`AutoMAX`

class AutoMAX(trainer, config, target)

Automated AUC optimisation using SMAC3.

Wraps a LibAUC Trainer with a SMAC3 HyperparameterOptimizationFacade to perform Bayesian hyperparameter search over the configuration space exposed by trainer.

Parameters:

trainer (callable) – A callable (typically an autopartial instance) that accepts a space argument and returns a Trainer. The .cs attribute of the trainer is used as the SMAC3 configuration space.
config (AutoMAXConfigration) – Run-level settings (number of trials, output directory, seed …).
target (str) – Name of the metric to maximise (e.g. "AUROC").

Example

from src.core import AutoMAX, AutoMAXConfigration

config = AutoMAXConfigration(
    name="my_search",
    n_trials=10,
    n_configs=1,
    SEED=42,
    output_directory="./automax_output",
    overwrite=True,
    deterministic=True,
)

tuner = AutoMAX(trainer=my_autopartial_trainer, config=config, target="AUROC")
tuner.optimize()

Constructor details

__init__(trainer, config, target): Initialises the SMAC3 scenario and HyperparameterOptimizationFacade. If config.overwrite is False and a previous run exists, the trial log is restored from state.pkl.

Methods

train(space, seed=0) → float

Train one model with the hyperparameter configuration space.

Called automatically by SMAC3 during optimisation; you rarely need to call this directly.

Parameters:

space – A SMAC3 / ConfigSpace Configuration object.
seed (int) – Random seed for this trial (passed in by SMAC3).

Returns:

Negative validation score (SMAC3 minimises cost).

Return type:

float

After training the method:

Appends the trial summary to self.log.
Persists self.log to <output_directory>/<name>/state.pkl.
Copies the model checkpoint directory to <experiment_name>_best whenever a new best score is found.

optimize() → Configuration

Run the full SMAC3 optimisation loop and print a trial-history table.

Resumes any interrupted trials before calling optimize().

Returns:: The incumbent (best) configuration found.
Return type:: Configuration

Trial history table — printed to stdout after optimisation:

----------------------------------------
Trial    Val AUROC
----------------------------------------
1        0.823456
2        0.851234
…
----------------------------------------

Attributes

Attribute	Description
`configspace`	The `ConfigurationSpace` taken from trainer.
`trainer`	The wrapped trainer callable.
`target`	Metric name string (e.g. `"AUROC"`).
`config`	The `AutoMAXConfigration` instance.
`smac`	The underlying `HyperparameterOptimizationFacade` instance.
`log`	List of per-trial summary dicts (`val`, `test`, `space`, …).
`finished`	Number of completed trials at initialisation time.
`best_score`	Running best validation score (`float('-inf')` at start).

AutoMAX

`AutoMAX`