Skip to content

augmentation_config

orchard.core.config.augmentation_config

Data Augmentation & Test-Time Augmentation (TTA) Schema.

Declarative schema for stochastic transformation pipeline. Synchronizes geometric and photometric noise used during training with TTA perturbations for calibrated model robustness.

Key Features
  • Geometric invariance: Horizontal flips and rotation for orientation generalization
  • Photometric consistency: Color jitter and scaling for acquisition variations
  • TTA ensemble strategy: Pixel shifts, scaling, and Gaussian blur for stable predictions through stochastic averaging
  • Validation guards: Domain-specific types ensuring physically plausible ranges for image classification

AugmentationConfig

Bases: BaseModel

Stochastic transformations for training and test-time augmentation (TTA).

Centralizes hyperparameters for geometric and photometric perturbations applied during training and inference phases.

Attributes:

Name Type Description
hflip Probability

Probability of horizontal flip during training (0.0-1.0).
rotation_angle RotationDegrees

Maximum rotation angle in degrees (0-360).
jitter_val NonNegativeFloat

Color jitter intensity for brightness, contrast, saturation.
min_scale Probability

Minimum scale factor for random resized crop (0.0-1.0).
tta_translate PixelShift

Pixel translation range for TTA ensemble.
tta_scale ZoomScale

Scale factor for TTA zoom augmentation.
tta_blur_sigma BlurSigma

Gaussian blur sigma for TTA smoothing.
tta_blur_kernel_size KernelSize

Gaussian blur kernel size for TTA (must be odd).