Edit on GitHub

Get Started: Experiment Tracking

Tools like Jupyter Notebooks are useful for rapid prototyping, but it's hard to keep track of changes and reproduce experiments. You can start using DVC to version your experiments without leaving your Jupyter Notebook. There are no logins, servers, databases, or UI to spin up. Every DVC experiment will be versioned without cluttering your repo, unlike saving each run to a separate directory or creating a Git branch for each.


All you need to start is a DVC repository and the DVCLive Python library installed:

$ pip install dvclive

In your Python code, you can start versioning your experiments in DVCLive's Live API or framework-specific callbacks with save_dvc_exp=True.

There are some examples below (other frameworks available):

from dvclive import Live
from dvclive.lightning import DVCLiveLogger

with Live(save_dvc_exp=True) as live:
    checkpoint = ModelCheckpoint(dirpath="mymodel")
    trainer = Trainer(
from dvclive import Live
from dvclive.huggingface import DVCLiveCallback

with Live(save_dvc_exp=True) as live:
        DVCLiveCallback(save_dvc_exp=True, live=live)
    live.log_artifact("mymodel", type="model")
from dvclive import Live
from dvclive.keras import DVCLiveCallback

with Live(save_dvc_exp=True) as live:
            DVCLiveCallback(save_dvc_exp=True, live=live)
    live.log_artifact("mymodel", type="model")
from dvclive import Live

with Live(save_dvc_exp=True) as live:
    live.log_param("epochs", NUM_EPOCHS)

    for epoch in range(NUM_EPOCHS):
        metrics = evaluate_model(...)
        for metric_name, value in metrics.items():
            live.log_metric(metric_name, value)

    live.log_artifact("model.pkl", type="model")

After this, each execution of the code will create a DVC experiment containing the results and the changes needed to reproduce it.

DVCLive will automatically log some metrics, parameters and plots from the ML Framework and any data tracked by DVC but you can also log additional info to be included in the experiment. live.log_artifact("mymodel", type="model") will track your model with DVC and enable managing it with Studio Model Registry.

Learn more about how DVCLive works


By following the steps above, you enable different options to monitor the training progress:

By default, DVCLive will generate or update a report displaying all the logged data.

If you pass report="notebook" to DVCLive, the report will be displayed and updated inside the output of the cell:

Notebook report

The DVC Extension for VS Code will also display all the data logged by DVCLive:

VS Code Report

If you want to share live updates with others or monitor while away from your machine, follow the instructions in Studio Live Experiments to display updates in the Studio web interface:

Studio Report


After you have run multiple experiments, you can compare the results:

You can use dvc exp show and dvc plots to compare and visualize metrics, parameters and plots across experiments.

$ dvc exp show
Experiment                 Created    train.loss   eval.loss   dice_multi   base_lr
workspace                  -            0.024942    0.013983        0.922   0.001
master                     05:26 PM      0.78426    0.054157      0.49599   0.1
├── 950c3b5 [bifid-says]   05:33 PM     0.024942    0.013983        0.922   0.001
├── 06090d7 [potty-sash]   05:31 PM     0.026193    0.015237      0.91494   0.01
└── d1ad0a9 [soupy-leak]   05:28 PM     0.075223    0.034786      0.49596   0.1
$ dvc plots diff $(dvc exp list --name-only)

plots diff

Inside the DVC Extension for VS Code, you can compare and visualize results using the Experiments and Plots views.

VS Code Comparison

Once you have shared the results to Studio, you can compare experiments against the entire repo history:

Studio view

Learn more about Comparing Experiments


🐛 Found an issue? Let us know! Or fix it:

Edit on GitHub

Have a question? Join our chat, we will help you:

Discord Chat