ClearML

Adding Task.init to an existing training script is enough for ClearML to start recording a run: console output, TensorBoard and Matplotlib plots, saved model files, GPU and CPU load, hyperparameters parsed from argparse, Click, or Hydra, the package list, and the git commit plus any uncommitted diff. Those records go to a ClearML server, either the free hosted tier or one you deploy yourself with Docker or Kubernetes, where a web UI compares runs side by side. Beyond tracking, the stack includes clearml-agent for queuing jobs onto remote GPU machines and cloning past experiments with changed settings, a pipeline controller, a dataset versioning CLI on top of S3, GCS, Azure, or NAS storage, hyperparameter optimization, and a sibling clearml-serving project built around Triton for model endpoints. Framework hooks cover PyTorch, TensorFlow, Keras, XGBoost, scikit-learn, and more. The Python SDK and agent are Apache 2.0, while the self-hosted server ships under the Server Side Public License, which matters if you plan to sell it as a service. Fits ML teams wanting an all-in-one MLOps stack on their own hardware.