API References¶

DATA¶

Dataloaders¶

`SimilarityDataLoader`	Special version of `DataLoader` which works with similarity samples.
`PairsSimilarityDataLoader`	DataLoader designed to work with data represented as `SimilarityPairSample`.
`GroupSimilarityDataLoader`	DataLoader designed to work with data represented as `SimilarityGroupSample`.

Datasets¶

SimilarityGroupDataset

Wrapper, which converts standard dataset of classification task into dataset, compatible with GroupSimilarityDataLoader.

Samples¶

`SimilarityGroupSample`	Represent groups of similar objects all of which should match with one-another within the group.
`SimilarityPairSample`	Represents a pair of objects, their similarity and relationship with other pairs.

DISTANCES¶

`Cosine`	Compute cosine similarities (and its interpretation as distances).
`DotProduct`	Compute dot product similarities (and its interpretation as distances).
`Euclidean`	Compute Euclidean distances (and its interpretation as similarities).
`Manhattan`	Compute Manhattan distances (and its interpretation as similarities).

EVAL¶

Counters¶

`AttachedMetric`	Attach batch-wise metric to `TrainableModel`
`Evaluator`	Calculate metrics on the whole datasets

Group metrics¶

`GroupMetric`	Base class for group metrics
`RetrievalRPrecision`	Compute the retrieval R-precision score for group based data

Pair metrics¶

`PairMetric`	Base class for metrics computation for pair based data
`RetrievalPrecision`	Calculates retrieval precision@k for pair based datasets
`RetrievalReciprocalRank`	Calculates retrieval reciprocal rank for pair based datasets

Samplers¶

`GroupSampler`	Perform selection of embeddings and targets for group based tasks.
`PairSampler`	Perform selection of embeddings and targets for pairs based tasks.

LOSSES¶

Base¶

`GroupLoss`	Base class for group losses.
`PairwiseLoss`	Base class for pairwise losses.

Implementations¶

`ArcFaceLoss`	Additive Angular Margin Loss as defined in https://arxiv.org/abs/1801.07698
`ContrastiveLoss`	Contrastive loss.
`MultipleNegativesRankingLoss`	Implement Multiple Negatives Ranking Loss as described in https://arxiv.org/pdf/1705.00652.pdf
`SoftmaxLoss`	Regular cross-entropy loss.
`TripletLoss`	Implements Triplet Loss as defined in https://arxiv.org/abs/1503.03832
`CircleLoss`	Implements Circle Loss as defined in https://arxiv.org/abs/2002.10857.
`FastAPLoss`	FastAP Loss
`CosFaceLoss`	Large Margin Cosine Loss as defined in https://arxiv.org/pdf/1801.09414.pdf
`CenterLoss`	Center Loss as defined in the paper "A Discriminative Feature Learning Approach for Deep Face Recognition" (http://ydwen.github.io/papers/WenECCV16.pdf) It aims to minimize the intra-class variations while keeping the features of different classes separable.

Extras¶

PytorchMetricLearningWrapper

Provide a simple wrapper to be able to use losses and miners from pytorch-metric-learning.

MAIN¶

Quaterion

Fine-tuning entry point

TRAIN¶

TrainableModel¶

TrainableModel

Base class for models to be trained.

Cache¶

`CacheConfig`	Determine cache settings.
`CacheType`	Available tensor devices to be used for caching.

UTILS¶

`TrainStage`	Handle train stage.
`get_triplet_mask`	Creates a 3D mask of valid triplets for the batch-all strategy.
`get_anchor_positive_mask`	Creates a 2D mask of valid anchor-positive pairs.
`get_anchor_negative_mask`	Creates a 2D mask of valid anchor-negative pairs.