ELFI is an open-source statistical software package for likelihood-free inference (ELFI) such as Approximate Bayesian Computation licensed under BSD3.
ELFI development team
Good to know
Here we describe some important concepts related to ELFI. These will help in understanding how to implement custom operations (such as simulators or summaries) and can potentially save the user from some pitfalls.
ELFI model
In ELFI, the priors, simulators, summaries, distances, etc. are called operations. ELFI provides a convenient syntax of combining these operations into a network that is called an ElfiModel, where each node represents an operation. Basically, the ElfiModel is a description of how different quantities needed in the inference are to be generated. The structure of the network is a directed acyclic graph (DAG).
Operations
Operations are functions (or more generally Python callables) in the nodes of the ELFI model. Those nodes that deal directly with data, e.g. priors, simulators, summaries and distances should return a numpy array of length batch_size that contains their output.
If your operation does not produce data wrapped to numpy arrays, you can use the elfi.tools.vectorize tool to achieve that. Note that sometimes it is required to specify which arguments to the vectorized function will be constants and at other times also specify the datatype (when automatic numpy array conversion does not produce desired result). It is always good to check that the output is sane using the node.generate method.
Reusing data
The OutputPool object can be used to store the outputs of any node in the graph. Note however that changing a node in the model will change the outputs of it’s child nodes. In Rejection sampling you can alter the child nodes of the nodes in the OutputPool and safely reuse the OutputPool with the modified model. This is especially handy when saving the simulations and trying out different summaries. BOLFI allows you to use the stored data as initialization data.
However passing a modified model with the OutputPool of the original model will produce biased results in other algorithms besides Rejection sampling. This is because more advanced algorithms learn from previous results. If the results change in some way, so will also the following parameter values and thus also their simulations and other nodes that depend on them. The Rejection sampling does not suffer from this because it always samples new parameter values directly from the priors, and therefore modified distance outputs have no effect to the parameter values of any later simulations.