lenskit.training.TrainingOptions#

class lenskit.training.TrainingOptions#

Options and context settings that govern model training.

retrain: bool = True#

Whether the model should retrain if it is already trained. If False, the component is allowed to skip training if it is already trained.

In the common case of training pipelines, this flag is examined by lenskit.pipeline.Pipeline.train(): if it is False, that method skips training any components that are already trained. Custom training code that wishes to avoid retraining models should check Trainable.is_trained() instead of assuming that individual components will respect this flag.

Note

This division of responsibility is to reduce the need for repetitive code: since implementing components seems to be a more common activity than logic that directly trains components (as opposed to pipelines) in ordinary LensKit use, making training code responsible for skipping retrain instead of requiring that of every component implementation allows individual implementations to be slightly simpler, without requiring separate options classes for pipeline and component training.

Changed in version 2026.1: Added the is_trained() method that implementers must now also provide.

device: str | None = None#: The device on which to train (e.g. 'cuda'). May be ignored if the model does not support the specified device.

rng: lenskit.random.RNGInput = None#: Random number generator to use for any randomness in the training process. This option contains any `SPEC 7`_-compatible random number generator specification; the random_generator() will convert that into a NumPy Generator.

environment: dict[str, str]#: Additional training environment variables to control training behavior. Variables and their meanings are defined by individual components. Variables in this option override system environment variables when fetched with envvar().

torch_profiler: torch.profiler.profile | None = None#: Torch profiler for profiling training options.

step_profiler()#: Signal to active profiler(s) that a new step has completed.

random_generator(*, type: Literal['numpy'] = 'numpy') → numpy.random.Generator#
random_generator(*, type: Literal['torch']) → torch.Generator: Obtain a random generator from the configured RNG or seed.

Note

Each call to this method will return a fresh generator from the same seed. Components should call it once at the beginning of their training procesess.

configured_device(*, gpu_default=False)#

Get the configured device, consulting environment variables and defaults if necessary. It looks for a device in the following order:

The device, if specified on this object.
The LK_DEVICE environment variable.
If CUDA is enabled and gpu_default is True, return “cuda”
The CPU.

Parameters:: gpu_default (bool) – Whether a CUDA GPU should be preferred if it is available and no device has been specified.
Return type:: str

env_var(name: str, default: str) → str#

env_var(name: str, default: str | None = None) → str | None

Fetch a training environment variable. Variables are first looked up in environment, then in os.environ.

lenskit.training.TrainingOptions#

This Page