physbo.blm package

Subpackages

• physbo.blm.basis package
  • Module contents
    • Fourier
• physbo.blm.core package
  • Module contents
    • Model
• physbo.blm.inf package
  • Submodules
    • physbo.blm.inf.exact module
  • Module contents
• physbo.blm.lik package
  • Module contents
    • Cov
    • Gauss
    • Linear
• physbo.blm.prior package
  • Module contents
    • gauss()

Module contents

Bayesian Linear Model

class physbo.blm.Model(lik, prior, options={})

Bases: object

Baysean Linear Model

prior

prior distribution of weights

Type:

physbo.blm.prior.Gauss

lik

kernel

Type:

physbo.blm.lik.Gauss

nbasis

number of features in random feature map

Type:

int

stats

auxially parameters for sampling

Type:

Tuple

method

sampling method

Type:

str

get_permutation_importance(X, t, n_perm: int, comm=None, split_features_parallel=False, rng=None)

Calculating permutation importance of model

Parameters:

• X (numpy.ndarray) – inputs

• t (numpy.ndarray) – target (label)

• n_perm (int) – number of permutations

• comm (MPI.Comm) – MPI communicator

Returns:

• numpy.ndarray – importance_mean

• numpy.ndarray – importance_std

get_post_fcov(X, Psi=None, diag=True)

calculates posterior covariance of model

Parameters:

• X (numpy.ndarray) – inputs

• Psi (numpy.ndarray) – feature maps (default: blm.lik.linear.basis.get_basis(X))

• diag (bool) – if True, returns only variances as a diagonal matrix (default: True)

Returns:

Returned shape is (num_points) if diag=true, (num_points, num_points) if diag=false, where num_points is the number of points in X.

Return type:

numpy.ndarray

See also

physbo.blm.inf.exact.get_post_fcov

get_post_fmean(X, Psi=None, w=None)

calculates posterior mean of model (function)

Parameters:

• X (numpy.ndarray) – inputs

• Psi (numpy.ndarray) – feature maps

• w (numpy.ndarray) – weight

See also

physbo.blm.inf.exact.get_post_fmean

get_post_params_mean()

calculates posterior mean of weights

Return type:

numpy.ndarray

See also

physbo.blm.inf.exact.get_post_params_mean

post_sampling(Xtest, Psi=None, N=1, alpha=1.0, rng=None)

draws samples of mean value of model

Parameters:

• Xtest (numpy.ndarray) – inputs

• Psi (numpy.ndarray) – feature maps (default: blm.lik.get_basis(Xtest))

• N (int) – number of samples (default: 1)

• alpha (float) – noise for sampling source

Return type:

numpy.ndarray

predict_sampling(Xtest, Psi=None, N=1, rng=None)

draws samples from model

Parameters:

• Xtest (numpy.ndarray) – inputs

• Psi (numpy.ndarray) – feature map (default: blm.lik.get_basis(Xtest))

• N (int) – number of samples (default: 1)

Return type:

numpy.ndarray (N x len(Xtest))

prepare(X, t, Psi=None)

initializes model by using the first training dataset

Parameters:

• X (numpy.ndarray) – inputs

• t (numpy.ndarray) – target (label)

• Psi (numpy.ndarray) – feature maps

See also

physbo.blm.inf.exact.prepare

sampling(w_mu=None, N=1, alpha=1.0, rng=None)

draws samples of weights

Parameters:

• blm (physbo.blm.core.Model) – model

• w_mu (numpy.ndarray) – mean of weight

• N (int) – the number of samples (default: 1)

• alpha (float) – noise for sampling source (default: 1.0)

Returns:

samples of weights

Return type:

numpy.ndarray

See also

physbo.blm.inf.exact.sampling

update_stats(x, t, psi=None)

updates model by using another training data

Parameters:

• x (numpy.ndarray) – input

• t (float) – target (label)

• psi (numpy.ndarray) – feature map

See also

physbo.blm.inf.exact.update_stats

class physbo.blm.Predictor(config, model=None)

Bases: BasePredictor

Predictor using Baysean linear model

blm

Type:

physbo.blm.core.Model

config

configuration

Type:

physbo.misc.SetConfig

Parameters:

• config (physbo.misc.SetConfig) – configuration

• model (physbo.gp.core.Model)

See also

physbo.predictor.BasePredictor

delete_stats()

resets model

draw_post_sample_params(training, alpha=1.0, objective_index=0, rng=None)

draws one sample of the posterior weight parameters

Together with evaluate_post_sample, this splits Thompson sampling into a random “draw” part and a deterministic “evaluate” part. Under MPI, the sample can be drawn on rank 0, broadcast, and evaluated on all ranks so that every rank scores its candidates with the same posterior sample.

Parameters:

• training (physbo.variable) – training dataset. If already trained, the model does not use this.

• alpha (float) – noise for sampling source (default: 1.0)

• objective_index (int) – Index of objective column to use when training.t is 2D (default: 0)

• rng (rng object, optional) – random number generator (default: global numpy.random state)

Returns:

sampled weight parameters (nbasis,)

Return type:

numpy.ndarray

evaluate_post_sample(w_hat, test, objective_index=0)

evaluates a posterior weight sample at test points (deterministic)

Parameters:

• w_hat (numpy.ndarray) – weight parameters drawn by draw_post_sample_params

• test (physbo.variable) – inputs

• objective_index (int) – Index of objective column to use when test.Z is 3D (default: 0)

Returns:

sampled function values at the test points

Return type:

numpy.ndarray

fit(training, num_basis=None, comm=None, objective_index=0, rng=None)

fit model to training dataset

Parameters:

• training (physbo.variable) – dataset for training

• num_basis (int) – the number of basis (default: self.config.predict.num_basis)

• comm (MPI.Comm) – MPI communicator

• objective_index (int) – Index of objective column to use when training.t is 2D (default: 0)

get_basis(X)

calculates feature maps Psi(X)

Parameters:

X (numpy.ndarray) – inputs

Returns:

Psi – feature maps

Return type:

numpy.ndarray

get_post_fcov(training, test, diag=True, objective_index=0)

calculates posterior variance-covariance matrix of model

Parameters:

• training (physbo.variable) – training dataset. If already trained, the model does not use this.

• test (physbo.variable) – inputs

• diag (bool) – If true, only variances (diagonal elements) are returned.

• objective_index (int) – Index of objective column to use when training.t is 2D (default: 0)

Returns:

Returned shape is (num_points) if diag=true, (num_points, num_points) if diag=false, where num_points is the number of points in test.

Return type:

numpy.ndarray

get_post_fmean(training, test, objective_index=0)

calculates posterior mean value of model

Parameters:

• training (physbo.variable) – training dataset. If already trained, the model does not use this.

• test (physbo.variable) – inputs

• objective_index (int) – Index of objective column to use when training.t is 2D (default: 0)

Returns:

Returned shape is (num_points), where num_points is the number of points in test.

Return type:

numpy.ndarray

get_post_params(training, test, objective_index=0)

calculates posterior weights

Parameters:

• training (physbo.variable) – training dataset. If already trained, the model does not use this.

• test (physbo.variable) – inputs (not used)

• objective_index (int) – Index of objective column to use when training.t is 2D (default: 0)

Return type:

numpy.ndarray

get_post_samples(training, test, N=1, alpha=1.0, objective_index=0, rng=None)

draws samples of mean values of model

Parameters:

• training (physbo.variable) – training dataset. If already trained, the model does not use this.

• test (physbo.variable) – inputs

• N (int) – number of samples (default: 1)

• alpha (float) – noise for sampling source (default: 1.0)

• objective_index (int) – Index of objective column to use when training.t is 2D (default: 0)

Return type:

numpy.ndarray

get_predict_samples(training, test, N=1, objective_index=0, rng=None)

draws samples of values of model

Parameters:

• training (physbo.variable) – training dataset. If already trained, the model does not use this.

• test (physbo.variable) – inputs

• N (int) – number of samples (default: 1)

• objective_index (int) – Index of objective column to use when training.t is 2D (default: 0)

Return type:

numpy.ndarray (N x len(test))

prepare(training, objective_index=0)

initializes model by using training data set

Parameters:

• training (physbo.variable) – dataset for training

• objective_index (int) – Index of objective column to use when training.t is 2D (default: 0)

update(training, test, objective_index=0)

updates the model.

If not yet initialized (prepared), the model will be prepared by training. Otherwise, the model will be updated by test.

Parameters:

• training (physbo.variable) – training dataset for initialization (preparation). If already prepared, the model ignore this.

• test (physbo.variable) – training data for update. If not prepared, the model ignore this.

• objective_index (int) – Index of objective column to use when training.t or test.t is 2D (default: 0)