DVCLive Documentation

This documentation provides the details about the dvclive Python API module, which can be imported regularly, for example:

from dvclive import Live

Basic Workflow

Initialize DVCLive

with Live() as live:

See Live() for details.

Log data

live.log_artifact("model.pt", type="model", name="gpt")

img = np.ones((500, 500, 3), np.uint8)
live.log_image("image.png", img)

live.log_metric("acc", 0.9)

params = {
    "num_classes": 10,
    "metrics": ["accuracy", "mae"],
    "optimizer": "adam"
}
live.log_params(params)

datapoints = [
    {"name": "petal_width", "importance": 0.4},
    {"name": "petal_length", "importance": 0.33},
    {"name": "sepal_width", "importance": 0.24},
    {"name": "sepal_length", "importance": 0.03}
]
live.log_plot(
    "iris", datapoints, x="importance", y="name",
    template="bar_horizontal", title="Iris Feature Importance"
)

y_true = [0, 0, 1, 1]
y_pred = [0.2, 0.5, 0.3, 0.8]
live.log_sklearn_plot("roc", y_true, y_score)

(Optionally) Update the step number

live.next_step()

See Live.next_step().

Under the hood, Live.next_step() calls Live.make_summary(), Live.make_dvcyaml(), and Live.make_report().

When access is enabled, updates will be sent to DVC Studio.

If you want to decouple the step update from the rest of the calls, you can manually modify the Live.step property and call Live.make_summary() / Live.make_dvcyaml() / Live.make_report().

Putting it all together

Joining the above snippets, you can include DVCLive in your training code:

# train.py

from dvclive import Live

with Live() as live:

    live.log_param("epochs", NUM_EPOCHS)

    for epoch in range(NUM_EPOCHS):
        train_model(...)
        metrics = evaluate_model(...)

        for metric_name, value in metrics.items():
            live.log_metric(metric_name, value)

        live.next_step()

    live.log_artifact(path, type="model", name=name)

Outputs

After you run your training code, all the logged data will be stored in the dvclive directory and tracked as a DVC experiment for analysis and comparison.

Run with DVC

Experimenting in Python interactively (like in notebooks) is great for exploration, but eventually you may need a more structured way to run reproducible experiments. By configuring DVC pipelines, you can run experiments with dvc exp run. Pipelines help you organize your ML workflow beyond a single notebook or script so you can modularize and parametrize your code. See how to setup a pipeline to work with DVCLive.