Gather and share insights
In the Neptune app, you can analyze runs and their metadata with customizable tools such as custom views, dashboards, and reports. Each of them becomes the most useful at different stages of the model training.
| Custom view | Dashboard | Report | |
|---|---|---|---|
| Displayed runs | Configurable with filters | Selected by a viewer | Selected by an author |
| Scope | One project | One project | Multiple projects and workspaces |
| Purpose | Filter, group, and organize runs | Analyze the ongoing experiments | Collaborate, document, and analyze the final results of experiments |
Custom view
Configure custom views to prepare your metadata for the analysis.
A custom view displays runs in a table where each logged metric is represented as a separate column. This layout can give you an idea about how the runs compare against each other even before you move to visualizations.
A custom view includes the following settings:
- Table configuration. You can add, remove, and rearrange columns.
- Query filters. Apply filters to gather runs that are worth comparing and analyzing in the next steps. Filter out runs that you find unnecessary.
- Groups. Organize runs into groups and compare them against each other.
You can apply custom views in other parts of the Neptune app, for example dashboards. As a result, you can narrow down the visualizations to the filtered runs.
- Ideal for filtering, grouping, and sorting runs.
- You can combine a custom view with other tools in the Neptune app to create focused visualizations.
Dashboard
Dashboard is a customized visaulization that stays persistent regardless of the selected runs.
Use dashboards to take a closer look at specific metrics and abstract everything else away. It's a quick and focused view, ideal when you need to repeatedly visualize and analyze specific metrics across multiple runs.
Dashboard configuration consists of widgets but it doesn't include runs. Users can apply a dashboard to whichever runs they select.
When you find an interesting run that you want to document and save, you can export a dashboard and the selected runs to a report. You can combine dashboards with custom views to break the analysis into smaller steps or create an iterative process.
- Ideal for analyzing the ongoing experiments.
- Compares experiments and metrics within one project.
- Applies to runs selected by dashboard viewers.
Report
Use reports for documenting insights and collaborating with your team members.
A report is a static representation of your analysis. You can refer to it anytime to view the captured stage of the training, selected runs, and additional context or explanations.
Reports allow you to perform an analysis in one place:
- Create multiple table views, each for a different cycle or data set.
- Export dashboards into a report and organize them into report sections. For example, a separate section for debugging, evaluation, or system metrics.
- Include runs from multiple projects or even workspaces. The report creator selects runs that the viewers see.
- Ideal for analyzing the final results of experiments.
- Compares experiments from multiple projects and workspaces.
- Includes runs selected by the report creator.
- Includes comments and extra resources such as links or images.