loss

Initialization for core loss modules.

Submodules are imported lazily (PEP 562) so that importing a single leaf module (e.g. stainedglass_core.loss.entropy) does not eagerly pull in siblings such as distillation/jigsaw/image_similarity and their optional dependencies (liger_kernel, segmentation_models_pytorch, the vision cloaks). This keeps lightweight, text-only consumers — such as the Megatron integration running in the NeMo container — importable without the full optional-dependency set. from stainedglass_core.loss import distillation continues to work on demand.

Modules:

Name	Description
augmented_lagrangian	Augmented-Lagrangian controller for a single scalar inequality constraint.
cloak	Module for cloak loss functions.
cosine	Module for cosine similarity and distance loss functions.
distillation	Module for distillation loss functions.
divergences	Module for f-divergence based loss functions.
entropy	Module for entropy-related loss functions, particularly for mutual information estimation.
image_similarity	Module for image-similarity privacy losses.
jigsaw	Module for Jigsaw loss functions.
means	Module for mean-related loss functions.
std	Module for standard deviation-related loss functions.
transform	Module for loss transformation utilities.

__dir__

__dir__() -> list[str]

Return the package's public attribute names.

Returns:

Type	Description
list[str]	The sorted list of names in __all__.

__getattr__

__getattr__(name: str) -> Any

Lazily import a loss submodule (PEP 562).

Parameters:

Name	Type	Description	Default
name	str	The attribute being accessed on the package.	required

Returns:

Type	Description
Any	The requested submodule.

Raises:

Type	Description
AttributeError	If name does not name a submodule of this package.