Build the classifier torch module: (theta, x) -> one logit

Three architectures behind one module, matching sbi's "resnet", "mlp" and "linear" classifiers. sbi normalizes between the hidden layers of its MLP (nn.LayerNorm by default); this one does not, keeping the trunk the same plain linear/relu stack every other estimator in the package uses (mlp_layers()).

Usage

nre_module(dim_x, dim_theta, classifier, hidden, n_blocks, embedding = NULL)