Actions are a way to interact with your running experiment.

Typical use cases for actions include:

  • hand-tuning hyperparameters, such as learning rate, in the middle of training your machine learning algorithm;
  • on-demand saving snapshot of current model weights to disk.

You can register actions in your code using Neptune Python Client Library. You can invoke registered actions in running experiments via CLI or using Neptune UI.

To register an action for your experiment, use method register_action, for example:

def save_model(path):
  return 'Saved in {}'.format(path)

ctx.register_action(name="save model", handler=save_model)

name is used to identify the action. handler is the callback that will be called upon the action invocation. It should receive single unicode/str argument and return unicode/str as the result.