sdm.prior#

HP

Base class for sampling hyperparameters.

Constant

Return a fixed hyperparameter.

Uniform

Sample real values uniformly between two bounds.

UniformInt

Sample integers uniformly between inclusive bounds.

LogUniform

Sample positive real values uniformly in logarithmic space.

LogUniformInt

Sample integers by truncating log-uniform values.

Choice

Choose uniformly among fixed options.

CorrelatedContext

Share sampling parameters within a scope.

Correlated

Sample numerical values from a shared, randomly sampled distribution.

Mixture

Select items using fixed or shared random weights.

DAG

Directed acyclic graph in CSR layout.

DAGSampler

Base class for sampling directed acyclic graphs.

RandomCauchyDAGSampler

Sample DAGs with Cauchy-distributed connectivity parameters.

MixtureDAGSampler

Pick one DAG sampler per call from a weighted pool.

should_drop

Decide whether to drop a dataset based on an ExtraTrees probe.

Root Samplers#

RootSampler

Base class for sampling values at nodes without parents.

NormalRootSampler

Sample root values from a normal distribution.

UniformRootSampler

Sample root values from a uniform distribution.

BallRootSampler

Sample root values uniformly within the unit ball.

CorrelatedRootSampler

Mix sampled root values to create correlated channels.

MixtureRootSampler

Pick one root sampler per call from a weighted pool.

Activations#

Activation

Base class for activation functions.

CallableActivation

Adapt a callable to the Activation interface.

StandardizedActivation

Apply an activation at a random center and scale.

ReLUPowActivation

Apply relu(x) ** exponent elementwise.

SignPowActivation

Apply sign(x) * abs(x) ** exponent elementwise.

InversePowActivation

Apply (abs(x) + 1e-3) ** -exponent elementwise.

IntegerPowActivation

Apply x ** exponent elementwise for an integer exponent.

MixtureActivation

Pick one activation per call from a weighted pool.

identity

Return the input values unchanged.

heaviside

Map positive values to 1 and others to 0 elementwise.

gaussian

Apply the Gaussian bump exp(-x ** 2) elementwise.

unit_indicator

Map values in [0, 1] to 1 and others to 0 elementwise.

softmax

Apply the softmax along the last dimension.

one_hot_argmax

One-hot encode the argmax along the last dimension.

argsort

Return sorting indices along the last dimension.

log_abs

Apply log(max(abs(x), 1e-6)) elementwise.

rank

Return zero-based ranks along the last dimension.

modulo

Return the fractional part x - floor(x) elementwise.

Functions#

RandomFunction

Base class for random functions on node values.

ParentMerger

Base class for combining parent values into node values.

RandomLinear

Apply a random linear function.

RandomQuadratic

Apply a random quadratic function.

RandomNN

Apply a random neural network.

RandomDiscretization

Replace each row with a linear transformation of its nearest center.

ConcatParentMerger

Concatenate parent values and apply random_fn.

MixtureFunction

Pick one random function per call from a weighted pool.

MixtureParentMerger

Pick one parent merger per call from a weighted pool.

RandomTree

Average the outputs of random symmetric trees.

RandomEMAssignment

Compute random soft cluster assignments and project them.

RandomGP

Approximate a Gaussian process with random Fourier features.

RandomProduct

Multiply the outputs of two sampled random functions.

ReduceParentMerger

Apply random_fn to each parent and reduce the outputs.

Matrix Samplers#

MatrixSampler

Base class for sampling random matrices.

GaussianMatrixSampler

Sample random Gaussian matrices.

PowerLawMatrixSampler

Sample Gaussian matrices with random power-law weights.

SingularValuesMatrixSampler

Sample matrices factored as U diag(w) V^T.

KernelMatrixSampler

Sample Laplace-kernel matrices between Gaussian-mixture anchor points.

ActivationMatrixSampler

Sample a base matrix and apply a random activation.

MixtureMatrixSampler

Pick one matrix sampler per call from a weighted pool.

Converters#

NumericalConverter

Base class for converting latent channels to numerical columns.

CategoricalConverter

Base class for converting latent channels to categorical columns.

IdentityConverter

Return single-channel values as numerical columns.

NearestNeighborConverter

Assign each row to its nearest sampled center using Minkowski distance.

KumaraswamyNumericalConverter

Apply Kumaraswamy warping after min-max scaling each column.

DistributionalMixtureConverter

Convert latent channels to numerical column values.

SoftmaxCategoricalConverter

Convert latent channels to category indices with softmax or argmax.

MixtureConverter

Pick one converter per column from a weighted pool.

Tabular Priors#

Schema

Specification of a dataset.

SchemaSampler

Sample Schema instances.

SCMPrior

Sample datasets from random structural causal models (SCMs).

Tabular Transforms#

PriorTransform

Base class for transforming synthetic feature and target tables.

Identity

Return features and targets unchanged.

Choice

Apply one randomly selected transform per dataset.

RejectNonfinite

Reject datasets with missing or nonfinite feature or target values.

Coarsen

Reduce selected feature columns to a few ordered levels.

CorrelateColumns

Blend random feature blocks toward shared Gaussian factors.

MaskBlock

Mask the same random rows in a fixed number of feature columns.

MaskByTarget

Mask feature rows selected by a target class or numerical tail.

MaskCategorical

Mask independent random rows in selected categorical feature columns.

MaskCensoring

Mask feature tails under MAR or self-masking MNAR censoring.

MaskLatent

Mask feature column blocks under an unobserved row factor.

MaskNearEmpty

Mask column groups to sparse, overlapping sets of observed rows.

MaskRandom

Independently mask numerical and categorical feature cells.

MaskRows

Mask the same random rows in a fraction of feature columns.

MaskStructured

Combine structured feature masking with conditional censoring.

RecodeHighCardinality

Recode selected finite feature columns as high-cardinality categories.

Relabel

Recompute the target from selected feature columns.

ScaleFeatures

Clip the outliers of each feature column and standardize it.

ScaleTarget

Clip the outliers of a numerical target and standardize it.

ShuffleClasses

Permute the classes of a categorical target.

ShuffleColumns

Permute the numerical and categorical feature columns together.

WarpTargetTails

Apply a random power or exponential warp to a numerical target.