AI-Nerl

 view release on metacpan or  search on metacpan

examples/digits/deep_digits.pl  view on Meta::CPAN

use PDL::IO::FITS;
use PDL::Constants 'E';
use lib 'lib';
use lib '../../lib';
use AI::Nerl;

use FindBin qw($Bin); 
chdir $Bin;

unless (-e "t10k-labels-idx1-ubyte.fits"){ die <<"NODATA";}
pull this data by running get_digits.sh
convert it to FITS by running idx_to_fits.pl
NODATA


my $images = rfits('t10k-images-idx3-ubyte.fits');
my $labels = rfits('t10k-labels-idx1-ubyte.fits');
my $y = identity(10)->range($labels->transpose)->sever;
say 't10k data loaded';

my $nerl = AI::Nerl->new(
   # type => image,dims=>[28,28],...
   scale_input => 1/256,
);

$nerl->init_network(l1 => 784, l3=>10, l2=>7);#method=batch,hidden=>12345,etc

my $prev_nerl = $nerl;
my $prev_cost = 10000;

examples/digits/digits.pl  view on Meta::CPAN

use PDL::IO::FITS;
use PDL::Constants 'E';
use lib 'lib';
use lib '../../lib';
use AI::Nerl;

use FindBin qw($Bin); 
chdir $Bin;

unless (-e "t10k-labels-idx1-ubyte.fits"){ die <<"NODATA";}
pull this data by running get_digits.sh
convert it to FITS by running idx_to_fits.pl
NODATA


my $images = rfits('t10k-images-idx3-ubyte.fits');
my $labels = rfits('t10k-labels-idx1-ubyte.fits');
my $y = identity(10)->range($labels->transpose)->sever;
$y *= 2;
$y -= 1;
say 't10k data loaded';

my $nerl = AI::Nerl->new(
   # type => image,dims=>[28,28],...
   scale_input => 1/256,
#   train_x => $images(0:99),
#   train_y => $y(0:99),
#   test_x => $images(8000:8999),
#   test_y => $y(8000:8999),
#   cv_x => $images(9000:9999),
#   cv_y => $y(9000:9999),

examples/digits/idx_to_fits.pl  view on Meta::CPAN

use PDL;
use PDL::NiceSlice;
use File::Slurp;
use PDL::IO::FITS;

use FindBin qw($Bin); 
chdir $Bin;

die 'filename' unless $ARGV[0];
my $img_filename = $ARGV[0];
my $img_data = read_file( $img_filename, binmode => ':raw' ) ;

my @header = map {ord}  split ('', substr ($img_data, 0, 4, ''));
my $numdims = $header[3];
my @dims = map {ord} split ('',substr($img_data, 0, 4*$numdims, ''));

#'IDX' format described here: http://yann.lecun.com/exdb/mnist/
for (0..$numdims-1){
   $dims[$_] = 256*$dims[4*$_+2] + $dims[4*$_+3];
}
@dims=@dims[0..$numdims-1];
#die join ' ',@dims;
#my @img_data = map{ord}split('',$img_data);
my $img_pdl = pdl(unpack('C*',$img_data));

use PDL::Graphics2D;

if(!defined($dims[1])){
   $img_pdl = $img_pdl->squeeze;
}
elsif ($dims[1]==28){ #images
   @dims = (28**2,$dims[0]);
   $img_pdl = $img_pdl->reshape(@dims)->transpose();
   #imag2d($img_pdl(3000)->reshape(28,28)/256);

lib/AI/Nerl.pm  view on Meta::CPAN

use PDL;
use AI::Nerl::Network;

# ABSTRACT: Neural networks with backpropagation.

# main_module

our $VERSION = .03;

#A Nerl is a mechanism to build neural networks?
#Give it training,test, and cv data?
#it settles on a learning rate and stuff?
#or maybe it's also a language for guided training?
#or maybe a visual gui thing?
#Not exactly sure. Maybe I'm tinkering with forces better left alone.
#That's a great excuse for failing horribly.


=head1 AI::Nerl - A sort of stackable neural network builder thing.

=head1 SYNOPSIS

lib/AI/Nerl/Network.pm  view on Meta::CPAN

=head1 COPYRIGHT

Copyright 2012 by Zach Morgan

This package is free software; you can redistribute it and/or modify it under the 
same terms as Perl itself.

=cut

# Simple nn with 1 hidden layer
# train with $nn->train(data,labels);

has scale_input => (
   is => 'ro',
   required => 0,
   isa => 'Num',
   default => 0,
);

# number of input,hidden,output neurons
has [qw/ l1 l2 l3 /] => (

 view all matches for this distribution
 view release on metacpan -  search on metacpan