Skip to content

Overview: experiment settings

The available settings for a particular experiment will differ based on the selected problem type and the selected experience level. For the most part, all problem types have certain common experiment settings while having specific problem type settings.

All common experiment settings can be view on the following page: common experiment settings.

All specific problem type settings can be viewed on the following pages:

Experience levels

By specifying your experience level with data science and machine learning, H2O Hydrogen Torch allows you to view specific settings when preparing a new experiment or prediction. Available experience levels are:

  • Novice
  • Skilled
  • Expert
  • Master

You can change the experience level anytime you view the settings for an experiment (in the Create experiment card). To select an experience level:

  1. In the Experience Level list, select the experience level you want to use.

Grid search enables you to select several options for certain grid search hyperparameters. When you turn Grid search On, several experiments start with all possible grid search hyperparameter combinations. After all grid search experiments are completed, you can compare between experiments and see which grid search hyperparameter is best for a given problem type experiment.

Note

  • H2O Hydrogen Torch is not AutoML, and therefore, enabling grid search will not lead to automatic hyperparameter tuning. Enabling grid search will allow for the selection of several options for a grid search hyperparameter.

  • The number of grid search experiments might be randomly sub-sampled based on the computational resources available.

You can turn grid search On or Off anytime you view the settings for an experiment (in the Create experiment card). To turn it On or Off consider the following instructions:

  1. In the H2O Hydrogen Torch navigation menu, click Create experiment.

  2. To enable grid search, click the Grid Search toggle.

Experiment statuses

The different experiment statuses are as follows:

  • Running: The experiment is running without any errors
  • Stopped: The experiment was stopped
  • Finished: The experiment was completed successfully
  • Failed: The experiment failed

  • OOM error: The experiment encountered a GPU out of memory error (OOM)

    Tip

    In an OOM error, reduce the batch size, image size, sequence length, or perhaps consider a more lightweight model.

  • Data error: The experiment encountered an error in reading the provided data

  • Metric error: The experiment encountered an error in the metric calculation

  • Augmentations error: The experiment encountered an error in custom augmentations

    Note

    An augmentation error will occur if the code snippet shared fails, where the code snippet could either be from the custom train/inference augmentations provided for an experiment.

  • Training error: NaN caught in loss during training. Please, reduce the learning rate, or disable mixed precision. Alternatively, gradient clipping may help to stabilize training

  • See logs: The experiment has failed. For the detailed error explanation, please, look at the experiment logs

Select experiment dataset

You can define the experiment dataset on the Create experiment card. Consider the following instructions:

  1. In the Dataset list, select the dataset you want to use.

Select problem type for experiment

When starting an experiment, you can define the problem type on the Create experiment card. On the Create experiment card, consider the following instructions:

  1. In the problem type list, select the problem type you want to use.

    Note

    Right after defining the problem type of an experiment, the settings available will change. The selected experience level determines the available settings for a particular problem type. To learn more, see Experience levels.

Optional: Select model type for experiment

  • Some problem types allow to choose between several modeling approaches by choosing a Model Type.

    Note

    This is an optional setting that is only available for certain problem types. When a model type is changed, different detailed experiment settings will become available for the selected problem type.

Import config from YAML

H2O Hydrogen Torch supports YAML config files' import and export functionality. Users can now download config settings of finished experiments, make changes, and re-upload them when starting a new experiment. To import a YAML file when starting a new experiment, consider the following instructions:

  1. In the H2O Hydrogen Torch navigation menu, click Create experiment.
  2. To enable the setting, click the Import config from YAML toggle.
  3. To upload the YAML file, click Browse....

Name experiment

You can define the experiment name on the Create experiment card when starting an experiment. On the Create experiment card, consider the following instructions:

  1. In the Experiment name box, enter an experiment name.

Run experiment

After defining the dataset, problem type, experiment name, and settings for an experiment, you can run the experiment as follows:

  1. On the Create experiment card, click Run experiment.


Back to top