API Reference / automl /

portfolio.py

Pick and rank the algorithms an AutoML run should try.

A wall-clock budget is spent one candidate at a time, so the order matters more than the list: whatever is tried first is what a short run gets to keep. This module queries the registry for every algorithm that fits the task, drops the ones that cannot run here, and returns them cheap-and-strong first -- linear and tree baselines, then kernel and instance methods, then ensembles, then boosting, then deep models.
python 
>>> from tuiml.automl.portfolio import build_portfolio
>>> names = [c.name for c in build_portfolio("classification")]
>>> "LogisticRegression" in names
True
>>> names.index("LogisticRegression") < names.index("XGBoostClassifier")
True

Classes

Candidate

class automl.portfolio.Candidate

One algorithm an AutoML run may try.

Parameters

name
str
Registry name, e.g. "RandomForestClassifier".
cls
type
The algorithm class itself.
tier
int
Cost tier index (lower is cheaper); see _TIERS.
tags
tuple of str
The registry tags the algorithm was registered with.

Functions

Func

build_portfolio

Line 155
build_portfolio(task: str, *, candidates: Optional[Iterable[Any]]=None, exclude: Optional[Iterable[str]]=None, exclude_tags: Optional[Iterable[str]]=None, include_wrappers: bool=False, max_candidates: Optional[int]=None) -> List[Candidate]

Return the ranked candidate algorithms for a task.

Overview

  1. Ask the registry for every component of the task's type.
  2. Drop namespaced wrapper keys (sklearn.SVC, weka.J48) unless
include_wrappers=True: they duplicate native algorithms and need an extra installed.
  1. Drop anything tagged with DEFAULT_EXCLUDED_TAGS, and anything
that cannot be constructed with its defaults in this environment.
  1. Sort by (tier, preference, name) -- a total order with no ties, so
two runs on the same install always try the same models in the same sequence.

Parameters

task
{"classification", "regression"}
Which kind of algorithm to collect.
candidates
iterable of str or type
An explicit list that overrides discovery entirely. Names are resolved through the registry; classes are used as given. The order is kept as written, since an explicit list is already a statement of preference.
exclude
iterable of str
Registry names to leave out.
exclude_tags
iterable of str
Tags to leave out, replacing DEFAULT_EXCLUDED_TAGS.
include_wrappers
bool = False
Keep namespaced wrapper entries such as "sklearn.SVC".
max_candidates
int
Truncate the ranked list to this many entries.

Returns

portfolio
list of Candidate
Ranked candidates, cheapest and most reliable first.

Raises

ValueError
If task is not a known task name.
python 
>>> from tuiml.automl.portfolio import build_portfolio
>>> portfolio = build_portfolio("regression", max_candidates=3)
>>> [c.name for c in portfolio]
['LinearRegression', 'BayesianLinearRegressor', 'SGDRegressor']

An explicit list is used verbatim:

json 
>>> [c.name for c in build_portfolio(
...     "classification", candidates=["SVC", "NaiveBayesClassifier"])]
['SVC', 'NaiveBayesClassifier']
Func

describe_portfolio

Line 296
describe_portfolio(portfolio: Sequence[Candidate]) -> List[Dict[str, Any]]

Render a portfolio as plain dicts, for logging or display.

Parameters

portfolio
sequence of Candidate
The candidates to describe.

Returns

rows
list of dict
One {"rank", "name", "tier", "tags"} dict per candidate.
python 
>>> from tuiml.automl.portfolio import build_portfolio, describe_portfolio
>>> describe_portfolio(build_portfolio("regression", max_candidates=1))
[{'rank': 1, 'name': 'LinearRegression', 'tier': 0, 'tags': ['linear', 'regression', 'interpretable']}]