voronoi

Functions:

Name	Description
`compute_voronoi_l2_distance`	Compute the Voronoi l2 distance is defined as half the distance between the embedding and the nearest neighbor in the embedding table.
`compute_voronoi_l2_distance_percentile`	Compute the Voronoi l2 distance is defined as the distance between the embedding and the nearest neighbor in the embedding table.

compute_voronoi_l2_distance ¶

compute_voronoi_l2_distance(
    embeddings_table_matrix: Tensor,
) -> torch.Tensor

Compute the Voronoi l2 distance is defined as half the distance between the embedding and the nearest neighbor in the embedding table.

Parameters:

Name	Type	Description	Default
`embeddings_table_matrix` ¶	`Tensor`	The weight of the embedding table.	required

Returns:

Type	Description
`torch.Tensor`	The Voronoi l2 distance matrix.

Added in version v1.4.0. To facilitate the computation of the Voronoi l2 distance.

compute_voronoi_l2_distance_percentile(
    embeddings_table_matrix: Tensor, percentile: float
) -> torch.Tensor

Compute the Voronoi l2 distance is defined as the distance between the embedding and the nearest neighbor in the embedding table.

Due to the memory intensive nature of calculating the quantiles of all the embeddings, we compute the distances on the GPU and

then transfer the distances to the CPU for quantile calculation.

Parameters:

Name	Type	Description	Default
`embeddings_table_matrix` ¶	`Tensor`	The weight of the embedding table.	required
`percentile` ¶	`float`	The percentile to compute.	required

Returns:

Type	Description
`torch.Tensor`	The Voronoi l2 distance at the given percentile.

Added in version v1.4.0. To facilitate the computation of the Voronoi l2 distance.