cosense3d.modules.losses package

Submodules

cosense3d.modules.losses.base_loss module

class cosense3d.modules.losses.base_loss.BaseLoss(reduction: str = 'mean', activation: str = 'none', loss_weight: float = 1.0)

Bases: Module

forward(preds: Tensor, targets: Tensor, weight: Tensor | None = None, avg_factor: int | None = None, reduction_override: str | None = None, *args, **kwargs) → Tensor

Parameters:

• preds – prediction tensor.

• targets – target tensor.

• weight – The weight of loss for each prediction. Defaults to None.

• avg_factor – Average factor that is used to average the loss. Defaults to None.

• reduction_override – The reduction method used to override the original reduction method of the loss. Defaults to None.

• args – additional arguments.

• kwargs –

Returns:

weighted loss.

loss(*args, **kwargs)

property name

training: bool

cosense3d.modules.losses.common module

cosense3d.modules.losses.common.cross_entroy_with_logits(preds, tgts, n_cls, weights=None, reduction='none')

cosense3d.modules.losses.common.focal_loss(preds, tgts, weights=None, reduction='none', gamma=2.0, alpha=0.25, use_sigmoid=True)

Parameters

preds: FloatTensor(…, n_cls) tgts: FloatTensor(…, n_cls) or LongTensor(…,) or LongTensor(…,1), largest label is background weights: same as preds or tgts ——-

cosense3d.modules.losses.common.indices_to_dense_vector(indices: Tensor, size: int, indices_value: float = 1.0, default_value: float = 0.0) → Tensor

Creates dense vector with indices set to specific value and rest to zeros.

This function exists because it is unclear if it is safe to use tf.sparse_to_dense(indices, [size], 1, validate_indices=False) with indices which are not ordered. This function accepts a dynamic size (e.g. tf.shape(tensor)[0])

Parameters:

• indices – 1d Tensor with integer indices which are to be set to indices_values.

• size – size of output Tensor.

• indices_value – values of elements specified by indices in the output vector.

• default_value – values of other elements in the output vector.

Returns:

dense 1D Tensor of shape [size] with indices set to indices_values and the rest set to default_value.

cosense3d.modules.losses.common.sigmoid_binary_cross_entropy(preds, tgts, weights=None, reduction='none')

Parameters

preds: Tensor(d1, …, dn) tgts: Tensor(d1, …, dn) weights. Tensor(d1, …, dn) reduction: str(‘none’ | ‘mean’ | ‘sum’) ——-

cosense3d.modules.losses.common.weighted_l1_loss(preds, targets, sigma=3.0, weights=None)

cosense3d.modules.losses.common.weighted_sigmoid_binary_cross_entropy(preds, tgts, weights=None, class_indices=None)

cosense3d.modules.losses.common.weighted_smooth_l1_loss(preds, targets, sigma=3.0, weights=None)

cosense3d.modules.losses.edl module

class cosense3d.modules.losses.edl.EDLLoss(n_cls: int, annealing_step: int, **kwargs)

Bases: BaseLoss

loss(preds, tgt, temp, n_cls_override=None)

training: bool

cosense3d.modules.losses.edl.edl_mse_loss(preds, tgt, n_cls, temp, annealing_step, model_label='edl')

Calculate evidential loss :param model_label: (str) a name to distinguish edl loss of different modules :param preds: (N, n_cls) the logits of each class :param tgt: (N,) labels with values from 0…(n_cls - 1) or (N, n_cls) :param n_cls: (int) number of classes, including background :param temp: current temperature for annealing of KL Divergence term of the loss :param annealing_step: maximum annealing step :return:

cosense3d.modules.losses.edl.evidence_to_conf_unc(evidence, edl=True)

cosense3d.modules.losses.edl.exp_evidence(y)

cosense3d.modules.losses.edl.kl_divergence(alpha, num_classes)

cosense3d.modules.losses.edl.loglikelihood_loss(y, alpha)

cosense3d.modules.losses.edl.mse_loss(y, alpha, epoch_num, num_classes, annealing_step)

cosense3d.modules.losses.edl.pred_to_conf_unc(preds, activation='relu', edl=True)

cosense3d.modules.losses.edl.relu_evidence(y)

cosense3d.modules.losses.edl.softplus_evidence(y)

cosense3d.modules.losses.focal_loss module

class cosense3d.modules.losses.focal_loss.FocalLoss(use_sigmoid: bool = True, gamma: float = 2.0, alpha: float = 0.25, activated: bool = False, bg_idx: int | None = None, **kwargs)

Bases: BaseLoss

loss(pred: Tensor, target: Tensor, *args, **kwargs)

Parameters:

• pred – prediction.

• target – ground truth targets.

• args –

• kwargs –

Returns:

training: bool

class cosense3d.modules.losses.focal_loss.GaussianFocalLoss(alpha: float = 2.0, gamma: float = 4.0, reduction: str = 'mean', loss_weight: float = 1.0)

Bases: BaseLoss

GaussianFocalLoss is a variant of focal loss.

More details can be found in the paper Code is modified from kp_utils.py # noqa: E501 Please notice that the target in GaussianFocalLoss is a gaussian heatmap, not 0/1 binary target.

loss(pred: Tensor, target: Tensor)

Focal Loss for targets in gaussian distribution.

Parameters:

• pred – The prediction.

• target – The learning target of the prediction in gaussian distribution.

Returns:

loss result.

training: bool

class cosense3d.modules.losses.focal_loss.QualityFocalLoss(use_sigmoid: bool = True, beta: float = 2.0, activated: bool = False, **kwargs)

Bases: BaseLoss

loss(pred: Tensor, target: Tensor)

Forward function.

Parameters:

• pred – Predicted joint representation of classification and quality (IoU) estimation with shape (N, C), C is the number of classes.

• target – Target category label with shape (N,) and target quality label with shape (N,).

Returns:

loss result.

training: bool

cosense3d.modules.losses.focal_loss.py_focal_loss_with_prob(pred: Tensor, target: Tensor, gamma: float = 2.0, alpha: float = 0.25)

PyTorch version of Focal Loss. Different from py_sigmoid_focal_loss, this function accepts probability as input.

Parameters:

• pred – The prediction probability with shape (N, C), C is the number of classes.

• target – The learning label of the prediction.

• gamma – The gamma for calculating the modulating factor. Defaults to 2.0.

• alpha – A balanced form for Focal Loss. Defaults to 0.25.

Returns:

loss result.

cosense3d.modules.losses.focal_loss.py_sigmoid_focal_loss(pred: Tensor, target: Tensor, gamma: float = 2.0, alpha: float = 0.25)

PyTorch version of Focal Loss. Different from py_sigmoid_focal_loss, this function accepts probability as input.

Parameters:

• pred – The prediction probability with shape (N, C), C is the number of classes.

• target – The learning label of the prediction.

• gamma – The gamma for calculating the modulating factor. Defaults to 2.0.

• alpha – A balanced form for Focal Loss. Defaults to 0.25.

Returns:

loss result.

cosense3d.modules.losses.focal_loss.quality_focal_loss(pred: ~torch.Tensor, target: (<class 'torch.Tensor'>, ), beta: float = 2.0) → Tensor

Quality Focal Loss (QFL) is from Generalized Focal Loss: Learning Qualified and Distributed Bounding Boxes for Dense Object Detection.

Parameters:

• pred – Predicted joint representation of classification and quality (IoU) estimation with shape (N, C), C is the number of classes.

• target – Target category label with shape (N,) and target quality label with shape (N,).

• beta – The beta parameter for calculating the modulating factor. Defaults to 2.0.

Returns:

Loss tensor with shape (N,).

cosense3d.modules.losses.focal_loss.quality_focal_loss_with_prob(pred: ~torch.Tensor, target: (<class 'torch.Tensor'>, ), beta: float = 2.0) → Tensor

Quality Focal Loss (QFL) is from Generalized Focal Loss: Learning Qualified and Distributed Bounding Boxes for Dense Object Detection.

Parameters:

• pred – Predicted joint representation of classification and quality (IoU) estimation with shape (N, C), C is the number of classes.

• target – Target category label with shape (N,) and target quality label with shape (N,).

• beta – The beta parameter for calculating the modulating factor. Defaults to 2.0.

Returns:

Loss tensor with shape (N,).

cosense3d.modules.losses.iou_loss module

class cosense3d.modules.losses.iou_loss.GIoULoss(eps: float = 1e-07, **kwargs)

Bases: BaseLoss

loss(pred, target)

training: bool

class cosense3d.modules.losses.iou_loss.IoULoss(mode: str = 'log', eps: float = 1e-06, **kwargs)

Bases: BaseLoss

loss(pred, target)

training: bool

cosense3d.modules.losses.l1_loss module

class cosense3d.modules.losses.l1_loss.L1Loss(reduction: str = 'mean', activation: str = 'none', loss_weight: float = 1.0)

Bases: BaseLoss

loss(pred, target)

training: bool

class cosense3d.modules.losses.l1_loss.SmoothL1Loss(beta: float = 1.0, **kwargs)

Bases: BaseLoss

loss(pred, target)

training: bool

cosense3d.modules.losses.vanilla_seg_loss module

class cosense3d.modules.losses.vanilla_seg_loss.VanillaSegLoss(d_weights, s_weights, d_coe, s_coe, l_weights=50, **kwargs)

Bases: Module

forward(static_pred=None, dynamic_pred=None, static_gt=None, dynamic_gt=None)

Perform loss function on the prediction.

Parameters

output_dictdict

The dictionary contains the prediction.

gt_dictdict

The dictionary contains the groundtruth.

Returns

Loss dictionary.

training: bool

Module contents

cosense3d.modules.losses.build_loss(type, **kwargs)