quaterion.eval.samplers.base_sampler module

class BaseSampler(sample_size=-1, device: device | str | None = None, log_progress: bool = True)

Bases: object

Sample part of embeddings and targets to perform metric calculation on a part of the data

Sampler allows reducing amount of time and resources to calculate a distance matrix. Instead of calculation of squared matrix with shape (num_embeddings, num_embeddings), it selects embeddings and computes matrix of a rectangle shape.

Args:

sample_size: amount of objects to select.

reset()

Reset accumulated state if any

sample(dataset: Sized, metric: BaseMetric, model: SimilarityModel) → Tuple[Tensor, Tensor]

Sample objects and labels to calculate metrics

Parameters:

• dataset – Sized object, like list, tuple, torch.utils.data.Dataset, etc. to sample

• metric – metric instance to compute final labels representation

• model – model to encode objects

Returns:

Tensor, Tensor – metrics labels and computed distance matrix