float predict( InputArray _inputs, OutputArray _outputs, int ) const
{
if( !trained )
CV_Error( CV_StsError, "The network has not been trained or loaded" );
Mat inputs = _inputs.getMat();
int type = inputs.type(), l_count = layer_count();
int n = inputs.rows, dn0 = n;
CV_Assert( (type == CV_32F || type == CV_64F) && inputs.cols == layer_sizes[0] );
int noutputs = layer_sizes[l_count-1];
Mat outputs;
int min_buf_sz = 2*max_lsize;
int buf_sz = n*min_buf_sz;
if( buf_sz > max_buf_sz )
{
dn0 = max_buf_sz/min_buf_sz;
dn0 = std::max( dn0, 1 );
buf_sz = dn0*min_buf_sz;
}
cv::AutoBuffer<double> _buf(buf_sz+noutputs);
double* buf = _buf;
if( !_outputs.needed() )
{
CV_Assert( n == 1 );
outputs = Mat(n, noutputs, type, buf + buf_sz);
}
else
{
_outputs.create(n, noutputs, type);
outputs = _outputs.getMat();
}
int dn = 0;
for( int i = 0; i < n; i += dn )
{
dn = std::min( dn0, n - i );
Mat layer_in = inputs.rowRange(i, i + dn);
Mat layer_out( dn, layer_in.cols, CV_64F, buf);
scale_input( layer_in, layer_out );
layer_in = layer_out;
for( int j = 1; j < l_count; j++ )
{
double* data = buf + ((j&1) ? max_lsize*dn0 : 0);
int cols = layer_sizes[j];
layer_out = Mat(dn, cols, CV_64F, data);
Mat w = weights[j].rowRange(0, layer_in.cols);
gemm(layer_in, w, 1, noArray(), 0, layer_out);
calc_activ_func( layer_out, weights[j] );
layer_in = layer_out;
}
layer_out = outputs.rowRange(i, i + dn);
scale_output( layer_in, layer_out );
}
if( n == 1 )
{
int maxIdx[] = {0, 0};
minMaxIdx(outputs, 0, 0, 0, maxIdx);
return (float)(maxIdx[0] + maxIdx[1]);
}
return 0.f;
}
float CvANN_MLP::predict( const CvMat* _inputs, CvMat* _outputs ) const
{
CV_FUNCNAME( "CvANN_MLP::predict" );
__BEGIN__;
double* buf;
int i, j, n, dn = 0, l_count, dn0, buf_sz, min_buf_sz;
if( !layer_sizes )
CV_ERROR( CV_StsError, "The network has not been initialized" );
if( !CV_IS_MAT(_inputs) || !CV_IS_MAT(_outputs) ||
!CV_ARE_TYPES_EQ(_inputs,_outputs) ||
CV_MAT_TYPE(_inputs->type) != CV_32FC1 &&
CV_MAT_TYPE(_inputs->type) != CV_64FC1 ||
_inputs->rows != _outputs->rows )
CV_ERROR( CV_StsBadArg, "Both input and output must be floating-point matrices "
"of the same type and have the same number of rows" );
if( _inputs->cols != layer_sizes->data.i[0] )
CV_ERROR( CV_StsBadSize, "input matrix must have the same number of columns as "
"the number of neurons in the input layer" );
if( _outputs->cols != layer_sizes->data.i[layer_sizes->cols - 1] )
CV_ERROR( CV_StsBadSize, "output matrix must have the same number of columns as "
"the number of neurons in the output layer" );
n = dn0 = _inputs->rows;
min_buf_sz = 2*max_count;
buf_sz = n*min_buf_sz;
if( buf_sz > max_buf_sz )
{
dn0 = max_buf_sz/min_buf_sz;
dn0 = MAX( dn0, 1 );
buf_sz = dn0*min_buf_sz;
}
buf = (double*)cvStackAlloc( buf_sz*sizeof(buf[0]) );
l_count = layer_sizes->cols;
for( i = 0; i < n; i += dn )
{
CvMat hdr[2], _w, *layer_in = &hdr[0], *layer_out = &hdr[1], *temp;
dn = MIN( dn0, n - i );
cvGetRows( _inputs, layer_in, i, i + dn );
cvInitMatHeader( layer_out, dn, layer_in->cols, CV_64F, buf );
scale_input( layer_in, layer_out );
CV_SWAP( layer_in, layer_out, temp );
for( j = 1; j < l_count; j++ )
{
double* data = buf + (j&1 ? max_count*dn0 : 0);
int cols = layer_sizes->data.i[j];
cvInitMatHeader( layer_out, dn, cols, CV_64F, data );
cvInitMatHeader( &_w, layer_in->cols, layer_out->cols, CV_64F, weights[j] );
cvGEMM( layer_in, &_w, 1, 0, 0, layer_out );
calc_activ_func( layer_out, _w.data.db + _w.rows*_w.cols );
CV_SWAP( layer_in, layer_out, temp );
}
cvGetRows( _outputs, layer_out, i, i + dn );
scale_output( layer_in, layer_out );
}
__END__;
return 0.f;
}
