view release on metacpan or  search on metacpan

lib/AI/ML/NeuralNetwork.pm  view on Meta::CPAN

			$var->{"dz$i"} = $var->{"da$i"} * $functions->{ $self->{"l$i"}{dfunc} }->($var->{"Z$i"}) ;
			$var->{"dw$i"} = (1 / $m) * ( $var->{"dz$i"} x T($var->{"A$j"}) );
			$var->{"db$i"} = (1 / $m) * sum( $var->{"dz$i"} , 0 );
			$var->{"da$j"} = T($self->{"l$i"}{w}) x $var->{"dz$i"} if $j >= 1;

            $self->{"l$i"}{w} = $self->{"l$i"}{w} - ( $alpha * $var->{"dw$i"} ); 
			$self->{"l$i"}{b} = $self->{"l$i"}{b} - ( $alpha * $var->{"db$i"} ); 


            if($iter == 100){
                $aux = $var->{"dz$i"};
                #$aux->save_csv("/tmp/DZ$i.csv");
                $aux = $var->{"dw$i"};
                #$aux->save_csv("/tmp/DW$i.csv");
                $aux = $var->{"db$i"};
                #$aux->save_csv("/tmp/DB$i.csv");
                #if ($j>=1){$aux = $var->{"da$j"};
                #$aux->save_csv("/tmp/da$j.m");}

                $aux = $self->{"l$i"}{w};
                #$aux->save_csv("/tmp/W$i.csv");
                $aux = $self->{"l$i"}{b};
                #$aux->save_csv("/tmp/B$i.csv");
            }
	    }
	}
    $self->{grads} = %$var if exists $opts{grads};
}


=head2 gradient_checking

=cut
sub gradient_checking {
    my ($self, $x, $y) = @_;
    my ($params, $grads, %dims) = $self->_get_params_grads();
    #print STDERR Dumper($params);
    #print STDERR Dumper($grads);
    #print STDERR Dumper(%dims);

    #my $n = $params->rows;
    #my $m = $params->columns;
    #print STDERR "elements:$n,$m\nParams vector\n";
    #for my $i (0..$n-1){
    #    print STDERR "$i:" .$params->get_element($i,0)."\n";
    #}
    #print STDERR "Grads vector\n";

    #for my $j (0..$n-1){
    #    print STDERR $params->get_element($j,0)."\n";
    #}
  
    #my $epsilon = 1e-7;
    #my $J_plus = Math::Lapack::Matrix->zeros($n,1);
    #my $J_minus = Math::Lapack::Matrix->zeros($n,1);
    #my $grad_aprox = Math::Lapack::Matrix->zeros($n,1);
    
    #for my $i (0..$n-1){
    #    $theta_plus = $params;
    #    $theta_plus->set_element($i,0) = $theta_plus->get_element($i,0) + $epsilon;       
    #    $J_plus($i,0) = _forward_prop_n($x, $y, _vector_to_hash($theta_plus, $n, %dims));
    #    
    #    $theta_minus = $params;
    #    $theta_minus->set_element($i,0) = $theta_minus->get_element($i,0) - $epsilon;       
    #    $J_minus($i,0) = _forward_prop_n($x, $y, _vector_to_hash($theta_minus, $n));

    #    $grad_aprox($i,0) = ($J_plus($i,0) - $j_minus($i,0)) / (2*$epsilon);
    #}

} 
    

=head2 _vector_to_hash

=cut
sub _vector_to_hash {
    my ($vector, $n, %dims) = @_;
    my $size = $vector->rows;
    my $pos = 0;
    my ($n_values, $weight, $bias);
    my %hash = {};
    
    for my $i (1..$n-1){
        $n_values = $dims{"w$i"}{rows} * $dims{"w$i"}{cols};
        $weight = $vector->slice( row_range => [$pos, $pos+$n_values-1] );
        $hash{"l$i"}{w} = $weight->reshape($dims{"w$i"}{rows}, $dims{"w$i"}{cols});         
        $pos += $n_values;

        $n_values = $dims{"b$i"}{rows} * $dims{"b$i"}{cols};
        $bias = $vector->reshape( row_range => [$pos, $pos+$n_values-1]);
        $hash{"l$i"}{b} = $bias->reshape($dims{"b$i"}{rows},$dims{"b$i"}{cols});
    
        $pos += $n_values;
    }
    return %hash;
}


=head2 _get_params_grads


=cut
sub _get_params_grads {
    my ($self) = @_;
    
    my ($matrix, $params, $grads, $n, %dims);

    my ($r, $c);
    $n = $self->{layers};

    $matrix = $self->{"l1"}{w};
    $dims{"w1"}{rows} = $matrix->rows; 
    $dims{"w1"}{cols} = $matrix->columns;
 ($r, $c) = $matrix->shape;
print STDERR "New dimension shape: $r,$c\n";
    $params = $matrix->reshape($matrix->rows * $matrix->columns, 1);
    ($r, $c) = $params->shape;
    print STDERR "$r,$c\n";
    
    $matrix = $self->{grads}{"dw1"};
    $grads = $matrix->reshape($matrix->rows * $matrix->columns, 1);
    for my $i (1..$n-1){
        print STDERR "layer: $i\n";
        if( $i > 1 ){