| 1 | # Licensed to the Apache Software Foundation (ASF) under one | ||
| 2 | # or more contributor license agreements. See the NOTICE file | ||
| 3 | # distributed with this work for additional information | ||
| 4 | # regarding copyright ownership. The ASF licenses this file | ||
| 5 | # to you under the Apache License, Version 2.0 (the | ||
| 6 | # "License"); you may not use this file except in compliance | ||
| 7 | # with the License. You may obtain a copy of the License at | ||
| 8 | # | ||
| 9 | # http://www.apache.org/licenses/LICENSE-2.0 | ||
| 10 | # | ||
| 11 | # Unless required by applicable law or agreed to in writing, | ||
| 12 | # software distributed under the License is distributed on an | ||
| 13 | # "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY | ||
| 14 | # KIND, either express or implied. See the License for the | ||
| 15 | # specific language governing permissions and limitations | ||
| 16 | # under the License. | ||
| 17 | |||
| 18 | use strict; | ||
| 19 | use warnings; | ||
| 20 | package AI::MXNet::Gluon::Contrib::NN::BasicLayers; | ||
| 21 | |||
| 22 | =head1 NAME | ||
| 23 | |||
| 24 | AI::MXNet::Gluon::Contrib::NN::BasicLayers - An additional collection of Gluon's building blocks. | ||
| 25 | =cut | ||
| 26 | |||
| 27 | use AI::MXNet::Function::Parameters; | ||
| 28 | package AI::MXNet::Gluon::NN::Concurrent; | ||
| 29 | use AI::MXNet::Gluon::Mouse; | ||
| 30 | extends 'AI::MXNet::Gluon::NN::Sequential'; | ||
| 31 | |||
| 32 | =head1 NAME | ||
| 33 | |||
| 34 | AI::MXNet::Gluon::NN::Concurrent - Lays Blocks concurrently. | ||
| 35 | =cut | ||
| 36 | |||
| 37 | =head1 DESCRIPTION | ||
| 38 | |||
| 39 | Lays Blocks concurrently. | ||
| 40 | |||
| 41 | This block feeds its input to all children blocks, and | ||
| 42 | produces the output by concatenating all the children blocks' outputs | ||
| 43 | on the specified axis. | ||
| 44 | |||
| 45 | Example: | ||
| 46 | |||
| 47 | $net = nn->Concurrent(); | ||
| 48 | # use net's name_scope to give children blocks appropriate names. | ||
| 49 | $net->name_scope(sub { | ||
| 50 | $net->add(nn->Dense(10, activation=>'relu')); | ||
| 51 | $net->add(nn->Dense(20)); | ||
| 52 | $net->add(nn->Identity()); | ||
| 53 | }); | ||
| 54 | |||
| 55 | Parameters | ||
| 56 | ---------- | ||
| 57 | axis : int, default -1 | ||
| 58 | The axis on which to concatenate the outputs. | ||
| 59 | =cut | ||
| 60 | has 'axis' => (is => 'rw', isa => 'Int', default => -1); | ||
| 61 | method python_constructor_arguments() { ['axis'] } | ||
| 62 | |||
| 63 | method forward(GluonInput $x) | ||
| 64 | { | ||
| 65 | return AI::MXNet::NDArray->concat((map { $_->($x) } $self->_children->values), dim=>$self->axis); | ||
| 66 | } | ||
| 67 | |||
| 68 | __PACKAGE__->register('AI::MXNet::Gluon::NN'); | ||
| 69 | |||
| 70 | package AI::MXNet::Gluon::NN::HybridConcurrent; | ||
| 71 | use AI::MXNet::Gluon::Mouse; | ||
| 72 | extends 'AI::MXNet::Gluon::NN::HybridSequential'; | ||
| 73 | |||
| 74 | =head1 NAME | ||
| 75 | |||
| 76 | AI::MXNet::Gluon::NN::HybridConcurrent - Lays HubridBlocks concurrently. | ||
| 77 | =cut | ||
| 78 | |||
| 79 | =head1 DESCRIPTION | ||
| 80 | |||
| 81 | Lays HybridBlocks concurrently. | ||
| 82 | |||
| 83 | This block feeds its input to all children blocks, and | ||
| 84 | produces the output by concatenating all the children blocks' outputs | ||
| 85 | on the specified axis. | ||
| 86 | |||
| 87 | Example: | ||
| 88 | |||
| 89 | $net = nn->HybridConcurrent(); | ||
| 90 | # use net's name_scope to give children blocks appropriate names. | ||
| 91 | $net->name_scope(sub { | ||
| 92 | $net->add(nn->Dense(10, activation=>'relu')); | ||
| 93 | $net->add(nn->Dense(20)); | ||
| 94 | $net->add(nn->Identity()); | ||
| 95 | }); | ||
| 96 | |||
| 97 | Parameters | ||
| 98 | ---------- | ||
| 99 | axis : int, default -1 | ||
| 100 | The axis on which to concatenate the outputs. | ||
| 101 | =cut | ||
| 102 | has 'axis' => (is => 'rw', isa => 'Int', default => -1); | ||
| 103 | method python_constructor_arguments() { ['axis'] } | ||
| 104 | |||
| 105 | method hybrid_forward(GluonClass $F, GluonInput $x) | ||
| 106 | { | ||
| 107 | return $F->concat((map { $_->($x) } $self->_children->values), dim=>$self->axis); | ||
| 108 | } | ||
| 109 | |||
| 110 | __PACKAGE__->register('AI::MXNet::Gluon::NN'); | ||
| 111 | |||
| 112 | package AI::MXNet::Gluon::NN::Identity; | ||
| 113 | use AI::MXNet::Gluon::Mouse; | ||
| 114 | extends 'AI::MXNet::Gluon::HybridBlock'; | ||
| 115 | |||
| 116 | =head1 NAME | ||
| 117 | |||
| 118 | AI::MXNet::Gluon::NN::Identity - Block that passes through the input directly. | ||
| 119 | =cut | ||
| 120 | |||
| 121 | =head1 DESCRIPTION | ||
| 122 | |||
| 123 | Block that passes through the input directly. | ||
| 124 | |||
| 125 | This block can be used in conjunction with HybridConcurrent | ||
| 126 | block for residual connection. | ||
| 127 | |||
| 128 | Example: | ||
| 129 | |||
| 130 | $net = nn->HybridConcurrent(); | ||
| 131 | # use net's name_scope to give child Blocks appropriate names. | ||
| 132 | $net->name_scope(sub { | ||
| 133 | $net->add(nn->Dense(10, activation=>'relu')); | ||
| 134 | $net->add(nn->Dense(20)); | ||
| 135 | $net->add(nn->Identity()); | ||
| 136 | }); | ||
| 137 | =cut | ||
| 138 | |||
| 139 | method hybrid_forward(GluonClass $F, GluonInput $x) | ||
| 140 | { | ||
| 141 | return $x; | ||
| 142 | } | ||
| 143 | |||
| 144 | __PACKAGE__->register('AI::MXNet::Gluon::NN'); | ||
| 145 | |||
| 146 | package AI::MXNet::Gluon::NN::SparseEmbedding; | ||
| 147 | use AI::MXNet::Gluon::Mouse; | ||
| 148 | extends 'AI::MXNet::Gluon::Block'; | ||
| 149 | |||
| 150 | =head1 NAME | ||
| 151 | |||
| 152 | AI::MXNet::Gluon::NN::SparseEmbedding - Turns non-negative integers (indexes/tokens) into dense vectors. | ||
| 153 | =cut | ||
| 154 | |||
| 155 | =head1 DESCRIPTION | ||
| 156 | |||
| 157 | Turns non-negative integers (indexes/tokens) into dense vectors | ||
| 158 | of fixed size. eg. [4, 20] -> [[0.25, 0.1], [0.6, -0.2]] | ||
| 159 | |||
| 160 | This SparseBlock is designed for distributed training with extremely large | ||
| 161 | input dimension. Both weight and gradient w.r.t. weight are AI::MXNet::NDArray::RowSparse. | ||
| 162 | |||
| 163 | Parameters | ||
| 164 | ---------- | ||
| 165 | input_dim : int | ||
| 166 | Size of the vocabulary, i.e. maximum integer index + 1. | ||
| 167 | output_dim : int | ||
| 168 | Dimension of the dense embedding. | ||
| 169 | dtype : Dtype, default 'float32' | ||
| 170 | Data type of output embeddings. | ||
| 171 | weight_initializer : Initializer | ||
| 172 | Initializer for the embeddings matrix. | ||
| 173 | =cut | ||
| 174 | |||
| 175 | has 'input_dim' => (is => 'ro', isa => 'Int', required => 1); | ||
| 176 | has 'output_dim' => (is => 'ro', isa => 'Int', required => 1); | ||
| 177 | has 'dtype' => (is => 'ro', isa => 'Dtype', default => 'float32'); | ||
| 178 | has 'weight_initializer' => (is => 'ro', isa => 'Maybe[Initializer]'); | ||
| 179 | method python_constructor_arguments() { [qw/input_dim output_dim dtype weight_initializer/] } | ||
| 180 | |||
| 181 | sub BUILD | ||
| 182 | { | ||
| 183 | my $self = shift; | ||
| 184 | $self->_kwargs({ | ||
| 185 | input_dim => $self->input_dim, | ||
| 186 | output_dim => $self->output_dim, | ||
| 187 | dtype => $self->dtype, | ||
| 188 | sparse_grad => 1 | ||
| 189 | }); | ||
| 190 | $self->weight($self->params->get('weight', shape=>[$self->input_dim, $self->output_dim], | ||
| 191 | init=>$self->weight_initializer, dtype=>$self->dtype, | ||
| 192 | grad_stype=>'row_sparse', stype=>'row_sparse')); | ||
| 193 | } | ||
| 194 | |||
| 195 | method forward(GluonInput $x) | ||
| 196 | { | ||
| 197 | my $weight = $self->weight->row_sparse_data($x); | ||
| 198 | return AI::MXNet::NDArray->Embedding($x, $weight, { name=>'fwd', %{ $self->_kwargs } }); | ||
| 199 | } | ||
| 200 | |||
| 201 | use overload '""' => sub { | ||
| 202 | my $self = shift; | ||
| 203 | $self->_class_name.'('.$self->input_dim.' -> '.$self->input_dim.', '.$self->dtype.')'; | ||
| 204 | }; | ||
| 205 | |||
| 206 | __PACKAGE__->register('AI::MXNet::Gluon::NN'); | ||
| 207 | |||
| 208 | 1; |