Retrieval Post-Processing

IDistancesPostprocessor

class oml.interfaces.retrieval.IDistancesPostprocessor

Bases: object

This is a parent class for the classes which apply some postprocessing after query-to-gallery distance matrix has been calculated. For example, we may want to apply one of re-ranking techniques.

process(distances: Tensor, queries: Any, galleries: Any) → Tensor

This method takes all the needed variables and returns the modified matrix of distances, where some distances are replaced with new ones.

Parameters

• distances – Matrix with the shape of [Q, G]

• queries – Queries in the amount of Q

• galleries – Galleries in the amount of G

Returns

An updated distances matrix with the shape of [Q, G]

PairwisePostprocessor

class oml.retrieval.postprocessors.pairwise.PairwisePostprocessor

Bases: IDistancesPostprocessor, ABC

This postprocessor allows us to re-estimate the distances between queries and top-n galleries closest to them. It creates pairs of queries and galleries and feeds them to a pairwise model.

process(distances: Tensor, queries: Any, galleries: Any) → Tensor

Parameters

• distances – Matrix with the shape of [Q, G]

• queries – Queries in the amount of Q

• galleries – Galleries in the amount of G

Returns

Distance matrix with the shape of [Q, G],

where top_n minimal values in each row have been updated by the pairwise model, other distances are shifted by a margin to keep the relative order.

inference(queries: Any, galleries: Any, ii_top: Tensor, top_n: int) → Tensor

Depends on the exact types of queries/galleries this method may be implemented differently.

Parameters

• queries – Queries in the amount of Q

• galleries – Galleries in the amount of G

• ii_top – Indices of the closest galleries with the shape of [Q, top_n]

• top_n – Number of the closest galleries to re-rank

Returns

An updated distance matrix with the shape of [Q, G]

PairwiseEmbeddingsPostprocessor

class oml.retrieval.postprocessors.pairwise.PairwiseEmbeddingsPostprocessor(top_n: int, pairwise_model: IPairwiseModel, num_workers: int, batch_size: int, verbose: bool = False, use_fp16: bool = False, is_query_key: str = 'is_query', is_gallery_key: str = 'is_gallery', embeddings_key: str = 'embeddings')

Bases: PairwisePostprocessor

__init__(top_n: int, pairwise_model: IPairwiseModel, num_workers: int, batch_size: int, verbose: bool = False, use_fp16: bool = False, is_query_key: str = 'is_query', is_gallery_key: str = 'is_gallery', embeddings_key: str = 'embeddings')

Parameters

• top_n – Model will be applied to the num_queries * top_n pairs formed by each query and top_n most relevant galleries.

• pairwise_model – Model which is able to take two embeddings as inputs and estimate the distance (not in a strictly mathematical sense) between them.

• num_workers – Number of workers in DataLoader

• batch_size – Batch size that will be used in DataLoader

• verbose – Set True if you want to see progress bar for an inference

• use_fp16 – Set True if you want to use half precision

• is_query_key – Key to access a binary mask indicates queries in case of using process_by_dict

• is_gallery_key – Key to access a binary mask indicates galleries in case of using process_by_dict

• embeddings_key – Key to access embeddings in case of using process_by_dict

inference(queries: Tensor, galleries: Tensor, ii_top: Tensor, top_n: int) → Tensor

Parameters

• queries – Queries representations with the shape of [Q, *]

• galleries – Galleries representations with the shape of [G, *]

• ii_top – Indices of the closest galleries with the shape of [Q, top_n]

• top_n – Number of the closest galleries to re-rank

Returns

Updated distance matrix with the shape of [Q, G]

PairwiseImagesPostprocessor

class oml.retrieval.postprocessors.pairwise.PairwiseImagesPostprocessor(top_n: int, pairwise_model: IPairwiseModel, transforms: Union[Compose, Compose], num_workers: int = 0, batch_size: int = 128, verbose: bool = True, use_fp16: bool = False, is_query_key: str = 'is_query', is_gallery_key: str = 'is_gallery', paths_key: str = 'paths')

Bases: PairwisePostprocessor

__init__(top_n: int, pairwise_model: IPairwiseModel, transforms: Union[Compose, Compose], num_workers: int = 0, batch_size: int = 128, verbose: bool = True, use_fp16: bool = False, is_query_key: str = 'is_query', is_gallery_key: str = 'is_gallery', paths_key: str = 'paths')

Parameters

• top_n – Model will be applied to the num_queries * top_n pairs formed by each query and its top_n most relevant galleries.

• pairwise_model – Model which is able to take two images as inputs and estimate the distance (not in a strictly mathematical sense) between them.

• transforms – Transforms that will be applied to an image

• num_workers – Number of workers in DataLoader

• batch_size – Batch size that will be used in DataLoader

• verbose – Set True if you want to see progress bar for an inference

• use_fp16 – Set True if you want to use half precision

• is_query_key – Key to access a binary mask indicates queries in case of using process_by_dict

• is_gallery_key – Key to access a binary mask indicates galleries in case of using process_by_dict

• paths_key – Key to access paths to images in case of using process_by_dict

inference(queries: List[Path], galleries: List[Path], ii_top: Tensor, top_n: int) → Tensor

Parameters

• queries – Paths to queries with the length of Q

• galleries – Paths to galleries with the length of G

• ii_top – Indices of the closest galleries with the shape of [Q, top_n]

• top_n – Number of the closest galleries to re-rank

Returns

Updated distance matrix with the shape of [Q, G]