bayesflow-adapter

You build a BayesFlow data preprocessing pipeline by creating an adapter that maps simulator output into inference variables, summary variables, and inference conditions. The adapter also applies custom transformations and standardizes data using analytical prior moments.

Standardization using prior moments prevents distribution shift by computing scaling factors from analytical prior distributions rather than the simulated data itself. This ensures the preprocessing adapter maintains consistent data distributions across training and inference.

Yes, you can apply custom element-wise or cross-key transformations to simulator output. The adapter supports both imperative and declarative specifications, allowing you to define feature engineering steps directly within the preprocessing pipeline.

The best way to map simulator output is using a BayesFlow adapter to organize keys into inference variables, summary variables, and inference conditions. This schema mapping ensures the data is correctly formatted for downstream neural network consumption.

You do not need to choose exclusively; the adapter supports both imperative and declarative specifications. You can define schema mapping, transformations, and standardization steps using whichever programming style fits your workflow.

Your BayesFlow network receives incorrectly formatted data when the preprocessing pipeline lacks proper schema mapping. Creating an adapter to organize simulator output into inference variables, summary variables, and conditions ensures correct formatting.