Skip to content

module adabmDCA.training_control

Shared lifecycle primitives for DCA training routines.

The numerical algorithms live in adabmDCA.training. This module owns the algorithm-independent parts of a run: counters, limits, history, progress reporting, cancellation, and checkpoint scheduling.

Global Variables

  • HISTORY_KEYS

class StopReason

Reason why a training strategy stopped.

class TrainingCancelled

Internal cancellation signal raised by the shared controller.

class TrainingLimits

Step budgets of a training run; None means no limit.

Attributes:

  • max_gradient_steps: Largest number of accepted parameter updates.
  • max_structure_steps: Largest number of graph activations or decimations.

method __init__

__init__(
    max_gradient_steps: 'int | None' = None,
    max_structure_steps: 'int | None' = None
) → None

class TrainingCounters

Progress counters of a training run.

Attributes:

  • gradient_steps: Accepted parameter updates.
  • structure_steps: Graph activations or decimations.
  • sweeps: Monte Carlo sweeps performed, including diagnostics.
  • stage: Current training phase.

method __init__

__init__(
    gradient_steps: 'int' = 0,
    structure_steps: 'int' = 0,
    sweeps: 'int' = 0,
    stage: 'str' = 'optimization'
) → None

class TrainingMetrics

Common metrics emitted after a meaningful training step.

method __init__

__init__(
    pearson: 'Any',
    slope: 'Any',
    pearson_val: 'Any',
    slope_val: 'Any',
    density: 'Any',
    elapsed_time: 'float',
    ll_train: 'Any' = None,
    ll_val: 'Any' = None,
    entropy: 'Any' = None,
    extra: 'dict[str, Any]' = <factory>
) → None

method as_record

as_record(epoch: 'int') → dict[str, Any]

class CheckpointStore

Persistence interface used by the training controller.

method check

check(updates: 'int') → bool

method log

log(record: 'dict[str, Any]') → None

method save

save(**snapshot: 'Any') → None

class StageProgress

Transient stage update; separate from committed metric records.

method __init__

__init__(
    stage: 'str',
    kind: 'str',
    gradient_steps: 'int',
    current: 'int | None' = None,
    total: 'int | None' = None,
    details: 'dict[str, Any]' = <factory>
) → None

class TrainingController

Coordinate one training run without owning its numerical updates.

method __init__

__init__(
    limits: 'TrainingLimits | None' = None,
    checkpoint: 'CheckpointStore | None' = None,
    observer: 'RecordObserver | None' = None,
    stage_observer: 'StageObserver | None' = None,
    is_cancelled: 'CancellationHook | None' = None
) → None

method add_gradient_steps

add_gradient_steps(count: 'int', sweeps_per_step: 'int' = 0) → None

method add_structure_step

add_structure_step(sweeps: 'int' = 0) → None

method begin_stage

begin_stage(stage: 'str', **metadata: 'Any') → None

Publish a phase change without coupling algorithms to a renderer.

method check_cancellation

check_cancellation() → None

method finalize

finalize(snapshot: 'Mapping[str, Any] | None' = None) → None

Persist the final state once, without duplicating a history row.

method gradient_limit_reached

gradient_limit_reached() → bool

method history_changed

history_changed() → None

Propagate a retroactive history change (resume, rollback) to the saved history.

method record

record(
    metrics: 'TrainingMetrics',
    epoch: 'int',
    snapshot: 'Mapping[str, Any] | None' = None
) → dict[str, Any]

Append and publish exactly one metrics record.

method remaining_gradient_steps

remaining_gradient_steps() → int | None

method report_stage_event

report_stage_event(stage: 'str', **details: 'Any') → None

method report_stage_progress

report_stage_progress(
    stage: 'str',
    current: 'int',
    total: 'int',
    **details: 'Any'
) → None

Report completed work without modifying log rows or counters.

method save_snapshot

save_snapshot(snapshot: 'Mapping[str, Any]') → None

Persist an explicit phase-boundary snapshot when configured.

method set_stop_reason

set_stop_reason(reason: 'StopReason') → StopReason

method structure_limit_reached

structure_limit_reached() → bool

This file was automatically generated via lazydocs.