package ICC::Support::nNET; use strict; use Carp; our $VERSION = 0.31; # revised 2016-05-17 # # Copyright © 2004-2018 by William B. Birkett # add development directory use lib 'lib'; # inherit from Shared use parent qw(ICC::Shared); # use POSIX math use POSIX (); # enable static variables use feature 'state'; # list of valid kernel types my @types = qw(CODE ICC::Support::rbf); # create new nNET object # hash may contain pointers to header, kernel, matrix, offset or init # kernel is a reference to an array of kernel objects or CODE references # matrix is a 2D array reference or Math::Matrix object # offset is a 1D array reference # hash keys are: ('header', 'kernel', 'matrix', 'offset', 'init') # parameters: ([ref_to_attribute_hash]) # returns: (ref_to_object) sub new { # get object class my $class = shift; # local variable my ($code); # create empty nNET object my $self = [ {}, # object header [], # kernel array [], # matrix matrix [] # offset vector ]; # if there are parameters if (@_) { # if one parameter, a hash reference if (@_ == 1 && ref($_[0]) eq 'HASH') { # make new nNET object from attribute hash _new_from_hash($self, shift()); # initialize object (if CODE reference defined) (defined($code = $self->[0]{'init'}) && &$code); } else { # error croak('nNET parameter must be a hash reference'); } } # bless object bless($self, $class); # return object reference return($self); } # initialize object # calls 'init' CODE reference, if any # used when retrieving an nNET object using Storable sub init { # get object reference my $self = shift(); # local variable my ($code); # initialize object (if CODE reference defined) (defined($code = $self->[0]{'init'}) && &$code); } # fit nNET object to data # determines optimum 'matrix' and 'offset' arrays # kernel nodes are not modified by this method # uses LAPACK dgelsd function to perform a least-squares fit # fitting is done with or without offset, according to offset_flag # fitting is done to output or input-output difference, according to diff_mode_flag # input and output are 2D array references or Math::Matrix objects # parameters: (ref_to_input_array, ref_to_output_array, [offset_flag, [diff_mode_flag]]) # returns: (dgelsd_info_value) sub fit { # get parameters my ($self, $in, $out, $oflag, $dflag) = @_; # local variables my ($dif, $info, $ab); # resolve offset flag $oflag = 0 if (! defined($oflag)); # verify input array (ref($in) eq 'ARRAY' && ref($in->[0]) eq 'ARRAY' && ! ref($in->[0][0])) || (UNIVERSAL::isa($in, 'Math::Matrix')) || croak('fit input not a 2-D array reference'); # verify output array (ref($out) eq 'ARRAY' && ref($out->[0]) eq 'ARRAY' && ! ref($out->[0][0])) || (UNIVERSAL::isa($out, 'Math::Matrix')) || croak('fit output not a 2-D array reference'); # verify array dimensions ($#{$in} == $#{$out}) || croak('fit input and output arrays have different number of rows'); # if difference mode if ($dflag) { # verify array dimensions ($#{$in->[0]} == $#{$out->[0]}) || croak('fit input and output arrays have different number of columns'); # for each row for my $i (0 .. $#{$in}) { # for each column for my $j (0 .. $#{$in->[0]}) { # compute output-input difference $dif->[$i][$j] = $out->[$i][$j] - $in->[$i][$j]; } } } # fit the matrix (hidden values to output or difference values) ($info, $ab) = ICC::Support::Lapack::nNET_fit(_hidden2($self, $in), $dflag ? $dif : $out, $oflag); # check result carp('fit failed - bad parameter when calling dgelsd') if ($info < 0); carp('fit failed - SVD algorithm failed to converge') if ($info > 0); # initialize matrix $self->[2] = []; # for each output for my $i (0 .. $#{$out->[0]}) { # for each kernel node for my $j (0 .. $#{$self->[1]}) { # set matrix element (transposing) $self->[2][$i][$j] = $ab->[$j][$i]; } } # if offset flag if ($oflag) { # set offset $self->[3] = [@{$ab->[$#{$self->[1]} + 1]}]; } else { # no offset undef($self->[3]); } # if difference flag if ($dflag) { # for each row for my $i (0 .. $#{$self->[2]}) { # for each column for my $j (0 .. $#{$self->[2]}) { # add identity matrix element $self->[2][$i][$j + $#{$self->[1]} + 1] = ($i == $j) ? 1 : 0; } } } # return info value return($info); } # get/set reference to header hash # parameters: ([ref_to_new_hash]) # returns: (ref_to_hash) sub header { # get object reference my $self = shift(); # if there are parameters if (@_) { # if one parameter, a hash reference if (@_ == 1 && ref($_[0]) eq 'HASH') { # set header to new hash $self->[0] = {%{shift()}}; } else { # error croak('parameter must be a hash reference'); } } # return reference return($self->[0]); } # get/set kernel array reference # parameters: ([ref_to_array]) # returns: (ref_to_array) sub kernel { # get object reference my $self = shift(); # if one parameter supplied if (@_ == 1) { # get parameter my $array = shift; # if an array reference if (ref($array) eq 'ARRAY') { # initialize array $self->[1] = []; # for each array element for my $i (0 .. $#{$array}) { # if array element is a valid kernel type if (grep {ref($array->[$i]) eq $_} @types) { # add array element $self->[1][$i] = $array->[$i]; } else { # wrong data type croak('invalid nNET kernel array element'); } } } else { # wrong data type croak('nNET kernel attribute must be an array reference'); } } elsif (@_) { # error croak('too many parameters'); } # return kernel array reference return($self->[1]); } # get/set reference to matrix # parameters: ([ref_to_new_array]) # returns: (ref_to_array) sub matrix { # get object reference my $self = shift(); # if there are parameters if (@_) { # if one parameter, a reference to 2D array if (@_ == 1 && ref($_[0]) eq 'ARRAY' && ref($_[0][0]) eq 'ARRAY') { # set object element $self->[2] = Storable::dclone(shift()); # if one parameter, a reference to Math::Matrix object } elsif (@_ == 1 && UNIVERSAL::isa($_[0], 'Math::Matrix')) { # set object element $self->[2] = Storable::dclone([@{shift()}]); } else { # wrong data type croak('nNET matrix attribute must be an array reference or Math::Matrix object'); } } # return matrix reference return($self->[2]); } # get/set reference to offset array # parameters: ([ref_to_new_array]) # returns: (ref_to_array) sub offset { # get object reference my $self = shift(); # if there are parameters if (@_) { # if one parameter, a reference to an array of scalars if (@_ == 1 && ref($_[0]) eq 'ARRAY' && @{$_[0]} == grep {! ref()} @{$_[0]}) { # set object element $self->[3] = [@{shift()}]; } else { # wrong data type croak('nNET offset attribute must be an array reference'); } } # return offset reference return($self->[3]); } # transform data # supported input types: # parameters: (list, [hash]) # parameters: (vector, [hash]) # parameters: (matrix, [hash]) # parameters: (Math::Matrix_object, [hash]) # parameters: (structure, [hash]) # returns: (same_type_as_input) sub transform { # set hash value (0 or 1) my $h = ref($_[-1]) eq 'HASH' ? 1 : 0; # if input a 'Math::Matrix' object if (@_ == $h + 2 && UNIVERSAL::isa($_[1], 'Math::Matrix')) { # call matrix transform &_trans2; # if input an array reference } elsif (@_ == $h + 2 && ref($_[1]) eq 'ARRAY') { # if array contains numbers (vector) if (! ref($_[1][0]) && @{$_[1]} == grep {Scalar::Util::looks_like_number($_)} @{$_[1]}) { # call vector transform &_trans1; # if array contains vectors (2-D array) } elsif (ref($_[1][0]) eq 'ARRAY' && @{$_[1]} == grep {ref($_) eq 'ARRAY' && Scalar::Util::looks_like_number($_->[0])} @{$_[1]}) { # call matrix transform &_trans2; } else { # call structure transform &_trans3; } # if input a list (of numbers) } elsif (@_ == $h + 1 + grep {Scalar::Util::looks_like_number($_)} @_) { # call list transform &_trans0; } else { # error croak('invalid transform input'); } } # inverse transform # note: number of undefined output values must equal number of defined input values # note: the input and output vectors contain the final solution on return # hash key 'init' specifies initial value vector # parameters: (input_vector, output_vector, [hash]) # returns: (RMS_error_value) sub inverse { # get parameters my ($self, $in, $out, $hash) = @_; # local variables my ($i, $j, @si, @so, $init); my ($int, $jac, $mat, $delta); my ($max, $elim, $dlim, $accum, $error); # initialize indices $i = $j = -1; # build slice arrays while validating input and output arrays ((grep {$i++; defined() && push(@si, $i)} @{$in}) == (grep {$j++; ! defined() && push(@so, $j)} @{$out})) || croak('wrong number of undefined values'); # get init array $init = $hash->{'init'}; # for each undefined output value for my $i (@so) { # set to supplied initial value or 0.5 $out->[$i] = defined($init->[$i]) ? $init->[$i] : 0.5; } # set maximum loop count $max = $hash->{'inv_max'} || 10; # loop error limit $elim = $hash->{'inv_elim'} || 1E-6; # set delta limit $dlim = $hash->{'inv_dlim'} || 0.5; # create empty solution matrix $mat = Math::Matrix->new([]); # compute initial transform values ($jac, $int) = jacobian($self, $out, $hash); # solution loop for (1 .. $max) { # for each input for my $i (0 .. $#si) { # for each output for my $j (0 .. $#so) { # copy Jacobian value to solution matrix $mat->[$i][$j] = $jac->[$si[$i]][$so[$j]]; } # save residual value to solution matrix $mat->[$i][$#si + 1] = $in->[$si[$i]] - $int->[$si[$i]]; } # solve for delta values $delta = $mat->solve; # for each output value for my $i (0 .. $#so) { # add delta (limited using hyperbolic tangent) $out->[$so[$i]] += POSIX::tanh($delta->[$i][0]/$dlim) * $dlim; } # compute updated transform values ($jac, $int) = jacobian($self, $out, $hash); # initialize error accumulator $accum = 0; # for each input for my $i (0 .. $#si) { # accumulate delta squared $accum += ($in->[$si[$i]] - $int->[$si[$i]])**2; } # compute RMS error $error = sqrt($accum/@si); # if error less than limit last if ($error < $elim); } # update input vector with final values @{$in} = @{$int}; # return return($error); } # compute Jacobian matrix # parameters: (input_vector, [hash]) # returns: (Jacobian_matrix, [output_vector]) sub jacobian { # get parameters my ($self, $in, $hash) = @_; # local variables my ($jac, $out); # check if ICC::Support::Lapack module is loaded state $lapack = defined($INC{'ICC/Support/Lapack.pm'}); # compute hidden Jacobian and output ($jac, $out) = _hidden3($self, $in); # if ICC::Support::Lapack module is loaded if ($lapack) { # if output values wanted if (wantarray) { # return Jacobian and output return(bless(ICC::Support::Lapack::mat_xplus($self->[2], $jac), 'Math::Matrix'), ICC::Support::Lapack::matf_vec_trans($out, $self->[2], $self->[3])); } else { # return Jacobian only return(bless(ICC::Support::Lapack::mat_xplus($self->[2], $jac), 'Math::Matrix')); } } else { croak('method not yet implemented'); } } # print object contents to string # format is an array structure # parameter: ([format]) # returns: (string) sub sdump { # get parameters my ($self, $p) = @_; # local variables my ($s, $fmt); # resolve parameter to an array reference $p = defined($p) ? ref($p) eq 'ARRAY' ? $p : [$p] : []; # get format string $fmt = defined($p->[0]) && ! ref($p->[0]) ? $p->[0] : 'undef'; # set string to object ID $s = sprintf("'%s' object, (0x%x)\n", ref($self), $self); # return return($s); } # transform list # parameters: (object_reference, list, [hash]) # returns: (list) sub _trans0 { # local variables my ($self, $hash, @out); # get object reference $self = shift(); # get optional hash $hash = pop() if (ref($_[-1]) eq 'HASH'); # compute output using '_trans1' @out = @{_trans1($self, \@_, $hash)}; # return return(@out); } # transform vector # parameters: (object_reference, vector, [hash]) # returns: (vector) sub _trans1 { # get parameters my ($self, $in, $hash) = @_; # local variables my ($out); # check if ICC::Support::Lapack module is loaded state $lapack = defined($INC{'ICC/Support/Lapack.pm'}); # if ICC::Support::Lapack module is loaded if ($lapack) { # call the BLAS dgemv function return(ICC::Support::Lapack::matf_vec_trans(_hidden($self, $in), $self->[2], $self->[3])); } else { croak('method not yet implemented'); } } # transform matrix (2-D array -or- Math::Matrix object) # parameters: (object_reference, matrix, [hash]) # returns: (matrix) sub _trans2 { # get parameters my ($self, $in, $hash) = @_; # local variables my ($out); # check if ICC::Support::Lapack module is loaded state $lapack = defined($INC{'ICC/Support/Lapack.pm'}); # if ICC::Support::Lapack module is loaded if ($lapack) { # call the BLAS dgemm function $out = ICC::Support::Lapack::matf_mat_trans(_hidden2($self, $in), $self->[2], $self->[3]); } else { croak('method not yet implemented'); } # return return(UNIVERSAL::isa($in, 'Math::Matrix') ? bless($out, 'Math::Matrix') : $out); } # transform structure # parameters: (object_reference, structure, [hash]) # returns: (structure) sub _trans3 { # get parameters my ($self, $in, $hash) = @_; # transform the array structure _crawl($self, $in, my $out = [], $hash); # return return($out); } # recursive transform # array structure is traversed until scalar arrays are found and transformed # parameters: (ref_to_object, ref_to_input_array, ref_to_output_array, hash) sub _crawl { # get parameters my ($self, $in, $out, $hash) = @_; # if input is a vector (reference to a scalar array) if (@{$in} == grep {! ref()} @{$in}) { # transform input vector and copy to output @{$out} = @{_trans1($self, $in, $hash)}; } else { # for each input element for my $i (0 .. $#{$in}) { # if an array reference if (ref($in->[$i]) eq 'ARRAY') { # transform next level _crawl($self, $in->[$i], $out->[$i] = [], $hash); } else { # error croak('invalid transform input'); } } } } # compute hidden node output vector # parameters: (ref_to_object, ref_to_input_vector) # returns: (ref_to_output_vector) sub _hidden { # get parameters my ($self, $in) = @_; # local variables my ($array, $node, $out); # get kernel array $array = $self->[1]; # for each node for my $i (0 .. $#{$array}) { # get node $node = $array->[$i]; # if a code reference if (ref($node) eq 'CODE') { # call subroutine $out->[$i] = &$node($in); # else a kernel object } else { # call transform method $out->[$i] = $node->transform($in); } } # if array rows < matrix columns (difference mode) if ($#{$array} < $#{$self->[2][0]}) { # append input values push(@{$out}, @{$in}); } # return return($out); } # compute hidden node output matrix # parameters: (ref_to_object, ref_to_array_of_input_vectors) # returns: (ref_to_array_of_output_vectors) sub _hidden2 { # get parameters my ($self, $in) = @_; # local variables my ($array, $node, $out); # get kernel array $array = $self->[1]; # initialize output array $out = []; # for each input row for my $i (0 .. $#{$in}) { # for each node for my $j (0 .. $#{$array}) { # get node $node = $array->[$j]; # if a code reference if (ref($node) eq 'CODE') { # call subroutine $out->[$i][$j] = &$node($in->[$i]); # else a kernel object } else { # call transform method $out->[$i][$j] = $node->transform($in->[$i]); } } # if array rows < matrix columns (difference mode) if ($#{$array} < $#{$self->[2][0]}) { # append input values push(@{$out->[$i]}, @{$in->[$i]}); } } # return return($out); } # compute hidden node Jacobian matrix # parameters: (ref_to_object, ref_to_input_vector) # returns: (ref_to_Jacobian_matrix, [ref_to_output_vector]) sub _hidden3 { # get parameters my ($self, $in) = @_; # local variables my ($array, $node, $jac, $out); # get kernel array $array = $self->[1]; # for each node for my $i (0 .. $#{$array}) { # get node $node = $array->[$i]; # if a code reference if (ref($node) eq 'CODE') { # if output requested if (wantarray) { # compute numerical Jacobian $jac->[$i] = _numjac($node, $in); # call subroutine $out->[$i] = &$node($in); } else { # compute numerical Jacobian $jac->[$i] = _numjac($node, $in); } # else a kernel object } else { # if output requested if (wantarray) { # call jacobian method ($jac->[$i], $out->[$i]) = $node->jacobian($in); } else { # call jacobian method $jac->[$i] = $node->jacobian($in); } } } # if array rows < matrix columns (difference mode) if ($#{$array} < $#{$self->[2][0]}) { # for each row for my $i (0 .. $#{$self->[2]}) { # for each column for my $j (0 .. $#{$self->[2]}) { # add identity matrix element $jac->[$i + $#{$self->[1]} + 1][$j] = $i == $j ? 1 : 0; } } } # if output vector requested if (wantarray) { # if array rows < matrix columns (difference mode) if ($#{$array} < $#{$self->[2][0]}) { # append input values push(@{$out}, @{$in}); } # return return($jac, $out); } else { # return return($jac); } } # compute numerical Jacobian # parameters: (code_reference, input_vector) # output: (Jacobian_vector) sub _numjac { # get parameters my ($node, $in) = @_; # local variables my ($delta, $ind, $out, $jac); # set delta value $delta = 1E-12; # compute nominal output $out = &$node($in); # for each input for my $i (0 .. $#{$in}) { # copy input values $ind = [@{$in}]; # add input delta $ind->[$i] += $delta; # compute slope $jac->[$i] = (&$node($ind) - $out)/$delta; } # return Jacobian return($jac); } # make new nNET object from attribute hash # hash may contain pointers to header, kernel, matrix, offset or init # hash keys are: ('header', 'kernel', 'matrix', 'offset', 'init') # object elements not specified in the hash are unchanged # parameters: (ref_to_object, ref_to_attribute_hash) sub _new_from_hash { # get parameters my ($self, $hash) = @_; # local variables my ($array, $code); # for each attribute for my $attr (keys(%{$hash})) { # if 'header' if ($attr eq 'header') { # if reference to hash if (ref($hash->{$attr}) eq 'HASH') { # set object element $self->[0] = {%{$hash->{$attr}}}; } else { # wrong data type croak('nNET header attribute must be a hash reference'); } # if 'kernel' } elsif ($attr eq 'kernel') { # if an array reference if (ref($hash->{$attr}) eq 'ARRAY') { # get array $array = $hash->{$attr}; # for each array element for my $i (0 .. $#{$array}) { # if array element is a valid kernel type if (grep {ref($array->[$i]) eq $_} @types) { # add array element $self->[1][$i] = $array->[$i]; } else { # wrong data type croak('invalid nNET kernel array element'); } } } else { # wrong data type croak('nNET kernel attribute must be an array reference'); } # if 'matrix' } elsif ($attr eq 'matrix') { # if reference to 2D array if (ref($hash->{$attr}) eq 'ARRAY' && ref($hash->{$attr}[0]) eq 'ARRAY') { # set object element $self->[2] = Storable::dclone($hash->{$attr}); # if reference to Math::Matrix object } elsif (UNIVERSAL::isa($hash->{$attr}, 'Math::Matrix')) { # set object element $self->[2] = Storable::dclone([@{$hash->{$attr}}]); } else { # wrong data type croak('nNET matrix attribute must be a 2-D array reference or Math::Matrix object'); } # if 'offset' } elsif ($attr eq 'offset') { # if reference to an array of scalars if (ref($hash->{$attr}) eq 'ARRAY' && @{$hash->{$attr}} == grep {! ref()} @{$hash->{$attr}}) { # set object element $self->[3] = [@{$hash->{$attr}}]; } else { # wrong data type croak('nNET offset attribute must be an array reference'); } # if 'init' } elsif ($attr eq 'init') { # if a CODE reference if (ref($hash->{$attr}) eq 'CODE') { # set object element $self->[0]{'init'} = $hash->{$attr}; } else { # wrong data type croak('nNET init attribute must be a CODE reference'); } } } } 1;