CvANN_MLP::calc_output_scale: scaling issue (Bug #558)

It is saying that it want to map mj..Mj to m..M, but it looks like the other way around.
Because of this, it is failing to go on second round with UPDATE_WEIGHT option. It just crashes after reporting out of range problem.

Here is my simple testing code, which is training for a XOR problem.

#include "ml.h" 
#include "cxcore.h" 
#include "cv.h" 
#include "stdio.h" 

int main(){
  //XOR input
  double in[]={ 0 ,0,
                1, 0,
                0, 1,
                1, 1};
  double out[]={ 0,
                 1,
                 1,
                 0};

  const int IVS = 2; // Input Vector Size
  const int OVS = 1; // Output Vector Size
  const int HN = 10; // Number of Hidden nodes
  const int NV= 4;   //Number of Training Vector

  int layer[] = { IVS, HN, OVS};

  [[CvMat]] *input =cvCreateMat( NV, IVS, CV_64FC1);
  [[CvMat]] *output =cvCreateMat( NV, OVS, CV_64FC1);
  [[CvMat]] *p_output =cvCreateMat( NV, OVS, CV_64FC1);
  [[CvMat]] *layersize =cvCreateMat( 1 , 3 , CV_32SC1);

  cvInitMatHeader(input, NV, IVS, CV_64FC1, in);
  cvInitMatHeader(output, NV, OVS, CV_64FC1, out);
  cvInitMatHeader(layersize, 1, 3, CV_32SC1, layer);

  [[CvANN]]_MLP train_model(layersize, [[CvANN]]_MLP::SIGMOID_SYM,1,1);
  std::cout<< " =========== =========== =========== =========== ==========="<<std::endl;
  std::cout<< "  * First Iteration with initialzation of weights"<<std::endl;
  std::cout<< " =========== =========== =========== =========== ==========="<<std::endl;
  int iter = train_model.train(  input,
                                 output,
                                 NULL,
                                 0,
                                 [[CvANN]]_MLP_TrainParams( cvTermCriteria ( CV_TERMCRIT_ITER |
                                                                         CV_TERMCRIT_EPS,
                                                                         5000,0.001),
                                   [[CvANN]]_MLP_TrainParams::RPROP,
                                   0.1,0.1),
                                 0
                                 //+ [[CvANN]]_MLP::NO_OUTPUT_SCALE
                              );

  std::cout << " * iteration :"<<iter<<std::endl;
  train_model.predict( input, p_output );
  for(int i=0; i<NV;i++){
    std::cout<< CV_MAT_ELEM(*input,double,i,0) << " ," << CV_MAT_ELEM(*input,double,i,1)
             << " : " << CV_MAT_ELEM(*p_output,double,i,0) <<std::endl;

  }
  train_model.save( "/scratch/eunyokim/src/Regina_TempCode/OpenCV_XOR_TEST/firstModel.xml");
  std::cout<< " =========== =========== =========== =========== ==========="<<std::endl;
  std::cout<< "  * Second Iteration without initialzation of weights"<<std::endl;
  std::cout<< " =========== =========== =========== =========== ==========="<<std::endl;

  int iter2 = train_model.train(  input,
                                 output,
                                 NULL,
                                 0,
                                 [[CvANN]]_MLP_TrainParams( cvTermCriteria ( CV_TERMCRIT_ITER |
                                                                         CV_TERMCRIT_EPS,
                                                                         5000,0.001),
                                   [[CvANN]]_MLP_TrainParams::RPROP,
                                   0.1,0.1),
                                 0
                                 +CvANN_MLP::UPDATE_WEIGHTS
                                 //+ [[CvANN]]_MLP::NO_OUTPUT_SCALE
                              );
  std::cout << " * iteration :"<<iter2<<std::endl;
  train_model.save( "/scratch/eunyokim/src/Regina_TempCode/OpenCV_XOR_TEST/secondModel.xml");

  train_model.predict( input, p_output );
  for(int i=0; i<NV;i++){
    std::cout<< CV_MAT_ELEM(*input,double,i,0) << " ," << CV_MAT_ELEM(*input,double,i,1)
             << " : " << CV_MAT_ELEM(*p_output,double,i,0) <<std::endl;

  }

}

And output looks like following:

h11. =========== =========== ===========

* First Iteration with initialzation of weights

h11. =========== =========== ===========

* iteration :17
0 ,0 : 0.00444597
1 ,0 : 1.01488
0 ,1 : 1.00985
1 ,1 : 0.00256477

h11. =========== =========== ===========

* Second Iteration without initialzation of weights

h11. =========== =========== ===========

[[OpenCV]] Error: One of arguments' values is out of range (Some of new output training vector
components run exceed the original range too much) in [[CvANN]]_MLP::calc_output_scale, file
/scratch/eunyokim/src/OpenCV/opencv/modules/ml/src/ann_mlp.cpp, line 641
terminate called after throwing an instance of 'cv::Exception'
what(): /scratch/eunyokim/src/OpenCV/opencv/modules/ml/src/ann_mlp.cpp:641: error: (-
211) Some of new output training vector components run exceed the original range too much
in function [[CvANN]]_MLP::calc_output_scale
Abort trap

But if I change the source code file of opencv/modules/ml/src/ann_mlp.cpp for the part of calc_output_scale. It looks fine at a glance. What I did is just simply switched scale and inv_scale variable.

if( reset_weights )
        for( j = 0; j < vcount; j++ )
        {
            // map mj..Mj to m..M
            double mj = scale[j*2], Mj = scale[j*2+1];
            double a, b;
            double delta = Mj - mj;
            if( delta < DBL_EPSILON )
                a = 1, b = (M + m - Mj - mj)*0.5;
            else
                a = (M - m)/delta, b = m - mj*a;
            //inv_scale[j*2] = a; inv_scale[j*2+1] = b;
            scale[j*2] = a; scale[j*2+1] = b;
            a = 1./a; b = -b*a;
            std::cout<<", a ="<<a<<" ,b = "<<b;
            //scale[j*2] = a; scale[j*2+1] = b;
            inv_scale[j*2] = a; inv_scale[j*2+1] = b;
        }

h11. =========== =========== ===========

* First Iteration with initialzation of weights

h11. =========== =========== ===========

prepare to train
count=4
fr0 :0 dr0 :0t(0,0) : 0
scaler0 :0 , scaler1: 0
fr0 :0 dr0 :1t(1,0) : 1
scaler0 :0 , scaler1: 1
fr0 :0 dr0 :1t(2,0) : 1
scaler0 :0 , scaler1: 1
fr0 :0 dr0 :0t(3,0) : 0
scaler0 :0 , scaler1: 1
mj = 0 ,Mj = 1 ,delta= 1, a =0.526316 ,b = 0.5 scaler0 :1.9 , scaler1: -0.95
* iteration :5000
0 ,0 : 0.381445
1 ,0 : 0.699498
0 ,1 : 0.72889
1 ,1 : -0.151035

h11. =========== =========== ===========

* Second Iteration without initialzation of weights

h11. =========== =========== ===========

prepare to train
count=4
fr0 :0 dr0 :0t(0,0) : 0
t : 0 scaler0 :1.9 , scaler1: -0.95
fr0 :0 dr0 :1t(1,0) : 1
t : 1 scaler0 :1.9 , scaler1: -0.95
fr0 :0 dr0 :1t(2,0) : 1
t : 1 scaler0 :1.9 , scaler1: -0.95
fr0 :0 dr0 :0t(3,0) : 0
t : 0 scaler0 :1.9 , scaler1: -0.95
* iteration :5000
0 ,0 : 0.381445
1 ,0 : 0.699498

It looks like working at the first glance, but it still has a problem, which is producing exactly same result at the second part to the first part. I am not sure if the UPDATE_WEIGHT option is working.

Thank you