ml-experiment-tracking

Use mlflow.start_run() as a context manager around training code, then call mlflow.log_params for hyperparameters, mlflow.log_metrics for evaluation scores, and mlflow.log_model to store the trained model. Set the tracking URI and experiment name before starting runs.

MLflow is self-hosted and open source with a built-in model registry supporting Staging and Production stages. Weights & Biases is a managed service with artifact lineage and collaborative dashboards. Both log params, metrics, and artifacts with similar APIs.

Call mlflow.log_metrics with a step argument inside the training loop, passing the epoch number as step. This records training curves over time rather than only final metrics, which helps identify overfitting early.

Register the model from a run using mlflow.register_model with the run URI, then use MlflowClient.transition_model_version_stage to move it to Staging and later Production. Archive existing production versions instead of deleting them for rollback.

Runs become irreproducible when random seeds, dataset versions, or code versions are not logged. Fix seeds for numpy, torch, and random, log the dataset hash and git commit with each run, and store requirements.txt as an artifact.