aimz: Flexible probabilistic impact modeling at scale
Overview
aimz is a Python library for flexible and scalable probabilistic impact modeling to assess the effects of interventions on outcomes of interest. Designed to work with user-defined models with probabilistic primitives, the library builds on NumPyro, JAX, Xarray, and Zarr to enable efficient inference workflows.
Features
- An intuitive API that combines ease of use from ML frameworks with the flexibility of probabilistic modeling.
- Scalable computation via parallelism and distributed data processing—no manual orchestration required.
- Variational inference as the primary inference engine, supporting custom optimization strategies and results.
- Support for interventional causal inference for modeling counterfactuals and causal relations.
Workflow
- Outline the model, considering the data generating process, latent variables, and causal relationships, if any.
- Translate the model into a kernel (i.e., a function) using NumPyro and JAX.
- Integrate the kernel into the provided API to train the model and perform inference.