export.drise package

Submodules

export.drise.exporter module

class export.drise.exporter.DRiseSaliencyObjectDetectionExporter(model: ObjectDetector, num_classes: int, batch_size: int = 1, name: str | None = None, num_masks: int = 1000, score_threshold: float = 0.5, criterion: Callable[[int, Batch], bool | Iterable[int]] | None = None)

Bases: Exporter

An exporter for D-RISE object detection saliency maps.

class ModelWrapper(model: ObjectDetector, num_classes: int, spec)

Bases: Module

A wrapper around an Armory object detector to make it work with the D-RISE implementation.

expand_scores(scores: Tensor, labels: Tensor) → Tensor

Expands the given batch of scores and labels into probabilities for all classes. The score for the assigned label remains as specified, with all other classes receiving the evenly divided remaining probability.

Parameters:

  • scores (torch.Tensor) – N-length tensor of predicted scores

  • labels (torch.Tensor) – N-length tensor of assigned labels

Returns:

(N, C) tensor of probabilities, where C is the total number of classes

Return type:

torch.Tensor

forward(images_pt: Tensor)

Invokes the wrapped model and returns the resulting boxes, class probabilities, and objectness as a tuple.

Parameters:

images_pt (torch.Tensor) – images_pt

training: bool
export_samples(batch_idx: int, batch: ObjectDetectionBatch, samples: Iterable[int]) → None

Export samples

Parameters:

  • batch_idx (int) – Batch index

  • batch (ObjectDetectionBatch) – Batch

  • samples (Iterable[int]) – Samples

spec = ImageSpec(dim=<ImageDimensions.CHW: 1>, scale=Scale(dtype=<DataType.FLOAT: 2>, max=1.0, mean=None, std=None))

export.drise.impl module

export.drise.impl.draw_box(image, box, color='red')

Draw box

Parameters:

  • image (_type_) – Image

  • box (_type_) – Box

  • color (str, optional) – Color, defaults to “red”

export.drise.impl.draw_contour(img_pt, cpu_map, thickness=1, levels=[0.95, 0.99], pixel_weight=True)

Draw contour

Parameters:

  • img_pt (_type_) – img_pt

  • cpu_map (_type_) – cpu_map

  • thickness (int, optional) – Thickness, defaults to 1

  • levels (list, optional) – Levels, defaults to [0.95, 0.99]

  • pixel_weight (bool, optional) – If true, uses the pixels that contain <level> of the total value (after subtracting the min) if false, <level> is quartiles. Defaults to True

export.drise.impl.draw_heatmap(image, saliency_map: Tensor, box)

Draw heatmap

Parameters:

  • image (_type_) – Image

  • saliency_map (torch.Tensor) – Saliency map

  • box (_type_) – Box

export.drise.impl.generate_masks(w=16, h=16, p=0.5, number_of_masks=5000)

Generate masks

Parameters:

  • w (int, optional) – Width, defaults to 16

  • h (int, optional) – Height defaults to 16

  • p (float, optional) – P, defaults to 0.5

  • number_of_masks (int, optional) – Number of masks, defaults to 5000

export.drise.impl.get_proposal_data(img_pt, model, w=16, h=16, p=0.5, number_of_masks=5000, batch_size=16, object_thres=0.5, device='cpu')

Get proposal data

Parameters:

  • img_pt (_type_) – img_pt

  • model (_type_) – Model

  • w (int, optional) – Width, defaults to 16

  • h (int, optional) – Height, defaults to 16

  • p (float, optional) – P, defaults to 0.5

  • number_of_masks (int, optional) – Number of masks, defaults to 5000

  • batch_size (int, optional) – Batch size, defaults to 16

  • object_thres (float, optional) – Object threshold, defaults to 0.5

  • device (str, optional) – Device, defaults to “cpu”

export.drise.impl.imgpath2tensor(directory, file_name)

Image path to tensor

Parameters:

  • directory (str) – Directory

  • file_name (str) – File name

export.drise.impl.make_saliency_img(img_pt, saliency_map, box, color='black', levels=[0.9, 0.95, 0.975, 0.99], contour=False)

Make saliency image

Parameters:

  • img_pt (_type_) – img_pt

  • saliency_map (_type_) – Saliency map

  • box (_type_) – Box

  • color (str, optional) – Color, defaults to “black”

  • levels (list, optional) – Levels, defaults to [0.9, 0.95, 0.975, 0.99]

  • contour (bool, optional) – Contour, defaults to False

export.drise.impl.make_saliency_map(img_pt, gt_box, gt_prob, masks, rand_offset_nums, boxes, cls_probs, objectiveness, device='cpu')

Make saliency map

Parameters:

  • img_pt (_type_) – img_pt

  • gt_box (_type_) – gt_box

  • gt_prob (_type_) – gt_prob

  • masks (_type_) – Masks

  • rand_offset_nums (_type_) – rand_offset_nums

  • boxes (_type_) – Box

  • cls_probs (_type_) – cls_probs

  • objectiveness (_type_) – Objectiveness

  • device (str, optional) – Device, defaults to “cpu”

export.drise.impl.masked_dataloader(img_pt, masks_all, rand_offset_nums, batch_size=4, device='cpu')

Masked dataloader

Parameters:

  • img_pt (_type_) – image_pt

  • masks_all (_type_) – masks_all

  • rand_offset_nums (_type_) – rand_offset_nums

  • batch_size (int, optional) – Batch size, defaults to 4

  • device (str, optional) – Device, defaults to “cpu”

Module contents