InstanceSpace

Build, inspect, persist and explore an instance-space pipeline

Syntax

space = InstanceSpace(metadata, options)
model = space.build()
result = space.explore(test_metadata, dataset_id=None)
space.save(path, secret_key=None)
space = InstanceSpace.load(path, secret_key=None)

See exact signatures and return types for all optional constructor and method arguments.

Description

InstanceSpace owns the training metadata, immutable configuration, stage runner, fitted model and exploration results. Its default stage set is the complete seven-stage pipeline. Additional stages declare their ordering explicitly with RunBefore and RunAfter.

build() returns a Model. model raises when a complete valid model is unavailable. run_until_stage returns a dictionary of available outputs; run_stage returns that stage's typed output record. A complete model requires the full pipeline. explore() applies trained artifacts and retains results in explore_results.

Examples

Follow the runnable Getting Started example for construction, fitting, plotting and exploration. Build and Explore shows stage inspection and checkpointing.

Parameters

Parameter	Meaning
`metadata`	Validated `Metadata` object containing training features and algorithm outcomes.
`options`	Validated `InstanceSpaceOptions`. Use `.default()` or `.from_dict()`.
`stages`	Optional stage classes. The default is the complete built-in pipeline. A subset must retain the prerequisites needed by the operations you request.
`additional_initial_inputs_type`	Optional NamedTuple type declaring extra plugin inputs.
`progress_reporter`	Optional progress reporter. The default reporter performs no external notifications.

Returns

build() returns Model; explore() returns ExploreResult. run_iter and explore_stage_iter yield intermediate stage records. File-loading convenience functions can return None on load failure; check before calling methods.

Resource Management

Call close() when an instance's executor is no longer needed. Keep outer application parallelism and per-model worker counts within the available machine resources.