loss_fn.detection package

Submodules

loss_fn.detection.base_detection_criteria module

class loss_fn.detection.base_detection_criteria.BaseDetectionCriteria(opts: Namespace, *args, **kwargs)[source]

Bases: BaseCriteria

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

Parameters:: opts – command line arguments

__init__(opts: Namespace, *args, **kwargs) → None[source]: Initializes internal Module state, shared by both nn.Module and ScriptModule.

classmethod add_arguments(parser: ArgumentParser) → ArgumentParser[source]: Add criterion-specific arguments to the parser.

loss_fn.detection.mask_rcnn_loss module

class loss_fn.detection.mask_rcnn_loss.MaskRCNNLoss(opts: Namespace, *args, **kwargs)[source]

Bases: BaseDetectionCriteria

Mask RCNN loss is computed inside the MaskRCNN model. This class is a wrapper to extract loss values for different heads (RPN, classification, etc.) and compute the weighted sum.

Parameters:: opts – command-line arguments

__init__(opts: Namespace, *args, **kwargs) → None[source]: Initializes internal Module state, shared by both nn.Module and ScriptModule.

classmethod add_arguments(parser: ArgumentParser) → ArgumentParser[source]: Add criterion-specific arguments to the parser.

forward(input_sample: Any, prediction: Dict[str, Tensor], *args, **kwargs) → Dict[str, Tensor][source]

Compute MaskRCNN loss.

Parameters:

input_sample – Input image tensor to the model.
prediction – Mapping of the Maskrcnn losses.

Shapes:: input_sample: This loss function does not care about input to the model. prediction: Dictionary containing scalar Mask RCNN loss values. Expected keys are

loss_classifier, loss_box_reg, loss_mask, loss_objectness, loss_rpn_box_reg.

Returns:: scalar) is returned. Output contains following keys: (total_loss, loss_classifier, loss_box_reg, loss_mask, loss_objectness, loss_rpn_box_reg).
Return type:: A mapping of (string

extra_repr() → str[source]

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.detection.ssd_multibox_loss module

class loss_fn.detection.ssd_multibox_loss.SSDLoss(opts: Namespace, *args, **kwargs)[source]

Bases: BaseDetectionCriteria

Loss for single shot multi-box object detection

Parameters:: opts – command-line arguments

__init__(opts: Namespace, *args, **kwargs) → None[source]: Initializes internal Module state, shared by both nn.Module and ScriptModule.

reset_unscaled_loss_values() → None[source]: Reset the unscaled coefficients for confidence and regression losses to small values

classmethod add_arguments(parser: ArgumentParser) → ArgumentParser[source]: Add criterion-specific arguments to the parser.

forward(input_sample: Any, prediction: Dict[str, Tensor], target: Dict[str, Tensor], *args, **kwargs) → Dict[str, Tensor][source]

Compute the SSD Loss

Parameters:

input_sample – Input image tensor to the model.
prediction – Model output. It is a mapping of the form (string: Tensor) containing two mandatory keys, i.e., scores and boxes
target – Ground truth labels. It is a mapping of the form (string: Tensor) containing two mandatory keys, i.e., box_labels and box_coordinates.

Shape:

input_sample: This loss function does not care about input to the model. prediction[“scores”]: Shape is [Batch size, number of anchors, number of classes] prediction[“boxes”]: Shape is [Batch size, number of anchors, 4] where 4 is the number of box coordinates

target[“box_labels”]: Shape is [Batch size, number of anchors] target[“box_coordinates”]: Shape is [Batch size, number of anchors, 4]

Returns:: scalar) is returned. Output contains following keys: (total_loss, reg_loss, cls_loss).
Return type:: A mapping of (string