DVCLive

DVCLive is a Python library for logging machine learning metrics and other metadata in simple file formats, which is fully compatible with DVC.

Documentation

• Get Started
• How it Works
• API Reference
• Integrations

---

Quickstart

| Python API Overview | PyTorch Lightning | Scikit-learn | Ultralytics YOLO v8 | |--------|--------|--------|--------| | <a href="https://colab.research.google.com/github/iterative/dvclive/blob/main/examples/DVCLive-Quickstart.ipynb"><img src="https://colab.research.google.com/assets/colab-badge.svg" /></a> | <a href="https://colab.research.google.com/github/iterative/dvclive/blob/main/examples/DVCLive-PyTorch-Lightning.ipynb"><img src="https://colab.research.google.com/assets/colab-badge.svg" /></a> | <a href="https://colab.research.google.com/github/iterative/dvclive/blob/main/examples/DVCLive-scikit-learn.ipynb"><img src="https://colab.research.google.com/assets/colab-badge.svg" /></a> | <a href="https://colab.research.google.com/github/iterative/dvclive/blob/main/examples/DVCLive-YOLO.ipynb"><img src="https://colab.research.google.com/assets/colab-badge.svg" /></a> |

Install dvclive

$ pip install dvclive

Initialize DVC Repository

$ git init
$ dvc init
$ git commit -m "DVC init"

Example code

Copy the snippet below into train.py for a basic API usage example:

import time
import random

from dvclive import Live

params = {"learning_rate": 0.002, "optimizer": "Adam", "epochs": 20}

with Live() as live:

    # log a parameters
    for param in params:
        live.log_param(param, params[param])

    # simulate training
    offset = random.uniform(0.2, 0.1)
    for epoch in range(1, params["epochs"]):
        fuzz = random.uniform(0.01, 0.1)
        accuracy = 1 - (2 ** - epoch) - fuzz - offset
        loss = (2 ** - epoch) + fuzz + offset

        # log metrics to studio
        live.log_metric("accuracy", accuracy)
        live.log_metric("loss", loss)
        live.next_step()
        time.sleep(0.2)

See Integrations for examples using DVCLive alongside different ML Frameworks.

Running

Run this a couple of times to simulate multiple experiments:

$ python train.py
$ python train.py
$ python train.py
...

Comparing

DVCLive outputs can be rendered in different ways:

DVC CLI

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.accuracy   train.loss   val.accuracy   val.loss   step   epochs
─────────────────────────────────────────────────────────────────────────────────────────────────────────────
workspace                  -                  6.0109      0.23311          6.062    0.24321      6   7
master                     08:50 PM                -            -              -          -      -   -
├── 4475845 [aulic-chiv]   08:56 PM           6.0109      0.23311          6.062    0.24321      6   7
├── 7d4cef7 [yarer-tods]   08:56 PM           4.8551      0.82012         4.5555   0.033533      4   5
└── d503f8e [curst-chad]   08:56 PM           4.9768     0.070585         4.0773    0.46639      4   5
─────────────────────────────────────────────────────────────────────────────────────────────────────────────

$ dvc plots diff $(dvc exp list --names-only) --open

DVC Extension for VS Code

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

While experiments are running, live updates will be displayed in both views.

DVC Studio

If you push the results to DVC Studio, you can compare experiments against the entire repo history:

You can enable Studio Live Experiments to see live updates while experiments are running.

---

Comparison to related technologies

DVCLive is an ML Logger, similar to:

• MLFlow
• Weights & Biases
• Neptune

The main differences with those ML Loggers are:

• DVCLive does not require any additional services or servers to run.
• DVCLive metrics, parameters, and plots are stored as plain text files that can be versioned by tools like Git or tracked as pointers to files in DVC storage.
• DVCLive can save experiments or runs as hidden Git commits.

You can then use different options to visualize the metrics, parameters, and plots across experiments.

---

Contributing

Contributions are very welcome. To learn more, see the Contributor Guide.

License

Distributed under the terms of the Apache 2.0 license, dvclive is free and open source software.