HyperDriveConfig Class
Configuration that defines a HyperDrive run.
HyperDrive configuration includes information about hyperparameter space sampling, termination policy, primary metric, resume from configuration, estimator, and the compute target to execute the experiment runs on.
Initialize the HyperDriveConfig.
- Inheritance
-
builtins.objectHyperDriveConfig
Constructor
HyperDriveConfig(hyperparameter_sampling, primary_metric_name, primary_metric_goal, max_total_runs, max_concurrent_runs=None, max_duration_minutes=10080, policy=None, estimator=None, run_config=None, resume_from=None, resume_child_runs=None, pipeline=None, debug_flag=None, custom_run_id=None)Parameters
- estimator
- MMLBaseEstimator
An estimator that will be called with sampled hyperparameters.
Specify only one of the following parameters: estimator, run_config,
or pipeline.
- policy
- EarlyTerminationPolicy
The early termination policy to use. If None - the default, no early termination policy will be used.
The MedianStoppingPolicy with delay_evaluation of 5
is a good termination policy to start with. These are conservative settings,
that can provide 25%-35% savings with no loss on primary metric (based on our evaluation data).
- primary_metric_goal
- PrimaryMetricGoal
Either PrimaryMetricGoal.MINIMIZE or PrimaryMetricGoal.MAXIMIZE. This parameter determines if the primary metric is to be minimized or maximized when evaluating runs.
- max_total_runs
- int
The maximum total number of runs to create. This is the upper bound; there may
be fewer runs when the sample space is smaller than this value.
If both max_total_runs and max_duration_minutes are specified, the
hyperparameter tuning experiment terminates when the first of these two thresholds
is reached.
- max_concurrent_runs
- int
The maximum number of runs to execute concurrently. If None, all runs are launched in parallel. The number of concurrent runs is gated on the resources available in the specified compute target. Hence, you need to ensure that the compute target has the available resources for the desired concurrency.
- max_duration_minutes
- int
The maximum duration of the HyperDrive run. Once this time is exceeded, any runs
still executing are cancelled. If both max_total_runs and
max_duration_minutes are specified, the hyperparameter tuning experiment
terminates when the first of these two thresholds is reached.
- resume_from
- HyperDriveRun or list[HyperDriveRun]
A hyperdrive run or a list of hyperdrive runs that will be inherited as data points to warm start the new run.
A hyperdrive child run or a list of hyperdrive child runs that will be resumed as new child runs of the new hyperdrive run.
- run_config
- ScriptRunConfig
An object for setting up configuration for script/notebook runs.
Specify only one of the following parameters: estimator, run_config,
or pipeline.
- pipeline
- Pipeline
A pipeline object for setting up configuration for pipeline runs.
The pipeline object will be called with the sample hyperparameters to submit pipeline runs.
Specify only one of the following parameters: estimator, run_config,
or pipeline.
- hyperparameter_sampling
- HyperParameterSampling
The hyperparameter space sampling definition.
- primary_metric_goal
- PrimaryMetricGoal
Either PrimaryMetricGoal.MINIMIZE or PrimaryMetricGoal.MAXIMIZE. This parameter determines if the primary metric is to be minimized or maximized when evaluating runs.
- max_total_runs
- int
The maximum total number of runs to create. This is the upper bound; there may be fewer runs when the sample space is smaller than this value.
- max_concurrent_runs
- int
The maximum number of runs to execute concurrently. If None, all runs are launched in parallel.
- max_duration_minutes
- int
The maximum duration of the HyperDrive run. Once this time is exceeded, any runs still executing are cancelled.
- policy
- EarlyTerminationPolicy
The early termination policy to use. If None - the default, no early termination policy will be used.
The <xref:azureml.train.hyperdrive.MedianTerminationPolicy> with delay_evaluation of 5
is a good termination policy to start with. These are conservative settings,
that can provide 25%-35% savings with no loss on primary metric (based on our evaluation data).
- estimator
- MMLBaseEstimator
An estimator that will be called with sampled hyper parameters.
Specify only one of the following parameters: estimator, run_config,
or pipeline.
- run_config
- ScriptRunConfig
An object for setting up configuration for script/notebook runs.
Specify only one of the following parameters: estimator, run_config,
or pipeline.
- resume_from
- HyperDriveRun | list[HyperDriveRun]
A hyperdrive run or a list of hyperdrive runs that will be inherited as data points to warm start the new run.
A hyperdrive child run or a list of hyperdrive child runs that will be resumed as new child runs of the new hyperdrive run.
- pipeline
- Pipeline
A pipeline object for setting up configuration for pipeline runs.
The pipeline object will be called with the sample hyperparameters to submit pipeline runs.
Specify only one of the following parameters: estimator, run_config,
or pipeline.
- debug_flag
Remarks
The example below shows creating a HyperDriveConfig object to use for hyperparameter tunning. In the example, the primary metric name matches a value logged in the training script.
hd_config = HyperDriveConfig(run_config=src,
hyperparameter_sampling=ps,
policy=early_termination_policy,
primary_metric_name='validation_acc',
primary_metric_goal=PrimaryMetricGoal.MAXIMIZE,
max_total_runs=4,
max_concurrent_runs=4)
For more information about working with HyperDriveConfig, see the tutorial Tune hyperparameters for your model.
Attributes
estimator
Return the estimator used in the HyperDrive run.
Value is None if the run uses a script run configuration or a pipeline.
Returns
The estimator.
Return type
pipeline
Return the pipeline used in the HyperDrive run.
Value is None if the run uses a script run configuration or estimator.
Returns
The pipeline.
Return type
run_config
Return the script/notebook configuration used in the HyperDrive run.
Value is None if the run uses an estimator or pipeline.
Returns
The run configuration.
Return type
source_directory
Feedback
Coming soon: Throughout 2024 we will be phasing out GitHub Issues as the feedback mechanism for content and replacing it with a new feedback system. For more information see: https://aka.ms/ContentUserFeedback.
Submit and view feedback for