Comparing runs
Select runs to compare through the runs table, then use the comparison views for visualization.
You can compare runs based on parameters, metrics, and other model-building metadata.
Neptune has the following comparison views:
  • Charts: compare learning curves for metrics or losses
  • Images: compare images across runs
  • Parallel coordinates: see parameters and metrics displayed on a parallel coordinates plot
  • Side-by-side: see the difference between runs in a table format

Selecting runs to compare

In the runs table:
  • In the Id column, toggle the eye icon to select the runs you want to compare.
Selecting Runs for comparison
  • In the left pane, select Compare runs to access the comparison views.
    • To see both the runs table and comparison view at once, select Horizontal split.
In the Horizontal split mode, you can update the selected runs and see the change live.
Choosing the Runs UI mode to Horizontal split
  • Switch between Charts, Side-by-side, and Parallel coordinates to see different run comparisons.

Editing the legend

To see more information about runs included in charts or comparison views, you can extend the legend with data from additional fields.
  • In the runs table, drag a column to the grey pinned columns area on the left.
Result: The fields of the pinned column are now included in the legend.
Adding columns to chart legend

Comparison options

Charts

Use Charts comparison to compare learning curves for metrics or losses.
You can:
  • filter out metrics that contain a particular string by typing it into Filter charts
  • hover over a chart to see metric value at a particular step
  • zoom in and out by selecting an X or Y range of the chart
Compare metrics or losses on Charts

Images

Use images comparison to see the difference between images you logged for selected Runs.
You can:
  • compare images for computer vision problems such as classification, object detection, segmentation, and so on
  • or compare custom metrics logged as images
Compare images across runs

Parallel coordinates

Use Parallel coordinates comparison to see how model parameters affect the metrics.
It shows all the parameters and metrics that are different between selected Runs.
You can:
  • explore different parameter and metric ranges by selecting a range of a metric/parameter column of interest
  • download the comparison as a CSV table by clicking the Export button
  • download the comparison as PNG of the parallel coordinates plot by clicking the Download as PNG button
Parallel coordinates comparison plot

Side-by-side

Use Side-by-side comparison to see the details of selected Runs in a table format.
Each column is one Run and each row is a single Field of that Run like a parameter or metric.
You can:
  • See the difference-only view of parameters and metrics by ticking the box Rows with diff only
  • See where the metrics and parameters changed by ticking the box Show cell changes. Choose which Run is a reference by clicking the Pin icon.
Side-by-side comparison of Runs

Artifact

Use Artifact comparison to see the difference between datasets, models, and other artifacts that you version in Neptune.
For Runs where artifacts are different, each artifact is displayed.
You can:
  • Choose for which Runs you want to compare artifacts (you can only compare two artifacts at the same time)
  • See the difference in Hash (md5 signature) of artifacts
  • Compare the size of the artifact
  • Browse the artifacts folder
Compare dataset artifacts in the Neptune UI
To compare objects with the Artifacts comparison you need to log them as artifacts first.

Notebooks

To be able to log and compare Jupyter Notebooks, you need to have the Neptune+Jupyter Integration setup.
You can compare different notebooks (or different checkpoints of the same notebook) logged to a project.
To do this,
  • Go to the Notebooks tab and click on Compare:
  • Select the notebooks or checkpoints you want to compare from the dropdown, and click on Compare:
The comparison shows you differences in the code, markdown, cell outputs, and execution order:

What's next?