Training

Parameter	Type	Default	Meaning / values
config.training.num_trials	int	10	Number of Optuna trials; positive.
config.training.trial_selection	TrialSelection	'best'	Trial to use: 'best', 'last', or a non-negative trial ID.
config.training.validation_ratio	float	0.2	Fraction of data reserved for validation; range [0, 1).
config.training.batch_size	int	64	Training batch size; positive.
config.training.epochs	int	1000	Maximum number of training epochs; positive.
config.training.learning_rate	float	0.001	Initial learning rate; positive.
config.training.dtype	torch.dtype \| None	None	PyTorch dtype for training; None uses the model default.
config.training.gradient_clip_norm	float \| None	None	Optional upper bound for the gradient norm.
config.training.monitor_metric	str \| None	'selection'	Metric used for model selection and early stopping.
config.training.early_stopping_enabled	bool	True	Enable early stopping.
config.training.early_stopping	dict[str, Any]	{'patience': 100, 'delta': 0.0001}	Early stopping settings, including patience and delta.
config.training.lr_scheduler_enabled	bool	True	Enable the learning rate scheduler.
config.training.lr_scheduler	dict[str, Any]	{'patience': 30, 'factor': 0.1, 'threshold': 1e-05}	Learning rate scheduler settings, including patience and factor.

Nested training settings

config.training.early_stopping contains patience (epochs without improvement) and delta (minimum improvement). config.training.lr_scheduler contains patience, factor (learning rate multiplier), and threshold (minimum improvement). Their defaults are shown in the table above.