farukcankaya’s gists

farukcankaya / multimodal_aug_color_train.py

Created October 20, 2022 23:27

farukcankaya / multimodal_aug_color_jitter_apply.py

Created October 20, 2022 16:51

	aug = InstanceColorJitterAugmentation(lambda img: ImageEnhance.Color(img).enhance(10))
	augmentations = T.AugmentationList([aug])
	aug_input = MultiModalAugInput(image, annos=d['annotations'])
	transforms = augmentations(aug_input)
	augmented_image = aug_input.image

farukcankaya / multimodal_aug_color_jitter.py

Created October 20, 2022 16:45

	class InstanceColorJitterTransform(MultiModalTransform):
	def __init__(self, color_operation: Callable, instance_rate: float, min_count_to_apply: int) -> None:
	if not callable(color_operation):
	raise ValueError("color_operation parameter should be callable")
	super().__init__()
	self._set_attributes(locals())

	def apply_multi_modal(self, img, annos, *args):
	instance_count = len(annos)
	apply_count = max(self.min_count_to_apply, int(instance_count * self.instance_rate))

farukcankaya / multimodal_aug_x.py

Last active October 20, 2022 13:29

	def transform(self, tfm: Transform) -> None:
	if isinstance(tfm, MultiModalTransform):
	self.image, self.annos = \
	tfm.apply_multi_modal(self.image, self.annos)
	else:
	super().transform(tfm)

farukcankaya / multimodal_aug_transform.py

Created October 20, 2022 08:20

	class Transform:
	@abstractmethod
	def apply_image(self, img: np.ndarray):

	@abstractmethod
	def apply_coords(self, coords: np.ndarray):

	def apply_segmentation(self, segmentation: np.ndarray) -> np.ndarray:
	return self.apply_image(segmentation)

farukcankaya / multimodal_aug_aug_input.py

Created October 20, 2022 07:41

	class AugInput:
	def transform(self, tfm: Transform) -> None:
	self.image = tfm.apply_image(self.image)
	if self.boxes is not None:
	self.boxes = tfm.apply_box(self.boxes)
	if self.sem_seg is not None:
	self.sem_seg = tfm.apply_segmentation(self.sem_seg)

farukcankaya / multimodal_aug_augmentation.py

Last active October 20, 2022 07:39

	class Augmentation:
	def __call__(self, aug_input) -> Transform:
	args = _get_aug_input_args(self, aug_input)
	tfm = self.get_transform(*args)
	aug_input.transform(tfm)
	return tfm

farukcankaya / multimodal_aug_dataset_mapper.py

Created October 20, 2022 00:48

	# from: https://github.com/facebookresearch/detectron2/blob/998c4e15220a0db9f172b1e7ecf376a59b873f1d/detectron2/data/dataset_mapper.py#L163-L165
	aug_input = T.AugInput(image, sem_seg=sem_seg_gt)
	transforms = self.augmentations(aug_input)
	image, sem_seg_gt = aug_input.image, aug_input.sem_seg

farukcankaya / detectron2_models.csv

Created October 3, 2022 21:36

@META_ARCH	@BACKBONE	@PROPOSAL_GENERATOR	@ROI_HEADS	@ROI_MASK_HEAD
GeneralizedRCNN	build_resnet_backbone	RPN	RROIHeads	MaskRCNNConvUpsampleHead
PanopticFPN	build_resnet_fpn_backbone	RRPN	Res5ROIHeads
ProposalNetwork	build_retinanet_resnet_fpn_backbone		CascadeROIHeads
RetinaNet			StandardROIHeads
SemanticSegmentor

farukcankaya / detectron2_visuzalization.py

Created October 3, 2022 19:03

check https://gist.github.com/farukcankaya/b1a461c0f269a7c8ae57dcf702b0da4c for VisualizationHook

	def do_train_visualization(visualizer, model, inputs):
	training_mode = model.training
	model.eval()

	outputs = visualizer.inference(model, inputs)
	visualizer.process(inputs, outputs)

	model.train(training_mode)

	with EventStorage(start_iter) as storage:

Faruk Cankaya farukcankaya