pdmlabs.optimization.hyperopt_adapter

pdmlabs.optimization.hyperopt_adapter#

Hyperopt TPE optimizer adapter for PdMLabs.

All hyperopt imports are lazy so this module is importable even when hyperopt is not installed. Install with:

pip install pdmlabs[hyperopt]

The adapter uses hyperopt.fmin with tpe.suggest (Tree-structured Parzen Estimator) as the single entry point. initial_random is forwarded to TPE as n_startup_jobs via functools.partial, so the first initial_random evaluations are pure random before the density estimator kicks in. max_evals is always the hard cap on total evaluations regardless of n_startup_jobs.

Space conversion rules (all preserve exact candidate sets via hp.choice):

  • list[int | float | bool | mixed] -> hp.choice(name, values)

  • list[str] -> hp.choice(name, values)

  • rv_frozen -> hp.uniform(name, ppf(0.01), ppf(0.99))

Parallelism note#

TPE is a strictly sequential density-estimation algorithm. Each candidate is proposed based on all previous observed results; there is no batch- proposal step (unlike GPyOpt’s local_penalization). hyperopt.fmin does not support multi-process parallelism without external infrastructure (Spark / MongoDB).

n_jobs is accepted for API consistency but ignored. A warning is emitted when n_jobs > 1, directing users to optimizer='mango', 'gpyopt', or 'smac' for parallel HPO.

Classes

HyperoptAdapter()

Adapter for Hyperopt TPE via hyperopt.fmin.
class pdmlabs.optimization.hyperopt_adapter.HyperoptAdapter#

Bases: BaseOptimizerAdapter

Adapter for Hyperopt TPE via hyperopt.fmin.

Requires hyperopt >= 0.3.0 (pip install pdmlabs[hyperopt]).

Parallelism#

Hyperopt’s TPE is strictly sequential; n_jobs > 1 emits a warning and falls back to a single worker. Use optimizer='mango', 'gpyopt', or 'smac' for parallel HPO.

maximize(param_space: dict, objective_fn, n_iterations: int, n_jobs: int, initial_random: int, constraint_fn=None) → dict#

Maximise objective_fn using Hyperopt TPE.

minimize(param_space: dict, objective_fn, n_iterations: int, n_jobs: int, initial_random: int, constraint_fn=None) → dict#

Minimise objective_fn using Hyperopt TPE.

supports_categorical: bool = True#
Contents