climb.tool.impl.data_suite.third_party.uq360.models.noise_models package

Submodules

climb.tool.impl.data_suite.third_party.uq360.models.noise_models.heteroscedastic_noise_models module

class climb.tool.impl.data_suite.third_party.uq360.models.noise_models.heteroscedastic_noise_models.GaussianNoise(cuda=False)

Bases: Module, AbstractNoiseModel

N(y_true | f_mu(x, w), f_sigma^2(x, w))

get_noise_var(log_var_pred)

Return the current estimate of noise variance

loss(y_true=None, mu_pred=None, log_var_pred=None, reduce_mean=True)

computes -1 * ln N (y_true | mu_pred, softplus(log_var_pred)) :param y_true: :param mu_pred: :param log_var_pred:

Returns:

climb.tool.impl.data_suite.third_party.uq360.models.noise_models.heteroscedastic_noise_models.transform(a)

climb.tool.impl.data_suite.third_party.uq360.models.noise_models.homoscedastic_noise_models module

class climb.tool.impl.data_suite.third_party.uq360.models.noise_models.homoscedastic_noise_models.GaussianNoiseFixedPrecision(std_y=1.0, cuda=False)

Bases: Module, AbstractNoiseModel

N(y_true | f(x, w), sigma_y**2); known sigma_y

get_noise_var()

Return the current estimate of noise variance

loss(y_pred=None, y_true=None)

computes -1 * ln N (y_pred | y_true, sigma_y**2) :param y_pred: :param y_true: :return:

class climb.tool.impl.data_suite.third_party.uq360.models.noise_models.homoscedastic_noise_models.GaussianNoiseGammaPrecision(a0=6, b0=6, cuda=False)

Bases: Module, AbstractNoiseModel

N(y_true | f(x, w), lambda^-1); lambda ~ Gamma(a, b). Uses a variational approximation; q(lambda) = Gamma(ahat, bhat)

get_noise_var()

Return the current estimate of noise variance

kl()

loss(y_pred=None, y_true=None)

computes -1 * E_q(lambda)[ln N (y_pred | y_true, lambda^-1)], where q(lambda) = Gamma(ahat, bhat) :param y_pred: :param y_true: :return:

climb.tool.impl.data_suite.third_party.uq360.models.noise_models.homoscedastic_noise_models.transform(a)

climb.tool.impl.data_suite.third_party.uq360.models.noise_models.noisemodel module

class climb.tool.impl.data_suite.third_party.uq360.models.noise_models.noisemodel.AbstractNoiseModel(*argv, **kwargs)

Bases: ABC

Abstract class. All noise models inherit from here.

abstract get_noise_var(*argv, **kwargs)

Return the current estimate of noise variance

abstract loss(*argv, **kwargs)

Compute loss given predictions and groundtruth labels

Module contents