rspamd/utils/fann_train.pl

#!/usr/bin/env perl

# This script is a very simple prototype to learn fann from rspamd logs
# For now, it is intended for internal use only

use strict;
use warnings FATAL => 'all';
use AI::FANN qw(:all);
use Getopt::Std;

my %sym_idx; # Symbols by index
my %sym_names; # Symbols by name
my $num = 1; # Number of symbols
my @spam;
my @ham;
my $max_samples = -1;
my $split = 1;
my $preprocessed = 0; # ouptut is in format <score>:<0|1>:<SYM1,...SYMN>
my $score_spam = 12;
my $score_ham = -6;

sub process {
    my ($input, $spam, $ham) = @_;
    my $samples = 0;

    while(<$input>) {
        if (!$preprocessed) {
            if (/^.*rspamd_task_write_log.*: \[(-?\d+\.?\d*)\/(\d+\.?\d*)\]\s*\[(.+)\].*$/) {
                if ($1 > $score_spam) {
                    $_ = "$1:1: $3";
                }
                elsif ($1 < $score_ham) {
                    $_ = "$1:0: $3\n";
                }
                else {
                    # Out of boundary
                    next;
                }
            }
            else {
                # Not our log message
                next;
            }
        }

        $_ =~ /^(-?\d+\.?\d*):([01]):\s*(\S.*)$/;

        my $is_spam = 0;

        if ($2 == 1) {
            $is_spam = 1;
        }

        my @ar = split /,/, $3;
        my %sample;

        foreach my $sym (@ar) {
            chomp $sym;
            if (!$sym_idx{$sym}) {
                $sym_idx{$sym} = $num;
                $sym_names{$num} = $sym;
                $num++;
            }

            $sample{$sym_idx{$sym}} = 1;
        }

        if ($is_spam) {
            push @{$spam}, \%sample;
        }
        else {
            push @{$ham}, \%sample;
        }

        $samples++;
        if ($max_samples > 0 && $samples > $max_samples) {
            return;
        }
    }
}

# Shuffle array
sub fisher_yates_shuffle
{
    my $array = shift;
    my $i = @$array;

    while ( --$i ) {
        my $j = int rand( $i + 1 );
        @$array[$i, $j] = @$array[$j, $i];
    }
}

# Train network
sub train {
    my ($ann, $sample, $result) = @_;

    my @row;

    for (my $i = 1; $i < $num; $i++) {
        if ($sample->{$i}) {
            push @row, 1;
        }
        else {
            push @row, 0;
        }
    }

    #print "@row -> @{$result}\n";

    $ann->train(\@row, \@{$result});
}

sub test {
    my ($ann, $sample) = @_;

    my @row;

    for (my $i = 1; $i < $num; $i++) {
        if ($sample->{$i}) {
            push @row, 1;
        }
        else {
            push @row, 0;
        }
    }

    my $ret = $ann->run(\@row);

    return $ret;
}

my %opts;
getopts('o:i:s:n:t:hpS:H:', \%opts);

if ($opts{'h'}) {
    print "$0 [-i input] [-o output] [-s scores] [-n max_samples] [-S spam_score] [-H ham_score] [-ph]\n";
    exit;
}

my $input = *STDIN;

if ($opts{'i'}) {
    open($input, '<', $opts{'i'}) or die "cannot open $opts{i}";
}

if ($opts{'n'}) {
    $max_samples = $opts{'n'};
}

if ($opts{'t'}) {
    # Test split
    $split = $opts{'t'};
}
if ($opts{'p'}) {
    $preprocessed = 1;
}

if ($opts{'H'}) {
    $score_ham = $opts{'H'};
}

if ($opts{'S'}) {
    $score_spam = $opts{'S'};
}

# ham_prob, spam_prob
my @spam_out = (1);
my @ham_out = (0);

process($input, \@spam, \@ham);
fisher_yates_shuffle(\@spam);
fisher_yates_shuffle(\@ham);

my $nspam = int(scalar(@spam) / $split);
my $nham = int(scalar(@ham) / $split);

my $ann = AI::FANN->new_standard($num - 1, ($num + 2) / 2, 1);

my @train_data;
# Train ANN
for (my $i = 0; $i < $nham; $i++) {
    push @train_data, [ $ham[$i], \@ham_out ];
}

for (my $i = 0; $i < $nspam; $i++) {
    push @train_data, [ $spam[$i], \@spam_out ];
}

fisher_yates_shuffle(\@train_data);

foreach my $train_row (@train_data) {
    train($ann, @{$train_row}[0], @{$train_row}[1]);
}

print "Trained $nspam SPAM and $nham HAM samples\n";

# Now run fann
if ($split > 1) {
    my $sample = 0.0;
    my $correct = 0.0;
    for (my $i = $nham; $i < $nham * $split; $i++) {
        my $ret = test($ann, $ham[$i]);
        #print "@{$ret}\n";
        if (@{$ret}[0] < 0.5) {
            $correct++;
        }
        $sample++;
    }

    print "Tested $sample HAM samples, correct matched: $correct, rate: ".($correct / $sample)."\n";

    $sample = 0.0;
    $correct = 0.0;

    for (my $i = $nspam; $i < $nspam * $split; $i++) {
        my $ret = test($ann, $spam[$i]);
        #print "@{$ret}\n";
        if (@{$ret}[0] > 0.5) {
            $correct++;
        }
        $sample++;
    }

    print "Tested $sample SPAM samples, correct matched: $correct, rate: ".($correct / $sample)."\n";
}

if ($opts{'o'}) {
    $ann->save($opts{'o'}) or die "cannot save ann into $opts{o}";
}

if ($opts{'s'}) {
    open(my $scores, '>',
        $opts{'s'}) or die "cannot open score file $opts{'s'}";
    print $scores "{";
    for (my $i = 1; $i < $num; $i++) {
        my $n = $i - 1;
        if ($i != $num - 1) {
            print $scores "\"$sym_names{$i}\":$n,";
        }
        else {
            print $scores "\"$sym_names{$i}\":$n}\n";
        }
    }
}