    if(outputs[best] < outputs[n]) best = n;
  return(best);
}


int normalize_X2mlp(inputs)
   float *inputs;
{
  int n;
  float new_inputs[NRAW_X2mlp];

static float means[NINPUTS_X2mlp] = {
          4.055730e-01,   4.734620e-01, };

static float sdevs[NINPUTS_X2mlp] = {
          1.496270e-01,   2.022970e-01, };

  for(n = 0; n < NINPUTS_X2mlp; n++){
    new_inputs[n] = (inputs[n]- means[n]) / sdevs[n];
  }

  for(n = 0; n < NINPUTS_X2mlp; n++){
    inputs[n] = new_inputs[n];
  }
  return((int) NINPUTS_X2mlp);
}
#undef NRAW_X2mlp
#undef NINPUTS_X2mlp
#undef NH1_X2mlp
#undef NCLASSES_X2mlp


_________________________________________________________________________
Get Your Private, Free E-mail from MSN Hotmail at http://www.hotmail.com.