You can optimize hyper-parameters to improve the performance of your models.
When you click the HPO button in the Pipeline Parameters panel, the Hyper-Parameter Optimizer dialog is displayed.
The Hyper-Parameter Optimizer dialog layout is shown in the image below.
You can enter information for Hyper-Parameter Optimization in the Settings Panel, and check the results in the Result Panel. Moreover, you can check the state of the Hyper-Parameter Optimizer above the Result Panel.
The state is classified into 5 different states.
To optimize the hyper-parameters, enter the necessary information in the settings panel.
- To activate the Run button, you must enter all items in Settings Panel except Early Stopping Rounds.
- Your input is not saved after you click the Close button.
|Select the component with the target you want to optimize.
|The last expression of a target component is automatically selected as the target.
|The Target Expression displays all target values.
|Select one sampler(tuning method) among Bayesian, Grid Search, and Random Search.
|Assign a type to each target. Choose Min if you want to minimize the target value and Max if you want to maximize it.
|Enter the number of iterations.
|Early Stopping Rounds
|(Optional) Set the number of iterations to stop optimization after a target value no longer needs optimization.
|Select one or more pipeline parameters. And enter a Range and choose Distribution for all selected pipeline parameters.
- Range : Enter a range of the target value.
- Distribution : Choose a distribution among Discrete Uniform, Float, Int, Loguniform, or Uniform.
- Interval : Enter Interval instead of selecting Distribution if Grid Search is selected as Sampler.
- If Run is clicked, the hyper-parameter optimization will run from the beginning, regardless of the Ready/Completed/Stopped/Failed status.
- It may be necessary to select one of the Best Trials before clicking Apply if you have multiple Target Values.
- Click the Run button to run the Hyper-Parameter Optimizer.
- As executions take place, you can view the target value changing in real time in the Graph and Log tabs in the Result Panel.
- Iteration Graph: A real-time graph of target values based on the number of executions
- Log Table: A real-time table of target and parameter values based on the number of executions
- Parameter Importance: A graph illustrating the relative importance of pipeline parameters (available only when there are two or more parameters)
- Pareto Front Plot: Graphical representation of the target values (available only when there are two or three target values)
- After the hyper-parameter optimization is complete, apply the optimized value by clicking Apply.
Updated 9 months ago