class DetectEigenFaces(Processor):
def __init__(self, weights, measure, thresh, eigenfaces,
mean_face, offsets=[0, 0]):
super(DetectEigenFaces, self).__init__()
self.offsets = offsets
self.class_names = list(weights.keys()) + ['Face not found']
self.colors = lincolor(len(self.class_names))
self.croped_images = None
# detection
self.detect = HaarCascadeFrontalFace()
self.square = SequentialProcessor()
self.square.add(pr.SquareBoxes2D())
self.square.add(pr.OffsetBoxes2D(offsets))
self.clip = pr.ClipBoxes2D()
self.crop = pr.CropBoxes2D()
self.face_detector = EigenFaceDetector(weights, measure, thresh,
eigenfaces, mean_face)
# drawing and wrapping
self.draw = pr.DrawBoxes2D(self.class_names, self.colors,
weighted=True, with_score=False)
self.wrap = pr.WrapOutput(['image', 'boxes2D'])
def call(self, image):
boxes2D = self.detect(image.copy())['boxes2D']
boxes2D = self.square(boxes2D)
boxes2D = self.clip(image, boxes2D)
self.cropped_images = self.crop(image, boxes2D)
for cropped_image, box2D in zip(self.cropped_images, boxes2D):
box2D.class_name = self.face_detector(cropped_image)
image = self.draw(image, boxes2D)
return self.wrap(image, boxes2D)
def __init__(self, size, num_classes, generator=None):
super(ProcessGrayImage, self).__init__()
self.size = size
self.process = SequentialProcessor([pr.ExpandDims(-1)])
if generator is not None:
self.process.add(pr.ImageDataProcessor(generator))
self.process.add(PreprocessImage((size, size), mean=None))
self.process.add(pr.ExpandDims(-1))
self.add(pr.UnpackDictionary(['image', 'label']))
self.add(pr.ExpandDomain(self.process))
self.add(
pr.SequenceWrapper({0: {
'image': [size, size, 1]
}}, {1: {
'label': [num_classes]
}}))
def __init__(self, renderer, image_paths, num_occlusions, split=pr.TRAIN):
super(DomainRandomizationProcessor, self).__init__()
self.copy = pr.Copy()
self.render = pr.Render(renderer)
self.augment = RandomizeRenderedImage(image_paths, num_occlusions)
preprocessors = [pr.ConvertColorSpace(pr.RGB2BGR), pr.NormalizeImage()]
self.preprocess = SequentialProcessor(preprocessors)
self.split = split
def __init__(self, class_names, prior_boxes, variances=[.1, .2]):
super(ShowBoxes, self).__init__()
self.deprocess_boxes = SequentialProcessor([
pr.DecodeBoxes(prior_boxes, variances),
pr.ToBoxes2D(class_names, True),
pr.FilterClassBoxes2D(class_names[1:])
])
self.denormalize_boxes2D = pr.DenormalizeBoxes2D()
self.draw_boxes2D = pr.DrawBoxes2D(class_names)
self.show_image = pr.ShowImage()
def __init__(self, estimate_pose, offsets, valid_class_names, draw=True):
super(PIX2POSE_ROCK, self).__init__()
self.estimate_pose = estimate_pose
self.object_sizes = self.estimate_pose.object_sizes
self.postprocess_boxes = SequentialProcessor([
pr.FilterClassBoxes2D(valid_class_names),
pr.SquareBoxes2D(),
pr.OffsetBoxes2D(offsets)
])
self.clip = pr.ClipBoxes2D()
self.crop = pr.CropBoxes2D()
self.unwrap = pr.UnwrapDictionary(['pose6D', 'points2D', 'points3D'])
self.draw_boxes2D = pr.DrawBoxes2D(valid_class_names)
self.cube_points3D = build_cube_points3D(*self.object_sizes)
self.draw_pose6D = pr.DrawPose6D(self.cube_points3D,
self.estimate_pose.camera.intrinsics)
self.draw = draw
self.wrap = pr.WrapOutput(['image', 'poses6D'])
def __init__(self,
handsegnet,
posenet,
posepriornet,
viewpointnet,
image_size=320,
crop_shape=(256, 256),
num_keypoints=21):
super(DetectHandKeypoints, self).__init__()
self.preprocess_image = SequentialProcessor([
pr.NormalizeImage(),
pr.ResizeImage((image_size, image_size)),
pr.ExpandDims(0)
])
postprocess_segmentation = PostProcessSegmentation(
image_size, crop_shape)
self.localize_hand = pr.Predict(handsegnet,
postprocess=postprocess_segmentation)
self.resize_scoremaps = ResizeScoreMaps(crop_shape)
self.merge_dictionaries = MergeDictionaries()
self.wrap_input = WrapToDictionary(['hand_side'])
self.predict_keypoints2D = pr.Predict(posenet)
self.predict_keypoints3D = pr.Predict(posepriornet)
self.predict_keypoints_angles = pr.Predict(viewpointnet)
self.postprocess_keypoints = PostProcessKeypoints()
self.resize = pr.ResizeImage(shape=crop_shape)
self.extract_2D_keypoints = ExtractKeypoints()
self.transform_keypoints = TransformKeypoints()
self.draw_keypoint = pr.DrawKeypoints2D(num_keypoints,
normalized=True,
radius=4)
self.denormalize = pr.DenormalizeImage()
self.wrap = pr.WrapOutput(['image', 'keypoints2D', 'keypoints3D'])
self.expand_dims = pr.ExpandDims(axis=0)
self.draw_boxes = pr.DrawBoxes2D(['hand'], [[0, 1, 0]])
def test_controlmap_reduction_and_keep():
pipeline = SequentialProcessor()
pipeline.add(ControlMap(Sum(), [1, 2], [1], {2: 0}))
assert pipeline(2, 5, 10) == (10, 2, 5 + 10)
def __init__(self, detector, keypoint_estimator, radius=3):
super(ProbabilisticKeypointPrediction, self).__init__()
# face detector
RGB2GRAY = pr.ConvertColorSpace(pr.RGB2GRAY)
self.detect = pr.Predict(detector, RGB2GRAY, pr.ToBoxes2D(['face']))
# creating pre-processing pipeline for keypoint estimator
preprocess = SequentialProcessor()
preprocess.add(pr.ResizeImage(keypoint_estimator.input_shape[1:3]))
preprocess.add(pr.ConvertColorSpace(pr.RGB2GRAY))
preprocess.add(pr.NormalizeImage())
preprocess.add(pr.ExpandDims(0))
preprocess.add(pr.ExpandDims(-1))
# creating post-processing pipeline for keypoint esimtator
# postprocess = SequentialProcessor()
# postprocess.add(ToNumpyArray())
# postprocess.add(pr.Squeeze(1))
# keypoint estimator predictions
self.estimate_keypoints = PredictMeanDistribution(
keypoint_estimator, preprocess)
# self.estimate_keypoints = pr.Predict(
# keypoint_estimator, preprocess, postprocess)
# used for drawing up keypoints in original image
self.change_coordinates = pr.ChangeKeypointsCoordinateSystem()
self.denormalize_keypoints = pr.DenormalizeKeypoints()
self.crop_boxes2D = pr.CropBoxes2D()
self.num_keypoints = len(keypoint_estimator.output_shape)
self.draw = pr.DrawKeypoints2D(self.num_keypoints, radius, False)
self.draw_boxes2D = pr.DrawBoxes2D(['face'], colors=[[0, 255, 0]])
self.wrap = pr.WrapOutput(['image', 'boxes2D'])
def __init__(self, model, draw=True):
super(GMMKeypoints, self).__init__()
self.num_keypoints = len(model.output_shape)
preprocess = SequentialProcessor()
preprocess.add(pr.ResizeImage(model.input_shape[1:3]))
preprocess.add(pr.ConvertColorSpace(pr.RGB2GRAY))
preprocess.add(pr.NormalizeImage())
preprocess.add(pr.ExpandDims(0))
preprocess.add(pr.ExpandDims(-1))
self.estimate_keypoints = PredictDistributions(model, preprocess)
self.to_grid = ToProbabilityGrid(GRID)
self.draw = draw
self.draw_probabilities = DrawProbabilities(self.num_keypoints)
self.wrap = pr.WrapOutput(['image', 'probabilities', 'distributions'])
import paz.processors as pr
from tensorflow.keras.preprocessing.image import ImageDataGenerator
from tensorflow.keras.utils import get_file
# let's download a test image and put it inside our PAZ directory
IMAGE_URL = ('https://github.com/oarriaga/altamira-data/releases/download'
'/v0.9/image_augmentation.png')
filename = os.path.basename(IMAGE_URL)
image_fullpath = get_file(filename, IMAGE_URL, cache_subdir='paz/tutorials')
# we load the original image and display it
image = load_image(image_fullpath)
show_image(image)
# We construct a data augmentation pipeline using the built-in PAZ processors:
augment = SequentialProcessor()
augment.add(pr.RandomContrast())
augment.add(pr.RandomBrightness())
augment.add(pr.RandomSaturation())
# We can now apply our pipeline as a normal function:
for _ in range(5):
image = load_image(image_fullpath)
# use it as a normal function
image = augment(image)
show_image(image)
# We can add to our sequential pipeline other function anywhere i.e. arg 0:
augment.insert(0, pr.LoadImage())
for _ in range(5):
# now we don't load the image every time.
x = self.preprocess(x)
distributions = self.model(x)
keypoints = np.zeros((self.num_keypoints, 2))
for arg, distribution in enumerate(distributions):
keypoints[arg] = distribution.mean()
return keypoints
def draw_circles(image, points, color=GREEN, radius=3):
for point in points:
draw_circle(image, point, color, radius)
return image
if __name__ == '__main__':
from facial_keypoints import FacialKeypoints
from paz.backend.image import show_image
from paz.abstract import SequentialProcessor
data_manager = FacialKeypoints('dataset/', 'train')
datasets = data_manager.load_data()
augment_keypoints = SequentialProcessor()
augment_keypoints.add(pr.RandomKeypointRotation())
augment_keypoints.add(pr.RandomKeypointTranslation())
for arg in range(100):
original_image = datasets[0]['image'].copy()
kp = datasets[0]['keypoints'].copy()
original_image, kp = augment_keypoints(original_image, kp)
original_image = draw_circles(original_image, kp.astype('int'))
show_image(original_image.astype('uint8'))
# The augment boxes pipeline
class AugmentBoxes(SequentialProcessor):
def __init__(self, mean=pr.BGR_IMAGENET_MEAN):
super(AugmentBoxes, self).__init__()
self.add(pr.ToImageBoxCoordinates())
self.add(pr.Expand(mean=mean))
self.add(pr.RandomSampleCrop())
self.add(pr.RandomFlipBoxesLeftRight())
self.add(pr.ToNormalizedBoxCoordinates())
# We now visualize our current box augmentation
# For that we build a quick pipeline for drawing our boxes
draw_boxes = SequentialProcessor([
pr.ControlMap(pr.ToBoxes2D(class_names, False), [1], [1]),
pr.ControlMap(pr.DenormalizeBoxes2D(), [0, 1], [1], {0: 0}),
pr.DrawBoxes2D(class_names),
pr.ShowImage()
])
# Let's test it our box data augmentation pipeline
augment_boxes = AugmentBoxes()
print('Box augmentation examples')
for _ in range(10):
image = P.image.load_image(image_fullpath)
image, boxes = augment_boxes(image, box_data.copy())
draw_boxes(P.image.resize_image(image, (300, 300)), boxes)
# There is also some box-preprocessing that is required.
# Mostly we must match our boxes to a set of default (prior) boxes.
# Then we must encode them and expand the class label to a one-hot vector.
class FlipBoxesLeftRight(Processor):
def __init__(self):
super(FlipBoxesLeftRight, self).__init__()
def call(self, image, boxes):
width = image.shape[1]
boxes[:, [0, 2]] = width - boxes[:, [2, 0]]
image = image[:, ::-1]
return image, boxes
data = [{
'value_A': np.array([[1.0, 2.0, 3.0, 4.0]]),
'value_B': np.array([[1.0, 1.1, 1.2], [2.0, 2.1, 2.2]])
}]
processor = SequentialProcessor()
processor.add(pr.UnpackDictionary(['value_A', 'value_B']))
processor.add(FlipBoxesLeftRight())
processor.add(
pr.SequenceWrapper({0: {
'value_A': [1, 4]
}}, {1: {
'value_B': [2, 3]
}}))
sequence = ProcessingSequence(processor, 1, data)
for _ in range(10):
batch = sequence.__getitem__(0)
value_A, value_B = batch[0]['value_A'][0], batch[1]['value_B'][0]
print(value_B)
# first we transform our numpy array into our built-in ``Box2D`` messages
to_boxes2D = pr.ToBoxes2D(class_names)
denormalize = pr.DenormalizeBoxes2D()
boxes2D = to_boxes2D(box_data)
image = load_image(image_fullpath)
boxes2D = denormalize(image, boxes2D)
draw_boxes2D = pr.DrawBoxes2D(class_names)
show_image(draw_boxes2D(image, boxes2D))
# As you can see, we were not able to put everything as a
# ``SequentialProcessor``. This is because we are dealing with 2 inputs:
# ``box_data`` and ``image``. We can join them into a single processor
# using ``pr.ControlMap`` wrap. ``pr.ControlMap`` allows you to select which
# arguments (``intro_indices``) are passed to your processor, and also where
# you should put the output of your processor (``outro_indices``).
draw_boxes = SequentialProcessor()
draw_boxes.add(pr.ControlMap(to_boxes2D, intro_indices=[1], outro_indices=[1]))
draw_boxes.add(pr.ControlMap(pr.LoadImage(), [0], [0]))
draw_boxes.add(pr.ControlMap(denormalize, [0, 1], [1], keep={0: 0}))
draw_boxes.add(pr.DrawBoxes2D(class_names))
draw_boxes.add(pr.ShowImage())
# now you have everything in a single packed function that loads and draws!
draw_boxes(image_fullpath, box_data)
# Also note if one of your function is ``eating`` away one input that you
# wish to keep in your pipeline, you can use the ``keep`` dictionary to
# explicitly say which of your inputs you wish to keep and where it should
# be located. This is represented respectively by the ``key`` and the
# ``value`` of the ``keep`` dictionary.
def test_controlmap_parallelization_in_different_order():
pipeline = SequentialProcessor()
pipeline.add(ControlMap(MultiplyByFactor(2.0), [1], [1]))
pipeline.add(ControlMap(MultiplyByFactor(3.0), [0], [0]))
assert pipeline(10, 5) == (10 * 3.0, 5 * 2.0)
def test_controlmap_reduction_and_retention():
pipeline = SequentialProcessor()
pipeline.add(ControlMap(Sum(), [1, 5], [3]))
assert pipeline(2, 5, 10, 6, 7, 8) == (2, 10, 6, 5 + 8, 7)
def test_controlmap_reduction_and_flip():
pipeline = SequentialProcessor()
pipeline.add(ControlMap(Sum(), [1, 2], [1]))
pipeline.add(ControlMap(MultiplyByFactor(0.5), [0], [0]))
pipeline.add(ControlMap(AddConstantToVector(0.1), [0, 1], [1, 0]))
assert pipeline(2, 5, 10) == ((5 + 10) + 0.1, (2 * 0.5) + 0.1)
def test_controlmap_reduction_and_selection_to_arg_2():
pipeline = SequentialProcessor()
pipeline.add(ControlMap(Sum(), [1, 2], [1]))
assert pipeline(2, 5, 10) == (2, 5 + 10)
def test_deterministic_and_stochastic_in_sequential_processor():
function = SequentialProcessor()
function.add(NormalAdd())
function.add(RandomAdd(probability=1.0))
assert function(2.0) == 4.0
