Note

Access to this page requires authorization. You can try signing in or .

Access to this page requires authorization. You can try .

BanditPolicy Class

Defines an early termination policy based on slack criteria and a frequency and delay interval for evaluation.

Constructor

BanditPolicy(*, delay_evaluation: int = 0, evaluation_interval: int = 0, slack_amount: float = 0, slack_factor: float = 0)

Keyword-Only Parameters

Name Description
delay_evaluation
int

Number of intervals by which to delay the first evaluation. Defaults to 0.

Default value: 0
evaluation_interval
int

Interval (number of runs) between policy evaluations. Defaults to 0.

Default value: 0
slack_amount

Absolute distance allowed from the best performing run. Defaults to 0.

Default value: 0
slack_factor

Ratio of the allowed distance from the best performing run. Defaults to 0.

Default value: 0

Examples

Configuring BanditPolicy early termination of a hyperparameter sweep on a Command job.


 from azure.ai.ml import command

 job = command(
 inputs=dict(kernel="linear", penalty=1.0),
 compute=cpu_cluster,
 environment=f"{job_env.name}:{job_env.version}",
 code="./scripts",
 command="python scripts/train.py --kernel $kernel --penalty $penalty",
 experiment_name="sklearn-iris-flowers",
 )

 # we can reuse an existing Command Job as a function that we can apply inputs to for the sweep configurations
 from azure.ai.ml.sweep import Uniform

 job_for_sweep = job(
 kernel=Uniform(min_value=0.0005, max_value=0.005),
 penalty=Uniform(min_value=0.9, max_value=0.99),
 )

 from azure.ai.ml.sweep import BanditPolicy

 sweep_job = job_for_sweep.sweep(
 sampling_algorithm="random",
 primary_metric="best_val_acc",
 goal="Maximize",
 max_total_trials=8,
 max_concurrent_trials=4,
 early_termination_policy=BanditPolicy(slack_factor=0.15, evaluation_interval=1, delay_evaluation=10),
 )

Feedback

Was this page helpful?