clearml-docs/docs/fundamentals/hpo.md

title
Hyperparameter Optimization

What is HyperParameter Optimization?

Hyperparameters are variables that directly control the behaviors of training algorithms, and have a significant effect on the performance of the resulting machine learning models. Finding the hyperparameter values that yield the best performing models can be complicated. Manually adjusting hyperparameters over the course of many training trials can be slow and tedious. Luckily, hyperparameter optimization can be automated and boosted using ClearML's HyperParameterOptimizer class.

ClearML's HyperParameter Optimization

ClearML provides the HyperParameterOptimizer class, which takes care of the entire optimization process for users in with a simple interface.

ClearML's approach to hyperparameter optimization is scalable, easy to set up and to manage, and it makes it easy to compare results.

Workflow

The diagram above demonstrates the typical flow of hyperparameter optimization where the parameters of a base task are optimized:

1. Configure an Optimization Task with a base task whose parameters will be optimized, and a set of parameter values to test
2. Clone the base task. Each clone's parameter is overridden with a value from the optimization task
3. Enqueue each clone for execution by a ClearML Agent
4. The Optimization Task records and monitors the cloned tasks' configuration and execution details, and returns a summary of the optimization results

Supported Optimizers

The HyperParameterOptimizer class contains ClearML’s hyperparameter optimization modules. Its modular design enables using different optimizers, including existing software frameworks, enabling simple, accurate, and fast hyperparameter optimization.

• Optuna - automation.optuna.optuna.OptimizerOptuna. Optuna is the default optimizer in ClearML. It makes use of different samplers such as grid search, random, bayesian, and evolutionary algorithms. For more information, see the Optuna documentation.
• BOHB - automation.hpbandster.bandster.OptimizerBOHB. BOHB performs robust and efficient hyperparameter optimization at scale by combining the speed of Hyperband searches with the guidance and guarantees of convergence of Bayesian Optimization. For more information about HpBandSter BOHB, see the HpBandSter documentation.
• Random uniform sampling of hyperparameters - automation.optimization.RandomSearch.
• Full grid sampling strategy of every hyperparameter combination - Grid search automation.optimization.GridSearch.
• Custom - automation.optimization.SearchStrategy - Use a custom class and inherit from the ClearML automation base strategy class

Defining a hyperparameter optimization search example

1. Import ClearML's automation modules:

from clearml.automation import UniformParameterRange, UniformIntegerParameterRange
from clearml.automation import HyperParameterOptimizer
from clearml.automation.optuna import OptimizerOptuna

1. Initialize the Task, which will be stored in ClearML Server when the code runs. After the code runs at least once, it can be reproduced, and the parameters can be tuned:

from clearml import Task

task = Task.init(project_name='Hyper-Parameter Optimization',
               task_name='Automatic Hyper-Parameter Optimization',
               task_type=Task.TaskTypes.optimizer,
               reuse_last_task_id=False)

1. Define the optimization configuration and resources budget:

optimizer = HyperParameterOptimizer(
      # specifying the task to be optimized, task must be in system already so it can be cloned
      base_task_id=TEMPLATE_TASK_ID,  
      # setting the hyper-parameters to optimize
      hyper_parameters=[
          UniformIntegerParameterRange('number_of_epochs', min_value=2, max_value=12, step_size=2),
          UniformIntegerParameterRange('batch_size', min_value=2, max_value=16, step_size=2),
          UniformParameterRange('dropout', min_value=0, max_value=0.5, step_size=0.05),
          UniformParameterRange('base_lr', min_value=0.00025, max_value=0.01, step_size=0.00025),
          ],
      # setting the objective metric we want to maximize/minimize
      objective_metric_title='accuracy',
      objective_metric_series='total',
      objective_metric_sign='max',  

      # setting optimizer  
      optimizer_class=OptimizerOptuna,
  
      # configuring optimization parameters
      execution_queue='default',  
      max_number_of_concurrent_tasks=2,  
      optimization_time_limit=60., 
      compute_time_limit=120, 
      total_max_jobs=20,  
      min_iteration_per_job=15000,  
      max_iteration_per_job=150000,  
      )

For more information about HyperParameterOptimizer and supported optimization modules, see the HyperParameterOptimizer class reference.

Tutorial

Check out the Hyperparameter Optimization tutorial for a step-by-step guide.