nsl.configs.make_adv_reg_config

Creates an nsl.configs.AdvRegConfig object.

nsl.configs.make_adv_reg_config(
    multiplier=attr.fields(AdvRegConfig).multiplier.default,
    feature_mask=attr.fields(AdvNeighborConfig).feature_mask.default,
    adv_step_size=attr.fields(AdvNeighborConfig).adv_step_size.default,
    adv_grad_norm=attr.fields(AdvNeighborConfig).adv_grad_norm.default,
    clip_value_min=attr.fields(AdvNeighborConfig).clip_value_min.default,
    clip_value_max=attr.fields(AdvNeighborConfig).clip_value_max.default,
    pgd_iterations=attr.fields(AdvNeighborConfig).pgd_iterations.default,
    pgd_epsilon=attr.fields(AdvNeighborConfig).pgd_epsilon.default,
    random_init=attr.fields(AdvNeighborConfig).random_init.default
)

Used in the notebooks

Used in the tutorials
Adversarial regularization for image classification

Args
multiplier	multiplier to adversarial regularization loss. Defaults to 0.2.
feature_mask	mask w/ values in [0, 1] applied on the gradient. Its shape should be the same as (or broadcastable to) that of the input features. If the input features are in a collection (e.g. list or dictionary), this field should also be a collection of the same structure. Input features corresponding to mask values of 0.0 are not perturbed. Setting this field to None is equivalent to setting a mask value of 1.0 for all input features.
adv_step_size	step size to find the adversarial sample. Defaults to 0.001.
adv_grad_norm	type of tensor norm to normalize the gradient. Input will be converted to NormType when applicable (e.g., a value of 'l2' will be converted to nsl.configs.NormType.L2). Defaults to L2 norm.
clip_value_min	minimum value to clip the features after perturbation. The shape should be the same as (or broadcastable to) input features. If the input features are in a collection (e.g. list or dictionary), this field should also be a collection with the same structure. An omitted or None-valued entry in the collection indicates no constraint on the corresponding feature.
clip_value_max	maximum value to clip the feature after perturbation. (See clip_value_min for the structure and shape limitations.)
pgd_iterations	number of attack iterations for Projected Gradient Descent (PGD) attack. Defaults to 1, which resembles the Fast Gradient Sign Method (FGSM) attack.
pgd_epsilon	radius of the epsilon ball to project back to. Only used in Projected Gradient Descent (PGD) attack.
random_init	Apply a random perturbation before FGSM/PGD steps. Default set to False for no random initialization being applied.

Returns
An nsl.configs.AdvRegConfig object.