Delphi

A version control system and standard runtime for Bayesian models in science.

Installation

Optionally, modify pyproject.toml to specify whether to use the GPU (default) or CPU runtime for jax.

[tool.poetry.dependencies]
jaxlib = {source = "jax-gpu"} # or "jax-cpu"

Then, install the package with poetry install.

Usage

Depositing, retreiving, and editing models

The Delphi class acts as a repository for Bayesian models, which can be deposited into its store using the Delphi.submit(ModelClass) method. Delphi automatically finds the file containing the model class and stores it under a unique entry within its models directory. This unique hash (model_id) is returned and can be used to retrieve the model class from the store later using the Delphi[model_id] syntax.

For any modifications to the model, Delphi.edit(model_id) copies a temporary copy of the model file into the scratch directory, where it can be modified. Once the modifications are complete, Delphi.submit(ModelClass, delete_file=True) will store the new version of the model in the models directory and delete the temporary copy in the scratch directory.

Operations on models

• Delphi.run(model.predict, data) runs MCMC sampling against the model and returns the final MCMC object containing sampled parameters and diagnostics.
• Delphi.mix([model_id1, model_id2, ...], alpha=1.0) returns a new predictor (dictionary of variable name => numpyro.dist.Distribution) using Dirichlet(alpha) as the mixing prior.
• Delphi.compare([model_id1, model_id2, ...], data, alpha=1.0) uses Delphi.mix to compute the posterior mixing distribution based on the observations in data, which can be used to calculate the relative marginal likelihood of each model (see Kamari et al.).

Model API

This API is currently in a state of flux and may be subject to breaking change.

To be used with Delphi, your model simply needs to inherit from delphi.model.Model and implement the following methods:

• predict: Given a set of inputs, return a set of predictions as a dictionary of variable name => numpyro.dist.Distribution.
• observables: Given a (partially incomplete) dataset, return a dictionary of variable name => observations, with np.nan for missing values.