Skip to main content

Experiments

Record every training run from your Python code: the params it used, each metric at every step, the files it produced and the model it trained. Compare runs side by side, see their training curves, and register the best run's model in the Model Registry.

Tracking works from anywhere your code runs:

  • In a Strongly workspace or job. pip install strongly is already there and no key is needed: the platform signs every call in as you.
  • Outside the platform (a laptop, CI, another cloud). Set STRONGLY_API_KEY (a REST API key from Profile > Security > API Keys) and STRONGLY_API_URL (your platform's address, such as https://strongly.example.com).

Experiments and runs​

An experiment groups the runs of one modelling problem, such as "churn-model". A run is one training attempt in it: one set of params, its metrics and its files. A run can contain nested runs, for example one per fold of a cross-validation or one per trial of a search.

In this section​

  • Tracking runs: params, metrics at every step, artifacts of any size, models and nested runs.
  • Autolog: log training without logging code, for scikit-learn, XGBoost, LightGBM, PyTorch Lightning, Keras, Transformers and more.
  • Compare and register: the Experiments pages, sharing, registering a run's model, and the client API.

AutoML runs​

An AutoML job is tracked the same way: the job is a run, each model it trained is a run in it with its params and metrics, and the job's leaderboard, feature importance and training summary are its artifacts. The job's run is running while the job trains and ends with it: completed, failed, or cancelled when the job is stopped.