Explore ML Experiments

The projects dashboard in DVC Studio contains all your projects. Click on a project name to open the project table, which contains:

• Git history and live experiments of the project
• Display preferences
• Buttons to visualize and compare experiments.
• Button to export project data.

Git history and live experiments

Branches and commits in your Git repository are displayed along with the corresponding models, metrics, hyperparameters, and DVC-tracked files.

Experiments that you push using the dvc exp push command as well as any live experiments that you send using DVCLive are displayed in a special experiment row nested under the parent Git commit. More details of how live experiments are displayed can be found here.

To manually check for updates in your repository, use the Reload button 🔄 located above the project table.

Nested branches

When a Git branch (e.g., feature-branch-1) is created from another branch (e.g., main), two possibilities exist:

• feature-branch-1 is still active (contains commits that are not present in main). This can happen if the user has pushed new commits to this branch and

  • either hasn't merged it into main yet
  • or has merged it into main but has continued to push more new commits to it after the merger.

  Since the branch now contains new unique commits, the project table will display both main and feature-branch-1 separately. feature-branch-1 will show the new commits that are not part of main while all the merged commits will be shown inside main.

• feature-branch-1 is inactive (does not contain any commits that are not present in main). This can happen in two cases:

  • if the user has not pushed any new commits to feature-branch-1
  • if the user has merged feature-branch-1 into main and has not pushed any new commits to it after the merger.

  Since the branch does not contain any new unique commits, DVC Studio considers feature-branch-1 as "nested" within main and does not display it as a separate branch. This helps to keep the project table concise and reduce clutter that can accumulate over time when inactive branches are not cleaned from the Git repository. After all, those inactive branches usually carry no new information for the purpose of managing experiments. If you would like to display all commits of such an inactive branch, use the Commits on branch = feature-branch-1 display filter.

Display preferences

The table contains buttons to specify filters and other preferences regarding which commits and columns to display.

Filters:

Click on the Filters button to specify which rows you want to show in the project table.

There are two types of filters:

• Quick filters (highlighted in orange above): Use the quick filter buttons to

  • Show only DVC experiments
  • Show only selected experiments
  • Toggle hidden commits (include or exclude hidden commits in the project table)

• Custom filters (highlighted in purple above): Filter commits by one or more of the following fields:

  • Column values (values of metrics, hyperparameters, etc.) and their deltas

  • Git related fields such as Git branch, commit message, tag and author

    The Branch filter displays only the specified branch and its commits.

    On the other hand, the Commits on branch filter will also display branches inside which the specified branch is nested.

    More details on nested branches

    When a Git branch is nested inside another branch, the project table does not display the nested branch. If feature-branch-1 is nested within main, feature-branch-1 is NOT displayed in the project table even if you apply the Branch = feature-brach-1 filter.

    In this case, if you would like to filter for commits in feature-branch-1, you should use the Commits on branch = feature-branch-1 filter. This will display the main branch with commits that were merged from feature-branch-1 into main. A hint is present to indicate that even though the commits appear inside main, they are part of the nested branch feature-branch-1.

    

  • The Custom filters can be un-applied without deletion, allowing you to create the filters once and toggle them on and off as needed.

Columns:

Select the columns you want to display and hide the rest.

If your project is missing some required columns or includes columns that you do not want, refer to the following troubleshooting sections:

• Project does not contain the columns that I want
• Project contains columns that I did not import

To reorder the columns, click and drag them in the table or from the Columns dropdown.

Columns menu and goals: Click on the column header to open a context menu with actions such as sorting and filtering the project table by the column's values.

For metrics, you can also specify goals, which indicate whether an increase or a decrease in the metric's value is desirable. Once a goal is set, the metric's values for all rows are compared against the value in the baseline row. Values that are better (higher or lower, depending on the goal) than that in the baseline row are highlighted in green, with the best one shown with a green border. Values that are worse than that in the baseline row are marked in pink.

Hide commits:

Commits can be hidden from the project table in the following ways:

• DVC Studio auto-hides irrelevant commits: DVC Studio identifies commits where metrics, files and hyperparameters did not change and hides them automatically.

• DVC Studio auto-hides commits that contain [skip studio] in the commit message: This is particularly useful if your workflow creates multiple commits per experiment and you would like to hide all those commits except the final one.

  For example, suppose you create a Git commit with hyper-parameter changes for running a new experiment, and your training CI job creates a new Git commit with the experiment results (metrics and plots). You may want to hide the first commit and only display the second commit, which has the new values for the hyper-parameters as well as experiment results. For this, you can use the string [skip studio] in the commit message of the first commit.

• Hide commits and branches manually: This can be useful if there are commits that do not add much value in your project. To hide a commit or branch, click on the 3-dot menu next to the commit or branch name and click on Hide commit or Hide branch.

  

• Unhide commits: You can unhide commits as needed, so that you don't lose any experimentation history. To display all hidden commits, click on the Show hidden commits toggle (refer filters). This will display all hidden commits, with a hidden (closed eye) indicator.

  

  To unhide any commit, click on the 3-dot menu for that commit and click on Show commit.

  

Delta mode

For metrics, models and files columns with numeric values, you can display either the absolute values or their delta (difference) from the baseline row. To toggle between these two options, use the Delta mode button.

Save changes:

Whenever you make any changes to your project's columns, commits or filters, a notification to save or discard your changes is displayed at the top of the project table. Saved changes remain intact even after you log out of DVC Studio and log back in later.

Visualize and compare experiments

Use the following buttons to visualize and compare experiments:

• Plots: Open the Plots pane and display plots for the selected commits.
• Trends: Generate trend charts to see how the metrics have changed over time.
• Compare: Compare experiments side by side.

These buttons appear above your project table as shown below.

On smaller screens, the buttons might appear without text labels, as shown below.

Export project data

The button to export data from the project table to CSV is present next to the Delta mode button.

Below is an example of the downloaded CSV file.