Skip to content

Customizing the density estimator

sbi allows to specify a custom density estimator for each of the implemented methods. For all options, check the API reference here.

Changing the type of density estimator

One option is using one of the preconfigured density estimators by passing a string in the density_estimator keyword argument to the inference object (NPE or NLE), e.g., “maf” for a Masked Autoregressive Flow, of “nsf” for a Neural Spline Flow with default hyperparameters.

import torch

from sbi.inference import NPE, NRE
from sbi.utils import BoxUniform

prior = BoxUniform(torch.zeros(2), torch.ones(2))
inference = NPE(prior=prior, density_estimator="maf")

In the case of NRE, the argument is called classifier:

inference = NRE(prior=prior, classifier="resnet")

Changing hyperparameters of density estimators

Alternatively, you can use a set of utils functions to configure a density estimator yourself, e.g., use a MAF with hyperparameters chosen for your problem at hand.

Here, because we want to use N*P*E, we specifiy a neural network targeting the posterior (using the utils function posterior_nn). In this example, we will create a neural spline flow ('nsf') with 60 hidden units and 3 transform layers:

# For SNLE: likelihood_nn(). For SNRE: classifier_nn()
from sbi.neural_nets import posterior_nn

density_estimator_build_fun = posterior_nn(
    model="nsf", hidden_features=60, num_transforms=3
)
inference = NPE(prior=prior, density_estimator=density_estimator_build_fun)

It is also possible to pass an embedding_net to posterior_nn() to automatically learn summary statistics from high-dimensional simulation outputs. You can find a more detailed tutorial on this in 04_embedding_networks.

Building new density estimators from scratch

Finally, it is also possible to implement your own density estimator from scratch, e.g., including embedding nets to preprocess data, or to a density estimator architecture of your choice.

For this, the density_estimator argument needs to be a function that takes theta and x batches as arguments to then construct the density estimator after the first set of simulations was generated. Our factory functions in sbi/neural_nets/factory.py return such a function.

The returned density_estimator object needs to be a subclass of DensityEstimator, which requires to implement three methods:

  • log_prob(input, condition, **kwargs): Return the log probabilities of the inputs given a condition or multiple i.e. batched conditions.
  • loss(input, condition, **kwargs): Return the loss for training the density estimator.
  • sample(sample_shape, condition, **kwargs): Return samples from the density estimator.

See more information on the Reference API page.