NAME
Peta::NN::Inference - load a trained model and get answers from it
VERSION
version 0.2610090
SYNOPSIS
use Peta::NN::Inference;
my $degree = Peta::NN::Inference->load('ces-adjective-degree.model');
print scalar $degree->predict('chytřejší'); # comparative
my $convert = Peta::NN::Inference->load('ces-adjective-convert.model');
print scalar $convert->predict('chytrý', 'superlative'); # nejchytřejší
my @all = $convert->predict_all(\@adjectives, 'comparative');
my ($answer, $confidence) = $degree->predict('nejistý');
# every answer the model considers, most probable first
printf "%-12s %.3f\n", @$_ for @{ $degree->distribution('lepší') };
# one verdict for several strings together
my $language = $identify->pooled([ split ' ', $sentence ]);DESCRIPTION
This is the inference leg of Peta::NN. A model file is written by the training leg (Peta::NN::Model->export); reading and using it needs only this module and Peta::NN::Codec.
Parameters
A model may take parameters after the string: as many as it was trained with, each one of the values it was trained with. A model whose parameters have names takes them by name, predict($noun, gender => 'neuter'); given lists the names. (answers, which pipelines and fused models call, takes the values in their order.) They are opaque: the model has learned what to do for 'dative', not what a dative is. parameters lists the known values per position. The wrong number of parameters, or a value the model never saw, is an error; nonsense that is well-formed is answered like anything else.
Certainty
predict gives the best answer and, in list context, its probability. distribution gives every answer with its probability. pooled combines the distributions of several strings into one, for class models.
Engine
engine says what the arithmetic runs on: pdl when this perl has PDL, plain otherwise or when PETA_NN_ENGINE=plain is set.
Files
A model file is read as plain data (nothing in it is turned into an object, so reading runs no code from the file) and checked in full before use: the format marker, every field's type, that each layer fits its neighbours and the labels, and that every weight is a finite number. A file that fails is refused with the reason. By convention model files end in .model.
Weights are stored with 32 bits each, or 8 (signed bytes and a scale per array), as the exporter chose; in memory they are ordinary numbers either way. info returns what the file says of itself and what follows from its contents.
METHODS
load
Peta::NN::Inference->load($file): the model of a model file. The file is checked in full first.
new
Peta::NN::Inference->new($data): a model from model data, which is checked first.
predict
predict($string, @parameters): the answer; in list context also its confidence.
predict_all
predict_all(\@strings, @parameters): the answers, in order.
answers
answers(\@strings, @parameters): [answer, confidence] for each string.
probabilities
probabilities(\@windows): for each window of token indices the probability of every label, in the order of labels.
pool
pool(\@probabilities): what pooled answers, from the probabilities of the strings themselves.
decisions
decisions(\@windows): [index of the best label, its probability] for each window of token indices. For callers that build the windows themselves, as Peta::NN::Fused does.
distribution
distribution($string, @parameters): every answer the model considers, as a list of [answer, probability], most probable first. For a rewrite model, one such list per character.
pooled
pooled(\@strings, @parameters): one distribution for several strings together, for a class model.
data
The model data the object was made from, with the weights as they were stored.
kind
class, edit or rewrite.
labels
The answers the model can give.
given
The names of the parameters, in their order. Empty for a model without parameters, and for one whose parameters were given no names.
values_from
values_from(@arguments): what a caller gave after the string, as the parameters' values in their order: taken by name where the parameters have names, as they come where they have none.
pooled_for
pooled_for(\@strings, @values): pooled with the parameters' values in their order.
parameters
The values the model knows for each of its parameters: one sorted list per position.
n_params
The number of weights.
info
What there is to know about the model, for a human: what the file says of itself and what follows from its contents.
FUNCTIONS
engine
pdl or plain: what the arithmetic runs on.
end_window
end_window($config, \@chars, @tokens): the window of the whole-string kinds. Used by the training leg to read a string as the inference leg does.
char_window
char_window($config, \@chars, $position, @tokens): the window around one character, for rewrite.
param_tokens
param_tokens($params, @values): the tokens of a call's parameters.
pack_weights
pack_weights(\@weights, $bits): a weight array as a model file stores it.
read_file
read_file($file, $format): a Storable file as plain data, refused unless it carries the format marker.
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.