| 1 | package AI::XGBoost::Booster; | ||
| 2 | |||
| 3 | 1 | ▶ | use strict; |
|
Subroutine
BEGIN: called CC=1 SCAR=0.0
| |||
| 4 | 1 | ▶ | use warnings; |
|
Subroutine
BEGIN: called CC=1 SCAR=0.0
| |||
| 5 | 1 | ▶ | use utf8; |
|
Subroutine
BEGIN: called CC=1 SCAR=0.0
| |||
| 6 | |||
| 7 | our $VERSION = '0.11'; # VERSION | ||
| 8 | |||
| 9 | # ABSTRACT: XGBoost main class for training, prediction and evaluation | ||
| 10 | |||
| 11 | 1 | ▶ | use Moose; |
|
Subroutine
BEGIN: called CC=1 SCAR=0.0
| |||
| 12 | 1 | ▶ | use AI::XGBoost::CAPI qw(:all); |
|
Subroutine
BEGIN: called CC=1 SCAR=2.6
| |||
| 13 | use namespace::autoclean; | ||
| 14 | |||
| 15 | has _handle => ( is => 'rw', | ||
| 16 | init_arg => undef, ); | ||
| 17 | |||
| 18 | sub update { | ||
| 19 | my $self = shift; | ||
| 20 | my %args = @_; | ||
| 21 | my ( $iteration, $dtrain ) = @args{qw(iteration dtrain)}; | ||
| 22 | XGBoosterUpdateOneIter( $self->_handle, $iteration, $dtrain->handle ); | ||
| 23 | return $self; | ||
| 24 | } | ||
| 25 | |||
| 26 | sub boost { | ||
| 27 | my $self = shift; | ||
| 28 | my %args = @_; | ||
| 29 | my ( $dtrain, $grad, $hess ) = @args{qw(dtrain grad hess)}; | ||
| 30 | XGBoosterBoostOneIter( $self->_handle, $dtrain, $grad, $hess ); | ||
| 31 | return $self; | ||
| 32 | } | ||
| 33 | |||
| 34 | sub predict { | ||
| 35 | my $self = shift; | ||
| 36 | my %args = @_; | ||
| 37 | my $data = $args{'data'}; | ||
| 38 | my $result = XGBoosterPredict( $self->_handle, $data->handle ); | ||
| 39 | my $result_size = scalar @$result; | ||
| 40 | my $matrix_rows = $data->num_row; | ||
| 41 | if ( $result_size != $matrix_rows && $result_size % $matrix_rows == 0 ) { | ||
| 42 | my $col_size = $result_size / $matrix_rows; | ||
| 43 | return [ map { [ @$result[ $_ * $col_size .. $_ * $col_size + $col_size - 1 ] ] } 0 .. $matrix_rows - 1 ]; | ||
| 44 | } | ||
| 45 | return $result; | ||
| 46 | } | ||
| 47 | |||
| 48 | sub set_param { | ||
| 49 | my $self = shift; | ||
| 50 | my ( $name, $value ) = @_; | ||
| 51 | XGBoosterSetParam( $self->_handle, $name, $value ); | ||
| 52 | return $self; | ||
| 53 | } | ||
| 54 | |||
| 55 | sub set_attr { | ||
| 56 | my $self = shift; | ||
| 57 | my ( $name, $value ) = @_; | ||
| 58 | XGBoosterSetAttr( $self->_handle, $name, $value ); | ||
| 59 | return $self; | ||
| 60 | } | ||
| 61 | |||
| 62 | sub get_attr { | ||
| 63 | my $self = shift; | ||
| 64 | my ($name) = @_; | ||
| 65 | XGBoosterGetAttr( $self->_handle, $name ); | ||
| 66 | } | ||
| 67 | |||
| 68 | sub get_score { | ||
| 69 | my $self = shift; | ||
| 70 | my %args = @_; | ||
| 71 | my ( $fmap, $importance_type ) = @args{qw(fmap importance_type)}; | ||
| 72 | |||
| 73 | if ( $importance_type eq "weight" ) { | ||
| 74 | my @trees = $self->get_dump; | ||
| 75 | } else { | ||
| 76 | |||
| 77 | } | ||
| 78 | |||
| 79 | } | ||
| 80 | |||
| 81 | sub get_dump { | ||
| 82 | my $self = shift; | ||
| 83 | return XGBoosterDumpModelEx( $self->_handle, "", 1, "text" ); | ||
| 84 | } | ||
| 85 | |||
| 86 | sub attributes { | ||
| 87 | my $self = shift; | ||
| 88 | return { map { $_ => $self->get_attr($_) } @{ XGBoosterGetAttrNames( $self->_handle ) } }; | ||
| 89 | } | ||
| 90 | |||
| 91 | sub TO_JSON { | ||
| 92 | my $self = shift; | ||
| 93 | my $trees = XGBoosterDumpModelEx( $self->_handle, "", 1, "json" ); | ||
| 94 | return "[" . join( ',', @$trees ) . "]"; | ||
| 95 | } | ||
| 96 | |||
| 97 | sub BUILD { | ||
| 98 | my $self = shift; | ||
| 99 | my $args = shift; | ||
| 100 | $self->_handle( XGBoosterCreate( [ map { $_->handle } @{ $args->{'cache'} } ] ) ); | ||
| 101 | } | ||
| 102 | |||
| 103 | sub DEMOLISH { | ||
| 104 | my $self = shift(); | ||
| 105 | XGBoosterFree( $self->_handle ); | ||
| 106 | } | ||
| 107 | |||
| 108 | __PACKAGE__->meta->make_immutable(); | ||
| 109 | |||
| 110 | 1; | ||
| 111 | |||
| 112 | __END__ | ||
| 113 | |||
| 114 | =pod | ||
| 115 | |||
| 116 | =encoding utf-8 | ||
| 117 | |||
| 118 | =head1 NAME | ||
| 119 | |||
| 120 | AI::XGBoost::Booster - XGBoost main class for training, prediction and evaluation | ||
| 121 | |||
| 122 | =head1 VERSION | ||
| 123 | |||
| 124 | version 0.11 | ||
| 125 | |||
| 126 | =head1 SYNOPSIS | ||
| 127 | |||
| 128 | use 5.010; | ||
| 129 | use aliased 'AI::XGBoost::DMatrix'; | ||
| 130 | use AI::XGBoost qw(train); | ||
| 131 | |||
| 132 | # We are going to solve a binary classification problem: | ||
| 133 | # Mushroom poisonous or not | ||
| 134 | |||
| 135 | my $train_data = DMatrix->From(file => 'agaricus.txt.train'); | ||
| 136 | my $test_data = DMatrix->From(file => 'agaricus.txt.test'); | ||
| 137 | |||
| 138 | # With XGBoost we can solve this problem using 'gbtree' booster | ||
| 139 | # and as loss function a logistic regression 'binary:logistic' | ||
| 140 | # (Gradient Boosting Regression Tree) | ||
| 141 | # XGBoost Tree Booster has a lot of parameters that we can tune | ||
| 142 | # (https://github.com/dmlc/xgboost/blob/master/doc/parameter.md) | ||
| 143 | |||
| 144 | my $booster = train(data => $train_data, number_of_rounds => 10, params => { | ||
| 145 | objective => 'binary:logistic', | ||
| 146 | eta => 1.0, | ||
| 147 | max_depth => 2, | ||
| 148 | silent => 1 | ||
| 149 | }); | ||
| 150 | |||
| 151 | # For binay classification predictions are probability confidence scores in [0, 1] | ||
| 152 | # indicating that the label is positive (1 in the first column of agaricus.txt.test) | ||
| 153 | my $predictions = $booster->predict(data => $test_data); | ||
| 154 | |||
| 155 | say join "\n", @$predictions[0 .. 10]; | ||
| 156 | |||
| 157 | =head1 DESCRIPTION | ||
| 158 | |||
| 159 | Booster objects control training, prediction and evaluation | ||
| 160 | |||
| 161 | Work In Progress, the API may change. Comments and suggestions are welcome! | ||
| 162 | |||
| 163 | =head1 METHODS | ||
| 164 | |||
| 165 | =head2 update | ||
| 166 | |||
| 167 | Update one iteration | ||
| 168 | |||
| 169 | =head3 Parameters | ||
| 170 | |||
| 171 | =over 4 | ||
| 172 | |||
| 173 | =item iteration | ||
| 174 | |||
| 175 | Current iteration number | ||
| 176 | |||
| 177 | =item dtrain | ||
| 178 | |||
| 179 | Training data (AI::XGBoost::DMatrix) | ||
| 180 | |||
| 181 | =back | ||
| 182 | |||
| 183 | =head2 boost | ||
| 184 | |||
| 185 | Boost one iteration using your own gradient | ||
| 186 | |||
| 187 | =head3 Parameters | ||
| 188 | |||
| 189 | =over 4 | ||
| 190 | |||
| 191 | =item dtrain | ||
| 192 | |||
| 193 | Training data (AI::XGBoost::DMatrix) | ||
| 194 | |||
| 195 | =item grad | ||
| 196 | |||
| 197 | Gradient of your objective function (Reference to an array) | ||
| 198 | |||
| 199 | =item hess | ||
| 200 | |||
| 201 | Hessian of your objective function, that is, second order gradient (Reference to an array) | ||
| 202 | |||
| 203 | =back | ||
| 204 | |||
| 205 | =head2 predict | ||
| 206 | |||
| 207 | Predict data using the trained model | ||
| 208 | |||
| 209 | =head3 Parameters | ||
| 210 | |||
| 211 | =over 4 | ||
| 212 | |||
| 213 | =item data | ||
| 214 | |||
| 215 | Data to predict | ||
| 216 | |||
| 217 | =back | ||
| 218 | |||
| 219 | =head2 set_param | ||
| 220 | |||
| 221 | Set booster parameter | ||
| 222 | |||
| 223 | =head3 Example | ||
| 224 | |||
| 225 | $booster->set_param('objective', 'binary:logistic'); | ||
| 226 | |||
| 227 | =head2 set_attr | ||
| 228 | |||
| 229 | Set a string attribute | ||
| 230 | |||
| 231 | =head2 get_attr | ||
| 232 | |||
| 233 | Get a string attribute | ||
| 234 | |||
| 235 | =head2 get_score | ||
| 236 | |||
| 237 | Get importance of each feature | ||
| 238 | |||
| 239 | =head3 Parameters | ||
| 240 | |||
| 241 | =over 4 | ||
| 242 | |||
| 243 | =item importance_type | ||
| 244 | |||
| 245 | Type of importance. Valid values: | ||
| 246 | |||
| 247 | =over 4 | ||
| 248 | |||
| 249 | =item weight | ||
| 250 | |||
| 251 | Number of times a feature is used to split the data across all trees | ||
| 252 | |||
| 253 | =item gain | ||
| 254 | |||
| 255 | Average gain of the feature when it is used in trees | ||
| 256 | |||
| 257 | =item cover | ||
| 258 | |||
| 259 | Average coverage of the feature when it is used in trees | ||
| 260 | |||
| 261 | =back | ||
| 262 | |||
| 263 | =item fmap | ||
| 264 | |||
| 265 | Name of feature map file | ||
| 266 | |||
| 267 | =back | ||
| 268 | |||
| 269 | =head2 get_dump | ||
| 270 | |||
| 271 | =head2 attributes | ||
| 272 | |||
| 273 | Returns all attributes of the booster as a HASHREF | ||
| 274 | |||
| 275 | =head2 TO_JSON | ||
| 276 | |||
| 277 | Serialize the booster to JSON. | ||
| 278 | |||
| 279 | This method is to be used with the option C<convert_blessed> from L<JSON>. | ||
| 280 | (See L<https://metacpan.org/pod/JSON#OBJECT-SERIALISATION>) | ||
| 281 | |||
| 282 | Warning: this API is subject to changes | ||
| 283 | |||
| 284 | =head2 BUILD | ||
| 285 | |||
| 286 | Use new, this method is just an internal helper | ||
| 287 | |||
| 288 | =head2 DEMOLISH | ||
| 289 | |||
| 290 | Internal destructor. This method is called automatically | ||
| 291 | |||
| 292 | =head1 AUTHOR | ||
| 293 | |||
| 294 | Pablo Rodríguez González <pablo.rodriguez.gonzalez@gmail.com> | ||
| 295 | |||
| 296 | =head1 COPYRIGHT AND LICENSE | ||
| 297 | |||
| 298 | Copyright (c) 2017 by Pablo Rodríguez González. | ||
| 299 | |||
| 300 | =cut | ||