Components

A soma model is a frozen foundation-model encoder followed by a trained downstream model: an aggregator for a set or a decoder for a grid, then a task head. Each part is a {name, params} leaf, chosen from a registry independently of the others. soma.components(kind) lists the registered entries of a kind with their params and defaults, and Extending soma shows how to add your own.

• Encoders: the frozen tile-, slide- and patient-level encoders

• Modeling: which downstream path fits your features and task

• Aggregators: the MIL aggregators that pool a bag of tile features

• Decoders: the decoders that map a dense feature grid back to pixels

The task heads are listed in Tasks.