Dashboards

In Experiments, compare runs side by side. This dashboard shows how three training changes affect accuracy and loss compared with a baseline:

A baseline, RandAugment, mixup, and a higher learning rate, compared through accuracy and loss.

The run selector on the left supplies the data. The dashboard on the right presents it in panels, grouped into sections. You can explore similar runs using the Hub demo project template.

Select runs#

Search or filter in the run selector to find runs. The eye buttons control which runs appear in the panels.

Filtering the run list does not change the comparison. Here, all four runs remain in the charts while the list shows only learning rate experiments.

Keep a baseline selected to see whether a change improves the result. Compare runs measured under the same conditions.

Choose a dashboard#

A dashboard is a saved layout of panels. Training comparison is the layout used here; you might keep another for latency and memory on target hardware. Switch layouts using the selector above the panels. Start with +, or use Duplicate to adapt an existing dashboard.

Organize the comparison#

Sections group related panels. This example keeps accuracy and loss together so you can compare training progress with the final outcome.

Add panel belongs to the section. Choose a chart or table to add to it.

Use Add section for another group of panels. Panel widths let you place related charts side by side or give a larger comparison the full width.

Read the training curves#

The accuracy curves show how quickly each run improves and where it levels off.

RandAugment and mixup finish above the baseline here. The higher learning rate finishes below it. The loss panel beside this chart provides another view of how training progressed.

The accuracy panel is a Line chart using val_accuracy. Edit panel lets you change the metric and smoothing:

The settings on the right update the chart preview on the left.

This view uses 80% smoothing to make the trends easier to follow. The raw curve remains visible so you can still see variation in the measurements.

Compare the final results#

The bar chart below the curves compares the last recorded accuracy from each run.

RandAugment has the highest final accuracy here, with mixup close behind.

This Bar chart uses val_accuracy with latest as its value. Choose the value that answers your question: the highest accuracy seen during training may differ from the last recorded accuracy.

For other comparisons, use a Scatter plot to explore accuracy against latency, Parallel coordinates for several measurements, or Table for exact values.

Save changes keeps the layout. Share in the dashboard actions shares the comparison with your selected runs.

A shared comparison includes the selected runs, so teammates can inspect the same results.

Apply changes a panel in your draft. Save changes saves the dashboard for your team. On a run or model’s embedded dashboard, the button is Save.

Next steps#