Augmentation

Parameter	Type	Default	Meaning / values
config.augmentation.mask_transforms.use_mask_transform	bool	True	Enable the default mask transform probabilities.
config.augmentation.mask_transforms.mask_transform_probs	dict[int \| str, Any]	{}	Probabilities for global, local, and class-specific transforms.
config.augmentation.mask_transforms.mask_transform_params	dict[int \| str, dict[str, Any]]	{}	Parameter ranges for individual transforms.
config.augmentation.mask_transforms.priorities	list[int] \| tuple[int, ...] \| None	None	Class priority when transformed masks overlap.
config.augmentation.mask_transforms.local_as_global	bool	False	Apply local transforms jointly to all anomaly classes.
config.augmentation.mask_transforms.padding_factor	int	2	Scale factor for the temporary transform canvas.
config.augmentation.random_offset_enabled	bool	True	Enable random offsets during training.
config.augmentation.random_offset_max_fraction	float	1.0	Maximum offset as a fraction of available space; range [0, 1].
config.augmentation.random_offset_foreground_threshold	float	0.001	Foreground threshold used for training offsets.

Masks use nearest-neighbor interpolation; jointly transformed images use linear interpolation. mask_transform_probs and mask_transform_params accept global or local transform names and class IDs. With use_mask_transform=True, the default anomaly size (3, 64, 64) gives these effective values; global elastic parameters change with anomaly_size.

Transform defaults

Transform	Default probability	Effective default parameters
zoom	1.0	{'min_zoom': 0.9, 'max_zoom': 0.9}
stretch	1.0	{'min_stretch': 1.0, 'max_stretch': 1.2}
rotate	0.0	{'max_rotation': 5.0}
elastic	1.0	{'sigma': (13, 13), 'magnitude': (13, 13)}
local_dilate	0.0	{'min_iterations': 0, 'max_iterations': 2}
local_stretch	0.0	{'min_stretch': 0.95, 'max_stretch': 1.05}
local_rotate	0.0	{'max_rotation': 5.0}
local_elastic	0.0	{'sigma': 30, 'magnitude': 20}