training_config

orchard.core.config.training_config

Optimization & Regularization Configuration Schema.

Declarative schema for training lifecycle, orchestrating optimization landscape and regularization boundaries. Synchronizes learning rate dynamics (Cosine Annealing) with data augmentation (Mixup) and precision policies (AMP).

Key Features

• Optimization dynamics: LR schedules and momentum for loss landscape navigation
• Regularization suite: Label smoothing and Mixup for generalization
• Precision & stability: AMP and gradient clipping for hardware throughput and stability
• Reproducibility: Global seeding for experiment replication (determinism in HardwareConfig)

Strict boundary validation (probability ranges, LR bounds) prevents unstable training states before first batch processing.

TrainingConfig

Bases: BaseModel

Optimization landscape and regularization strategies.

Validates training hyperparameters and provides structure for reproducibility, optimization, regularization, and scheduling.

Attributes:

Name	Type	Description
seed	int	Random seed for reproducibility.
batch_size	BatchSize	Training samples per batch (1-128).
epochs	PositiveInt	Maximum training epochs.
patience	NonNegativeInt	Early stopping patience in epochs.
use_tqdm	bool	Enable progress bar display.
optimizer_type	Literal['sgd', 'adamw']	Optimizer algorithm ('sgd' or 'adamw').
learning_rate	LearningRate	Initial learning rate (1e-8 to 1.0).
min_lr	LearningRate	Minimum learning rate for scheduler.
momentum	Momentum	SGD momentum coefficient.
weight_decay	WeightDecay	L2 regularization strength.
grad_clip	GradNorm | None	Maximum gradient norm for clipping.
label_smoothing	SmoothingValue	Label smoothing factor (0.0-0.3).
mixup_alpha	NonNegativeFloat	Mixup interpolation coefficient.
mixup_epochs	NonNegativeInt	Number of epochs to apply mixup.
use_tta	bool	Enable test-time augmentation.
scheduler_type	Literal['cosine', 'plateau', 'step', 'none']	LR scheduler type ('cosine', 'plateau', 'step', 'none').
monitor_metric	Literal['auc', 'accuracy', 'f1']	Metric driving checkpointing and early stopping ('auc', 'accuracy', or 'f1').
scheduler_patience	NonNegativeInt	ReduceLROnPlateau patience epochs (ignored by other schedulers).
scheduler_factor	Probability	LR reduction factor (plateau factor and StepLR gamma).
step_size	PositiveInt	StepLR decay period in epochs.
use_amp	bool	Enable Automatic Mixed Precision training.
criterion_type	Literal['cross_entropy', 'focal']	Loss function ('cross_entropy' or 'focal').
weighted_loss	bool	Enable class-frequency loss weighting.
focal_gamma	NonNegativeFloat	Focal loss focusing parameter.

validate_amp()

Validate AMP compatibility with batch size.

Raises:

Type	Description
OrchardConfigError	If AMP enabled with batch_size < 4.

Returns:

Type	Description
'TrainingConfig'	Validated TrainingConfig instance.

Source code in orchard/core/config/training_config.py

@model_validator(mode="after")
def validate_amp(self) -> "TrainingConfig":
    """
    Validate AMP compatibility with batch size.

    Raises:
        OrchardConfigError: If AMP enabled with batch_size < 4.

    Returns:
        Validated TrainingConfig instance.
    """
    if self.use_amp and self.batch_size < 4:
        raise OrchardConfigError(
            "AMP enabled with very small batch size (<4) can cause NaN gradients."
        )
    return self