Experimenter

mllabs._experimenter.Experimenter

Executes and manages a Pipeline experiment on a single dataset.

Splits data using sp (outer) and optionally sp_v (inner), then runs Stage builds and Head experiments fold-by-fold.

Parameters:

Name	Type	Description	Default
data		Input dataset (pandas DataFrame, polars DataFrame, or numpy array).	required
path	str | Path	Directory for persisting experiment artifacts.	required
data_names	list[str]	Column names override.	None
sp		Outer splitter (sklearn splitter API). Default ShuffleSplit(n_splits=1, random_state=1).	ShuffleSplit(n_splits=1, random_state=1)
sp_v		Inner splitter for nested cross-validation. None disables.	None
splitter_params	dict	Maps splitter keyword args to column names in data, e.g. {'y': 'target'}.	None
title	str	Human-readable experiment title.	None
data_key	str	Identifier verified on :meth:load to prevent data mismatch.	None
cache_maxsize	int	Stage output cache size in bytes. Default 4 GB.	4 * 1024 ** 3
logger		Logger instance. Default DefaultLogger(level=['info', 'progress']).	DefaultLogger(level=['info', 'progress'])

Attributes:

Name	Type	Description
pipeline	Pipeline	The pipeline being experimented on.
node_objs	dict	{node_name: StageObj | HeadObj}.
cache	DataCache	Shared LRU cache.
collectors	dict	Registered :class:~mllabs.collector.Collector instances.
trainers	dict	Registered :class:~mllabs._trainer.Trainer instances.
status	str	'open' or 'closed'.

create(data, path, data_names=None, sp=ShuffleSplit(n_splits=1, random_state=1), sp_v=None, splitter_params=None, title=None, data_key=None, cache_maxsize=4 * 1024 ** 3, logger=DefaultLogger(level=['info', 'progress']), aug_data=None) staticmethod

load(filepath, data, data_key=None, aug_data=None) staticmethod

Load a saved Experimenter from disk.

Parameters:

Name	Type	Description	Default
filepath	str | Path	Path to the experiment directory (contains __exp.pkl).	required
data		Dataset to attach. Must match the original data shape.	required
data_key	str	If the saved experiment has a data_key, this must match.	None

Returns:

Name	Type	Description
Experimenter		Restored experimenter with all nodes, collectors, and
		trainers reloaded.

Raises:

Type	Description
ValueError	If data_key does not match the saved value.

set_grp(name, role=None, processor=None, edges=None, method=None, parent=None, adapter=None, params=None, desc=None, exist='diff')

set_node(name, grp, processor=None, edges=None, method=None, adapter=None, params=None, desc=None, exist='diff')

rename_grp(name_from, name_to)

remove_grp(name)

remove_node(name)

노드를 제거

Parameters:

Name	Type	Description	Default
name		제거할 노드 이름	required

Raises:

Type	Description
ValueError	노드가 존재하지 않거나, 자식 노드가 있는 경우

build(nodes=None, rebuild=False)

Build Stage nodes.

Parameters:

Name	Type	Description	Default
nodes		Node query — None (all stages), list, or regex str.	None
rebuild	bool	If True, rebuild already-built nodes.	False

exp(nodes=None)

Run Head nodes and invoke all matching Collectors.

Parameters:

Name	Type	Description	Default
nodes		Node query — None (all heads), list, or regex str.	None

reset_nodes(nodes)

Reset nodes to init state.

Removes node objects, clears cache entries, and resets Collector and Trainer data for the affected nodes.

Parameters:

Name	Type	Description	Default
nodes	list[str]	Node names to reset.	required

show_error_nodes(nodes=None, traceback=False)

Print nodes in error state.

Parameters:

Name	Type	Description	Default
nodes		Node query to filter. None checks all nodes.	None
traceback	bool	Include full traceback in output.	False

finalize(nodes)

Release memory for built Head nodes (built → finalized).

Disk artifacts are preserved so nodes can be reloaded.

Parameters:

Name	Type	Description	Default
nodes		Node query for Head nodes to finalize.	required

reinitialize(nodes)

close_exp()

Finalize all built nodes and mark the experiment as closed.

Collector data is preserved. After this call, :attr:status is 'closed' and no further builds or experiments are permitted until :meth:reopen_exp is called.

reopen_exp()

Reopen a closed experiment and rebuild Stage nodes.

Clears all node objects, sets status back to 'open', then calls :meth:build.

add_collector(collector, exist='skip')

Register a Collector and immediately collect from built Head nodes.

Parameters:

Name	Type	Description	Default
collector	Collector	Collector instance to register.	required
exist	str	'skip' (default) returns existing if already registered; 'error' raises.	'skip'

Returns:

Name	Type	Description
Collector		The registered collector.

get_collector(name)

remove_collector(name)

collect(collector, nodes=None, exist='skip')

Run a Collector ad-hoc over already-built Head nodes.

Parameters:

Name	Type	Description	Default
collector	Collector	Collector instance to run.	required
nodes		Node query — None (all heads), list, or regex str.	None
exist	str	'skip' (default) skips nodes already collected.	'skip'

Returns:

Name	Type	Description
Collector		The same collector after collection.

add_trainer(name, data=None, splitter='same', splitter_params=None, exist='skip', aug_data=None)

Create and register a Trainer.

Parameters:

Name	Type	Description	Default
name	str	Trainer name.	required
data		Dataset for the Trainer. None → use Experimenter's data.	None
splitter		Splitter to use. 'same' reuses sp_v; pass a sklearn splitter object for a custom split strategy; None trains on the full dataset.	'same'
splitter_params	dict	Column mappings for the splitter. Must be None when splitter='same'.	None
exist	str	'skip' (default) returns existing if name already registered; 'error' raises.	'skip'

Returns:

Name	Type	Description
Trainer		The newly created (or existing) Trainer.

get_trainer(name)

remove_trainer(name)

get_node_output(node, idx, v=None)

get_node_train_output(node, idx, v=None)

get_node_valid_output(node, idx, v=None)

process_ext(data, node, idx)

get_n_splits()

get_n_splits_inner()

desc_status()

desc_spec()

desc_pipeline(max_depth=None, direction='TD')

desc_node(node_name, direction='TD', show_params=False)

mllabs._experimenter.DataCache

LRU cache for Stage node outputs, keyed by (node, type, fold_idx).

Capacity is measured in bytes using nbytes / memory_usage.

Parameters:

Name	Type	Description	Default
maxsize	int	Maximum cache capacity in bytes. Default 4 GB.	4 * 1024 ** 3

get_data(node, typ, idx)

put_data(node, typ, idx, data)

clear_nodes(nodes)