nnpd.core.runner
================

.. py:module:: nnpd.core.runner

.. autoapi-nested-parse::

   Plan first; train and cache models; analyze saved models with separate analysis settings.



Attributes
----------

.. autoapisummary::

   nnpd.core.runner.TRAINING_FILE
   nnpd.core.runner.ANALYSIS_FILE
   nnpd.core.runner.KEEP_SETTINGS
   nnpd.core.runner.STAGES


Classes
-------

.. autoapisummary::

   nnpd.core.runner.Job


Functions
---------

.. autoapisummary::

   nnpd.core.runner.expand_jobs
   nnpd.core.runner.infeasible
   nnpd.core.runner.check_feasible
   nnpd.core.runner.plan
   nnpd.core.runner.analyze_run
   nnpd.core.runner.save_references
   nnpd.core.runner.execute
   nnpd.core.runner.restore
   nnpd.core.runner.verify_store


Module Contents
---------------

.. py:data:: TRAINING_FILE
   :value: 'settings_training.py'


.. py:data:: ANALYSIS_FILE
   :value: 'settings_analysis.py'


.. py:data:: KEEP_SETTINGS
   :value: 3


.. py:data:: STAGES
   :value: ('train', 'run', 'analyze')


.. py:class:: Job

   .. py:attribute:: settings
      :type:  dict


   .. py:attribute:: member
      :type:  dict


   .. py:attribute:: changed
      :type:  str | None


   .. py:attribute:: workload
      :type:  dict


   .. py:property:: id
      :type: str



   .. py:method:: as_dict() -> dict


.. py:function:: expand_jobs(training_settings: dict, experiment: nnpd.core.api.Experiment, analysis_settings: dict | None = None, analysis: nnpd.core.api.Analysis | None = None) -> list[Job]

   Validate *every* training alternative, and the analysis of each, before anything runs.


.. py:function:: infeasible(jobs: list[Job]) -> dict[str, list[str]]

   Configurations whose analysis would exceed the configured limits, with the reasons.


.. py:function:: check_feasible(jobs: list[Job]) -> None

   Refuse the whole launch before anything is computed if any configuration cannot be analyzed.


.. py:function:: plan(profile: str | None = None, settings_dir: str | pathlib.Path = '.') -> list[Job]

   What `run` would train, with each model's analysis workload; computes nothing.


.. py:function:: analyze_run(context: nnpd.core.api.Context) -> dict

   Compute the selected metrics and plots of one run, replacing earlier ones. Never trains.


.. py:function:: save_references(context: nnpd.core.api.Context) -> None

.. py:function:: execute(stage: str = 'run', *, profile: str | None = None, settings_dir: str | pathlib.Path = '.') -> list[pathlib.Path]

   Train, analyze, or both, from the settings files in settings_dir; the CLI does the same.

   'train' trains the models of settings_training.py and resets their analysis.
   'analyze' analyzes the latest trained models with settings_analysis.py; it never trains.
   'run' does both. 'jobs': torchrun workers process disjoint runs; 'ddp': workers train
   each model together and only rank zero analyzes. After success, runs of earlier settings
   are removed, the cache keeps what the last KEEP_SETTINGS trainings used, and the settings
   files are copied into the output folder.


.. py:function:: restore(run_dir: str | pathlib.Path, *, device: str = 'cpu') -> nnpd.core.api.Context

   Open a saved run, with the analysis settings it was analyzed with, on any supported device.


.. py:function:: verify_store(root: str | pathlib.Path) -> int

   Recompute all saved artifact checksums; raises on the first damaged file.