cascade.core.CausalNetwork

class cascade.core.CausalNetwork(n_vars, n_particles, n_covariates, n_layers, hidden_dim, latent_dim, dropout, beta, scaffold_mod, sparse_mod, acyc_mod, latent_mod, lik_mod, kernel_mod, scaffold_kws=None, sparse_kws=None, acyc_kws=None, latent_kws=None, lik_kws=None, kernel_kws=None, design=None)

Bases: LightningModule, Module

Causal discovery neural network

Parameters:

• n_vars (int) – Number of variables to model

• n_particles (int) – Number of SVGD particles

• n_covariates (int) – Dimension of covariates

• n_layers (int) – Number of MLP layers in the structural equations

• hidden_dim (int) – MLP hidden layer dimension in the structural equations

• latent_dim (int) – Dimension of the latent variable

• dropout (float) – Dropout rate

• beta (float) – KL weight of the latent variable

• scaffold_mod (str) – Scaffold graph module, must be one of {“Edgewise”, “Bilinear”}

• sparse_mod (str) – Sparse prior module, must be one of {“L1”, “ScaleFree”}

• acyc_mod (str) – Acyclic prior module, must be one of {“TrExp”, “SpecNorm”, “LogDet”}

• latent_mod (str) – Latent module, must be one of {“NilLatent”, “EmbLatent”, “GCNLatent”}

• lik_mod (str) – Causal likelihood module, must be one of {“Normal”, “NegBin”}

• kernel_mod (str) – SVGD kernel module, must be one of {“KroneckerDelta”, “RBF”}

• scaffold_kws (Mapping[str, Any] | None) – Keyword arguments to the scaffold graph module, see Edgewise or Bilinear for details

• sparse_kws (Mapping[str, Any] | None) – Keyword arguments to the sparse prior module, see L1 or ScaleFree for details

• acyc_kws (Mapping[str, Any] | None) – Keyword arguments to the acyclic prior module, see TrExp, SpecNorm, or LogDet for details

• latent_kws (Mapping[str, Any] | None) – Keyword arguments to the latent module, see NilLatent, EmbLatent, or GCNLatent for details

• lik_kws (Mapping[str, Any] | None) – Keyword arguments to the causal likelihood module, see Normal or NegBin for details

• kernel_kws (Mapping[str, Any] | None) – Keyword arguments to the SVGD kernel module, see KroneckerDelta or RBF for details

• design (IntervDesign | None) – Optional intervention design module, see IntervDesign for details

Methods

backward	Implementation of the main SVGD logic
cascade	Cascade pass of the model
compute_kernel	Compute the SVGD kernel
compute_lik	Compute likelihood terms from a minibatch
compute_prior	Compute the prior energy terms
explain	Explanation pass of the model
forward	Forward pass of the model
predict_step	Prediction step for a minibatch
prune	Prune the scaffold and structural equations accordingly
reset_parameters
reset_properties
set_design	Set the design module
training_step	Training step for a minibatch
validation_step	Validation step for a minibatch

Attributes

EXP_AVG
fit_stage	Prediction mode, see FitStage for details
fixed_vars	Fixed variables during counterfactual prediction
predict_mode	Prediction mode, see PredictMode for details
prefit	Whether to run prefit on the covariates only
topo_gens	Topological generations of the causal graph