/ 
Peta-NN-0.2610090
/ Peta::NN::Backend

NAME

Peta::NN::Backend - the engines Peta::NN can compute on

VERSION

version 0.2610090

SYNOPSIS

my $net = Peta::NN->new(..., backend => 'pdl');     # or 'plain', 'gpu', 'auto'

print join ' ', Peta::NN::Backend::available();     # gpu pdl plain

The default is plain, or $ENV{PETA_NN_BACKEND} when set.

FUNCTIONS

names

The names of all backends, whether this perl has them or not.

available

The names of the backends that work in this perl, on this machine.

try

my $backend = Peta::NN::Backend::try('gpu') or warn $@;

A backend object, or undef when its engine is not in this perl or finds nothing to run on; $@ then says why.

create

A backend object for a name or for auto, which takes pdl if it loads and plain otherwise: that is where a network that is left the choice starts, and it settles on one of the available backends when it is trained ("train" in Peta::NN). Dies if the backend is not available.

THE BACKEND INTERFACE

A backend owns the tensors (a batch of rows: inputs, activations, weights, gradients) and implements these operations on whole batches. Tensors are opaque to the caller.

new

A backend object. Dies when the engine is missing.

name

The backend's name.

tensor

tensor(\@flat, $cols): rows of $cols numbers, row after row.

tokens

tokens(\@flat, $per_row): rows of token indices, for embed.

flat

A tensor as a flat Perl list.

zeros_like

A tensor of zeros in the shape of another.

affine

affine($X, $W, $b): X W^T + b.

affine_grad

affine_grad($X, $W, $dY, $need_dx): (dW, db, dX), summed over the batch.

activate

activate($kind, $X): relu, tanh or sigmoid of every value.

activate_grad

activate_grad($kind, $Y, $dY): the gradient, in terms of the output $Y.

embed

embed($E, $tokens): table rows, concatenated per sample.

embed_grad

embed_grad($E, $tokens, $dX): the gradient of the table.

softmax_ce

softmax_ce($logits, \@classes, \@weights): (summed loss, dlogits). The weights, one per row, are optional.

mse

mse($out, \@flat_targets, \@weights): (summed loss, dout).

decay

decay($p, $factor): every value times $factor, in place.

sgd_update

sgd_update($p, $g, $velocity, $lr, $momentum, $scale), in place.

adam_update

adam_update($p, $g, $m, $v, $rate, $b1, $b2, $eps, $scale), in place.

argmax_rows

argmax_rows($tensor, $cols): the index of the largest value in each row. Implemented here for every backend.

AUTHOR

PetaMem s.r.o. <info@petamem.com>

COPYRIGHT

Copyright (c) 2026 PetaMem s.r.o.

LICENSE

This package is free software, dual-licensed under the Artistic License 2.0 and the BSD 2-Clause License. See the LICENSE file of the distribution.