loss_fn package

Subpackages

• loss_fn.classification package
  • Submodules
  • loss_fn.classification.base_classification_criteria module
    • BaseClassificationCriteria
      • BaseClassificationCriteria.__init__()
      • BaseClassificationCriteria.add_arguments()
      • BaseClassificationCriteria.forward()
  • loss_fn.classification.binary_cross_entropy module
    • BinaryCrossEntropy
      • BinaryCrossEntropy.__init__()
      • BinaryCrossEntropy.add_arguments()
      • BinaryCrossEntropy.extra_repr()
  • loss_fn.classification.cross_entropy module
    • CrossEntropy
      • CrossEntropy.__init__()
      • CrossEntropy.add_arguments()
      • CrossEntropy.extra_repr()
  • Module contents
• loss_fn.detection package
  • Submodules
  • loss_fn.detection.base_detection_criteria module
    • BaseDetectionCriteria
      • BaseDetectionCriteria.__init__()
      • BaseDetectionCriteria.add_arguments()
  • loss_fn.detection.mask_rcnn_loss module
    • MaskRCNNLoss
      • MaskRCNNLoss.__init__()
      • MaskRCNNLoss.add_arguments()
      • MaskRCNNLoss.forward()
      • MaskRCNNLoss.extra_repr()
  • loss_fn.detection.ssd_multibox_loss module
    • SSDLoss
      • SSDLoss.__init__()
      • SSDLoss.reset_unscaled_loss_values()
      • SSDLoss.add_arguments()
      • SSDLoss.forward()
      • SSDLoss.extra_repr()
  • Module contents
• loss_fn.distillation package
  • Submodules
  • loss_fn.distillation.base_distillation module
    • BaseDistillationCriteria
      • BaseDistillationCriteria.__init__()
      • BaseDistillationCriteria.add_arguments()
      • BaseDistillationCriteria.forward()
  • loss_fn.distillation.hard_distillation module
    • HardDistillationLoss
      • HardDistillationLoss.__init__()
      • HardDistillationLoss.add_arguments()
      • HardDistillationLoss.extra_repr()
  • loss_fn.distillation.soft_kl_distillation module
    • SoftKLLoss
      • SoftKLLoss.__init__()
      • SoftKLLoss.add_arguments()
      • SoftKLLoss.extra_repr()
  • Module contents
• loss_fn.multi_modal_img_text package
  • Submodules
  • loss_fn.multi_modal_img_text.base_multi_modal_img_text_criteria module
    • BaseMultiModalImageTextCriteria
      • BaseMultiModalImageTextCriteria.__init__()
      • BaseMultiModalImageTextCriteria.add_arguments()
  • loss_fn.multi_modal_img_text.contrastive_loss_clip module
    • ContrastiveLossClip
      • ContrastiveLossClip.__init__()
      • ContrastiveLossClip.forward()
    • gather_features()
  • Module contents
• loss_fn.segmentation package
  • Submodules
  • loss_fn.segmentation.base_segmentation_criteria module
    • BaseSegmentationCriteria
      • BaseSegmentationCriteria.__init__()
      • BaseSegmentationCriteria.add_arguments()
  • loss_fn.segmentation.cross_entropy module
    • SegCrossEntropy
      • SegCrossEntropy.__init__()
      • SegCrossEntropy.add_arguments()
      • SegCrossEntropy.forward()
      • SegCrossEntropy.extra_repr()
  • Module contents
• loss_fn.utils package
  • Submodules
  • loss_fn.utils.build_helper module
    • build_cls_teacher_from_opts()
  • loss_fn.utils.class_weighting module
    • compute_class_weights()
  • Module contents

Submodules

loss_fn.base_criteria module

class loss_fn.base_criteria.BaseCriteria(opts: Namespace, *args, **kwargs)

Bases: Module, ABC

Base class for defining loss functions. Sub-classes must implement compute_loss function.

Parameters:

opts – command line arguments

__init__(opts: Namespace, *args, **kwargs) → None

Initializes internal Module state, shared by both nn.Module and ScriptModule.

classmethod add_arguments(parser: ArgumentParser) → ArgumentParser

Add criterion-specific arguments to the parser.

abstract forward(input_sample: Any, prediction: Any, target: Any, *args, **kwargs) → Any

Compute the loss.

Parameters:

• input_sample – Input to the model.

• prediction – Model’s output

• target – Ground truth labels

extra_repr() → str

Set the extra representation of the module

To print customized extra information, you should re-implement this method in your own modules. Both single-line and multi-line strings are acceptable.

loss_fn.composite_loss module

class loss_fn.composite_loss.CompositeLoss(opts: Namespace, *args, **kwargs)

Bases: BaseCriteria

Combines different loss functions and returns the weighted sum of these losses. loss_category and loss_weight are two mandatory keys that allows us to combine different losses and compute their weighted sum. The loss_category specifies the category of a loss function and is a string (e.g., classification). The loss_weight specifies the contribution of a loss function and is a float value (e.g., 1.0). The sum of `loss_weight`s corresponding to different loss functions is not required to be 1.

Parameters:

opts – command-line arguments

Example:: # Example yaml config for combining classification and neural_augmentation loss function is given below. # Please note that configuration for each loss function should start with - in composite_loss.

loss:

category: “composite_loss” composite_loss:

• loss_category: “classification” loss_weight: 1.0 classification:

  name: “cross_entropy” cross_entropy:

  label_smoothing: 0.1

• loss_category: “neural_augmentation” loss_weight: 1.0 neural_augmentation:

  perceptual_metric: “psnr” target_value: [ 40, 10 ] curriculum_method: “cosine”

__init__(opts: Namespace, *args, **kwargs) → None

Initializes internal Module state, shared by both nn.Module and ScriptModule.

classmethod build_composite_loss_fn(opts: Namespace, *args, **kwargs) → Tuple[Mapping[str, BaseCriteria], Mapping[str, float]]

Build loss functions from command line arguments and loss registry

Parameters:

opts – command-line arguments

Returns:

A tuple of two dictionaries. The first dictionary, task_loss_fn_mapping, contains information about loss function category and module. The second dictionary, task_loss_wts_mapping contains the information about loss function category and weight.

classmethod add_arguments(parser: ArgumentParser) → ArgumentParser

Add criterion-specific arguments to the parser.

forward(input_sample: Any, prediction: Any, target: Any, *args, **kwargs) → Any

Compute the weighted sum of different loss functions.

Parameters:

• input_sample – Input to the model.

• prediction – Model’s output

• target – Ground truth labels

Returns:

scalar loss value) with total_loss as a mandatory key. The other keys corresponds to loss category names and their values contain category-specific scalar loss values. total_loss is weighted sum of these category-specific losses.

Return type:

A mapping of the form (str

train(mode: bool = True) → None

Sets the loss functions in training mode.

eval() → None

Sets the loss functions in evaluation mode.

loss_fn.neural_augmentation module

class loss_fn.neural_augmentation.NeuralAugmentation(opts: Namespace, *args, **kwargs)

Bases: BaseCriteria

Compute the augmentation loss, as described in the RangeAugment paper.

Parameters:

opts – command line arguments

__init__(opts: Namespace, *args, **kwargs) → None

Initializes internal Module state, shared by both nn.Module and ScriptModule.

classmethod add_arguments(parser: ArgumentParser) → ArgumentParser

Add criterion-specific arguments to the parser.

forward(input_sample: Tensor | Mapping[str, Tensor | List[Tensor]], prediction: Mapping[str, Tensor], *args, **kwargs) → Tensor

Compute the loss between input and augmented image, as described in RangeAugment paper.

Parameters:

• input_sample – Input sample can either be a Tensor or a dictionary with mandatory key “image”. In case of a dictionary, the values can be a Tensor or list of Tensors.

• prediction – Output of augmentation model. Mapping of (string: Tensor) with augmented_tensor as the required key.

Shapes:

input_sample:

• Tensor: The shape of input tensor is [N, C, H, W]

• Mapping[str, Tensor]: The shape of tensor is [N, C, H, W]

• Mapping[str, List[Tensor]]: The length of List is N, and the shape of each tensor is [1, C, H, W]

prediction: The shape of prediction[“augmented_tensor”] is [N, C, H, W]

Returns:

A scalar loss value

…note:

During validation or evaluation, neural augmentation loss is not computed and 0 is returned

extra_repr() → str

Set the extra representation of the module

To print customized extra information, you should re-implement this method in your own modules. Both single-line and multi-line strings are acceptable.

loss_fn.neural_augmentation.linear_curriculum(start: int, end: int, period: int) → Tensor

This function implements linear curriculum

Parameters:

• start – the starting value for the set of points

• end – the ending value for the set of points

• period – size of the constructed tensor

Returns:

A float tensor of length period

loss_fn.neural_augmentation.cosine_curriculum(start: int, end: int, period: int) → Tensor

This function implements cosine curriculum :param start: the starting value for the set of points :param end: the ending value for the set of points :param period: size of the constructed tensor

Returns:

A float tensor of length period

Module contents

loss_fn.build_loss_fn(opts: Namespace, category: str | None = '', *args, **kwargs) → BaseCriteria

Helper function to build loss function from command-line arguments.

Parameters:

• opts – command-line arguments

• category – Optional task category (e.g., classification). Specifying category may be useful for building composite loss functions. See loss_fns.composite_loss.CompositeLoss.build_composite_loss_fn function for an example

Returns:

Loss function module

loss_fn.add_loss_fn_arguments(parser: ArgumentParser) → ArgumentParser

This method gets a parser object, and for every loss that is registered in the LOSS_REGISTRY adds its arguments to it.