opr.miners package

Module for miners.

opr.miners.batch_hard_triplet_miner

Batch hard triplet miner implementation.

Code adopted from repository: https://github.com/jac99/MinkLocMultimodal, MIT License

class opr.miners.batch_hard_triplet_miner.BatchHardTripletMiner(distance: BaseDistance)[source]

Bases: Module

Batch hard triplet miner.

Original idea is taken from repository: https://github.com/jac99/MinkLocMultimodal, MIT License

forward(embeddings: Tensor, positives_mask: Tensor, negatives_mask: Tensor) → Tuple[Tensor, Tensor, Tensor][source]

Mine hard triplets from given batch of embeddings. For each element in batch triplet will be mined.

Parameters:

  • embeddings (Tensor) – Model output embeddings.

  • positives_mask (Tensor) – Binary mask of positive elements in batch.

  • negatives_mask (Tensor) – Binary mask of negative elements in batch.

Returns:

Hard triplets tuple (a, p, n).

Return type:

Tuple[Tensor, Tensor, Tensor]

property stats: Dict[str, float]

Return statistics of last forward pass.