ml-experiment-tracker

Use mlflow.set_experiment to create an experiment, then mlflow.start_run to log parameters with log_param, metrics with log_metrics, and models with framework-specific log_model calls. Enable mlflow.autolog for automatic logging in supported frameworks.

MLflow provides a self-hostable tracking server with a built-in model registry and stage transitions. Weights & Biases offers hosted dashboards, Bayesian hyperparameter sweeps, and rich media logging like tables, images, and confusion matrices.

Run dvc init in your git repository, then dvc add on data files to create .dvc pointer files committed to git. Push actual data to remote storage with dvc remote add and dvc push, and define pipelines in dvc.yaml.

Set seeds for random, NumPy, and torch including torch.cuda.manual_seed_all, then enable cudnn deterministic mode and disable benchmark. Also log the git commit hash, dependency versions, and all hyperparameters for each run.

Define a sweep configuration with a search method like bayes, a target metric, and parameter distributions, then create it with wandb.sweep. Launch agents with wandb.agent to execute training runs that sample configurations automatically.

Common causes include unset random seeds, unversioned datasets, unpinned dependencies, and unlogged hyperparameters. Verify the git commit, data hash, environment lockfile, and seed configuration match the original run.