Skip to content

module adabmDCA.ptt.config

Configuration for PTT snapshot retention, reservoirs and mixing analysis.

Global Variables

  • DEFAULT_PTT_N_SWEEPS
  • DEFAULT_PTT_N_CHAINS
  • DEFAULT_PTT_SAMPLER
  • DEFAULT_PTT_MAX_EPOCHS
  • PTT_OPTIMIZERS
  • PTT_ACTIVATIONS

class PTTConfig

Settings of Parallel Trajectory Tempering, for training and sampling.

Pass an instance as ptt= to train_model (or through TrainingConfig(ptt=...)) to train with PTT. The defaults suit most families; see docs/ptt.md for the method. Round counts are exchange rounds: one pass of replica exchanges followed by the local sweeps.

Attributes:

  • swaps: Exchange-round factor: each update runs int(swaps * sqrt(active replicas)) rounds.
  • target_acceptance: Swap acceptance below which the current endpoint is kept as a new snapshot in the ladder.
  • min_acceptance: Swap acceptance below which a mixing check is run.
  • max_replicas: Active ladder size above which the reservoir is refreshed and the ladder compressed.
  • target_replicas: Active replicas kept after compression.
  • reservoir_size: Size of the reservoir population; None means 10 * n_chains. Must be at least n_chains.
  • full_sampler: Use the full historical ladder when collecting a new reservoir.
  • equilibration_rounds: Extra rounds after a snapshot is inserted or replaced.
  • initialization_rounds: Warmup rounds before the first update.
  • mixing_chains: Chains per replica in the autocorrelation mixing experiment.
  • mixing_initial_rounds: Initial trajectory length of that experiment.
  • mixing_thermalization_rounds: Warmup before a mixing measurement and after a recovery.
  • mixing_max_rounds: Round budget of one mixing check; exceeding it is a mixing failure, which triggers recovery.
  • mixing_window_factor: Required trajectory length relative to the longest correlation time (autocorrelation method).
  • mixing_method: Training mixing check: "renewal" (the ladder is repopulated by fresh configurations) or "autocorrelation".
  • renewal_tolerance: Fraction of old configurations, per replica, below which the ladder counts as renewed.
  • max_recoveries: Largest number of learning-rate halvings after mixing failures.
  • min_learning_rate: Learning-rate floor of recovery.
  • optimizer: "adaptive" (steps bounded by a KL trust region, training paused while the endpoint chains lag behind the model) or "sgd" (fixed learning rate).
  • trust_radius: Adaptive optimizer: largest predicted KL divergence of one update.
  • lag_tolerance: Adaptive optimizer: largest lag of the endpoint chains, in population standard deviations, before training pauses.
  • lag_horizon: Adaptive optimizer: memory of past updates, in updates.
  • lag_pause_rounds: Adaptive optimizer: longest pause, in rounds.
  • validation_stop: Stop when the validation log-likelihood plateaus instead of at the target Pearson; requires a validation alignment.
  • validation_window: Updates per window of the plateau test, which compares the medians of the last two windows.
  • validation_min_gain: Stop when the median gain between the two windows is below this value (log-likelihood per site).
  • activation: eaDCA: "fixed" activates activation_fraction of the inactive couplings per block; "adaptive" activates at most that fraction, only significant candidates within a KL budget.
  • activation_kl_share: Adaptive activation: share of trust_radius for the first update of the new couplings; halved by each recovery.
  • activation_significance: Adaptive activation: standard errors by which a candidate's pair frequency must differ from the data; 0 disables the test.

Example: 

from adabmDCA import PTTConfig, train_model
train_model("family.fasta", alphabet="rna", ptt=PTTConfig(validation_stop=True),
            validation_path="validation.fasta", output_dir="model")

method __init__

__init__(
    swaps: 'int' = 1,
    target_acceptance: 'float' = 0.25,
    min_acceptance: 'float' = 0.1,
    max_replicas: 'int' = 2,
    target_replicas: 'int' = 2,
    reservoir_size: 'int | None' = None,
    full_sampler: 'bool' = False,
    equilibration_rounds: 'int' = 10,
    initialization_rounds: 'int' = 100,
    mixing_chains: 'int' = 100,
    mixing_initial_rounds: 'int' = 100,
    mixing_thermalization_rounds: 'int' = 1000,
    mixing_max_rounds: 'int' = 20000,
    mixing_window_factor: 'float' = 20.0,
    mixing_method: 'str' = 'renewal',
    renewal_tolerance: 'float' = 0.01,
    max_recoveries: 'int' = 3,
    min_learning_rate: 'float' = 1e-08,
    optimizer: 'str' = 'adaptive',
    trust_radius: 'float' = 0.01,
    lag_tolerance: 'float' = 0.25,
    lag_horizon: 'int' = 200,
    lag_pause_rounds: 'int' = 100,
    validation_stop: 'bool' = False,
    validation_window: 'int' = 200,
    validation_min_gain: 'float' = 0.0,
    activation: 'str' = 'fixed',
    activation_kl_share: 'float' = 0.5,
    activation_significance: 'float' = 3.0
) → None

This file was automatically generated via lazydocs.