Ensembles¶

Ensembles take multiple models and combine them into a single model.

BaseEnsemble¶

from powerful_benchmarker.ensembles import BaseEnsemble
BaseEnsemble(normalize_embeddings=True, use_trunk_output=False)

normalize_embeddings: Perform L2 normalization if True. The specific details are determined by the child class.
use_trunk_output: Use the output of the trunk of the ensemble. The specific details are determined by the child class.

Loads models given a list of split scheme folders, and returns a list containing the loaded models.

get_list_of_models(model_factory, model_args, model_name, factory_kwargs, split_folders, device)

Returns a single trunk and embedder, given a list of trunks and embedders. Must be implemented by the child class.

create_ensemble_model(list_of_trunks, list_of_embedders)

Returns a trunk that outputs the concatenation of multiple trunk models.

Returns an embedder that outputs the concatenation of multiple embedder models.

The trunk's output can be passed into the embedder.

from powerful_benchmarker.ensembles import ConcatenateEmbeddings
ConcatenateEmbeddings(**kwargs)