How AutoML works

AutoML iterates through experiments, trials, and training tasks to automatically search for the optimal hyperparameter combination.

The following figure shows how AutoML works.

After you set the hyperparameter value ranges, search algorithm, and stop conditions, AutoML passes them to the backend as one experiment.

1. The experiment generates hyperparameter combinations based on the configured algorithm. Each combination corresponds to one trial.

   Note

   You can run multiple trials concurrently to increase speed, but this consumes more resources per unit of time.

2. Each trial corresponds to one hyperparameter combination and one or more compute tasks. A task can be a DLC task, which uses general computing resources and Lingjun resources, or a MaxCompute task, which uses MaxCompute computing resources. Each task follows the billing, configuration, and usage logic of its own service.

3. After a trial starts, AutoML repeatedly checks the task metrics.

4. The experiment stops when it meets a stop condition. Stop conditions include the maximum number of searches, the search algorithm stop condition, and the completion of computing for all combinations.

5. AutoML returns the results. The results can be hyperparameter combinations or the best model of each trial. To obtain the best model, set a model storage path. You can also view the results in the log data.

Based on the preceding mechanism, complete the required configurations before you start an experiment. Configure the basic experiment settings, trial settings, DLC or MaxCompute task settings, and hyperparameter search settings.