DVCLive

What DVCLive is

DVCLive is a free, open source Python API module for recording machine learning parameters, metrics, artifacts, images, and plots. It saves run data in a `dvclive` directory and tracks it as a DVC experiment, so runs can be analyzed and compared. Its tracking workflow uses Git and DVC rather than a central database. For large artifacts, DVCLive can use DVC to cache the file and create a `.dvc` metadata file tracked in Git. Optional system monitoring records GPU, CPU, RAM, and disk usage. The listed framework integrations cover Fast.ai, Hugging Face, Keras, LightGBM, MMCV, Optuna, PyTorch, PyTorch Lightning, Scikit-learn, TensorFlow, XGBoost, and Ultralytics YOLO v8. Optional dependencies are available for image logging, plots, Scikit-learn plots, Markdown reports, and framework integrations. The latest version requires Python 3.9 or later. Each run is saved to the same path, overwriting results already stored there. DVCLive can also work with DVC pipelines for structured, reproducible machine learning experiments. DVC directs users to its Discord community, GitHub issues, and forum for help. lakeFS has taken over stewardship and active development of DVC and is committed to maintaining DVCLive; DVC says the project will remain open source without paywalled features or restricted access.

Who it suits

DVCLive suits Python machine learning practitioners who want to log experiment data and compare runs through DVC and Git. It can also suit teams using its listed framework integrations or DVC pipelines, provided they account for its same-path overwrite behavior.

Limits

The latest version requires Python 3.9 or later. Each run uses the same path and overwrites previous results there; optional dependencies are needed for some logging, reporting, and framework integration features.

Getting started

1. Install Python 3.9 or later for the latest DVCLive version.
2. Add DVCLive to a Python machine learning workflow.
3. Install optional dependencies for the image, plot, report, or framework integration features you plan to use.
4. Track run data through Git and DVC.
5. Use a DVC pipeline if you want to run structured experiments.

Verdict

Choose DVCLive if you want a free Python library for recording and comparing machine learning runs within a Git and DVC workflow. Its framework integrations and pipeline use can support varied experiment setups. Plan your run paths carefully because each run overwrites prior results saved to the same location.