Experiment details

Basic information

1. Go to the AutoML page.

   1. Log on to the PAI console.

   2. In the left-side navigation pane, click Workspaces. On the Workspaces page, click the name of the workspace that you want to manage.

   3. In the left-side navigation pane, choose Model Training > AutoML.

2. On the experiment list page, click an experiment name to go to the Experiment Details page.

3. On the Basic Information tab of the Experiment Details page, you can view the basic configuration of the experiment, Trial configuration, execution configuration, search configuration, and the execution progress of the Trial.

   • In the Basic Settings section, you can view details such as the experiment ID and status. An experiment can have one of the following statuses:

     • CREATED: The system is initializing the experiment. This process includes creating management processes, parsing configuration modules, and submitting tasks to backend services such as Deep Learning Containers (DLC) or MaxCompute. At this stage, no trials are running.

     • FINISHED: The experiment enters this state after all its trials have completed and at least one trial is FINISHED.

     • FAILED: The experiment enters this state if all its trials fail. The experiment also enters the FAILED state if you manually stop all its trials.

     • RUNNING: The experiment is RUNNING as long as at least one of its trials is running.

     • TERMINATING: The experiment enters this temporary state after you manually stop it.

     • USER_CANCELED: The experiment enters this state after you manually stop it and all processes have terminated.

     • NO_MORE_TRIAL: The experiment has reached the maximum number of trials defined in its configuration and will not generate new ones. After the last running trial completes, the experiment status changes from NO_MORE_TRIAL to FINISHED.

     • TUNER_NO_MORE_TRIAL tuner: The tuner has stopped generating new trials because it lacks sufficient information to determine the next set of hyperparameters to test. In this case, the experiment status is TUNER_NO_MORE_TRIAL tuner.

   • In the Trial Configuration, Execution Configuration, and Search Configuration sections, you can view the experiment's parameter configurations.

   • In the Trial Execution Progress section, you can view the trials' execution progress and status statistics.

Trial list

1. Go to the Experiment Details page. For more information, see View basic information.

2. The Trial List tab displays a list of all trials automatically generated for the experiment.

   The experiment uses the configured algorithm to generate hyperparameter combinations, creating one trial for each combination. Each trial corresponds to a task, which can be a Deep Learning Containers (DLC) task or one or more MaxCompute tasks, depending on the experiment's execution configuration. This underlying task then executes the trial.

Trial run details

The Trial List tab shows the details of each trial, including its execution status, final metric, and hyperparameter combination.

• Status: A trial can be in one of the following states:

  • RUNNING: The trial is in progress.

  • FINISHED: The trial completed successfully.

  • FAILED: The trial failed to complete.

  • USER_CANCELED: You manually stopped the trial.

  • EARLY_STOPPED: The trial was stopped early.

• Final Metric: Based on the metric weight configuration in the experiment, the system uses the weighted sum as the final evaluation metric.

• Hyperparameter Combination: The configured algorithm automatically generates a unique hyperparameter combination for each trial.

You can identify the optimal hyperparameter combination by comparing the Final Metric of each trial against the configured Optimization Direction.

Logs

If an experiment or a trial fails, you can use logs to efficiently troubleshoot the failure.

Experiment logs

1. Go to the Experiment Details page. For more information, see View basic information.

2. Click the Log button on the right side of the page. Alternatively, on the Basic Information tab, in the Basic Settings section, click the Log button next to Status.

3. In the Experiment Log panel, review the log details. Two types of logs are available:

   • nnimanager.log

     This log records experiment management details, such as start and stop events, and error messages.

   • dispatcher.log

     This log records task scheduling and resource management details, such as trial assignment and resource allocation.

Trial logs

1. Go to the Experiment Details page. For more information, see View basic information.

2. On the Trial List tab, find the target trial and click Log in the Actions column.

3. In the Trial Logs panel, review the trial's log details.

   A trial generates the following three types of logs:

   • trial.log

     This log records the detailed execution process of the trial.

   • stdout

     This log contains the standard output of the trial, including print statements and other outputs. It helps you understand debugging information and execution results.

   • stderr

     This log contains the standard error from the trial, including exceptions, error messages, and warnings. When a trial encounters an error, this log provides detailed descriptions to help you troubleshoot the issue.