lib/Peta/NN/Layer/Dense.pm

filter: Statementstmt 100% 23 / 23 Branchbran 100% 4 / 4 Conditioncond n/a MC/DCmcdc n/a Subroutinesub 100% 9 / 9 Podpod 100% 8 / 8 totaltotal 100% 44 / 44 CC 3 SCAR 3.1
CC
3
Cov
100%
init
2.0
param_cols
2.0
BEGIN
1.0
CRAP 1.4
« lib/Peta/NN/Layer/Activation.pm ↑ Summary lib/Peta/NN/Layer/Embed.pm »
1 package Peta::NN::Layer::Dense;
2 # ABSTRACT: fully connected layer
3  
4 # Fully connected layer: Y = X W^T + b, for a whole batch of rows X. The
5 # weights are a backend tensor of one row per output.
6  
7 27 ▶use v5.36;
Subroutine BEGIN: called CC=1 SCAR=0.0
8  
9 27 our $VERSION = '0.2610090';
10  
11 302 ▶sub new ($class, %arg) {
Subroutine new: called CC=1 SCAR=0.0
Pod documented
12 302 return bless { n_out => $arg{out}, need_dx => 1 }, $class;
13 }
14  
15 3195 ▶sub type ($self) { return 'dense' }
Subroutine type: called CC=1 SCAR=0.0
Pod documented
16  
17 # Glorot uniform: keeps the signal's variance about constant through the layer
18 # for tanh and sigmoid, and is close enough to He's for the small ReLU nets here.
19 301 ▶sub init ($self, $n_in, $rng, $backend) {
Subroutine init: called CC=2 SCAR=6.9
Pod documented
20 301 ▶ die "dense cannot read tokens; put an embed layer first\n" if ref $n_in;
Branch: 1 branch
BranchTrueFalseTotal
if ref $n_in 1 300 301
21 300 my $n_out = $self->{n_out};
22 300 my $limit = sqrt(6 / ($n_in + $n_out));
23 300 $self->{backend} = $backend;
24 300 $self->{n_in} = $n_in;
25 300 $self->{W} = $backend->tensor([ map { (2 * $rng->uniform - 1) * $limit } 1 .. $n_in * $n_out ], $n_in);
26 300 $self->{b} = $backend->tensor([ (0.0) x $n_out ], $n_out);
27 300 return $n_out;
28 }
29  
30 34949 ▶sub param_names ($self) { return qw(W b) }
Subroutine param_names: called CC=1 SCAR=0.0
Pod documented
31  
32 # Numbers per row of a parameter tensor: W has one row per output.
33 8008 ▶sub param_cols ($self, $name) { return $name eq 'W' ? $self->{n_in} : $self->{n_out} }
Subroutine param_cols: called CC=2 SCAR=6.9
Pod documented
Branch: 1 branch
BranchTrueFalseTotal
$name eq 'W' ? : 4004 4004 8008
34  
35 191094 ▶sub forward ($self, $X) {
Subroutine forward: called CC=1 SCAR=0.0
Pod documented
36 191094 $self->{X} = $X;
37 191094 return $self->{backend}->affine($X, $self->{W}, $self->{b});
38 }
39  
40 # Leaves the batch's gradient in gW and gb and returns the gradient with
41 # respect to the input. The first layer of a net has nobody to hand that to,
42 # so the net clears need_dx there and the backend skips it.
43 186200 ▶sub backward ($self, $dY) {
Subroutine backward: called CC=1 SCAR=0.0
Pod documented
44 186200 my @grad = $self->{backend}->affine_grad($self->{X}, $self->{W}, $dY, $self->{need_dx});
45 186200 @$self{qw(gW gb)} = @grad[ 0, 1 ];
46 186200 return $grad[2];
47 }
48  
49 20 ▶sub spec ($self) { return [ dense => $self->{n_out} ] }
Subroutine spec: called CC=1 SCAR=0.0
Pod documented
50  
51 27 1;
52  
53 __END__
« lib/Peta/NN/Layer/Activation.pm ↑ Summary lib/Peta/NN/Layer/Embed.pm »