img2nn.pl

#!/usr/bin/perl

# neural net to represent an image by mapping (x,y) => (r,g,b)
# James Stanley 2021
# inspired by https://news.ycombinator.com/item?id=28448626

use strict;
use warnings;

use AI::FANN qw(:all);
use GD;

my $usage = "usage: img2nn IMGFILE\n";
my $imgfile = shift or die $usage;

my $img = GD::Image->new($imgfile)
    or die "can't open $imgfile: $!\n";
my ($w,$h) = $img->getBounds;

my @train;

for my $y (0 .. $h-1) {
    for my $x (0 .. $w-1) {
        my $col = $img->getPixel($x,$y);
        my ($r,$g,$b) = $img->rgb($col);
        $r = $r/255;
        $g = $g/255;
        $b = $b/255;
        push @train, [($x-$w/2)/($w/2),($y-$h/2)/($h/2)], [$r,$g,$b];
    }
}

my $traindata = AI::FANN::TrainData->new(@train);

my $ann = AI::FANN->new_standard(2, 3,15,3, 3);

$ann->hidden_activation_function(FANN_GAUSSIAN_SYMMETRIC);
$ann->output_activation_function(FANN_ELLIOT);

my $max_epochs = 40;
my $epochs_between_reports = 100;
my $desired_error = 0;

saveimg($ann, 0);
for my $i (1..1000) {
    $max_epochs = int((4**($i/250)))+1; # get frames more frequently at the start, while more is changing
    $ann->train_on_data($traindata, $max_epochs, $epochs_between_reports, $desired_error);

    saveimg($ann, $i);
}

sub saveimg {
    my ($ann, $i) = @_;
    my ($w,$h) = (200,200);

    my $img = GD::Image->new($w,$h,1);

    for my $y (0 .. $h-1) {
        for my $x (0 .. $w-1) {
            my ($r,$g,$b) = @{ $ann->run([($x-$w/2)/($w/2),($y-$h/2)/($h/2)]) };
            $r = $r * 255;
            $g = $g * 255;
            $b = $b * 255;
            $r = 0 if $r < 0;
            $g = 0 if $g < 0;
            $b = 0 if $b < 0;
            $img->setPixel($x, $y, $img->colorAllocate($r,$g,$b));
        }
    }

    open(my $fh, '>', sprintf("anim/%04d.png", $i))
        or die "can't write under anim/: $!";
    print $fh $img->png;
    close $fh;
}