DVCLive

3.49.0 · active · verified Sat Apr 11

DVCLive is a Python library for logging machine learning metrics and other metadata. It is designed to be fully compatible with DVC (Data Version Control) and stores logged information in simple, human-readable file formats (like .tsv, .json, .yaml) that can be versioned by Git. It provides real-time experiment tracking and integrates with various ML frameworks, helping users maintain reproducible ML workflows. The current version is 3.49.0, and the library is actively developed with frequent releases.

Warnings

breaking Scikit-learn `probas_pred` argument change (v3.48.3 and `sklearn>=1.7`). Older DVCLive versions with `sklearn` 1.7+ might encounter `TypeError: missing a required argument: 'y_score'` when using `live.log_sklearn_plot()`. DVCLive 3.48.3 fixed this internally, so ensure your `dvclive` is updated if you use newer `sklearn`.
Fix: Upgrade DVCLive to version 3.48.3 or newer, or pin `scikit-learn` to a version older than 1.7.
breaking Dropped Catalyst ML framework integration (v3.47.0). Support for the Catalyst ML framework was removed in DVCLive 3.47.0.
Fix: Users relying on Catalyst integration must either pin `dvclive` to a version older than 3.47.0 or migrate their logging setup away from Catalyst's DVCLive callback.
gotcha Matplotlib `Figure` logging behavior change (v3.48.4). Previously, `live.log_image()` with a matplotlib figure might have implicitly logged the most recently active figure. Since 3.48.4, it strictly logs the `matplotlib.figure.Figure` instance explicitly provided as an argument. Make sure to pass the intended figure object.
Fix: Always explicitly pass the `matplotlib.figure.Figure` object you intend to log to `live.log_image()` rather than relying on global state.
gotcha DVC `live` section deprecation in `dvc.yaml` (DVC 3.0 / DVCLive ~3.0). The `live` section for DVCLive configuration in `dvc.yaml` was deprecated and is no longer the primary way to configure DVCLive. Configuration should primarily be done through the Python `Live` API.
Fix: Configure `dvclive.Live` instances directly in your Python code using its `__init__` parameters. Avoid relying on or defining the `live` section in `dvc.yaml`.
gotcha `save_dvc_exp` ignored in `dvc repro`. When `dvclive` runs as part of a `dvc repro` command, the `save_dvc_exp=True` argument to `Live()` is ignored. DVC experiments will not be automatically saved by `dvclive` in this context.
Fix: To explicitly save experiments when running within a DVC pipeline, use `dvc exp run` instead of `dvc repro`.

Install

pip install dvclive Core library
pip install dvclive[sklearn,image,huggingface,lightning,tf,fastai,optuna,xgb,lgbm,mmcv] With common ML framework integrations and features

Imports

Live
```
from dvclive import Live
```
The `dvclive.init()` function was an older pattern. The modern approach is to instantiate and use the `Live` class.

Quickstart

This quickstart demonstrates basic logging of parameters and metrics using `dvclive.Live` within a simulated training loop. Metrics and parameters will be saved in the `dvclive` directory, typically as `metrics.json`, `params.yaml`, and time-series `.tsv` files. Running this code multiple times will generate new experiment steps that can be tracked and compared with DVC.

import time
import random
from dvclive import Live

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

with Live() as live:
    # Log parameters
    for param in params:
        live.log_param(param, params[param])

    # Simulate training loop
    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 for the current step
        live.log_metric("accuracy", accuracy)
        live.log_metric("loss", loss)
        live.next_step()
        time.sleep(0.05) # Simulate work, shorten for quick demo

view raw JSON →