NAME
Peta::NN::Job - train a model until it meets given fidelity thresholds, at the smallest size that can
VERSION
version 0.2610090
SYNOPSIS
use Peta::NN::Job;
my $job = Peta::NN::Job->new(
model => { kind => 'edit', side => 'both', window => 7 },
pairs => \@pairs, # [input, output, parameters...]
group => sub ($pair) { $lemma_of{ $pair->[0] } },
always_train => sub ($pair) { $is_exception{ $pair->[0] } },
subsets => {
exceptions => { of => 'train', where => sub ($pair) { $is_exception{ $pair->[0] } } },
'to positive' => { where => sub ($pair) { $pair->[2] eq 'positive' } }, # by a parameter
},
fidelity => { all => 0.96, exceptions => 1.00, 'to positive' => 0.85 },
budget => { ms => 1, seconds => 600 },
search => { scale => [8, 256], depth => [1, 2] },
);
my $model = $job->run;
print $job->report;
$model->export(...) if $job->result->{met};DESCRIPTION
fidelity names what must hold: all is the share of held-out pairs answered exactly; any other name is a subset from subsets, measured on the held-out part, or with of => 'train' on the training part (what the model was shown and must retain).
A job trains one model until it meets them. Training watches the distance to the thresholds and goes on for as long as the model gets closer. A model that has stopped getting closer is first trained with more weight on what it must retain and has not, and then given wider hidden layers, which changes none of its answers, and trained on. It is not thrown away for a larger one.
search gives the scale, the smallest and the largest width; start, the width to begin at, the smallest unless given; grow, the factor by which a stuck model is widened (2); and the depths to try, the next one only if the one before could not get there. shape turns the starting width and a depth into a layer list; the default is an embedding and depth hidden layers.
train is passed on to training. Its epochs is the most one stage trains before the job looks at the model again, its patience the number of epochs a model may go without getting closer before it counts as stuck.
budget limits the cost: params, ms (per item, timed on the inference leg), runs (stages of training) and seconds (wall clock for the job).
run returns the model. If it met everything it is trained once more from nothing on a second seed, at its size, to confirm; result says whether it met and was confirmed, and carries the one measurement on the test part.
METHODS
new
model, pairs and fidelity are required. subsets, group, always_train, split, shape, search, train, budget, seed and backend are optional.
run
run(progress => sub ($stage) { ... }): trains, and returns the model: the one that meets the thresholds, or the closest it got.
result
What the run came to: met, confirmed, the model's size, cost and fidelity, the fidelity on the test part, the number of stages and epochs and the time.
attempts
Every stage of training, in the order it was run: its name, the width, the epochs it trained, the fidelity it reached, and how it ended.
report
The stages and the outcome as text.
part
part('train'), part('validation'), part('test'): the pairs of a part of the split.
contradictions
[the pair, the output that was kept for its input] for every pair that was set aside because its input already had another answer.
indistinct
[pair, another answer, a pair that has it] for every training pair the model reads exactly as it reads a pair with a different answer. Such pairs are not counted in what must be retained: no training can get them right together, only a wider window.
FUNCTIONS
default_shape
default_shape($scale, $depth): the layers for a starting width and a depth unless shape says otherwise.
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.