jijzepttools.blackbox_optimization.fm_config

Configuration classes for Factorization Machine optimizers.

Classes

AdamConfig	Adam optimizer configuration for FM training.
LBFGSConfig	L-BFGS optimizer configuration for FM training.

Module Contents

class AdamConfig

Adam optimizer configuration for FM training.

n_epochs

Number of training epochs.

Type:

int, default=1000

batch_size

Batch size for mini-batch training.

Type:

int, default=32

lr

Learning rate.

Type:

float, default=0.001

weight_decay

Weight decay (L2 regularization).

Type:

float, default=0.0

early_stopping

Whether to use early stopping.

Type:

bool, default=True

es_min_epochs

Minimum number of epochs before early stopping can trigger.

Type:

int, default=10

es_patience

Patience for early stopping (number of epochs without improvement).

Type:

int, default=10

es_threshold

Threshold of loss relative improvement for early stopping.

Type:

float, default=5e-4

es_verbose

Whether to print early stopping message.

Type:

bool, default=False

n_epochs: int = 1000

batch_size: int = 32

lr: float = 0.001

weight_decay: float = 0.0

early_stopping: bool = True

es_min_epochs: int = 10

es_patience: int = 10

es_threshold: float = 0.0005

es_verbose: bool = False

class LBFGSConfig

L-BFGS optimizer configuration for FM training.

L-BFGS is a quasi-Newton method that uses full-batch optimization. It typically converges faster than Adam for small to medium datasets.

max_iter

Maximum number of iterations per optimization step.

Type:

int, default=20

lr

Learning rate (step size). For L-BFGS, 1.0 is often optimal.

Type:

float, default=1.0

history_size

Number of previous iterations to store for approximating the inverse Hessian.

Type:

int, default=10

line_search_fn

Line search function. “strong_wolfe” is recommended for stability. Set to None to disable line search.

Type:

str or None, default=”strong_wolfe”

max_iter: int = 20

lr: float = 1.0

history_size: int = 10

line_search_fn: str | None = 'strong_wolfe'