Peta-NN 0.2610090

a small neural network, defined, trained and saved in Perl

Small neural networks that each do one limited thing with a string, trained
in Perl and put together. The target is the class of language tasks that
were solved with rule sets 10-15 years ago (inflection, language
identification, diacritics, word classes): a model learns from pairs a rule
set or a lexicon produces, ships as a file of a few kilobytes, and also
answers for the words the rules never listed.

Training small models and composing them is the main way of working. A
chain of models is itself a model; its parts stay parts, so one can be
improved alone; and fused, a whole batch passes all of them on a graphics
card without coming back to Perl in between. Why it is built this way:
What Peta::NN is for.

MODULES

    Peta::NN                       a small neural network, defined, trained and saved in Perl
    Peta::NN::Backend              the engines Peta::NN can compute on
    Peta::NN::Backend::PDL         Peta::NN on PDL ndarrays
    Peta::NN::Backend::Plain       Peta::NN on plain Perl arrays
    Peta::NN::Backend::WebGPU      Peta::NN on the graphics card
    Peta::NN::Chain                models put together into a model
    Peta::NN::Codec                characters to token indices, string pairs to edit labels
    Peta::NN::Data                 records with named fields, to train models on and measure them by
    Peta::NN::Fused                micro models fused into one, as they are, the seams inside
    Peta::NN::Inference            load a trained model and get answers from it
    Peta::NN::Job                  train a model until it meets given fidelity thresholds, at the smallest size that can
    Peta::NN::Layer::Activation    relu, tanh and sigmoid
    Peta::NN::Layer::Dense         fully connected layer
    Peta::NN::Layer::Embed         learned vectors for token indices
    Peta::NN::Model                train a micro model from string pairs, export it as a model file
    Peta::NN::Optimizer            SGD with momentum, and Adam
    Peta::NN::Parallel             independent pieces of work on several cores
    Peta::NN::Pipeline             models in series, and routed by a classifier
    Peta::NN::RNG                  seeded random numbers that are the same on every perl

INSTALLATION

    perl Makefile.PL
    make
    make test
    make install

Perl 5.36 or later. PDL makes training and inference faster and
Parallel::ForkManager lets training runs go in parallel; neither is needed.

DOCUMENTATION

    docs/getting_started.md      a first model, a second, the two as one
    docs/index.md                the guides, the reference, the examples

    perldoc Peta::NN::Data
    perldoc Peta::NN::Model
    perldoc Peta::NN::Chain

LICENSE

    This package is free software, dual-licensed under the
    Artistic License 2.0 and the BSD 2-Clause License.
    https://opensource.org/licenses/Artistic-2.0
    https://opensource.org/licenses/BSD-2-Clause

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