Skip to content

module adabmDCA.training_config

Canonical configuration values and validation for DCA training.

Global Variables

  • DEFAULT_PTT_MAX_EPOCHS
  • DEFAULT_PTT_N_CHAINS
  • DEFAULT_PTT_N_SWEEPS
  • DEFAULT_PTT_SAMPLER
  • MODEL_TYPES
  • SAMPLERS
  • DEFAULT_MODEL_TYPE
  • DEFAULT_ALPHABET
  • DEFAULT_LEARNING_RATE
  • DEFAULT_N_SWEEPS
  • DEFAULT_SAMPLER
  • DEFAULT_N_CHAINS
  • DEFAULT_TARGET_PEARSON
  • DEFAULT_MAX_EPOCHS
  • DEFAULT_L2_REGULARIZATION
  • DEFAULT_SEED
  • DEFAULT_CLUSTERING_SEQID
  • DEFAULT_ACTIVATION_STEPS
  • DEFAULT_ACTIVATION_FRACTION
  • DEFAULT_TARGET_DENSITY
  • DEFAULT_DECIMATION_RATE
  • DEFAULT_DEVICE
  • DEFAULT_DTYPE
  • DEFAULT_CHECKPOINT_INTERVAL
  • SPARSE_CHECKPOINT_INTERVAL
  • DEFAULT_INNER_GRADIENT_STEPS
  • EDGE_DEFAULT_PSEUDOCOUNT
  • EDGE_EMPIRICAL_PSEUDOCOUNT

class ConfigurationError

Raised when a training configuration is internally inconsistent.

class TrainingConfig

All settings of one training run, validated on construction.

Pass it as config= to train_model; it then takes precedence over the function's individual keyword arguments. Settings that do not apply to the chosen model_type are ignored.

Attributes:

  • model_type: "bmDCA" (fully connected), "eaDCA" (coupling activation), "edDCA" (decimation) or "edgeDCA" (edge activation).
  • alphabet: "protein", "rna", "dna" or an ordered custom token string.
  • learning_rate: Step size of the parameter updates (ignored by edgeDCA).
  • n_sweeps: Monte Carlo sweeps per update.
  • sampler: "metropolized_gibbs", "gibbs" or "metropolis".
  • n_chains: Number of Markov chains.
  • target_pearson: Training stops when the Pearson correlation between data and model connected correlations reaches this value.
  • max_epochs: Step budget: gradient steps for bmDCA, graph steps for the sparse models. Superseded by the two limits below.
  • max_gradient_steps: Optional limit on parameter updates.
  • max_structure_steps: Optional limit on graph activations or decimations.
  • pseudocount: Pseudocount of the training statistics; None means 1/Meff (0.1 for edgeDCA).
  • l2_regularization: Strength of the L2 penalty on the couplings.
  • seed: Random seed.
  • clustering_seqid: Sequence identity above which sequences share weight.
  • no_reweighting: Give every training sequence the same weight.
  • activation_steps: eaDCA: gradient updates between graph activations.
  • activation_fraction: eaDCA: fraction of the inactive couplings activated per graph update (the upper limit for PTT adaptive activation).
  • target_density: edDCA: coupling density at which decimation stops.
  • decimation_rate: edDCA: fraction of active couplings removed per decimation.
  • device: "auto", "cpu", "cuda" or "mps".
  • dtype: "float32", "float64" or "bfloat16" (CUDA/MPS; CUDA only with PTT).
  • use_wandb: Log to Weights & Biases.
  • checkpoint_interval: Updates between saved checkpoints; None uses the default.
  • inner_gradient_steps: edDCA: largest number of gradient updates to re-converge the model after each decimation.
  • edge_empirical_pseudocount: edgeDCA: small pseudocount of the target statistics, which avoids zero frequencies.
  • ptt: PTTConfig to train with PTT, or None for PCD.

Raises:

  • ConfigurationError: If a setting is invalid or unsupported for the model.

Example: 

config = TrainingConfig(model_type="eaDCA", alphabet="rna",
                        ptt=PTTConfig(activation="adaptive"))
train_model("family.fasta", config=config, output_dir="model")

method __init__

__init__(
    model_type: 'str' = 'bmDCA',
    alphabet: 'str' = 'protein',
    learning_rate: 'float' = 0.01,
    n_sweeps: 'int' = 10,
    sampler: 'str' = 'metropolized_gibbs',
    n_chains: 'int' = 2048,
    target_pearson: 'float' = 0.95,
    max_epochs: 'int' = 50000,
    max_gradient_steps: 'int | None' = None,
    max_structure_steps: 'int | None' = None,
    pseudocount: 'float | None' = None,
    l2_regularization: 'float' = 0.0,
    seed: 'int' = 0,
    clustering_seqid: 'float' = 0.8,
    no_reweighting: 'bool' = False,
    activation_steps: 'int' = 10,
    activation_fraction: 'float' = 0.001,
    target_density: 'float' = 0.02,
    decimation_rate: 'float' = 0.01,
    device: 'str' = 'auto',
    dtype: 'str' = 'float32',
    use_wandb: 'bool' = False,
    checkpoint_interval: 'int | None' = 500,
    inner_gradient_steps: 'int' = 10000,
    edge_empirical_pseudocount: 'float' = 1e-06,
    ptt: 'PTTConfig | None' = None
) → None

property empirical_pseudocount

Pseudocount used for target statistics before regularization.

property limits

Resolve the legacy epoch limit into explicit model-aware limits.

property resolved_checkpoint_interval

Checkpoint interval in updates, with the default filled in.

method as_dict

as_dict() → dict[str, Any]

Return a serializable snapshot suitable for run metadata.

method resolve_pseudocount

resolve_pseudocount(effective_size: 'float') → float

Return the pseudocount to use for a training set of effective_size (Meff).

Args:

  • effective_size: Sum of the training sequence weights.

Returns: pseudocount if set; otherwise 0.1 for edgeDCA and 1/Meff for the other models.

Raises:

  • ConfigurationError: If effective_size is not positive.

This file was automatically generated via lazydocs.