The current implementation of INITIALIZE does not actually initialize any Loom models, it only prepares database tables to store outcomes of loom analyses.
Instead, INITIALIZE should perform a single subsample annealing pass through the data, which will ensure that models are actually initialized and queries can be run against them. Model improvement can then be achieved using additional SMC passes through the dataset with ANALYZE.
The current implementation of INITIALIZE does not actually initialize any Loom models, it only prepares database tables to store outcomes of loom analyses.
https://github.com/probcomp/bayeslite/blob/39e96af568d8f11528583b263042b4711e0cd5ff/src/metamodels/loom_metamodel.py#L309-L324
Instead, INITIALIZE should perform a single subsample annealing pass through the data, which will ensure that models are actually initialized and queries can be run against them. Model improvement can then be achieved using additional SMC passes through the dataset with ANALYZE.