cvnets.models.detection package

Subpackages

• cvnets.models.detection.utils package
  • Submodules
  • cvnets.models.detection.utils.rcnn_utils module
    • FastRCNNConvFCHead
      • FastRCNNConvFCHead.__init__()
    • RPNHead
      • RPNHead.__init__()
      • RPNHead.forward()
    • MaskRCNNHeads
      • MaskRCNNHeads.__init__()
    • MaskRCNNPredictor
      • MaskRCNNPredictor.__init__()
    • FastRCNNPredictor
      • FastRCNNPredictor.__init__()
      • FastRCNNPredictor.forward()
    • replace_syncbn_with_syncbnfp32()
  • Module contents

Submodules

cvnets.models.detection.base_detection module

class cvnets.models.detection.base_detection.BaseDetection(opts: Namespace, encoder: BaseImageEncoder, *args, **kwargs)

Bases: BaseAnyNNModel

Base class for the task of object detection

Parameters:

• opts – Command-line arguments

• encoder – Image-encoder model (e.g., MobileNet or ResNet)

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

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

classmethod add_arguments(parser: ArgumentParser) → ArgumentParser

Add model specific arguments

static reset_layer_parameters(layer: Module, opts: Namespace) → None

Initialize weights of a given layer

classmethod build_model(opts: Namespace, *args, **kwargs) → BaseAnyNNModel

Build a model from command-line arguments. Sub-classes must implement this method

Parameters:

opts – Command-line arguments

…note::

This function is typically implemented in the base class for each task and implementation is reused by all models in that task.

cvnets.models.detection.base_detection.check_feature_map_output_channels(config: Dict, layer_name: str) → int

cvnets.models.detection.mask_rcnn module

class cvnets.models.detection.mask_rcnn.MaskRCNNEncoder(opts: Namespace, encoder: BaseImageEncoder, output_strides: List, projection_channels: int, encoder_lr_multiplier: float | None = 1.0, *args, **kwargs)

Bases: Module

__init__(opts: Namespace, encoder: BaseImageEncoder, output_strides: List, projection_channels: int, encoder_lr_multiplier: float | None = 1.0, *args, **kwargs) → None

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

get_augmented_tensor() → Tensor

forward(x: Tensor) → Dict[str, Tensor]

Defines the computation performed at every call.

Should be overridden by all subclasses.

Note

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.

get_trainable_parameters(weight_decay: float = 0.0, no_decay_bn_filter_bias: bool = False, *args, **kwargs) → Tuple[List, List]

class cvnets.models.detection.mask_rcnn.MaskRCNNDetector(opts, encoder: BaseImageEncoder, *args, **kwargs)

Bases: BaseDetection

This class implements a Mask RCNN style object detector <https://arxiv.org/abs/1703.06870>

Parameters:

• opts – command-line arguments

• encoder (BaseImageEncoder) – Encoder network (e.g., ResNet or MobileViT)

__init__(opts, encoder: BaseImageEncoder, *args, **kwargs) → None

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

update_layer_norm_eps()

set_norm_layer_opts()

reset_norm_layer_opts(default_norm)

classmethod add_arguments(parser: ArgumentParser) → ArgumentParser

Add model specific arguments

reset_generalized_rcnn_transform(height, width)

get_trainable_parameters(weight_decay: float = 0.0, no_decay_bn_filter_bias: bool = False, *args, **kwargs) → Tuple[List, List]

Get parameters for training along with the learning rate.

Parameters:

• weight_decay – weight decay

• no_decay_bn_filter_bias – Do not decay BN and biases. Defaults to False.

Returns:

Returns a tuple of length 2. The first entry is a list of dictionary with three keys (params, weight_decay, param_names). The second entry is a list of floats containing learning rate for each parameter.

Note

Kwargs may contain module_name. To avoid multiple arguments with the same name, we pop it and concatenate with encoder or head name

forward(x: Dict, *args, **kwargs) → Tuple[Tensor, ...] | Tuple[Any, ...] | Dict

Implement the model-specific forward function in sub-classes.

predict(x: Tensor, *args, **kwargs) → DetectionPredTuple

Predict the bounding boxes given an image tensor

dummy_input_and_label(batch_size: int) → Dict

Create dummy input and labels for CI/CD purposes.

cvnets.models.detection.ssd module

class cvnets.models.detection.ssd.SingleShotMaskDetector(opts, encoder: BaseImageEncoder)

Bases: BaseDetection

This class implements a Single Shot Object Detector

Parameters:

• opts – command-line arguments

• encoder (BaseImageEncoder) – Encoder network (e.g., ResNet or MobileViT)

coordinates = 4

__init__(opts, encoder: BaseImageEncoder) → None

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

classmethod add_arguments(parser: ArgumentParser) → ArgumentParser

Add model specific arguments

static reset_layers(module) → None

static process_anchors_ar(anchor_ar: List) → List

get_backbone_features(x: Tensor) → Dict[str, Tensor]

ssd_forward(end_points: Dict[str, Tensor], device: device | None = device(type='cpu'), *args, **kwargs) → Tuple[Tensor, Tensor, Tensor] | Tuple[Tensor, ...]

forward(x: Tensor | Dict) → Tuple[Tensor, ...] | Tuple[Any, ...] | Dict

Implement the model-specific forward function in sub-classes.

predict(x: Tensor, *args, **kwargs) → DetectionPredTuple

Predict the bounding boxes given an image tensor

postprocess_detections(boxes: Tensor, scores: Tensor) → List[DetectionPredTuple]

Post process detections, including NMS

dummy_input_and_label(batch_size: int) → Dict

Create dummy input and labels for CI/CD purposes.

Module contents

class cvnets.models.detection.DetectionPredTuple(labels, scores, boxes, masks)

Bases: tuple

boxes

Alias for field number 2

labels

Alias for field number 0

masks

Alias for field number 3

scores

Alias for field number 1