huggingface

StainedGlassTrainer

Bases: Trainer, Generic[M, NLP, NL]

Simple subclass of transformers.Trainer to add StainedGlass-specific optimizers and loss functions.

Contains all the features of the original Trainer class, including training and evaluation loops for PyTorch, optimized for 🤗 Transformers.

TODO: include examples for model vs model_init, specifying optimizer with the init, etc...

__init__

__init__(model: NoisyModel[M, NLP, NL], *args: Any, alpha: float, save_only_noise_layer: bool = False, **kwargs: Any) -> None

Initialize a StainedGlassTrainer.

For additional arguments, see the documentation for transformers.Trainer: https://huggingface.co/docs/transformers/main_classes/trainer

Parameters:

Name	Type	Description	Default
model	NoisyModel[M, NLP, NL]	The model to train.	required
alpha	float	The alpha parameter for the noise loss function.	required
save_only_noise_layer	bool	Whether to save only the noise layer when saving the model of type NoisyTransformerModel.	False
*args	Any	Arguments to pass to the superclass.	required
**kwargs	Any	Keyword arguments to pass to the superclass.	required

compute_loss

compute_loss(model: Module, inputs: Any, return_outputs: bool = False) -> torch.Tensor | tuple[torch.Tensor, Any]

Compute the loss for training Stained Glass Transforms.

Parameters:

Name	Type	Description	Default
model	Module	The model to compute the loss for.	required
inputs	Any	The inputs to the model.	required
return_outputs	bool	Whether to return the outputs of the model alongside the loss.	False

Returns:

Type	Description
torch.Tensor | tuple[torch.Tensor, Any]	The loss for training Stained Glass Transforms.