lib/Peta/NN/Backend/PDL.pm

filter: Statementstmt 97.4% 77 / 79 Branchbran 83.3% 5 / 6 Conditioncond 50.0% 2 / 4 MC/DCmcdc 0.0% 0 / 2 Subroutinesub 100% 28 / 28 Podpod 100% 17 / 17 totaltotal 94.8% 129 / 136 CC 8 SCAR 4.5
CC
8
Cov
94%
mse
3.5
softmax_ce
3.0
activate
2.1
CRAP 1.6
« lib/Peta/NN/RNG.pm ↑ Summary lib/Peta/NN/Layer/Activation.pm »
1 package Peta::NN::Backend::PDL;
2 # ABSTRACT: Peta::NN on PDL ndarrays
3  
4 # The CPU backend on PDL: a tensor is an ndarray of dims (cols, rows), which
5 # is PDL's own order for "rows of cols numbers", so a flat row-after-row list
6 # maps onto it without rearranging. Every operation is a handful of whole-array
7 # expressions; the batch never passes through a Perl loop.
8  
9 8 ▶use v5.36;
Subroutine BEGIN: called CC=1 SCAR=0.0
10  
11 # PDL exports dozens of functions, flat() among them, which is a method here.
12 # They are imported into a package of their own; this one calls PDL by name.
13 {
14 package Peta::NN::Backend::PDL::Imports;
15 8 ▶ use PDL::LiteF;
Subroutine BEGIN: called CC=1 SCAR=0.0
16 }
17  
18 8 ▶use parent -norequire, 'Peta::NN::Backend';
Subroutine BEGIN: called CC=1 SCAR=0.0
19 8 ▶use Peta::NN::Backend;
Subroutine BEGIN: called CC=1 SCAR=0.0
20  
21 8 our $VERSION = '0.2610090';
22  
23 8 my $TINY = 1e-300; # keeps log() defined when a probability underflows
24  
25 77 ▶sub new ($class) { return bless {}, $class }
Subroutine new: called CC=1 SCAR=0.0
Pod documented
26  
27 1732 ▶sub name ($self) { return 'pdl' }
Subroutine name: called CC=1 SCAR=0.0
Pod documented
28  
29 25068 ▶sub tensor ($self, $flat, $cols) {
Subroutine tensor: called CC=1 SCAR=0.0
Pod documented
30 25068 return PDL->pdl(PDL::double(), $flat)->reshape($cols, @$flat / $cols);
31 }
32  
33 # Tokens stay a Perl list: all a backend does with them is build an index.
34 42344 ▶sub tokens ($self, $flat, $per_row) { return [ [@$flat], $per_row ] }
Subroutine tokens: called CC=1 SCAR=0.0
Pod documented
35  
36 16045 ▶sub flat ($self, $tensor) { return [ $tensor->list ] }
Subroutine flat: called CC=1 SCAR=0.0
Pod documented
37  
38 1627 ▶sub zeros_like ($self, $tensor) { return PDL->zeroes(PDL::double(), $tensor->dims) }
Subroutine zeros_like: called CC=1 SCAR=0.0
Pod documented
39  
40 # X is (in, batch), W is (in, out): X x W^T is (out, batch), and b of dims
41 # (out) broadcasts over the rows.
42 107642 ▶sub affine ($self, $X, $W, $b) {
Subroutine affine: called CC=1 SCAR=0.0
Pod documented
43 107642 return ($X x $W->transpose) + $b;
44 }
45  
46 104630 ▶sub affine_grad ($self, $X, $W, $dY, $need_dx) {
Subroutine affine_grad: called CC=2 SCAR=6.9
Pod documented
47 104630 my $gW = $dY->transpose x $X; # (in, out)
48 104630 my $gb = $dY->mv(1, 0)->sumover; # summed over the batch: (out)
49 104630 ▶ return ($gW, $gb) if !$need_dx;
Branch: 1 branch
BranchTrueFalseTotal
unless $need_dx 10809 93821 104630
50 93821 return ($gW, $gb, $dY x $W); # (in, batch)
51 }
52  
53 my %ACTIVATE = (
54 42560 ▶ relu => sub ($x) { $x * ($x > 0) },
Subroutine __ANON__: called CC=1 SCAR=0.0
55 11221 ▶ tanh => sub ($x) { PDL::tanh($x) },
Subroutine __ANON__: called CC=1 SCAR=0.0
56 202 ▶ sigmoid => sub ($x) { 1 / (1 + exp(-$x)) },
Subroutine __ANON__: called CC=1 SCAR=0.0
57 8 );
58  
59 my %ACTIVATE_GRAD = (
60 41513 ▶ relu => sub ($y, $dy) { $dy * ($y > 0) },
Subroutine __ANON__: called CC=1 SCAR=0.0
61 10802 ▶ tanh => sub ($y, $dy) { $dy * (1 - $y * $y) },
Subroutine __ANON__: called CC=1 SCAR=0.0
62 2 ▶ sigmoid => sub ($y, $dy) { $dy * $y * (1 - $y) },
Subroutine __ANON__: called CC=1 SCAR=0.0
63 8 );
64  
65 53983 ▶sub activate ($self, $kind, $X) {
Subroutine activate: called CC=2 SCAR=7.2
Pod documented
66 53983 ▶ my $f = $ACTIVATE{$kind} // die "unknown activation '$kind'\n";
Condition50%1/2
$ACTIVATE{$kind} // die("unknown activation '${kind}'\n")
l!l
539830
MC/DC0%0/1
$ACTIVATE{$kind} // die("unknown activation '${kind}'\n")
$ACTIVATE{$kind}
Truth table50%1/2
A
A= $ACTIVATE{$kind}
Aresult
11
00
67 53983 return $f->($X);
68 }
69  
70 52317 ▶sub activate_grad ($self, $kind, $Y, $dY) {
Subroutine activate_grad: called CC=2 SCAR=7.2
Pod documented
71 52317 ▶ my $f = $ACTIVATE_GRAD{$kind} // die "unknown activation '$kind'\n";
Condition50%1/2
$ACTIVATE_GRAD{$kind} // die("unknown activation '${kind}'\n")
l!l
523170
MC/DC0%0/1
$ACTIVATE_GRAD{$kind} // die("unknown activation '${kind}'\n")
$ACTIVATE_GRAD{$kind}
Truth table50%1/2
A
A= $ACTIVATE_GRAD{$kind}
Aresult
11
00
72 52317 return $f->($Y, $dY);
73 }
74  
75 # One row per token, a single 1 in the column of its index: (vocab, tokens).
76 # Lookup and its gradient are then both one matrix product.
77 83849 ▶sub _one_hot ($tokens, $vocab) {
Subroutine _one_hot: called CC=1 SCAR=0.0
78 83849 my $index = PDL->pdl(PDL::long(), $tokens->[0]);
79 83849 my $hot = PDL->zeroes(PDL::double(), $vocab, $index->nelem);
80 83849 (my $ones = $hot->index($index)) .= 1;
81 83849 return $hot;
82 }
83  
84 42344 ▶sub embed ($self, $E, $tokens) {
Subroutine embed: called CC=1 SCAR=0.0
Pod documented
85 42344 my ($dim, $vocab) = $E->dims;
86 42344 my $rows = _one_hot($tokens, $vocab) x $E; # (dim, tokens)
87 42344 my $per_row = $tokens->[1];
88 42344 return $rows->reshape($dim * $per_row, @{ $tokens->[0] } / $per_row);
89 }
90  
91 41505 ▶sub embed_grad ($self, $E, $tokens, $dX) {
Subroutine embed_grad: called CC=1 SCAR=0.0
Pod documented
92 41505 my ($dim, $vocab) = $E->dims;
93 41505 my $count = @{ $tokens->[0] };
94 41505 return _one_hot($tokens, $vocab)->transpose x $dX->copy->reshape($dim, $count);
95 }
96  
97 46452 ▶sub softmax_ce ($self, $logits, $classes, $weights = undef) {
Subroutine softmax_ce: called CC=3 SCAR=11.0
Pod documented
98 46452 my $class = PDL->pdl(PDL::long(), $classes);
99 46452 my $p = exp($logits - $logits->maximum->dummy(0));
100 46452 $p /= $p->sumover->dummy(0);
101 46452 my $row_loss = -log($p->index($class) + $TINY);
102 46452 (my $hit = $p->index($class)) -= 1;
103 46452 ▶ return ($row_loss->sum, $p) if !$weights;
Branch: 1 branch
BranchTrueFalseTotal
unless $weights 29892 16560 46452
104 16560 my $weight = PDL->pdl(PDL::double(), $weights); # one per row
105 16560 return (($row_loss * $weight)->sum, $p * $weight->dummy(0));
106 }
107  
108 7244 ▶sub mse ($self, $out, $targets, $weights = undef) {
Subroutine mse: called CC=3 SCAR=12.5
Pod documented
109 7244 my ($cols) = $out->dims;
110 7244 my $diff = $out - $self->tensor($targets, $cols);
111 7244 ▶ return (($diff * $diff)->sum / $cols, 2 * $diff / $cols) if !$weights;
Branch: 1 branch
BranchTrueFalseTotal
unless $weights 7244 0 7244
112 0 my $weight = PDL->pdl(PDL::double(), $weights);
113 0 return ((($diff * $diff)->sumover * $weight)->sum / $cols, 2 * $diff * $weight->dummy(0) / $cols);
114 }
115  
116 # Multiply every value by $factor, in place: weight decay.
117 60 ▶sub decay ($self, $P, $factor) {
Subroutine decay: called CC=1 SCAR=0.0
Pod documented
118 60 $P *= $factor;
119 60 return;
120 }
121  
122 600 ▶sub sgd_update ($self, $P, $G, $V, $lr, $momentum, $scale) {
Subroutine sgd_update: called CC=1 SCAR=0.0
Pod documented
123 600 $V *= $momentum;
124 600 $V -= ($lr * $scale) * $G;
125 600 $P += $V;
126 600 return;
127 }
128  
129 251299 ▶sub adam_update ($self, $P, $G, $M, $V, $rate, $b1, $b2, $eps, $scale) {
Subroutine adam_update: called CC=1 SCAR=0.0
Pod documented
130 251299 my $grad = $scale * $G;
131 251299 $M *= $b1;
132 251299 $M += (1 - $b1) * $grad;
133 251299 $V *= $b2;
134 251299 $V += (1 - $b2) * $grad * $grad;
135 251299 $P -= $rate * $M / (sqrt($V) + $eps);
136 251299 return;
137 }
138  
139 8 1;
140  
141 __END__
« lib/Peta/NN/RNG.pm ↑ Summary lib/Peta/NN/Layer/Activation.pm »