The Experiments Table

The experiments table is a customizable list of experiments associated with a project. From the experiments table, view experiment details, and to work with experiments (for example, reset, clone, enqueue experiments, and create tracking leaderboards to monitor experimentation). The experiments table auto refresh allows you to continually monitor experiment progress. Sometimes, it may be faster to locate experiments on the All projects page, because its experiments table contains all experiments.

Note

Archiving experiments and models To assist you in focusing on active experimentation, you can archive experiments and models. They appear in the archive area only, not in the active experiments and models tables. See Archiving.

image

Experiments table columns

The experiments table default and customizable columns are described in the following table.

Column Description Type
TYPE Type of experiment. ClearML supports multiple task types for experimentation, and a variety of workflows and use cases. Default
NAME Experiment name. Default
TAGS Descriptive, user-defined, color-coded tags assigned to experiments. Use tags to classify experiments, and filter the list. See tagging experiments. Default
STATUS Experiment state (status). See a list of the task states and state transtions. Default
PROJECT Experiment's project. Default
USER User creating or cloning the experiment. Default (hidden)
STARTED Elapsed time since the experiment started. To view the date and time, hover over the elapsed time. Default
UPDATED Elapsed time since the last update to the experiment. To view the date and time, hover over the elapsed time. Default
ITERATION Last or most recent iteration of the experiment. Default
DESCRIPTION A description of the experiment. For cloned experiments, the description indicates it is auto generated with a timestamp. Default (hidden)
RUN TIME The current / total running time of the experiment. Default (hidden)
Metrics Add metrics column (last, minimum, and / or maximum values). The metrics depend upon the experiments in the table. See adding metrics. Customizable
Hyperparameters Add hyperparameters. The hyperparameters depend upon the experiments in the table. See adding hyperparameters Customizable

Customizing the experiments table

You can customize the experiments table by showing / hiding default columns, adding metrics and hyperparameters, sorting, and filtering.

Use experiments table customization for various use cases, including:

  • Sorting models by metrics - Models are associated with the experiments that created them. To sort models by metrics, we use the experiments table. For each metric, use the last value, the minimal value, and / or the maximal value.

  • Tracking hyperparameters - Track hyperparameters by adding them as columns, and applying filters and sorting.

Changes are persistent (cached in the browser), and represented in the URL so that you can save customized settings in a browser bookmark and share it with other ClearML users to collaborate.

image

Adding metrics

Add metrics columns to the experiments table. The metrics depend upon the experiments in the table.

To add metrics:

  • Add metrics - Click image > + METRICS > Expand a metric > Select the LAST (value), MIN (minimal value), and / or MAX (maximal value) checkboxes.

Adding hyperparameters

Add hyperparameters to the experiments table. The hyperparameters depend upon the experiments in the table.

To add hyperparameters:

  • Add hyperparameters - Click image > HYPER PARAMETERS > Expand a section > Select the hyperparameter checkboxes.

Using other customization features

To use other customization features:

  • Show / hide columns - Click image > select or clear the checkboxes of columns to show or hide.

  • Filter columns - Type of experiment, experiment status (state), user

  • Sort columns - Metrics and hyperparameters, type of experiment, experiment name, start and last update elapsed time, and last iteration.

  • Dynamic column ordering - Drag a column title to a different position.

  • Column resizing - In the column heading, drag to a new size.

  • Column autofit - In the column heading, double click a resizer (column separator).

ClearML actions from the experiments table

The following table describes the ClearML features you can use from the experiments table, including the states that allow each feature.

ClearML Action Description States Valid for the Action State Transition
View details The ClearML Web UI provides two ways to view experiment details (in addition to the columns on the experiments table): the info panel (keep the experiments table in view) and the full screen details view (more easily work with all experiment details). Any state None
Manage a queue If an experiment is Pending in a queue, view the utilization of that queue, manage that queue (remove experiments and change the order of experiments), and view information about the worker(s) listening to the queue. See the Workers and queues page. Enqueued None
View a worker If an experiment is Running, view resource utilization, worker details, and queues to which a worker is listening. Running None
Share For ClearML Hosted Service users only, share an experiment and its model with a ClearML Hosted Service user in another workspace. Any state None
Archive To more easily work with active experiments, move an experiment to the archive. See Archiving. Any state None
Enqueue Add an experiment to queue for a worker or workers (listening to the queue) to execute. Draft Pending
Dequeue Remove an experiment from a queue. Pending Draft
Reset Delete the log and output from a previous run of an experiment (for example, before rerunning it). Completed, Aborted, or Failed Draft
Abort Manually terminate a Running experiment. Running Aborted
Publish Publish an experiment to prevent changes to its tracking data, inputs, and outputs. Published experiments are read-only. If an experiment is Published, its model is also Published (becomes read-only). Published experiments cannot be enqueued to execute; to enqueue a Published experiment (or tune it and then enqueue the tuned experiment), first clone it and then work with the newly cloned experiment. Completed, Aborted, or Failed. Published
Tags Tag experiments with color-coded labels to assist you in organizing your work. See tagging experiments. Any state None
Clone Make an exact, editable copy of an experiment (for example, to reproduce an experiment, but keep the original). Draft Newly Cloned Experiment is Draft
Move To organize your work and improve collaboration, move an experiment to another project. Any state None