peal.adaptors.clarc.RRClArCConfig

class peal.adaptors.clarc.RRClArCConfig

Bases: ClArCConfig

Config of RRClArC.

Parameters:

• gradient_target (str, default "all") – Which logits are differentiated for the RR penalty: "max", "attacked_class", "all" or "all_random" (random signs).

• mean_grad (bool, default False) – Spatially average the gradient before projecting it on the CAV.

• rrc_loss (str, default "l2") – Penalty form: "l2" (squared projection) or "cosine".

Methods

Attributes

adaptor_type	The type of adaptor that shall be used.
gradient_target
mean_grad
rrc_loss
model_path
base_dir
data
training
task

adaptor_type: str = 'RRClArC'

The type of adaptor that shall be used. This is necessary to know which pydantic class to use when loading from yaml.

gradient_target: str = 'all'

mean_grad: bool = False

rrc_loss: str = 'l2'