Built-in component catalogue#
Almost everything in a KonfAI config is a component referenced by name: a model, a loss, a metric, a transform, an augmentation, a scheduler, or a storage backend. This section is the catalogue of what ships in the box — these names are exactly what you reference in YAML: copy them into your config verbatim, with the constructor arguments listed alongside.
Note
These pages are generated from a source-level read of konfai/: class names,
constructor signatures and defaults are taken directly from the code.
The pages#
Start with Models (the network you train), then Losses & metrics (what you attach to its named outputs). Transforms and Augmentations cover the data pipeline, Schedulers the loss-weight and learning-rate schedules, and Storage backends & formats the on-disk formats.
How a name is resolved#
Every component name in a config is resolved by konfai.utils.utils.get_module
in one of two ways:
Form |
Example |
Resolves to |
|---|---|---|
bare name |
|
inside that kind’s package ( |
|
|
any importable module — an installed library or a local |
So the tables below list the bare name for built-ins; you are never limited
to them — any importable class that satisfies the same contract works via the
module:Class form. See Configuration model for the full
resolution rules and Extension points for how to write
your own.
How to discover a component’s parameters#
The tables give the key constructor arguments and defaults, but the exact,
always-current parameter set is whatever the class’s __init__ declares — the
reflection engine binds YAML keys directly to constructor parameter names. Two
ways to get the exhaustive list for any component:
Let KonfAI materialise the defaults. Reference the component in a config and run the workflow (or run with
KONFAI_CONFIG_MODE=default). KonfAI writes every resolved default back into the YAML file, giving you a complete, fully-expanded subtree to edit. (This is the same config-mutation behaviour that surprises new users — here it is a feature.)Read the signature. Bare names map to
konfai/metric/measure.py,konfai/data/transform.py,konfai/data/augmentation.py,konfai/models/**, andkonfai/metric/schedulers.py.
Next steps#
Configuration model — how names and
classpathare resolvedExtension points — writing your own model / loss / transform