AI-NNEasy
view release on metacpan or search on metacpan
lib/AI/NNEasy.hploo view on Meta::CPAN
my $out_sz = ref $out ? $out->{nodes} : $out ;
@layers = ($in_sz+$out_sz) if !@layers ;
foreach my $layers_i ( @layers ) {
$layers_i = $in_sz+$out_sz if $layers_i <= 0 ;
}
$conf ||= {} ;
my $decay = $$conf{decay} || 0 ;
my $nn_in = $this->_layer_conf( { decay=>$decay } , $in ) ;
my $nn_out = $this->_layer_conf( { decay=>$decay , activation_function=>'linear' } , $out ) ;
foreach my $layers_i ( @layers ) {
$layers_i = $this->_layer_conf( { decay=>$decay } , $layers_i ) ;
}
my $nn_conf = {random_connections=>0 , networktype=>'feedforward' , random_weights=>1 , learning_algorithm=>'backprop' , learning_rate=>0.1 , bias=>1} ;
foreach my $Key ( keys %$nn_conf ) { $$nn_conf{$Key} = $$conf{$Key} if exists $$conf{$Key} ;}
$this->{NN_ARGS} = [[ $nn_in , @layers , $nn_out ] , $nn_conf] ;
$this->{NN} = AI::NNEasy::NN->new( @{$this->{NN_ARGS}} ) ;
$this->{FILE} = $file ;
@out_types = (0,1) if !@out_types ;
@out_types = sort {$a <=> $b} @out_types ;
$this->{OUT_TYPES} = \@out_types ;
if ( $error_ok <= 0 ) {
my ($min_dif , $last) ;
my $i = -1 ;
foreach my $out_types_i ( @out_types ) {
++$i ;
if ($i > 0) {
my $dif = $out_types_i - $last ;
$min_dif = $dif if !defined $min_dif || $dif < $min_dif ;
}
$last = $out_types_i ;
}
$error_ok = $min_dif / 2 ;
$error_ok -= $error_ok*0.1 ;
}
$this->{ERROR_OK} = $error_ok ;
return $this ;
}
sub _layer_conf ($def,$conf) {
$def ||= {} ;
$conf = { nodes=>$conf } if !ref($conf) ;
foreach my $Key ( keys %$def ) { $$conf{$Key} = $$def{$Key} if !exists $$conf{$Key} ;}
my $layer_conf = {nodes=>1 , persistent_activation=>0 , decay=>0 , random_activation=>0 , threshold=>0 , activation_function=>'tanh' , random_weights=>1} ;
foreach my $Key ( keys %$layer_conf ) { $$layer_conf{$Key} = $$conf{$Key} if exists $$conf{$Key} ;}
return $layer_conf ;
}
sub reset_nn {
$this->{NN} = AI::NNEasy::NN->new( @{ $this->{NN_ARGS} } ) ;
}
sub load ($file) {
$file ||= $this->{FILE} ;
if ( -s $file ) {
open (my $fh, $file) ;
my $dump = join '' , <$fh> ;
close ($fh) ;
my $restored = thaw($dump) ;
if ($restored) {
my $fl = $this->{FILE} ;
%$this = %$restored ;
$this->{FILE} = $fl if $fl ;
return 1 ;
}
}
return ;
}
sub save ($file) {
$file ||= $this->{FILE} ;
my $dump = freeze( {%$this} ) ;
open (my $fh,">$this->{FILE}") ;
print $fh $dump ;
close ($fh) ;
}
sub learn ($in,$out,$n) {
$n ||= 100 ;
my $err ;
for (1..100) {
$this->{NN}->run($in) ;
$err = $this->{NN}->learn($out) ;
}
$err *= -1 if $err < 0 ;
return $err ;
}
*_learn_set_get_output_error = \&_learn_set_get_output_error_c ;
sub _learn_set_get_output_error_pl ($set , $error_ok , $ins_ok , $verbose) {
for (my $i = 0 ; $i < @$set ; $i+=2) {
$this->{NN}->run($$set[$i]) ;
$this->{NN}->learn($$set[$i+1]) ;
}
my ($err,$learn_ok,$print) ;
for (my $i = 0 ; $i < @$set ; $i+=2) {
$this->{NN}->run($$set[$i]) ;
my $er = $this->{NN}->RMSErr($$set[$i+1]) ;
$er *= -1 if $er < 0 ;
++$learn_ok if $er < $error_ok ;
$err += $er ;
$print .= join(' ',@{$$set[$i]}) ." => ". join(' ',@{$$set[$i+1]}) ." > $er\n" if $verbose ;
}
$err /= $ins_ok ;
return ( $err , $learn_ok , $print ) ;
}
sub[C] SV* _av_join( AV* av ) {
SV* ret = sv_2mortal(newSVpv("",0)) ;
int i ;
for (i = 0 ; i <= av_len(av) ; ++i) {
SV* elem = *av_fetch(av, i ,0) ;
if (i > 0) sv_catpv(ret , " ") ;
sv_catsv(ret , elem) ;
}
lib/AI/NNEasy.hploo view on Meta::CPAN
PUTBACK ;
call_method("run", G_DISCARD) ;
PUSHMARK(SP) ;
XPUSHs( nn );
XPUSHs( set_out );
PUTBACK ;
call_method("RMSErr", G_SCALAR) ;
SPAGAIN ;
ret = POPs ;
er = SvNV(ret) ;
if (er < 0) er *= -1 ;
if (er < error_ok) ++learn_ok ;
err += er ;
if ( verbose ) sv_catpvf(print_verbose , "%s => %s > %f\n" ,
SvPV( _av_join( OBJ_AV(set_in) ) , len) ,
SvPV( _av_join( OBJ_AV(set_out) ) , len) ,
er
) ;
}
err /= ins_ok ;
if (verbose) {
EXTEND(SP , 3) ;
ST(0) = sv_2mortal(newSVnv(err)) ;
ST(1) = sv_2mortal(newSViv(learn_ok)) ;
ST(2) = print_verbose ;
XSRETURN(3) ;
}
else {
EXTEND(SP , 2) ;
ST(0) = sv_2mortal(newSVnv(err)) ;
ST(1) = sv_2mortal(newSViv(learn_ok)) ;
XSRETURN(2) ;
}
}
sub learn_set (\@set,$ins_ok,$limit,$verbose) {
my $ins_sz = @set / 2 ;
$ins_ok ||= $ins_sz ;
my $err_static_limit = 15 ;
my $err_static_limit_positive ;
if ( ref($limit) eq 'ARRAY' ) {
($limit,$err_static_limit,$err_static_limit_positive) = @$limit ;
}
$limit ||= 30000 ;
$err_static_limit_positive ||= $err_static_limit/2 ;
my $error_ok = $this->{ERROR_OK} ;
my $check_diff_count = 1000 ;
my ($learn_ok,$counter,$err,$err_last,$err_count,$err_static, $reset_count1 , $reset_count2 ,$print) ;
$err_static = 0 ;
while ( ($learn_ok < $ins_ok) && ($counter < $limit) ) {
($err , $learn_ok , $print) = $this->_learn_set_get_output_error(\@set , $error_ok , $ins_ok , $verbose) ;
++$counter ;
if ( !($counter % 100) || $learn_ok == $ins_ok ) {
my $err_diff = $err_last - $err ;
$err_diff *= -1 if $err_diff < 0 ;
$err_count += $err_diff ;
++$err_static if $err_diff <= 0.00001 || $err > 1 ;
print "err_static = $err_static\n" if $verbose && $err_static ;
$err_last = $err ;
my $reseted ;
if ( $err_static >= $err_static_limit || ($err > 1 && $err_static >= $err_static_limit_positive) ) {
$err_static = 0 ;
$counter -= 2000 ;
$reseted = 1 ;
++$reset_count1 ;
if ( ( $reset_count1 + $reset_count2 ) > 2 ) {
$reset_count1 = $reset_count2 = 0 ;
print "** Reseting NN...\n" if $verbose ;
$this->reset_nn ;
}
else {
print "** Reseting weights due NULL diff...\n" if $verbose ;
$this->{NN}->init ;
}
}
if ( !($counter % $check_diff_count) ) {
$err_count /= ($check_diff_count/100) ;
print "ERR COUNT> $err_count\n" if $verbose ;
if ( !$reseted && $err_count < 0.001 ) {
$err_static = 0 ;
$counter -= 1000 ;
++$reset_count2 ;
if ( ($reset_count1 + $reset_count2) > 2 ) {
$reset_count1 = $reset_count2 = 0 ;
print "** Reseting NN...\n" if $verbose ;
$this->reset_nn ;
}
else {
print "** Reseting weights due LOW diff...\n" if $verbose ;
$this->{NN}->init ;
}
}
$err_count = 0 ;
}
if ( $verbose ) {
print "\nepoch $counter : error_ok = $error_ok : error = $err : err_diff = $err_diff : err_static = $err_static : ok = $learn_ok\n" ;
print $print ;
}
}
print "epoch $counter : error = $err : ok = $learn_ok\n" if $verbose > 1 ;
}
}
sub get_set_error (\@set,$ins_ok) {
my $ins_sz = @set / 2 ;
$ins_ok ||= $ins_sz ;
my $err ;
for (my $i = 0 ; $i < @set ; $i+=2) {
$this->{NN}->run($set[$i]) ;
my $er = $this->{NN}->RMSErr($set[$i+1]) ;
$er *= -1 if $er < 0 ;
$err += $er ;
}
$err /= $ins_ok ;
return $err ;
}
sub run ($in) {
$this->{NN}->run($in) ;
my $out = $this->{NN}->output() ;
return $out ;
}
sub run_get_winner {
my $out = $this->run(@_) ;
foreach my $out_i ( @$out ) {
$out_i = $this->out_type_winner($out_i) ;
}
return $out ;
}
sub out_type_winner ($val) {
my ($out_type , %err) ;
foreach my $types_i ( @{ $this->{OUT_TYPES} } ) {
my $er = $types_i - $val ;
lib/AI/NNEasy.hploo view on Meta::CPAN
## Our maximal error for the output calculation.
my $ERR_OK = 0.1 ;
## Create the NN:
my $nn = AI::NNEasy->new(
'xor.nne' , ## file to save the NN.
[0,1] , ## Output types of the NN.
$ERR_OK , ## Maximal error for output.
2 , ## Number of inputs.
1 , ## Number of outputs.
[3] , ## Hidden layers. (this is setting 1 hidden layer with 3 nodes).
) ;
## Our set of inputs and outputs to learn:
my @set = (
[0,0] => [0],
[0,1] => [1],
[1,0] => [1],
[1,1] => [0],
);
## Calculate the actual error for the set:
my $set_err = $nn->get_set_error(\@set) ;
## If set error is bigger than maximal error lest's learn this set:
if ( $set_err > $ERR_OK ) {
$nn->learn_set( \@set ) ;
## Save the NN:
$nn->save ;
}
## Use the NN:
my $out = $nn->run_get_winner([0,0]) ;
print "0 0 => @$out\n" ; ## 0 0 => 0
my $out = $nn->run_get_winner([0,1]) ;
print "0 1 => @$out\n" ; ## 0 1 => 1
my $out = $nn->run_get_winner([1,0]) ;
print "1 0 => @$out\n" ; ## 1 0 => 1
my $out = $nn->run_get_winner([1,1]) ;
print "1 1 => @$out\n" ; ## 1 1 => 0
## or just interate through the @set:
for (my $i = 0 ; $i < @set ; $i+=2) {
my $out = $nn->run_get_winner($set[$i]) ;
print "@{$set[$i]}) => @$out\n" ;
}
=> METHODS
==> new ( FILE , @OUTPUT_TYPES , ERROR_OK , IN_SIZE , OUT_SIZE , @HIDDEN_LAYERS , %CONF )
*> FILE
The file path to save the NN. Default: 'nneasy.nne'.
*> @OUTPUT_TYPES
An array of outputs that the NN can have, so the NN can find the nearest number in this
list to give your the right output.
*> ERROR_OK
The maximal error of the calculated output.
If not defined ERROR_OK will be calculated by the minimal difference between 2 types at
@OUTPUT_TYPES dived by 2:
@OUTPUT_TYPES = [0 , 0.5 , 1] ;
ERROR_OK = (1 - 0.5) / 2 = 0.25 ;
*> IN_SIZE
The input size (number of nodes in the inpute layer).
*> OUT_SIZE
The output size (number of nodes in the output layer).
*> @HIDDEN_LAYERS
A list of size of hidden layers. By default we have 1 hidden layer, and
the size is calculated by I<(IN_SIZE + OUT_SIZE)>. So, for a NN of
2 inputs and 1 output the hidden layer have 3 nodes.
*> %CONF
Conf can be used to define special parameters of the NN:
Default:
{networktype=>'feedforward' , random_weights=>1 , learning_algorithm=>'backprop' , learning_rate=>0.1 , bias=>1}
Options:
**> networktype
The type of the NN. For now only accepts I<'feedforward'>.
**> random_weights
Maximum value for initial weight.
**> learning_algorithm
Algorithm to train the NN. Accepts I<'backprop'> and I<'reinforce'>.
**> learning_rate
Rate used in the learning_algorithm.
**> bias
If true will create a BIAS node. Usefull when you have NULL inputs, like [0,0].
/*>
Here's a completly example of use:
my $nn = AI::NNEasy->new(
'xor.nne' , ## file to save the NN.
[0,1] , ## Output types of the NN.
0.1 , ## Maximal error for output.
2 , ## Number of inputs.
1 , ## Number of outputs.
[3] , ## Hidden layers. (this is setting 1 hidden layer with 3 nodes).
{random_connections=>0 , networktype=>'feedforward' , random_weights=>1 , learning_algorithm=>'backprop' , learning_rate=>0.1 , bias=>1} ,
) ;
And a simple example that will create a NN equal of the above:
my $nn = AI::NNEasy->new('xor.nne' , [0,1] , 0.1 , 2 , 1 ) ;
==> load
Load the NN if it was previously saved.
==> save
Save the NN to a file using L<Storable>.
==> learn (@IN , @OUT , N)
Learn the input.
*> @IN
The values of one input.
*> @OUT
The values of the output for the input above.
*> N
Number of times that this input should be learned. Default: 100
Example:
$nn->learn( [0,1] , [1] , 10 ) ;
==> learn_set (@SET , OK_OUTPUTS , LIMIT , VERBOSE)
Learn a set of inputs until get the right error for the outputs.
*> @SET
A list of inputs and outputs.
*> OK_OUTPUTS
Minimal number of outputs that should be OK when calculating the erros.
By default I<OK_OUTPUTS> should have the same size of number of different
inouts in the @SET.
*> LIMIT
Limit of interations when learning. Default: 30000
*> VERBOSE
If TRUE turn verbose method ON when learning.
==> get_set_error (@SET , OK_OUTPUTS)
Get the actual error of a set in the NN. If the returned error is bigger than
I<ERROR_OK> defined on I<new()> you should learn or relearn the set.
==> run (@INPUT)
Run a input and return the output calculated by the NN based in what the NN already have learned.
==> run_get_winner (@INPUT)
Same of I<run()>, but the output will return the nearest output value based in the
I<@OUTPUT_TYPES> defined at I<new()>.
For example an input I<[0,1]> learned that have
the output I<[1]>, actually will return something like 0.98324 as output and
not 1, since the error never should be 0. So, with I<run_get_winner()>
we get the output of I<run()>, let's say that is 0.98324, and find what output
is near of this number, that in this case should be 1. An output [0], will return
by I<run()> something like 0.078964, and I<run_get_winner()> return 0.
=> Samples
Inside the release sources you can find the directory ./samples where you have some
examples of code using this module.
=> INLINE C
Some functions of this module have I<Inline> functions writed in C.
I have made a C version only for the functions that are wild called, like:
AI::NNEasy::_learn_set_get_output_error
AI::NNEasy::NN::tanh
AI::NNEasy::NN::feedforward::run
AI::NNEasy::NN::backprop::hiddenToOutput
AI::NNEasy::NN::backprop::hiddenOrInputToHidden
AI::NNEasy::NN::backprop::RMSErr
What give to us the speed that we need to learn fast the inputs, but at the same time
be able to create flexible NN.
=> Class::HPLOO
I have used L<Class::HPLOO> to write fast the module, specially the XS support.
L<Class::HPLOO> enables this kind of syntax for Perl classes:
class Foo {
sub bar($x , $y) {
$this->add($x , $y) ;
}
sub[C] int add( int x , int y ) {
int res = x + y ;
return res ;
}
}
What make possible to write the module in 2 days! ;-P
=> Basics of a Neural Network
I<- This is just a simple text for lay pleople,
to try to make them to understand what is a Neural Network and how it works
without need to read a lot of books -.>
A NN is based in nodes/neurons and layers, where we have the input layer, the hidden layers and the output layer.
For example, here we have a NN with 2 inputs, 1 hidden layer, and 2 outputs:
Input Hidden Output
input1 ---->n1\ /---->n4---> output1
\ /
n3
/ \
input2 ---->n2/ \---->n5---> output2
Basically, when we have an input, let's say [0,1], it will active I<n2>, that will
active I<n3> and I<n3> will active I<n4> and I<n5>, but the link between I<n3> and I<n4> has a I<weight>, and
between I<n3> and I<n5> another I<weight>. The idea is to find the I<weights> between the
nodes that can give to us an output near the real output. So, if the output of [0,1]
is [1,1], the nodes I<output1> and I<output2> should give to us a number near 1,
let's say 0.98654. And if the output for [0,0] is [0,0], I<output1> and I<output2> should give to us a number near 0,
let's say 0.078875.
What is hard in a NN is to find this I<weights>. By default L<AI::NNEasy> uses
I<backprop> as learning algorithm. With I<backprop> it pastes the inputs through
the Neural Network and adjust the I<weights> using random numbers until we find
a set of I<weights> that give to us the right output.
The secret of a NN is the number of hidden layers and nodes/neurons for each layer.
Basically the best way to define the hidden layers is 1 layer of (INPUT_NODES+OUTPUT_NODES).
So, a layer of 2 input nodes and 1 output node, should have 3 nodes in the hidden layer.
This definition exists because the number of inputs define the maximal variability of
the inputs (N**2 for bollean inputs), and the output defines if the variability is reduced by some logic restriction, like
int the XOR example, where we have 2 inputs and 1 output, so, hidden is 3. And as we can see in the
logic we have 3 groups of inputs:
0 0 => 0 # false
0 1 => 1 # or
1 0 => 1 # or
1 1 => 1 # true
Actually this is not the real explanation, but is the easiest way to understand that
you need to have a number of nodes/neuros in the hidden layer that can give the
right output for your problem.
Other inportant step of a NN is the learning fase. Where we get a set of inputs
and paste them through the NN until we have the right output. This process basically
will adjust the nodes I<weights> until we have an output near the real output that we want.
Other important concept is that the inputs and outputs in the NN should be from 0 to 1.
So, you can define sets like:
0 0 => 0
0 0.5 => 0.5
0.5 0.5 => 1
1 0.5 => 0
1 1 => 1
But what is really recomended is to always use bollean values, just 0 or 1, for inputs and outputs,
since the learning fase will be faster and works better for complex problems.
=> SEE ALSO
L<AI::NNFlex>, L<AI::NeuralNet::Simple>, L<Class::HPLOO>, L<Inline>.
=> AUTHOR
Graciliano M. P. <gmpassos@cpan.org>
I will appreciate any type of feedback (include your opinions and/or suggestions). ;-P
Thanks a lot to I<Charles Colbourn <charlesc at nnflex.g0n.net>>, that is the
author of L<AI::NNFlex>, that 1st wrote it, since NNFlex was my starting point to
do this NN work, and 2nd to be in touch with the development of L<AI::NNEasy>.
=> COPYRIGHT
This program is free software; you can redistribute it and/or
modify it under the same terms as Perl itself.
( run in 0.542 second using v1.01-cache-2.11-cpan-99c4e6809bf )