Representation

class torchok.tasks.pairwise_task.PairwiseLearnTask(hparams: DictConfig, num_classes: int, backbone_name: str, pooling_name: Optional[str] = None, head_name: Optional[str] = None, neck_name: Optional[str] = None, backbone_params: Optional[dict] = None, neck_params: Optional[dict] = None, pooling_params: Optional[dict] = None, head_params: Optional[dict] = None, inputs: Optional[dict] = None)

Bases: ClassificationTask

Deep Metric Learning task for pairwise losses.

This task use ClassificationDataset in multilabel mode, to detect similar images. An example config for this task is torchok/examples/configs/pairwise_sop.yaml.

__init__(hparams: DictConfig, num_classes: int, backbone_name: str, pooling_name: Optional[str] = None, head_name: Optional[str] = None, neck_name: Optional[str] = None, backbone_params: Optional[dict] = None, neck_params: Optional[dict] = None, pooling_params: Optional[dict] = None, head_params: Optional[dict] = None, inputs: Optional[dict] = None)

Init PairwiseLearnTask.

Parameters

  • hparams – Hyperparameters that set in yaml file.

  • num_classes – number of all classes in train multilabel dataset.

  • backbone_name – name of the backbone architecture in the BACKBONES registry.

  • pooling_name – name of the backbone architecture in the POOLINGS registry.

  • head_name – name of the neck architecture in the HEADS registry.

  • neck_name – if present, name of the head architecture in the NECKS registry. Otherwise, model will be created without neck.

  • backbone_params – parameters for backbone constructor.

  • neck_params – parameters for neck constructor. in_channels will be set automatically based on backbone.

  • pooling_params – parameters for neck constructor. in_channels will be set automatically based on neck or backbone if neck is absent.

  • head_params – parameters for head constructor. in_channels will be set automatically based on neck.

  • inputs – information about input model shapes and dtypes.

forward_with_gt(batch: Dict[str, Tensor]) → Dict[str, Tensor]

Forward with ground truth labels.

Parameters

batch –

Dictionary with the following keys and values:

  • image (torch.Tensor):

    tensor of shape (B, C, H, W), representing input images.

  • target (torch.Tensor):

    tensor of shape (B), target class or labels per each image.

Returns

Dictionary with the following keys and values

  • ’emb1’: model forward method output.

  • ’emb2’: model forward method output same as emb1.

  • ’target’: target value batch if key target in batch, otherwise this key not in output dictionary.

  • ’R’: calculated relevance matrix if key target in batch, otherwise this key not in output dictionary.

calc_relevance_matrix(y: Tensor) → Tensor

Calculates binary relevance matrix given multi-label matrix y.

Parameters

  • y – Multi-label matrix of shape (N, L) representing labels for N samples, where L - number of classes.

  • 1 (Values are either 0 or) –

  • y1[i (where) –

  • class. (k] = 1 indicate that i-th sample belongs to k-th) –

Returns

Binary relevance matrix R of shape (N, M) where R[i, j] = 1 means that samples i and j are relevant to each other, dtype=float32.

class torchok.tasks.triplet_task.TripletLearnTask(hparams: DictConfig, **kwargs)

Bases: ClassificationTask

A class for triplet learning task.

__init__(hparams: DictConfig, **kwargs)

Init TripletLearnTask.

Parameters

hparams – Hyperparameters that set in yaml file.

forward_with_gt(batch: Dict[str, Union[Tensor, int]]) → Dict[str, Tensor]

Forward with ground truth labels.

validation_step(batch: Dict[str, Union[Tensor, int]], batch_idx: int) → Dict[str, Tensor]

Complete validation loop.