Visualize and Compare Experiments

You can visualize and compare experiments using using plots, images, charts, etc.

Display plots and images

You can visualize certain metrics of machine learning experiments as plots. Usual plot examples are AUC curves, loss functions, and confusion matrices, among others. This type of metrics files are created by users, or generated by user data processing code, and can be defined in dvc.yaml (plots field) for tracking (optional). Refer to the DVC plots documentation for details on how to add plots to your repositories.

Types of plots

DVC Studio can work with two types of plots files in your repository:

1. Data series files, which can be JSON, YAML, CSV or TSV. Data from these files will populate your AUC curves, loss functions and other metric plots.
2. Image files in JPEG, GIF, or PNG format. These images will be displayed directly in DVC Studio.

You can define multiple plots in a single repository. Below is an example snippet from a dvc.yaml file showing the evaluate stage of the DVC pipeline.

evaluate:
  cmd: python src/evaluate.py
  deps:
    - output/data.pkl
    - output/model.h5
    - src/evaluate.py
  metrics:
    - output/metrics.json:
        cache: false
  plots:
    - output/predictions.json:
        cache: false
        template: confusion
        x: actual
        y: predicted
    - output/misclassified_samples/:
        cache: false

As you can see,

• metrics from output/predictions.json will be rendered in a confusion matrix,
• images in the output/misclassified_samples/ directory will be displayed directly.

You can also specify a single image file (eg, output/misclassified_sample1.png).

How to generate plots

To generate the plots, select one or more experiments (represented by the commits), and click on the Show plots button.

The plots will appear in the plots pane. If you have selected more than one experiment, you can use the plots to compare them.

Generate trend charts

Click on the Trends button to generate a plot of how the metrics changed over the course of the different experiments. For each metric, the trend charts show how the metric changed from one commit to another. You can include one or more branches in the trend chart.

Compare experiments

To compare different experiments, select two experiments (represented by the commits), and click on the Compare button. The metrics, parameters and files in the selected experiments will be displayed side by side for easy comparison.