增加keras-cv模型及训练代码

Keras/keras-cv/keras_cv/layers/preprocessing/random_hue_test.py

+# Copyright 2022 The KerasCV Authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     https://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+import tensorflow as tf
+from absl.testing import parameterized
+
+from keras_cv import core
+from keras_cv.layers import preprocessing
+
+
+class RandomHueTest(tf.test.TestCase, parameterized.TestCase):
+    def test_preserves_output_shape(self):
+        image_shape = (4, 8, 8, 3)
+        image = tf.random.uniform(shape=image_shape) * 255.0
+
+        layer = preprocessing.RandomHue(factor=(0.3, 0.8), value_range=(0, 255))
+        output = layer(image)
+
+        self.assertEqual(image.shape, output.shape)
+        self.assertNotAllClose(image, output)
+
+    def test_adjust_no_op(self):
+        image_shape = (4, 8, 8, 3)
+        image = tf.random.uniform(shape=image_shape) * 255.0
+
+        layer = preprocessing.RandomHue(factor=(0.0, 0.0), value_range=(0, 255))
+        output = layer(image)
+        self.assertAllClose(image, output, atol=1e-5, rtol=1e-5)
+
+    def test_adjust_full_opposite_hue(self):
+        image_shape = (4, 8, 8, 3)
+        image = tf.random.uniform(shape=image_shape) * 255.0
+
+        layer = preprocessing.RandomHue(factor=(1.0, 1.0), value_range=(0, 255))
+        output = layer(image)
+
+        channel_max = tf.math.reduce_max(output, axis=-1)
+        channel_min = tf.math.reduce_min(output, axis=-1)
+        # Make sure the max and min channel are the same between input and output
+        # In the meantime, and channel will swap between each other.
+        self.assertAllClose(channel_max, tf.math.reduce_max(image, axis=-1))
+        self.assertAllClose(channel_min, tf.math.reduce_min(image, axis=-1))
+
+    @parameterized.named_parameters(
+        ("025", 0.25), ("05", 0.5), ("075", 0.75), ("100", 1.0)
+    )
+    def test_adjusts_all_values_for_factor(self, factor):
+        image_shape = (4, 8, 8, 3)
+        # Value range (0, 100)
+        image = tf.random.uniform(shape=image_shape) * 100.0
+
+        layer = preprocessing.RandomHue(factor=(factor, factor), value_range=(0, 255))
+        output = layer(image)
+        self.assertNotAllClose(image, output, atol=1e-5, rtol=1e-5)
+
+    def test_adjustment_for_non_rgb_value_range(self):
+        image_shape = (4, 8, 8, 3)
+        # Value range (0, 100)
+        image = tf.random.uniform(shape=image_shape) * 100.0
+
+        layer = preprocessing.RandomHue(factor=(0.0, 0.0), value_range=(0, 255))
+        output = layer(image)
+        self.assertAllClose(image, output, atol=1e-5, rtol=1e-5)
+
+        layer = preprocessing.RandomHue(factor=(0.3, 0.8), value_range=(0, 255))
+        output = layer(image)
+        self.assertNotAllClose(image, output)
+
+    def test_with_uint8(self):
+        image_shape = (4, 8, 8, 3)
+        image = tf.cast(tf.random.uniform(shape=image_shape) * 255.0, dtype=tf.uint8)
+
+        layer = preprocessing.RandomHue(factor=(0.0, 0.0), value_range=(0, 255))
+        output = layer(image)
+        self.assertAllClose(image, output, atol=1e-5, rtol=1e-5)
+
+        layer = preprocessing.RandomHue(factor=(0.3, 0.8), value_range=(0, 255))
+        output = layer(image)
+        self.assertNotAllClose(image, output)
+
+    def test_config(self):
+        layer = preprocessing.RandomHue(factor=(0.3, 0.8), value_range=(0, 255))
+        config = layer.get_config()
+        self.assertTrue(isinstance(config["factor"], core.UniformFactorSampler))
+        self.assertEqual(config["factor"].get_config()["lower"], 0.3)
+        self.assertEqual(config["factor"].get_config()["upper"], 0.8)
+        self.assertEqual(config["value_range"], (0, 255))

Keras/keras-cv/keras_cv/layers/preprocessing/random_jpeg_quality.py

+# Copyright 2022 The KerasCV Authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     https://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import tensorflow as tf
+
+from keras_cv.layers.preprocessing.base_image_augmentation_layer import (
+    BaseImageAugmentationLayer,
+)
+from keras_cv.utils import preprocessing
+
+
+@tf.keras.utils.register_keras_serializable(package="keras_cv")
+class RandomJpegQuality(BaseImageAugmentationLayer):
+    """Applies Random Jpeg compression artifacts to an image.
+
+    Performs the jpeg compression algorithm on the image.  This layer can used in order
+    to ensure your model is robust to artifacts introduced by JPEG compresion.
+
+    Args:
+        factor: 2 element tuple or 2 element list.  During augmentation, a random number
+        is drawn from the factor distribution.  This value is passed to
+        `tf.image.adjust_jpeg_quality()`.
+        seed: Integer. Used to create a random seed.
+
+    Usage:
+    ```python
+    layer = keras_cv.RandomJpegQuality(factor=(75, 100)))
+    (images, labels), _ = tf.keras.datasets.cifar10.load_data()
+    augmented_images = layer(images)
+    ```
+    """
+
+    def __init__(self, factor, seed=None, **kwargs):
+        super().__init__(**kwargs)
+        if isinstance(factor, (float, int)):
+            raise ValueError(
+                "RandomJpegQuality() expects factor to be a 2 element "
+                "tuple, list or a `keras_cv.FactorSampler`. "
+                "RandomJpegQuality() received `factor={factor}`."
+            )
+        self.seed = seed
+        self.factor = preprocessing.parse_factor(
+            factor, min_value=0, max_value=100, param_name="factor", seed=self.seed
+        )
+
+    def get_random_transformation(self, **kwargs):
+        return self.factor(dtype=tf.int32)
+
+    def augment_image(self, image, transformation=None, **kwargs):
+        jpeg_quality = transformation
+        return tf.image.adjust_jpeg_quality(image, jpeg_quality)
+
+    def augment_bounding_boxes(self, bounding_boxes, **kwargs):
+        return bounding_boxes
+
+    def augment_label(self, label, transformation=None, **kwargs):
+        return label
+
+    def augment_segmentation_mask(self, segmentation_mask, transformation, **kwargs):
+        return segmentation_mask
+
+    def get_config(self):
+        config = super().get_config()
+        config.update({"factor": self.factor, "seed": self.seed})
+        return config

Keras/keras-cv/keras_cv/layers/preprocessing/random_jpeg_quality_test.py

+# Copyright 2022 The KerasCV Authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     https://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import tensorflow as tf
+
+from keras_cv.layers import preprocessing
+
+
+class RandomJpegQualityTest(tf.test.TestCase):
+    def test_return_shapes(self):
+        layer = preprocessing.RandomJpegQuality(factor=[0, 100])
+
+        # RGB
+        xs = tf.ones((2, 512, 512, 3))
+        xs = layer(xs)
+        self.assertEqual(xs.shape, [2, 512, 512, 3])
+
+        # greyscale
+        xs = tf.ones((2, 512, 512, 1))
+        xs = layer(xs)
+        self.assertEqual(xs.shape, [2, 512, 512, 1])
+
+    def test_in_single_image(self):
+        layer = preprocessing.RandomJpegQuality(factor=[0, 100])
+
+        # RGB
+        xs = tf.cast(
+            tf.ones((512, 512, 3)),
+            dtype=tf.float32,
+        )
+
+        xs = layer(xs)
+        self.assertEqual(xs.shape, [512, 512, 3])
+
+        # greyscale
+        xs = tf.cast(
+            tf.ones((512, 512, 1)),
+            dtype=tf.float32,
+        )
+
+        xs = layer(xs)
+        self.assertEqual(xs.shape, [512, 512, 1])
+
+    def test_non_square_images(self):
+        layer = preprocessing.RandomJpegQuality(factor=[0, 100])
+
+        # RGB
+        xs = tf.ones((2, 256, 512, 3))
+        xs = layer(xs)
+        self.assertEqual(xs.shape, [2, 256, 512, 3])
+
+        # greyscale
+        xs = tf.ones((2, 256, 512, 1))
+        xs = layer(xs)
+        self.assertEqual(xs.shape, [2, 256, 512, 1])

Keras/keras-cv/keras_cv/layers/preprocessing/random_rotation.py

+# Copyright 2022 The KerasCV Authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     https://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import numpy as np
+import tensorflow as tf
+
+from keras_cv import bounding_box
+from keras_cv.layers.preprocessing.base_image_augmentation_layer import (
+    BaseImageAugmentationLayer,
+)
+from keras_cv.utils import preprocessing
+
+# In order to support both unbatched and batched inputs, the horizontal
+# and verticle axis is reverse indexed
+H_AXIS = -3
+W_AXIS = -2
+
+
+@tf.keras.utils.register_keras_serializable(package="keras_cv")
+class RandomRotation(BaseImageAugmentationLayer):
+    """A preprocessing layer which randomly rotates images during training.
+
+    This layer will apply random rotations to each image, filling empty space
+    according to `fill_mode`.
+
+    By default, random rotations are only applied during training.
+    At inference time, the layer does nothing. If you need to apply random
+    rotations at inference time, set `training` to True when calling the layer.
+
+    Input pixel values can be of any range (e.g. `[0., 1.)` or `[0, 255]`) and
+    of interger or floating point dtype. By default, the layer will output
+    floats.
+
+    Input shape:
+      3D (unbatched) or 4D (batched) tensor with shape:
+      `(..., height, width, channels)`, in `"channels_last"` format
+
+    Output shape:
+      3D (unbatched) or 4D (batched) tensor with shape:
+      `(..., height, width, channels)`, in `"channels_last"` format
+
+    Arguments:
+      factor: a float represented as fraction of 2 Pi, or a tuple of size 2
+        representing lower and upper bound for rotating clockwise and
+        counter-clockwise. A positive values means rotating counter clock-wise,
+        while a negative value means clock-wise. When represented as a single
+        float, this value is used for both the upper and lower bound. For
+        instance, `factor=(-0.2, 0.3)` results in an output rotation by a random
+        amount in the range `[-20% * 2pi, 30% * 2pi]`. `factor=0.2` results in
+        an output rotating by a random amount in the range
+        `[-20% * 2pi, 20% * 2pi]`.
+      fill_mode: Points outside the boundaries of the input are filled according
+        to the given mode (one of `{"constant", "reflect", "wrap", "nearest"}`).
+        - *reflect*: `(d c b a | a b c d | d c b a)` The input is extended by
+          reflecting about the edge of the last pixel.
+        - *constant*: `(k k k k | a b c d | k k k k)` The input is extended by
+          filling all values beyond the edge with the same constant value k = 0.
+        - *wrap*: `(a b c d | a b c d | a b c d)` The input is extended by
+          wrapping around to the opposite edge.
+        - *nearest*: `(a a a a | a b c d | d d d d)` The input is extended by
+          the nearest pixel.
+      interpolation: Interpolation mode. Supported values: `"nearest"`,
+        `"bilinear"`.
+      seed: Integer. Used to create a random seed.
+      fill_value: a float represents the value to be filled outside the
+        boundaries when `fill_mode="constant"`.
+      bounding_box_format: The format of bounding boxes of input dataset. Refer
+        https://github.com/keras-team/keras-cv/blob/master/keras_cv/bounding_box/converters.py
+        for more details on supported bounding box formats.
+      segmentation_classes: an optional integer with the number of classes in
+        the input segmentation mask. Required iff augmenting data with sparse
+        (non one-hot) segmentation masks. Include the background class in this
+        count (e.g. for segmenting dog vs background, this should be set to 2).
+    """
+
+    def __init__(
+        self,
+        factor,
+        fill_mode="reflect",
+        interpolation="bilinear",
+        seed=None,
+        fill_value=0.0,
+        bounding_box_format=None,
+        segmentation_classes=None,
+        **kwargs,
+    ):
+        super().__init__(seed=seed, force_generator=True, **kwargs)
+        self.factor = factor
+        if isinstance(factor, (tuple, list)):
+            self.lower = factor[0]
+            self.upper = factor[1]
+        else:
+            self.lower = -factor
+            self.upper = factor
+        if self.upper < self.lower:
+            raise ValueError(
+                "Factor cannot have negative values, " "got {}".format(factor)
+            )
+        preprocessing.check_fill_mode_and_interpolation(fill_mode, interpolation)
+        self.fill_mode = fill_mode
+        self.fill_value = fill_value
+        self.interpolation = interpolation
+        self.seed = seed
+        self.bounding_box_format = bounding_box_format
+        self.segmentation_classes = segmentation_classes
+
+    def get_random_transformation(self, **kwargs):
+        min_angle = self.lower * 2.0 * np.pi
+        max_angle = self.upper * 2.0 * np.pi
+        angle = self._random_generator.random_uniform(
+            shape=[1], minval=min_angle, maxval=max_angle
+        )
+        return {"angle": angle}
+
+    def augment_image(self, image, transformation, **kwargs):
+        return self._rotate_image(image, transformation)
+
+    def _rotate_image(self, image, transformation):
+        image = preprocessing.ensure_tensor(image, self.compute_dtype)
+        original_shape = image.shape
+        image = tf.expand_dims(image, 0)
+        image_shape = tf.shape(image)
+        img_hd = tf.cast(image_shape[H_AXIS], tf.float32)
+        img_wd = tf.cast(image_shape[W_AXIS], tf.float32)
+        angle = transformation["angle"]
+        output = preprocessing.transform(
+            image,
+            preprocessing.get_rotation_matrix(angle, img_hd, img_wd),
+            fill_mode=self.fill_mode,
+            fill_value=self.fill_value,
+            interpolation=self.interpolation,
+        )
+        output = tf.squeeze(output, 0)
+        output.set_shape(original_shape)
+        return output
+
+    def augment_bounding_boxes(
+        self, bounding_boxes, transformation, image=None, **kwargs
+    ):
+        if self.bounding_box_format is None:
+            raise ValueError(
+                "`RandomRotation()` was called with bounding boxes,"
+                "but no `bounding_box_format` was specified in the constructor."
+                "Please specify a bounding box format in the constructor. i.e."
+                "`RandomRotation(bounding_box_format='xyxy')`"
+            )
+        else:
+            bounding_boxes = bounding_box.convert_format(
+                bounding_boxes,
+                source=self.bounding_box_format,
+                target="xyxy",
+                images=image,
+            )
+        image_shape = tf.shape(image)
+        h = image_shape[H_AXIS]
+        w = image_shape[W_AXIS]
+        _, _, _, _, rest = tf.split(
+            bounding_boxes, [1, 1, 1, 1, bounding_boxes.shape[-1] - 4], axis=-1
+        )
+        # origin coordinates, all the points on the image are rotated around
+        # this point
+        origin_x, origin_y = tf.cast(w / 2, dtype=self.compute_dtype), tf.cast(
+            h / 2, dtype=self.compute_dtype
+        )
+        angle = transformation["angle"]
+        angle = -angle
+        # calculate coordinates of all four corners of the bounding box
+        point = tf.stack(
+            [
+                tf.stack([bounding_boxes[:, 0], bounding_boxes[:, 1]], axis=1),
+                tf.stack([bounding_boxes[:, 2], bounding_boxes[:, 1]], axis=1),
+                tf.stack([bounding_boxes[:, 2], bounding_boxes[:, 3]], axis=1),
+                tf.stack([bounding_boxes[:, 0], bounding_boxes[:, 3]], axis=1),
+            ],
+            axis=1,
+        )
+        # point_x : x coordinates of all corners of the bounding box
+        point_x = tf.gather(point, [0], axis=2)
+        # point_y : y cordinates of all corners of the bounding box
+        point_y = tf.gather(point, [1], axis=2)
+        # rotated bounding box coordinates
+        # new_x : new position of x coordinates of corners of bounding box
+        new_x = (
+            origin_x
+            + tf.multiply(
+                tf.cos(angle), tf.cast((point_x - origin_x), dtype=tf.float32)
+            )
+            - tf.multiply(
+                tf.sin(angle), tf.cast((point_y - origin_y), dtype=tf.float32)
+            )
+        )
+        # new_y : new position of y coordinates of corners of bounding box
+        new_y = (
+            origin_y
+            + tf.multiply(
+                tf.sin(angle), tf.cast((point_x - origin_x), dtype=tf.float32)
+            )
+            + tf.multiply(
+                tf.cos(angle), tf.cast((point_y - origin_y), dtype=tf.float32)
+            )
+        )
+        # rotated bounding box coordinates
+        out = tf.concat([new_x, new_y], axis=2)
+        # find readjusted coordinates of bounding box to represent it in corners
+        # format
+        min_cordinates = tf.math.reduce_min(out, axis=1)
+        max_cordinates = tf.math.reduce_max(out, axis=1)
+        bounding_boxes_out = tf.concat([min_cordinates, max_cordinates, rest], axis=1)
+        bounding_boxes_out = bounding_box.clip_to_image(
+            bounding_boxes_out,
+            bounding_box_format="xyxy",
+            images=image,
+        )
+        # cordinates cannot be float values, it is casted to int32
+        bounding_boxes_out = bounding_box.convert_format(
+            bounding_boxes_out,
+            source="xyxy",
+            target=self.bounding_box_format,
+            dtype=self.compute_dtype,
+            images=image,
+        )
+        return bounding_boxes_out
+
+    def augment_label(self, label, transformation, **kwargs):
+        return label
+
+    def augment_segmentation_mask(self, segmentation_mask, transformation, **kwargs):
+        # If segmentation_classes is specified, we have a dense segmentation mask.
+        # We therefore one-hot encode before rotation to avoid bad interpolation
+        # during the rotation transformation. We then make the mask sparse
+        # again using tf.argmax.
+        if self.segmentation_classes:
+            one_hot_mask = tf.one_hot(
+                tf.squeeze(segmentation_mask, axis=-1), self.segmentation_classes
+            )
+            rotated_one_hot_mask = self._rotate_image(one_hot_mask, transformation)
+            rotated_mask = tf.argmax(rotated_one_hot_mask, axis=-1)
+            return tf.expand_dims(rotated_mask, axis=-1)
+        else:
+            if segmentation_mask.shape[-1] == 1:
+                raise ValueError(
+                    "Segmentation masks must be one-hot encoded, or "
+                    "RandomRotate must be initialized with "
+                    "`segmentation_classes`. `segmentation_classes` was not "
+                    f"specified, and mask has shape {segmentation_mask.shape}"
+                )
+            rotated_mask = self._rotate_image(segmentation_mask, transformation)
+            # Round because we are in one-hot encoding, and we may have
+            # pixels with ambugious value due to floating point math for rotation.
+            return tf.round(rotated_mask)
+
+    def compute_output_shape(self, input_shape):
+        return input_shape
+
+    def get_config(self):
+        config = {
+            "factor": self.factor,
+            "fill_mode": self.fill_mode,
+            "fill_value": self.fill_value,
+            "interpolation": self.interpolation,
+            "bounding_box_format": self.bounding_box_format,
+            "segmentation_classes": self.segmentation_classes,
+            "seed": self.seed,
+        }
+        base_config = super().get_config()
+        return dict(list(base_config.items()) + list(config.items()))

Keras/keras-cv/keras_cv/layers/preprocessing/random_rotation_test.py

+# Copyright 2022 The KerasCV Authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     https://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+import numpy as np
+import tensorflow as tf
+
+from keras_cv import bounding_box
+from keras_cv.layers.preprocessing.random_rotation import RandomRotation
+
+
+class RandomRotationTest(tf.test.TestCase):
+    def test_random_rotation_output_shapes(self):
+        input_images = np.random.random((2, 5, 8, 3)).astype(np.float32)
+        expected_output = input_images
+        layer = RandomRotation(0.5)
+        actual_output = layer(input_images, training=True)
+        self.assertEqual(expected_output.shape, actual_output.shape)
+
+    def test_random_rotation_inference(self):
+        input_images = np.random.random((2, 5, 8, 3)).astype(np.float32)
+        expected_output = input_images
+        layer = RandomRotation(0.5)
+        actual_output = layer(input_images, training=False)
+        self.assertAllClose(expected_output, actual_output)
+
+    def test_config_with_custom_name(self):
+        layer = RandomRotation(0.5, name="image_preproc")
+        config = layer.get_config()
+        layer_reconstructed = RandomRotation.from_config(config)
+        self.assertEqual(layer_reconstructed.name, layer.name)
+
+    def test_unbatched_image(self):
+        input_image = np.reshape(np.arange(0, 25), (5, 5, 1)).astype(np.float32)
+        # 180 rotation.
+        layer = RandomRotation(factor=(0.5, 0.5))
+        output_image = layer(input_image)
+        expected_output = np.asarray(
+            [
+                [24, 23, 22, 21, 20],
+                [19, 18, 17, 16, 15],
+                [14, 13, 12, 11, 10],
+                [9, 8, 7, 6, 5],
+                [4, 3, 2, 1, 0],
+            ]
+        ).astype(np.float32)
+        expected_output = np.reshape(expected_output, (5, 5, 1))
+        self.assertAllClose(expected_output, output_image)
+
+    def test_augment_bbox_dict_input(self):
+        input_image = np.random.random((512, 512, 3)).astype(np.float32)
+        bboxes = tf.convert_to_tensor(
+            [[200, 200, 400, 400, 1], [100, 100, 300, 300, 2]]
+        )
+        input = {"images": input_image, "bounding_boxes": bboxes}
+        # 180 rotation.
+        layer = RandomRotation(factor=(0.5, 0.5), bounding_box_format="xyxy")
+        output_bbox = layer(input)
+        expected_output = np.asarray(
+            [[112.0, 112.0, 312.0, 312.0, 1], [212.0, 212.0, 412.0, 412.0, 2]],
+        )
+        expected_output = np.reshape(expected_output, (2, 5))
+        self.assertAllClose(expected_output, output_bbox["bounding_boxes"])
+
+    def test_output_dtypes(self):
+        inputs = np.array([[[1], [2]], [[3], [4]]], dtype="float64")
+        layer = RandomRotation(0.5)
+        self.assertAllEqual(layer(inputs).dtype, "float32")
+        layer = RandomRotation(0.5, dtype="uint8")
+        self.assertAllEqual(layer(inputs).dtype, "uint8")
+
+    def test_ragged_bounding_boxes(self):
+        input_image = np.random.random((2, 512, 512, 3)).astype(np.float32)
+        bboxes = tf.ragged.constant(
+            [
+                [[200, 200, 400, 400], [100, 100, 300, 300]],
+                [[200, 200, 400, 400]],
+            ],
+            dtype=tf.float32,
+        )
+        bboxes = bounding_box.add_class_id(bboxes)
+        input = {"images": input_image, "bounding_boxes": bboxes}
+        layer = RandomRotation(factor=(0.5, 0.5), bounding_box_format="xyxy")
+        output = layer(input)
+        expected_output = tf.ragged.constant(
+            [
+                [[112.0, 112.0, 312.0, 312.0, 0], [212.0, 212.0, 412.0, 412.0, 0]],
+                [[112.0, 112.0, 312.0, 312.0, 0]],
+            ],
+            ragged_rank=1,
+        )
+        self.assertAllClose(expected_output, output["bounding_boxes"])
+
+    def test_augment_sparse_segmentation_mask(self):
+        classes = 8
+
+        input_images = np.random.random((2, 20, 20, 3)).astype(np.float32)
+        # Masks are all 0s or 8s, to verify that when we rotate we don't do bad
+        # mask interpolation to either a 0 or a 7
+        masks = np.random.randint(2, size=(2, 20, 20, 1)) * (classes - 1)
+        inputs = {"images": input_images, "segmentation_masks": masks}
+
+        # Attempting to rotate a sparse mask without specifying classes fails.
+        bad_layer = RandomRotation(factor=(0.25, 0.25))
+        with self.assertRaisesRegex(ValueError, "masks must be one-hot"):
+            outputs = bad_layer(inputs)
+
+        # 90 degree rotation.
+        layer = RandomRotation(factor=(0.25, 0.25), segmentation_classes=classes)
+        outputs = layer(inputs)
+        expected_masks = np.rot90(masks, axes=(1, 2))
+        self.assertAllClose(expected_masks, outputs["segmentation_masks"])
+
+        # 45 degree rotation. Only verifies that no interpolation takes place.
+        layer = RandomRotation(factor=(0.125, 0.125), segmentation_classes=classes)
+        outputs = layer(inputs)
+        self.assertAllInSet(outputs["segmentation_masks"], [0, 7])
+
+    def test_augment_one_hot_segmentation_mask(self):
+        classes = 8
+
+        input_images = np.random.random((2, 20, 20, 3)).astype(np.float32)
+        masks = tf.one_hot(np.random.randint(classes, size=(2, 20, 20)), classes)
+        inputs = {"images": input_images, "segmentation_masks": masks}
+
+        # 90 rotation.
+        layer = RandomRotation(factor=(0.25, 0.25))
+        outputs = layer(inputs)
+
+        expected_masks = np.rot90(masks, axes=(1, 2))
+
+        self.assertAllClose(expected_masks, outputs["segmentation_masks"])

Keras/keras-cv/keras_cv/layers/preprocessing/random_saturation.py

+# Copyright 2022 The KerasCV Authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     https://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+import tensorflow as tf
+
+from keras_cv.layers.preprocessing.base_image_augmentation_layer import (
+    BaseImageAugmentationLayer,
+)
+from keras_cv.utils import preprocessing
+
+
+@tf.keras.utils.register_keras_serializable(package="keras_cv")
+class RandomSaturation(BaseImageAugmentationLayer):
+    """Randomly adjusts the saturation on given images.
+
+    This layer will randomly increase/reduce the saturation for the input RGB
+    images. At inference time, the output will be identical to the input.
+    Call the layer with `training=True` to adjust the saturation of the input.
+
+    Args:
+        factor: A tuple of two floats, a single float or `keras_cv.FactorSampler`.
+            `factor` controls the extent to which the image saturation is impacted.
+            `factor=0.5` makes this layer perform a no-op operation. `factor=0.0` makes
+            the image to be fully grayscale. `factor=1.0` makes the image to be fully
+            saturated.
+            Values should be between `0.0` and `1.0`. If a tuple is used, a `factor`
+            is sampled between the two values for every image augmented.  If a single
+            float is used, a value between `0.0` and the passed float is sampled.
+            In order to ensure the value is always the same, please pass a tuple with
+            two identical floats: `(0.5, 0.5)`.
+        seed: Integer. Used to create a random seed.
+    """
+
+    def __init__(self, factor, seed=None, **kwargs):
+        super().__init__(seed=seed, **kwargs)
+        self.factor = preprocessing.parse_factor(
+            factor,
+            min_value=0.0,
+            max_value=1.0,
+        )
+        self.seed = seed
+
+    def get_random_transformation(self, **kwargs):
+        return self.factor()
+
+    def augment_image(self, image, transformation=None, **kwargs):
+        # Convert the factor range from [0, 1] to [0, +inf]. Note that the
+        # tf.image.adjust_saturation is trying to apply the following math formula
+        # `output_saturation = input_saturation * factor`. We use the following
+        # method to the do the mapping.
+        # `y = x / (1 - x)`.
+        # This will ensure:
+        #   y = +inf when x = 1 (full saturation)
+        #   y = 1 when x = 0.5 (no augmentation)
+        #   y = 0 when x = 0 (full gray scale)
+
+        # Convert the transformation to tensor in case it is a float. When
+        # transformation is 1.0, then it will result in to divide by zero error, but
+        # it will be handled correctly when it is a one tensor.
+        transformation = tf.convert_to_tensor(transformation)
+        adjust_factor = transformation / (1 - transformation)
+        return tf.image.adjust_saturation(image, saturation_factor=adjust_factor)
+
+    def augment_bounding_boxes(self, bounding_boxes, transformation=None, **kwargs):
+        return bounding_boxes
+
+    def augment_label(self, label, transformation=None, **kwargs):
+        return label
+
+    def augment_segmentation_mask(self, segmentation_mask, transformation, **kwargs):
+        return segmentation_mask
+
+    def get_config(self):
+        config = {
+            "factor": self.factor,
+            "seed": self.seed,
+        }
+        base_config = super().get_config()
+        return dict(list(base_config.items()) + list(config.items()))

Keras/keras-cv/keras_cv/layers/preprocessing/random_saturation_test.py

+# Copyright 2022 The KerasCV Authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     https://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+import tensorflow as tf
+
+from keras_cv import core
+from keras_cv.layers import preprocessing
+
+
+class RandomSaturationTest(tf.test.TestCase):
+    def test_preserves_output_shape(self):
+        image_shape = (4, 8, 8, 3)
+        image = tf.random.uniform(shape=image_shape) * 255.0
+
+        layer = preprocessing.RandomSaturation(factor=(0.3, 0.8))
+        output = layer(image)
+
+        self.assertEqual(image.shape, output.shape)
+        self.assertNotAllClose(image, output)
+
+    def test_no_adjustment_for_factor_point_five(self):
+        image_shape = (4, 8, 8, 3)
+        image = tf.random.uniform(shape=image_shape) * 255.0
+
+        layer = preprocessing.RandomSaturation(factor=(0.5, 0.5))
+        output = layer(image)
+
+        self.assertAllClose(image, output, atol=1e-5, rtol=1e-5)
+
+    def test_adjust_to_grayscale(self):
+        image_shape = (4, 8, 8, 3)
+        image = tf.random.uniform(shape=image_shape) * 255.0
+
+        layer = preprocessing.RandomSaturation(factor=(0.0, 0.0))
+        output = layer(image)
+
+        channel_mean = tf.math.reduce_mean(output, axis=-1)
+        channel_values = tf.unstack(output, axis=-1)
+        # Make sure all the pixel has the same value among the channel dim, which is
+        # a fully gray RGB.
+        for channel_value in channel_values:
+            self.assertAllClose(channel_mean, channel_value, atol=1e-5, rtol=1e-5)
+
+    def test_adjust_to_full_saturation(self):
+        image_shape = (4, 8, 8, 3)
+        image = tf.random.uniform(shape=image_shape) * 255.0
+
+        layer = preprocessing.RandomSaturation(factor=(1.0, 1.0))
+        output = layer(image)
+
+        channel_mean = tf.math.reduce_min(output, axis=-1)
+        # Make sure at least one of the channel is 0.0 (fully saturated image)
+        self.assertAllClose(channel_mean, tf.zeros((4, 8, 8)))
+
+    def test_adjustment_for_non_rgb_value_range(self):
+        image_shape = (4, 8, 8, 3)
+        # Value range (0, 100)
+        image = tf.random.uniform(shape=image_shape) * 100.0
+
+        layer = preprocessing.RandomSaturation(factor=(0.5, 0.5))
+        output = layer(image)
+        self.assertAllClose(image, output, atol=1e-5, rtol=1e-5)
+
+        layer = preprocessing.RandomSaturation(factor=(0.3, 0.8))
+        output = layer(image)
+        self.assertNotAllClose(image, output)
+
+    def test_with_unit8(self):
+        image_shape = (4, 8, 8, 3)
+        image = tf.cast(tf.random.uniform(shape=image_shape) * 255.0, dtype=tf.uint8)
+
+        layer = preprocessing.RandomSaturation(factor=(0.5, 0.5))
+        output = layer(image)
+        self.assertAllClose(image, output, atol=1e-5, rtol=1e-5)
+
+        layer = preprocessing.RandomSaturation(factor=(0.3, 0.8))
+        output = layer(image)
+        self.assertNotAllClose(image, output)
+
+    def test_config(self):
+        layer = preprocessing.RandomSaturation(factor=(0.3, 0.8))
+        config = layer.get_config()
+        self.assertTrue(isinstance(config["factor"], core.UniformFactorSampler))
+        self.assertEqual(config["factor"].get_config()["lower"], 0.3)
+        self.assertEqual(config["factor"].get_config()["upper"], 0.8)

Keras/keras-cv/keras_cv/layers/preprocessing/random_sharpness.py

+# Copyright 2022 The KerasCV Authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     https://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+import tensorflow as tf
+
+from keras_cv.layers.preprocessing.base_image_augmentation_layer import (
+    BaseImageAugmentationLayer,
+)
+from keras_cv.utils import preprocessing
+
+
+@tf.keras.utils.register_keras_serializable(package="keras_cv")
+class RandomSharpness(BaseImageAugmentationLayer):
+    """Randomly performs the sharpness operation on given images.
+
+    The sharpness operation first performs a blur operation, then blends between the
+    original image and the blurred image.  This operation makes the edges of an image
+    less sharp than they were in the original image.
+
+    References:
+        - [PIL](https://pillow.readthedocs.io/en/stable/reference/ImageEnhance.html)
+
+    Args:
+        factor: A tuple of two floats, a single float or `keras_cv.FactorSampler`.
+            `factor` controls the extent to which the image sharpness is impacted.
+            `factor=0.0` makes this layer perform a no-op operation, while a value of
+            1.0 uses the sharpened result entirely.  Values between 0 and 1 result in
+            linear interpolation between the original image and the sharpened image.
+            Values should be between `0.0` and `1.0`.  If a tuple is used, a `factor` is
+            sampled between the two values for every image augmented.  If a single float
+            is used, a value between `0.0` and the passed float is sampled.  In order to
+            ensure the value is always the same, please pass a tuple with two identical
+            floats: `(0.5, 0.5)`.
+        value_range: the range of values the incoming images will have.
+            Represented as a two number tuple written [low, high].
+            This is typically either `[0, 1]` or `[0, 255]` depending
+            on how your preprocessing pipeline is setup.
+    """
+
+    def __init__(
+        self,
+        factor,
+        value_range,
+        seed=None,
+        **kwargs,
+    ):
+        super().__init__(seed=seed, **kwargs)
+        self.value_range = value_range
+        self.factor = preprocessing.parse_factor(factor)
+        self.seed = seed
+
+    def get_random_transformation(self, **kwargs):
+        return self.factor()
+
+    def augment_image(self, image, transformation=None, **kwargs):
+        image = preprocessing.transform_value_range(
+            image, original_range=self.value_range, target_range=(0, 255)
+        )
+        original_image = image
+
+        # Make image 4D for conv operation.
+        image = tf.expand_dims(image, axis=0)
+
+        # [1 1 1]
+        # [1 5 1]
+        # [1 1 1]
+        # all divided by 13 is the default 3x3 gaussian smoothing kernel.
+        # Correlating or Convolving with this filter is equivalent to performing a
+        # gaussian blur.
+        kernel = (
+            tf.constant(
+                [[1, 1, 1], [1, 5, 1], [1, 1, 1]], dtype=tf.float32, shape=[3, 3, 1, 1]
+            )
+            / 13.0
+        )
+
+        # Tile across channel dimension.
+        channels = tf.shape(image)[-1]
+        kernel = tf.tile(kernel, [1, 1, channels, 1])
+        strides = [1, 1, 1, 1]
+
+        smoothed_image = tf.nn.depthwise_conv2d(
+            image, kernel, strides, padding="VALID", dilations=[1, 1]
+        )
+        smoothed_image = tf.clip_by_value(smoothed_image, 0.0, 255.0)
+        smoothed_image = tf.squeeze(smoothed_image, axis=0)
+
+        # For the borders of the resulting image, fill in the values of the
+        # original image.
+        mask = tf.ones_like(smoothed_image)
+        padded_mask = tf.pad(mask, [[1, 1], [1, 1], [0, 0]])
+        padded_smoothed_image = tf.pad(smoothed_image, [[1, 1], [1, 1], [0, 0]])
+
+        result = tf.where(
+            tf.equal(padded_mask, 1), padded_smoothed_image, original_image
+        )
+        # Blend the final result.
+        result = preprocessing.blend(original_image, result, transformation)
+        result = preprocessing.transform_value_range(
+            result, original_range=(0, 255), target_range=self.value_range
+        )
+        return result
+
+    def augment_bounding_boxes(self, bounding_boxes, transformation, **kwargs):
+        return bounding_boxes
+
+    def augment_label(self, label, transformation=None, **kwargs):
+        return label
+
+    def augment_segmentation_mask(self, segmentation_mask, transformation, **kwargs):
+        return segmentation_mask
+
+    def get_config(self):
+        config = super().get_config()
+        config.update(
+            {"factor": self.factor, "value_range": self.value_range, "seed": self.seed}
+        )
+        return config

Keras/keras-cv/keras_cv/layers/preprocessing/random_sharpness_test.py

+# Copyright 2022 The KerasCV Authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     https://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+import tensorflow as tf
+
+from keras_cv.layers import preprocessing
+
+
+class RandomSharpnessTest(tf.test.TestCase):
+    def test_random_sharpness_preserves_output_shape(self):
+        img_shape = (50, 50, 3)
+        xs = tf.stack(
+            [2 * tf.ones(img_shape), tf.ones(img_shape)],
+            axis=0,
+        )
+
+        layer = preprocessing.RandomSharpness(0.0, value_range=(0, 255))
+        ys = layer(xs)
+
+        self.assertEqual(xs.shape, ys.shape)
+        self.assertAllClose(xs, ys)
+
+    def test_random_sharpness_blur_effect_single_channel(self):
+        xs = tf.expand_dims(
+            tf.constant(
+                [
+                    [0, 0, 0, 0, 0, 0, 0],
+                    [0, 0, 0, 0, 0, 0, 0],
+                    [0, 0, 0, 0, 0, 0, 0],
+                    [0, 0, 0, 1, 0, 0, 0],
+                    [0, 0, 0, 0, 0, 0, 0],
+                    [0, 0, 0, 0, 0, 0, 0],
+                    [0, 0, 0, 0, 0, 0, 0],
+                ]
+            ),
+            axis=-1,
+        )
+        xs = tf.expand_dims(xs, axis=0)
+
+        layer = preprocessing.RandomSharpness((1.0, 1.0), value_range=(0, 255))
+        ys = layer(xs)
+
+        self.assertEqual(xs.shape, ys.shape)
+
+        result = tf.expand_dims(
+            tf.constant(
+                [
+                    [0, 0, 0, 0, 0, 0, 0],
+                    [0, 0, 0, 0, 0, 0, 0],
+                    [0, 0, 1 / 13, 1 / 13, 1 / 13, 0, 0],
+                    [0, 0, 1 / 13, 5 / 13, 1 / 13, 0, 0],
+                    [0, 0, 1 / 13, 1 / 13, 1 / 13, 0, 0],
+                    [0, 0, 0, 0, 0, 0, 0],
+                    [0, 0, 0, 0, 0, 0, 0],
+                ]
+            ),
+            axis=-1,
+        )
+        result = tf.expand_dims(result, axis=0)
+
+        self.assertAllClose(ys, result)

Keras/keras-cv/keras_cv/layers/preprocessing/random_shear.py

+# Copyright 2022 The KerasCV Authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     https://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+import warnings
+
+import tensorflow as tf
+
+import keras_cv
+from keras_cv import bounding_box
+from keras_cv.layers.preprocessing.base_image_augmentation_layer import (
+    BaseImageAugmentationLayer,
+)
+from keras_cv.utils import preprocessing
+
+
+@tf.keras.utils.register_keras_serializable(package="keras_cv")
+class RandomShear(BaseImageAugmentationLayer):
+    """A preprocessing layer which randomly shears images during training.
+    This layer will apply random shearings to each image, filling empty space
+    according to `fill_mode`.
+    By default, random shears are only applied during training.
+    At inference time, the layer does nothing. If you need to apply random
+    shear at inference time, set `training` to True when calling the layer.
+    Input pixel values can be of any range and any data type.
+    Input shape:
+      3D (unbatched) or 4D (batched) tensor with shape:
+      `(..., height, width, channels)`, in `"channels_last"` format
+    Output shape:
+      3D (unbatched) or 4D (batched) tensor with shape:
+      `(..., height, width, channels)`, in `"channels_last"` format
+
+    Args:
+        x_factor: A tuple of two floats, a single float or a
+            `keras_cv.FactorSampler`. For each augmented image a value is sampled
+            from the provided range. If a float is passed, the range is interpreted as
+            `(0, x_factor)`.  Values represent a percentage of the image to shear over.
+             For example, 0.3 shears pixels up to 30% of the way across the image.
+             All provided values should be positive.  If `None` is passed, no shear
+             occurs on the X axis.
+             Defaults to `None`.
+        y_factor: A tuple of two floats, a single float or a
+            `keras_cv.FactorSampler`. For each augmented image a value is sampled
+            from the provided range. If a float is passed, the range is interpreted as
+            `(0, y_factor)`. Values represent a percentage of the image to shear over.
+            For example, 0.3 shears pixels up to 30% of the way across the image.
+            All provided values should be positive.  If `None` is passed, no shear
+            occurs on the Y axis.
+            Defaults to `None`.
+        interpolation: interpolation method used in the `ImageProjectiveTransformV3` op.
+             Supported values are `"nearest"` and `"bilinear"`.
+             Defaults to `"bilinear"`.
+        fill_mode: fill_mode in the `ImageProjectiveTransformV3` op.
+             Supported values are `"reflect"`, `"wrap"`, `"constant"`, and `"nearest"`.
+             Defaults to `"reflect"`.
+        fill_value: fill_value in the `ImageProjectiveTransformV3` op.
+             A `Tensor` of type `float32`. The value to be filled when fill_mode is
+             constant".  Defaults to `0.0`.
+        bounding_box_format: The format of bounding boxes of input dataset. Refer to
+             https://github.com/keras-team/keras-cv/blob/master/keras_cv/bounding_box/converters.py
+             for more details on supported bounding box formats.
+        seed: Integer. Used to create a random seed.
+    """
+
+    def __init__(
+        self,
+        x_factor=None,
+        y_factor=None,
+        interpolation="bilinear",
+        fill_mode="reflect",
+        fill_value=0.0,
+        bounding_box_format=None,
+        seed=None,
+        **kwargs,
+    ):
+        super().__init__(seed=seed, **kwargs)
+        if x_factor is not None:
+            self.x_factor = preprocessing.parse_factor(
+                x_factor, max_value=None, param_name="x_factor", seed=seed
+            )
+        else:
+            self.x_factor = x_factor
+        if y_factor is not None:
+            self.y_factor = preprocessing.parse_factor(
+                y_factor, max_value=None, param_name="y_factor", seed=seed
+            )
+        else:
+            self.y_factor = y_factor
+        if x_factor is None and y_factor is None:
+            warnings.warn(
+                "RandomShear received both `x_factor=None` and `y_factor=None`.  As a "
+                "result, the layer will perform no augmentation."
+            )
+        self.interpolation = interpolation
+        self.fill_mode = fill_mode
+        self.fill_value = fill_value
+        self.seed = seed
+        self.bounding_box_format = bounding_box_format
+
+    def get_random_transformation(self, **kwargs):
+        x = self._get_shear_amount(self.x_factor)
+        y = self._get_shear_amount(self.y_factor)
+        return (x, y)
+
+    def _get_shear_amount(self, constraint):
+        if constraint is None:
+            return None
+
+        invert = preprocessing.random_inversion(self._random_generator)
+        return invert * constraint()
+
+    def augment_image(self, image, transformation=None, **kwargs):
+        image = tf.expand_dims(image, axis=0)
+
+        x, y = transformation
+
+        if x is not None:
+            transform_x = RandomShear._format_transform(
+                [1.0, x, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0]
+            )
+            image = preprocessing.transform(
+                images=image,
+                transforms=transform_x,
+                interpolation=self.interpolation,
+                fill_mode=self.fill_mode,
+                fill_value=self.fill_value,
+            )
+
+        if y is not None:
+            transform_y = RandomShear._format_transform(
+                [1.0, 0.0, 0.0, y, 1.0, 0.0, 0.0, 0.0]
+            )
+            image = preprocessing.transform(
+                images=image,
+                transforms=transform_y,
+                interpolation=self.interpolation,
+                fill_mode=self.fill_mode,
+                fill_value=self.fill_value,
+            )
+
+        return tf.squeeze(image, axis=0)
+
+    def augment_label(self, label, transformation=None, **kwargs):
+        return label
+
+    def augment_bounding_boxes(
+        self, bounding_boxes, transformation, image=None, **kwargs
+    ):
+        if self.bounding_box_format is None:
+            raise ValueError(
+                "`RandomShear()` was called with bounding boxes,"
+                "but no `bounding_box_format` was specified in the constructor."
+                "Please specify a bounding box format in the constructor. i.e."
+                "`RandomShear(bounding_box_format='xyxy')`"
+            )
+        bounding_boxes = keras_cv.bounding_box.convert_format(
+            bounding_boxes,
+            source=self.bounding_box_format,
+            target="rel_xyxy",
+            images=image,
+            dtype=self.compute_dtype,
+        )
+        x, y = transformation
+        extended_bboxes, rest_axes = self._convert_to_extended_corners_format(
+            bounding_boxes
+        )
+        if x is not None:
+            extended_bboxes = self._apply_horizontal_transformation_to_bounding_box(
+                extended_bboxes, x
+            )
+        # apply vertical shear
+        if y is not None:
+            extended_bboxes = self._apply_vertical_transformation_to_bounding_box(
+                extended_bboxes, y
+            )
+
+        bounding_boxes = self._convert_to_four_coordinate(extended_bboxes, x, y)
+        # join rest of the axes with bbox axes
+        bounding_boxes = tf.concat(
+            [bounding_boxes, rest_axes],
+            axis=-1,
+        )
+        bounding_boxes = bounding_box.clip_to_image(
+            bounding_boxes, images=image, bounding_box_format="rel_xyxy"
+        )
+        # convert to universal output format
+        bounding_boxes = keras_cv.bounding_box.convert_format(
+            bounding_boxes,
+            source="rel_xyxy",
+            target=self.bounding_box_format,
+            images=image,
+            dtype=self.compute_dtype,
+        )
+        return bounding_boxes
+
+    def get_config(self):
+        config = super().get_config()
+        config.update(
+            {
+                "x_factor": self.x_factor,
+                "y_factor": self.y_factor,
+                "interpolation": self.interpolation,
+                "fill_mode": self.fill_mode,
+                "fill_value": self.fill_value,
+                "bounding_box_format": self.bounding_box_format,
+                "seed": self.seed,
+            }
+        )
+        return config
+
+    @staticmethod
+    def _format_transform(transform):
+        transform = tf.convert_to_tensor(transform, dtype=tf.float32)
+        return transform[tf.newaxis]
+
+    @staticmethod
+    def _convert_to_four_coordinate(extended_bboxes, x, y):
+        """convert from extended coordinates to 4 coordinates system"""
+        (
+            top_left_x,
+            top_left_y,
+            bottom_right_x,
+            bottom_right_y,
+            top_right_x,
+            top_right_y,
+            bottom_left_x,
+            bottom_left_y,
+        ) = tf.split(extended_bboxes, 8, axis=1)
+
+        # choose x1,x2 when x>0
+        def positive_case_x():
+            final_x1 = bottom_left_x
+            final_x2 = top_right_x
+            return final_x1, final_x2
+
+        # choose x1,x2 when x<0
+        def negative_case_x():
+            final_x1 = top_left_x
+            final_x2 = bottom_right_x
+            return final_x1, final_x2
+
+        if x is not None:
+            final_x1, final_x2 = tf.cond(
+                tf.less(x, 0), negative_case_x, positive_case_x
+            )
+        else:
+            final_x1, final_x2 = top_left_x, bottom_right_x
+
+        # choose y1,y2 when y > 0
+        def positive_case_y():
+            final_y1 = top_right_y
+            final_y2 = bottom_left_y
+            return final_y1, final_y2
+
+        # choose y1,y2 when y < 0
+        def negative_case_y():
+            final_y1 = top_left_y
+            final_y2 = bottom_right_y
+            return final_y1, final_y2
+
+        if y is not None:
+            final_y1, final_y2 = tf.cond(
+                tf.less(y, 0), negative_case_y, positive_case_y
+            )
+        else:
+            final_y1, final_y2 = top_left_y, bottom_right_y
+        return tf.concat(
+            [final_x1, final_y1, final_x2, final_y2],
+            axis=1,
+        )
+
+    @staticmethod
+    def _apply_horizontal_transformation_to_bounding_box(extended_bounding_boxes, x):
+        # create transformation matrix [1,4]
+        matrix = tf.stack([1.0, -x, 0, 1.0], axis=0)
+        # reshape it to [2,2]
+        matrix = tf.reshape(matrix, (2, 2))
+        # reshape unnormalized bboxes from [N,8] -> [N*4,2]
+        new_bboxes = tf.reshape(extended_bounding_boxes, (-1, 2))
+        # [[1,x`],[y`,1]]*[x,y]->[new_x,new_y]
+        transformed_bboxes = tf.reshape(
+            tf.einsum("ij,kj->ki", matrix, new_bboxes), (-1, 8)
+        )
+        return transformed_bboxes
+
+    @staticmethod
+    def _apply_vertical_transformation_to_bounding_box(extended_bounding_boxes, y):
+        # create transformation matrix [1,4]
+        matrix = tf.stack([1.0, 0, -y, 1.0], axis=0)
+        # reshape it to [2,2]
+        matrix = tf.reshape(matrix, (2, 2))
+        # reshape unnormalized bboxes from [N,8] -> [N*4,2]
+        new_bboxes = tf.reshape(extended_bounding_boxes, (-1, 2))
+        # [[1,x`],[y`,1]]*[x,y]->[new_x,new_y]
+        transformed_bboxes = tf.reshape(
+            tf.einsum("ij,kj->ki", matrix, new_bboxes), (-1, 8)
+        )
+        return transformed_bboxes
+
+    @staticmethod
+    def _convert_to_extended_corners_format(bounding_boxes):
+        """splits corner bboxes top left,bottom right to 4 corners top left,
+        bottom right,top right and bottom left"""
+        x1, y1, x2, y2, rest = tf.split(
+            bounding_boxes, [1, 1, 1, 1, bounding_boxes.shape[-1] - 4], axis=-1
+        )
+        new_bboxes = tf.concat(
+            [x1, y1, x2, y2, x2, y1, x1, y2],
+            axis=-1,
+        )
+        return new_bboxes, rest

Keras/keras-cv/keras_cv/layers/preprocessing/random_shear_test.py

+# Copyright 2022 The KerasCV Authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     https://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+import tensorflow as tf
+
+from keras_cv import bounding_box
+from keras_cv.layers import preprocessing
+
+classes = 10
+
+
+class RandomShearTest(tf.test.TestCase):
+    def test_aggressive_shear_fills_at_least_some_pixels(self):
+        img_shape = (50, 50, 3)
+        xs = tf.stack(
+            [2 * tf.ones(img_shape), tf.ones(img_shape)],
+            axis=0,
+        )
+        xs = tf.cast(xs, tf.float32)
+
+        fill_value = 0.0
+        layer = preprocessing.RandomShear(
+            x_factor=(3, 3), seed=0, fill_mode="constant", fill_value=fill_value
+        )
+        xs = layer(xs)
+
+        # Some pixels should be replaced with fill value
+        self.assertTrue(tf.math.reduce_any(xs[0] == fill_value))
+        self.assertTrue(tf.math.reduce_any(xs[0] == 2.0))
+        self.assertTrue(tf.math.reduce_any(xs[1] == fill_value))
+        self.assertTrue(tf.math.reduce_any(xs[1] == 1.0))
+
+    def test_return_shapes(self):
+        """test return dict keys and value pairs"""
+        xs = tf.ones((2, 512, 512, 3))
+        # randomly sample labels
+        ys_labels = tf.random.categorical(tf.math.log([[0.5, 0.5]]), 2)
+        ys_labels = tf.squeeze(ys_labels)
+        ys_labels = tf.one_hot(ys_labels, classes)
+
+        # randomly sample bounding boxes
+        ys_bounding_boxes = tf.random.uniform((2, 3, 7), 0, 1)
+
+        layer = preprocessing.RandomShear(
+            x_factor=(0.1, 0.3),
+            y_factor=(0.1, 0.3),
+            seed=0,
+            fill_mode="constant",
+            bounding_box_format="rel_xyxy",
+        )
+        # mixup on labels
+        outputs = layer(
+            {"images": xs, "labels": ys_labels, "bounding_boxes": ys_bounding_boxes}
+        )
+        xs, ys_labels, ys_bounding_boxes = (
+            outputs["images"],
+            outputs["labels"],
+            outputs["bounding_boxes"],
+        )
+        self.assertEqual(xs.shape, [2, 512, 512, 3])
+        self.assertEqual(ys_labels.shape, [2, 10])
+        self.assertEqual(ys_bounding_boxes.shape, [2, 3, 7])
+
+    def test_single_image_input(self):
+        """test for single image input"""
+        xs = tf.ones((512, 512, 3))
+        ys = tf.ones(shape=(5, 5))
+        inputs = {"images": xs, "bounding_boxes": ys}
+        layer = preprocessing.RandomShear(
+            x_factor=(3, 3),
+            seed=0,
+            fill_mode="constant",
+            bounding_box_format="rel_xyxy",
+        )
+        outputs = layer(inputs)
+        xs, ys_bounding_boxes = (
+            outputs["images"],
+            outputs["bounding_boxes"],
+        )
+        self.assertEqual(xs.shape, [512, 512, 3])
+        self.assertEqual(ys_bounding_boxes.shape, [5, 5])
+
+    def test_area(self):
+        """test for shear bbox transformation since new bbox will be
+        greater than old bbox"""
+        xs = tf.ones((512, 512, 3))
+        ys = tf.constant([[0.3, 0.4, 0.5, 0.6, 2], [0.9, 0.8, 1.0, 1.0, 3]])
+        inputs = {"images": xs, "bounding_boxes": ys}
+        layer = preprocessing.RandomShear(
+            x_factor=(0.3, 0.7),
+            y_factor=(0.4, 0.7),
+            seed=0,
+            fill_mode="constant",
+            bounding_box_format="rel_xyxy",
+        )
+        outputs = layer(inputs)
+        xs, ys_bounding_boxes = (
+            outputs["images"],
+            outputs["bounding_boxes"],
+        )
+        new_area = tf.math.multiply(
+            tf.abs(tf.subtract(ys_bounding_boxes[..., 2], ys_bounding_boxes[..., 0])),
+            tf.abs(tf.subtract(ys_bounding_boxes[..., 3], ys_bounding_boxes[..., 1])),
+        )
+        old_area = tf.math.multiply(
+            tf.abs(tf.subtract(ys[..., 2], ys[..., 0])),
+            tf.abs(tf.subtract(ys[..., 3], ys[..., 1])),
+        )
+        tf.debugging.assert_greater_equal(new_area, old_area)
+
+    def test_in_tf_function(self):
+        """test for class works with tf function"""
+        xs = tf.cast(
+            tf.stack(
+                [2 * tf.ones((4, 4, 3)), tf.ones((4, 4, 3))],
+                axis=0,
+            ),
+            tf.float32,
+        )
+        ys = tf.cast(
+            tf.stack(
+                [
+                    tf.constant([[0.3, 0.4, 0.5, 0.6], [0.9, 0.8, 1.0, 1.0]]),
+                    tf.constant([[0.3, 0.4, 0.5, 0.6], [0.9, 0.8, 1.0, 1.0]]),
+                ],
+                axis=0,
+            ),
+            tf.float32,
+        )
+        ys = bounding_box.add_class_id(ys)
+        layer = preprocessing.RandomShear(
+            x_factor=0.2, y_factor=0.2, bounding_box_format="rel_xyxy"
+        )
+
+        @tf.function
+        def augment(x, y):
+            return layer({"images": x, "bounding_boxes": y})
+
+        outputs = augment(xs, ys)
+        xs, ys = outputs["images"], outputs["bounding_boxes"]
+
+        # None of the individual values should still be close to 1 or 0
+        self.assertNotAllClose(xs, 1.0)
+        self.assertNotAllClose(xs, 2.0)
+
+        # No labels should still be close to their originals
+        self.assertNotAllClose(ys, 1.0)
+        self.assertNotAllClose(ys, 0.0)
+
+    def test_no_augmentation(self):
+        """test for no image and bbox augmenation when x_factor,y_factor is 0,0"""
+        xs = tf.cast(
+            tf.stack(
+                [2 * tf.ones((4, 4, 3)), tf.ones((4, 4, 3))],
+                axis=0,
+            ),
+            tf.float32,
+        )
+        ys = tf.cast(
+            tf.stack(
+                [
+                    tf.constant([[0.3, 0.4, 0.5, 0.6], [0.9, 0.8, 1.0, 1.0]]),
+                    tf.constant([[0.3, 0.4, 0.5, 0.6], [0.9, 0.8, 1.0, 1.0]]),
+                ],
+                axis=0,
+            ),
+            tf.float32,
+        )
+        ys = bounding_box.add_class_id(ys)
+        layer = preprocessing.RandomShear(
+            x_factor=0, y_factor=0, bounding_box_format="rel_xyxy"
+        )
+        outputs = layer({"images": xs, "bounding_boxes": ys})
+        output_xs, output_ys = outputs["images"], outputs["bounding_boxes"]
+        self.assertAllEqual(xs, output_xs)
+        self.assertAllEqual(ys, output_ys)
+
+    def test_bounding_box_x_augmentation(self):
+        xs = tf.cast(
+            tf.stack(
+                [2 * tf.ones((4, 4, 3)), tf.ones((4, 4, 3))],
+                axis=0,
+            ),
+            tf.float32,
+        )
+        ys = tf.cast(
+            tf.stack(
+                [
+                    tf.constant([[0.3, 0.4, 0.5, 0.6], [0.4, 0.8, 1.0, 1.0]]),
+                    tf.constant([[0.3, 0.4, 0.5, 0.6], [0.4, 0.8, 1.0, 1.0]]),
+                ],
+                axis=0,
+            ),
+            tf.float32,
+        )
+        ys = bounding_box.add_class_id(ys)
+        layer = preprocessing.RandomShear(
+            x_factor=0.5, y_factor=0, bounding_box_format="rel_xyxy"
+        )
+        outputs = layer({"images": xs, "bounding_boxes": ys})
+        _, output_ys = outputs["images"], outputs["bounding_boxes"]
+
+        # assert ys are unchanged
+        self.assertAllEqual(ys[..., 1], output_ys[..., 1])
+        self.assertAllEqual(ys[..., 3], output_ys[..., 3])
+
+        # assert xs are changed
+        self.assertNotAllClose(ys[..., 0], output_ys[..., 0])
+        self.assertNotAllClose(ys[..., 2], output_ys[..., 2])
+
+    def test_bounding_box_y_augmentation(self):
+        xs = tf.cast(
+            tf.stack(
+                [2 * tf.ones((4, 4, 3)), tf.ones((4, 4, 3))],
+                axis=0,
+            ),
+            tf.float32,
+        )
+        ys = tf.cast(
+            tf.stack(
+                [
+                    tf.constant([[0.3, 0.4, 0.5, 0.6], [0.9, 0.2, 1.0, 1.0]]),
+                    tf.constant([[0.3, 0.4, 0.5, 0.6], [0.9, 0.2, 1.0, 1.0]]),
+                ],
+                axis=0,
+            ),
+            tf.float32,
+        )
+        ys = bounding_box.add_class_id(ys)
+        layer = preprocessing.RandomShear(
+            x_factor=0, y_factor=0.5, bounding_box_format="rel_xyxy"
+        )
+        outputs = layer({"images": xs, "bounding_boxes": ys})
+        _, output_ys = outputs["images"], outputs["bounding_boxes"]
+        self.assertAllEqual(ys[..., 0], output_ys[..., 0])
+        self.assertNotAllClose(ys[..., 1], output_ys[..., 1])
+        self.assertAllEqual(ys[..., 2], output_ys[..., 2])
+        self.assertNotAllClose(ys[..., 3], output_ys[..., 3])
+
+    def test_rel_xyxy(self):
+        """test for shear bbox augmentation for relative xyxy bbox input"""
+        xs = tf.cast(
+            tf.stack(
+                [2 * tf.ones((4, 4, 3)), tf.ones((4, 4, 3))],
+                axis=0,
+            ),
+            tf.float32,
+        )
+        ys = tf.cast(
+            tf.stack(
+                [
+                    tf.constant([[0.3, 0.4, 0.5, 0.6], [0.9, 0.8, 1.0, 1.0]]),
+                    tf.constant([[0.3, 0.4, 0.5, 0.6], [0.9, 0.8, 1.0, 1.0]]),
+                ],
+                axis=0,
+            ),
+            tf.float32,
+        )
+        ys = bounding_box.add_class_id(ys)
+        layer = preprocessing.RandomShear(
+            x_factor=0, y_factor=0, bounding_box_format="rel_xyxy"
+        )
+        outputs = layer({"images": xs, "bounding_boxes": ys})
+        _, output_ys = outputs["images"], outputs["bounding_boxes"]
+        self.assertAllEqual(ys, output_ys)
+
+    def test_xyxy(self):
+        """test for shear bbox augmentation for xyxy format"""
+        xs = tf.cast(
+            tf.stack(
+                [2 * tf.ones((100, 100, 3)), tf.ones((100, 100, 3))],
+                axis=0,
+            ),
+            tf.float32,
+        )
+        ys = tf.cast(
+            tf.stack(
+                [
+                    tf.constant([[10.0, 20.0, 40.0, 50.0], [12.0, 22.0, 42.0, 54.0]]),
+                    tf.constant([[10.0, 20.0, 40.0, 50.0], [12.0, 22.0, 42.0, 54.0]]),
+                ],
+                axis=0,
+            ),
+            tf.float32,
+        )
+        ys = bounding_box.add_class_id(ys)
+        layer = preprocessing.RandomShear(
+            x_factor=0, y_factor=0, bounding_box_format="xyxy"
+        )
+        outputs = layer({"images": xs, "bounding_boxes": ys})
+        _, output_ys = outputs["images"], outputs["bounding_boxes"]
+        self.assertAllClose(ys, output_ys)
+
+    def test_clip_bounding_box(self):
+        """test for bbox clipping to image width and height"""
+        xs = tf.cast(
+            tf.stack(
+                [2 * tf.ones((4, 4, 3)), tf.ones((4, 4, 3))],
+                axis=0,
+            ),
+            tf.float32,
+        )
+        ys = tf.cast(
+            tf.stack(
+                [
+                    tf.constant([[0.0, 0.0, 40.0, 50.0], [0.0, 0.0, 42.0, 54.0]]),
+                    tf.constant([[0.0, 0.0, 40.0, 50.0], [0.0, 0.0, 42.0, 54.0]]),
+                ],
+                axis=0,
+            ),
+            tf.float32,
+        )
+        ys = bounding_box.add_class_id(ys)
+        ground_truth = tf.cast(
+            tf.stack(
+                [
+                    tf.constant([[0, 0, 4, 4, 0], [0, 0, 4, 4, 0]]),
+                    tf.constant([[0, 0, 4, 4, 0], [0, 0, 4, 4, 0]]),
+                ],
+                axis=0,
+            ),
+            tf.float32,
+        )
+        layer = preprocessing.RandomShear(
+            x_factor=0, y_factor=0, bounding_box_format="xyxy"
+        )
+        outputs = layer({"images": xs, "bounding_boxes": ys})
+        _, output_ys = outputs["images"], outputs["bounding_boxes"]
+        self.assertAllEqual(ground_truth, output_ys)
+
+    def test_dtype(self):
+        """test for output dtype is returned as standardize dtype"""
+        xs = tf.cast(
+            tf.stack(
+                [2 * tf.ones((4, 4, 3)), tf.ones((4, 4, 3))],
+                axis=0,
+            ),
+            tf.float32,
+        )
+        ys = tf.cast(
+            tf.stack(
+                [
+                    tf.constant([[10.0, 20.0, 40.0, 50.0], [12.0, 22.0, 42.0, 54.0]]),
+                    tf.constant([[10.0, 20.0, 40.0, 50.0], [12.0, 22.0, 42.0, 54.0]]),
+                ],
+                axis=0,
+            ),
+            tf.float32,
+        )
+        ys = bounding_box.add_class_id(ys)
+        layer = preprocessing.RandomShear(
+            x_factor=0, y_factor=0, bounding_box_format="xyxy"
+        )
+        outputs = layer({"images": xs, "bounding_boxes": ys})
+        _, output_ys = outputs["images"], outputs["bounding_boxes"]
+        self.assertEqual(layer.compute_dtype, output_ys.dtype)
+
+    def test_output_values(self):
+        """test to verify augmented bounding box output coordinate"""
+        xs = tf.cast(
+            tf.stack(
+                [2 * tf.ones((100, 100, 3)), tf.zeros((100, 100, 3))],
+                axis=0,
+            ),
+            tf.float32,
+        )
+        ys = tf.cast(
+            tf.stack(
+                [
+                    tf.constant([[10.0, 20.0, 40.0, 50.0], [12.0, 22.0, 42.0, 54.0]]),
+                    tf.constant([[10.0, 20.0, 40.0, 50.0], [12.0, 22.0, 42.0, 54.0]]),
+                ],
+                axis=0,
+            ),
+            tf.float32,
+        )
+        ys = bounding_box.add_class_id(ys)
+        true_ys = tf.cast(
+            tf.stack(
+                [
+                    tf.constant(
+                        [
+                            [7.60, 20.43, 39.04, 51.79, 0.0],
+                            [9.41, 22.52, 40.94, 55.88, 0.0],
+                        ]
+                    ),
+                    tf.constant(
+                        [
+                            [13.68, 22.51, 49.20, 59.05, 0],
+                            [16.04, 24.95, 51.940, 63.56, 0],
+                        ]
+                    ),
+                ],
+                axis=0,
+            ),
+            tf.float32,
+        )
+        layer = preprocessing.RandomShear(
+            x_factor=0.2, y_factor=0.2, bounding_box_format="xyxy", seed=1
+        )
+        outputs = layer({"images": xs, "bounding_boxes": ys})
+        _, output_ys = outputs["images"], outputs["bounding_boxes"]
+        self.assertAllClose(true_ys, output_ys, rtol=1e-02, atol=1e-03)

Keras/keras-cv/keras_cv/layers/preprocessing/randomly_zoomed_crop.py

+# Copyright 2022 The KerasCV Authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     https://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import tensorflow as tf
+
+from keras_cv import core
+from keras_cv.layers.preprocessing.base_image_augmentation_layer import (
+    BaseImageAugmentationLayer,
+)
+from keras_cv.utils import preprocessing
+
+
+@tf.keras.utils.register_keras_serializable(package="keras_cv")
+class RandomlyZoomedCrop(BaseImageAugmentationLayer):
+    """Randomly crops a part of an image and zooms it by a provided amount size.
+
+    This implementation takes a distortion-oriented approach, which means the
+    amount of distortion in the image is proportional to the `zoom_factor`
+    argument. To do this, we first sample a random value for `zoom_factor` and
+    `aspect_ratio_factor`. Further we deduce a `crop_size` which abides by the
+    calculated aspect ratio. Finally we do the actual cropping operation and
+    resize the image to `(height, width)`.
+
+    Args:
+        height: The height of the output shape.
+        width: The width of the output shape.
+        zoom_factor: A tuple of two floats, ConstantFactorSampler or
+            UniformFactorSampler. Represents the area relative to the original image
+            of the cropped image before resizing it to `(height, width)`.
+        aspect_ratio_factor: A tuple of two floats, ConstantFactorSampler or
+            UniformFactorSampler. Aspect ratio means the ratio of width to
+            height of the cropped image. In the context of this layer, the aspect ratio
+            sampled represents a value to distort the aspect ratio by.
+            Represents the lower and upper bound for the aspect ratio of the
+            cropped image before resizing it to `(height, width)`.  For most tasks, this
+            should be `(3/4, 4/3)`.  To perform a no-op provide the value `(1.0, 1.0)`.
+        interpolation: (Optional) A string specifying the sampling method for
+            resizing. Defaults to "bilinear".
+        seed: (Optional) Used to create a random seed. Defaults to None.
+    """
+
+    def __init__(
+        self,
+        height,
+        width,
+        zoom_factor,
+        aspect_ratio_factor,
+        interpolation="bilinear",
+        seed=None,
+        **kwargs,
+    ):
+        super().__init__(seed=seed, **kwargs)
+        self.height = height
+        self.width = width
+        self.aspect_ratio_factor = preprocessing.parse_factor(
+            aspect_ratio_factor,
+            min_value=0.0,
+            max_value=None,
+            param_name="aspect_ratio_factor",
+            seed=seed,
+        )
+        self.zoom_factor = preprocessing.parse_factor(
+            zoom_factor,
+            min_value=0.0,
+            max_value=None,
+            param_name="zoom_factor",
+            seed=seed,
+        )
+
+        self._check_class_arguments(height, width, zoom_factor, aspect_ratio_factor)
+
+        self.interpolation = interpolation
+        self.seed = seed
+
+    def get_random_transformation(
+        self, image=None, label=None, bounding_box=None, **kwargs
+    ):
+        zoom_factor = self.zoom_factor()
+        aspect_ratio = self.aspect_ratio_factor()
+
+        original_height = tf.cast(tf.shape(image)[-3], tf.float32)
+        original_width = tf.cast(tf.shape(image)[-2], tf.float32)
+
+        crop_size = (
+            tf.round(self.height / zoom_factor),
+            tf.round(self.width / zoom_factor),
+        )
+
+        new_height = crop_size[0] / tf.sqrt(aspect_ratio)
+
+        new_width = crop_size[1] * tf.sqrt(aspect_ratio)
+
+        height_offset = self._random_generator.random_uniform(
+            (),
+            minval=tf.minimum(0.0, original_height - new_height),
+            maxval=tf.maximum(0.0, original_height - new_height),
+            dtype=tf.float32,
+        )
+
+        width_offset = self._random_generator.random_uniform(
+            (),
+            minval=tf.minimum(0.0, original_width - new_width),
+            maxval=tf.maximum(0.0, original_width - new_width),
+            dtype=tf.float32,
+        )
+
+        new_height = new_height / original_height
+        new_width = new_width / original_width
+
+        height_offset = height_offset / original_height
+        width_offset = width_offset / original_width
+
+        return (new_height, new_width, height_offset, width_offset)
+
+    def call(self, inputs, training=True):
+        if training:
+            return super().call(inputs, training)
+        else:
+            inputs = self._ensure_inputs_are_compute_dtype(inputs)
+            inputs, meta_data = self._format_inputs(inputs)
+            output = inputs
+            # self._resize() returns valid results for both batched and
+            # unbatched
+            output["images"] = self._resize(inputs["images"])
+
+            return self._format_output(output, meta_data)
+
+    def augment_image(self, image, transformation, **kwargs):
+        image_shape = tf.shape(image)
+
+        height = tf.cast(image_shape[-3], tf.float32)
+        width = tf.cast(image_shape[-2], tf.float32)
+
+        image = tf.expand_dims(image, axis=0)
+        new_height, new_width, height_offset, width_offset = transformation
+
+        transform = RandomlyZoomedCrop._format_transform(
+            [
+                new_width,
+                0.0,
+                width_offset * width,
+                0.0,
+                new_height,
+                height_offset * height,
+                0.0,
+                0.0,
+            ]
+        )
+
+        image = preprocessing.transform(
+            images=image,
+            transforms=transform,
+            output_shape=(self.height, self.width),
+            interpolation=self.interpolation,
+            fill_mode="reflect",
+        )
+
+        return tf.squeeze(image, axis=0)
+
+    @staticmethod
+    def _format_transform(transform):
+        transform = tf.convert_to_tensor(transform, dtype=tf.float32)
+        return transform[tf.newaxis]
+
+    def _resize(self, image):
+        outputs = tf.keras.preprocessing.image.smart_resize(
+            image, (self.height, self.width)
+        )
+        # smart_resize will always output float32, so we need to re-cast.
+        return tf.cast(outputs, self.compute_dtype)
+
+    def _check_class_arguments(self, height, width, zoom_factor, aspect_ratio_factor):
+        if not isinstance(height, int):
+            raise ValueError("`height` must be an integer. Received height={height}")
+
+        if not isinstance(width, int):
+            raise ValueError("`width` must be an integer. Received width={width}")
+
+        if (
+            not isinstance(zoom_factor, (tuple, list, core.FactorSampler))
+            or isinstance(zoom_factor, float)
+            or isinstance(zoom_factor, int)
+        ):
+            raise ValueError(
+                "`zoom_factor` must be tuple of two positive floats"
+                " or keras_cv.core.FactorSampler instance. Received "
+                f"zoom_factor={zoom_factor}"
+            )
+
+        if (
+            not isinstance(aspect_ratio_factor, (tuple, list, core.FactorSampler))
+            or isinstance(aspect_ratio_factor, float)
+            or isinstance(aspect_ratio_factor, int)
+        ):
+            raise ValueError(
+                "`aspect_ratio_factor` must be tuple of two positive floats or "
+                "keras_cv.core.FactorSampler instance. Received "
+                f"aspect_ratio_factor={aspect_ratio_factor}"
+            )
+
+    def augment_target(self, augment_target, **kwargs):
+        return augment_target
+
+    def get_config(self):
+        config = super().get_config()
+        config.update(
+            {
+                "height": self.height,
+                "width": self.width,
+                "zoom_factor": self.zoom_factor,
+                "aspect_ratio_factor": self.aspect_ratio_factor,
+                "interpolation": self.interpolation,
+                "seed": self.seed,
+            }
+        )
+        return config
+
+    def _crop_and_resize(self, image, transformation, method=None):
+        image = tf.expand_dims(image, axis=0)
+        boxes = transformation
+
+        # See bit.ly/tf_crop_resize for more details
+        augmented_image = tf.image.crop_and_resize(
+            image,  # image shape: [B, H, W, C]
+            boxes,  # boxes: (1, 4) in this case; represents area
+            # to be cropped from the original image
+            [0],  # box_indices: maps boxes to images along batch axis
+            # [0] since there is only one image
+            (self.height, self.width),  # output size
+            method=method or self.interpolation,
+        )
+
+        return tf.squeeze(augmented_image, axis=0)

Keras/keras-cv/keras_cv/layers/preprocessing/randomly_zoomed_crop_test.py

+# Copyright 2022 The KerasCV Authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     https://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+import tensorflow as tf
+from absl.testing import parameterized
+
+from keras_cv.layers import preprocessing
+
+
+class RandomlyZoomedCropTest(tf.test.TestCase, parameterized.TestCase):
+    height, width = 300, 300
+    batch_size = 4
+    target_size = (224, 224)
+    seed = 42
+
+    def test_train_augments_image(self):
+        # Checks if original and augmented images are different
+
+        input_image_shape = (self.batch_size, self.height, self.width, 3)
+        image = tf.random.uniform(shape=input_image_shape, seed=self.seed)
+
+        layer = preprocessing.RandomlyZoomedCrop(
+            height=self.target_size[0],
+            width=self.target_size[1],
+            aspect_ratio_factor=(3 / 4, 4 / 3),
+            zoom_factor=(0.8, 1.0),
+            seed=self.seed,
+        )
+        output = layer(image, training=True)
+
+        input_image_resized = tf.image.resize(image, self.target_size)
+
+        self.assertNotAllClose(output, input_image_resized)
+
+    def test_grayscale(self):
+        input_image_shape = (self.batch_size, self.height, self.width, 1)
+        image = tf.random.uniform(shape=input_image_shape)
+
+        layer = preprocessing.RandomlyZoomedCrop(
+            height=self.target_size[0],
+            width=self.target_size[1],
+            aspect_ratio_factor=(3 / 4, 4 / 3),
+            zoom_factor=(0.8, 1.0),
+        )
+        output = layer(image, training=True)
+
+        input_image_resized = tf.image.resize(image, self.target_size)
+
+        self.assertAllEqual(output.shape, (4, 224, 224, 1))
+        self.assertNotAllClose(output, input_image_resized)
+
+    def test_preserves_image(self):
+        image_shape = (self.batch_size, self.height, self.width, 3)
+        image = tf.random.uniform(shape=image_shape)
+
+        layer = preprocessing.RandomlyZoomedCrop(
+            height=self.target_size[0],
+            width=self.target_size[1],
+            aspect_ratio_factor=(3 / 4, 4 / 3),
+            zoom_factor=(0.8, 1.0),
+        )
+
+        input_resized = tf.image.resize(image, self.target_size)
+        output = layer(image, training=False)
+
+        self.assertAllClose(output, input_resized)
+
+    @parameterized.named_parameters(
+        ("Not tuple or list", dict()),
+        ("Length not equal to 2", [1, 2, 3]),
+        ("Members not int", (2.3, 4.5)),
+        ("Single float", 1.5),
+    )
+    def test_height_errors(self, height):
+        with self.assertRaisesRegex(
+            ValueError,
+            "`height` must be an integer. Received height=(.*)",
+        ):
+            _ = preprocessing.RandomlyZoomedCrop(
+                height=height,
+                width=100,
+                aspect_ratio_factor=(3 / 4, 4 / 3),
+                zoom_factor=(0.8, 1.0),
+            )
+
+    @parameterized.named_parameters(
+        ("Not tuple or list", dict()),
+        ("Length not equal to 2", [1, 2, 3]),
+        ("Members not int", (2.3, 4.5)),
+        ("Single float", 1.5),
+    )
+    def test_width_errors(self, width):
+        with self.assertRaisesRegex(
+            ValueError,
+            "`width` must be an integer. Received width=(.*)",
+        ):
+            _ = preprocessing.RandomlyZoomedCrop(
+                height=100,
+                width=width,
+                aspect_ratio_factor=(3 / 4, 4 / 3),
+                zoom_factor=(0.8, 1.0),
+            )
+
+    @parameterized.named_parameters(
+        ("Single integer", 5),
+        ("Single float", 5.0),
+    )
+    def test_aspect_ratio_factor_errors(self, aspect_ratio_factor):
+        with self.assertRaisesRegex(
+            ValueError,
+            "`aspect_ratio_factor` must be tuple of two positive floats or "
+            "keras_cv.core.FactorSampler instance. Received aspect_ratio_factor=(.*)",
+        ):
+            _ = preprocessing.RandomlyZoomedCrop(
+                height=self.target_size[0],
+                width=self.target_size[1],
+                aspect_ratio_factor=aspect_ratio_factor,
+                zoom_factor=(0.8, 1.0),
+            )
+
+    @parameterized.named_parameters(
+        ("Single integer", 5),
+        ("Single float", 5.0),
+    )
+    def test_zoom_factor_errors(self, zoom_factor):
+        with self.assertRaisesRegex(
+            ValueError,
+            "`zoom_factor` must be tuple of two positive floats"
+            " or keras_cv.core.FactorSampler instance. Received "
+            "zoom_factor=(.*)",
+        ):
+            _ = preprocessing.RandomlyZoomedCrop(
+                height=self.target_size[0],
+                width=self.target_size[1],
+                aspect_ratio_factor=(3 / 4, 4 / 3),
+                zoom_factor=zoom_factor,
+            )

Keras/keras-cv/keras_cv/layers/preprocessing/solarization.py

+# Copyright 2022 The KerasCV Authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     https://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+import tensorflow as tf
+
+from keras_cv.layers.preprocessing.base_image_augmentation_layer import (
+    BaseImageAugmentationLayer,
+)
+from keras_cv.utils import preprocessing
+
+
+@tf.keras.utils.register_keras_serializable(package="keras_cv")
+class Solarization(BaseImageAugmentationLayer):
+    """Applies (max_value - pixel + min_value) for each pixel in the image.
+
+    When created without `threshold` parameter, the layer performs solarization to
+    all values. When created with specified `threshold` the layer only augments
+    pixels that are above the `threshold` value
+
+    Reference:
+    - [AutoAugment: Learning Augmentation Policies from Data](
+        https://arxiv.org/abs/1805.09501
+    )
+    - [RandAugment](https://arxiv.org/pdf/1909.13719.pdf)
+
+    Args:
+        value_range: a tuple or a list of two elements. The first value represents
+            the lower bound for values in passed images, the second represents the
+            upper bound. Images passed to the layer should have values within
+            `value_range`.
+        addition_factor: (Optional)  A tuple of two floats, a single float or a
+            `keras_cv.FactorSampler`. For each augmented image a value is sampled
+            from the provided range. If a float is passed, the range is interpreted as
+            `(0, addition_factor)`. If specified, this value is added to each pixel
+            before solarization and thresholding.  The addition value should be scaled
+            according to the value range (0, 255). Defaults to 0.0.
+        threshold_factor: (Optional)  A tuple of two floats, a single float or a
+            `keras_cv.FactorSampler`. For each augmented image a value is sampled
+            from the provided range. If a float is passed, the range is interpreted as
+            `(0, threshold_factor)`. If specified, only pixel values above this
+            threshold will be solarized.
+        seed: Integer. Used to create a random seed.
+
+    Usage:
+    ```python
+    (images, labels), _ = tf.keras.datasets.cifar10.load_data()
+    print(images[0, 0, 0])
+    # [59 62 63]
+    # Note that images are Tensor with values in the range [0, 255]
+    solarization = Solarization()
+    images = solarization(images)
+    print(images[0, 0, 0])
+    # [196, 193, 192]
+    ```
+
+    Call arguments:
+        images: Tensor of type int or float, with pixels in
+            range [0, 255] and shape [batch, height, width, channels]
+            or [height, width, channels].
+    """
+
+    def __init__(
+        self,
+        value_range,
+        addition_factor=0.0,
+        threshold_factor=0.0,
+        seed=None,
+        **kwargs
+    ):
+        super().__init__(seed=seed, **kwargs)
+        self.seed = seed
+        self.addition_factor = preprocessing.parse_factor(
+            addition_factor, max_value=255, seed=seed, param_name="addition_factor"
+        )
+        self.threshold_factor = preprocessing.parse_factor(
+            threshold_factor, max_value=255, seed=seed, param_name="threshold_factor"
+        )
+        self.value_range = value_range
+
+    def get_random_transformation(self, **kwargs):
+        return (self.addition_factor(), self.threshold_factor())
+
+    def augment_image(self, image, transformation=None, **kwargs):
+        (addition, threshold) = transformation
+        image = preprocessing.transform_value_range(
+            image, original_range=self.value_range, target_range=(0, 255)
+        )
+        result = image + addition
+        result = tf.clip_by_value(result, 0, 255)
+        result = tf.where(result < threshold, result, 255 - result)
+        result = preprocessing.transform_value_range(
+            result, original_range=(0, 255), target_range=self.value_range
+        )
+        return result
+
+    def augment_bounding_boxes(self, bounding_boxes, **kwargs):
+        return bounding_boxes
+
+    def augment_label(self, label, transformation=None, **kwargs):
+        return label
+
+    def augment_segmentation_mask(self, segmentation_mask, transformation, **kwargs):
+        return segmentation_mask
+
+    def get_config(self):
+        config = {
+            "threshold_factor": self.threshold_factor,
+            "addition_factor": self.addition_factor,
+            "value_range": self.value_range,
+            "seed": self.seed,
+        }
+        base_config = super().get_config()
+        return dict(list(base_config.items()) + list(config.items()))

Keras/keras-cv/keras_cv/layers/preprocessing/solarization_test.py

+# Copyright 2022 The KerasCV Authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     https://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+import tensorflow as tf
+from absl.testing import parameterized
+
+from keras_cv.layers.preprocessing.solarization import Solarization
+
+
+class SolarizationTest(tf.test.TestCase, parameterized.TestCase):
+    @parameterized.named_parameters(
+        ("0_255", 0, 255),
+        ("64_191", 64, 191),
+        ("127_128", 127, 128),
+        ("191_64", 191, 64),
+        ("255_0", 255, 0),
+    )
+    def test_output_values(self, input_value, expected_value):
+        solarization = Solarization(value_range=(0, 255))
+
+        self._test_input_output(
+            layer=solarization,
+            input_value=input_value,
+            expected_value=expected_value,
+            dtype=tf.uint8,
+        )
+
+    @parameterized.named_parameters(
+        ("0_245", 0, 245),
+        ("255_0", 255, 0),
+    )
+    def test_solarization_with_addition(self, input_value, output_value):
+        solarization = Solarization(addition_factor=(10.0, 10.0), value_range=(0, 255))
+        self._test_input_output(
+            layer=solarization,
+            input_value=input_value,
+            expected_value=output_value,
+            dtype=tf.float32,
+        )
+
+    @parameterized.named_parameters(
+        ("0_0", 0, 0),
+        ("64_64", 64, 64),
+        ("127_127", 127, 127),
+        ("191_64", 191, 64),
+        ("255_0", 255, 0),
+    )
+    def test_only_values_above_threshold_are_solarized(self, input_value, output_value):
+        solarization = Solarization(threshold_factor=(128, 128), value_range=(0, 255))
+
+        self._test_input_output(
+            layer=solarization,
+            input_value=input_value,
+            expected_value=output_value,
+            dtype=tf.uint8,
+        )
+
+    def _test_input_output(self, layer, input_value, expected_value, dtype):
+        input = tf.ones(shape=(2, 224, 224, 3), dtype=dtype) * input_value
+        expected_output = tf.clip_by_value(
+            (
+                tf.ones(shape=(2, 224, 224, 3), dtype=layer.compute_dtype)
+                * expected_value
+            ),
+            0,
+            255,
+        )
+
+        output = layer(input)
+
+        self.assertAllClose(output, expected_output)

Keras/keras-cv/keras_cv/layers/preprocessing/with_labels_test.py

+# Copyright 2022 The KerasCV Authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     https://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+import tensorflow as tf
+from absl.testing import parameterized
+
+from keras_cv.layers import preprocessing
+
+TEST_CONFIGURATIONS = [
+    ("AutoContrast", preprocessing.AutoContrast, {"value_range": (0, 255)}),
+    ("ChannelShuffle", preprocessing.ChannelShuffle, {}),
+    ("Equalization", preprocessing.Equalization, {"value_range": (0, 255)}),
+    (
+        "RandomCropAndResize",
+        preprocessing.RandomCropAndResize,
+        {
+            "target_size": (224, 224),
+            "crop_area_factor": (0.8, 1.0),
+            "aspect_ratio_factor": (3 / 4, 4 / 3),
+        },
+    ),
+    (
+        "RandomlyZoomedCrop",
+        preprocessing.RandomlyZoomedCrop,
+        {
+            "height": 224,
+            "width": 224,
+            "zoom_factor": (0.8, 1.0),
+            "aspect_ratio_factor": (3 / 4, 4 / 3),
+        },
+    ),
+    ("Grayscale", preprocessing.Grayscale, {}),
+    ("GridMask", preprocessing.GridMask, {}),
+    (
+        "Posterization",
+        preprocessing.Posterization,
+        {"bits": 3, "value_range": (0, 255)},
+    ),
+    (
+        "RandomColorDegeneration",
+        preprocessing.RandomColorDegeneration,
+        {"factor": 0.5},
+    ),
+    (
+        "RandomCutout",
+        preprocessing.RandomCutout,
+        {"height_factor": 0.2, "width_factor": 0.2},
+    ),
+    (
+        "RandomHue",
+        preprocessing.RandomHue,
+        {"factor": 0.5, "value_range": (0, 255)},
+    ),
+    (
+        "RandomChannelShift",
+        preprocessing.RandomChannelShift,
+        {"value_range": (0, 255), "factor": 0.5},
+    ),
+    (
+        "RandomColorJitter",
+        preprocessing.RandomColorJitter,
+        {
+            "value_range": (0, 255),
+            "brightness_factor": (-0.2, 0.5),
+            "contrast_factor": (0.5, 0.9),
+            "saturation_factor": (0.5, 0.9),
+            "hue_factor": (0.5, 0.9),
+            "seed": 1,
+        },
+    ),
+    (
+        "RandomGaussianBlur",
+        preprocessing.RandomGaussianBlur,
+        {"kernel_size": 3, "factor": (0.0, 3.0)},
+    ),
+    ("RandomJpegQuality", preprocessing.RandomJpegQuality, {"factor": (75, 100)}),
+    ("RandomSaturation", preprocessing.RandomSaturation, {"factor": 0.5}),
+    (
+        "RandomSharpness",
+        preprocessing.RandomSharpness,
+        {"factor": 0.5, "value_range": (0, 255)},
+    ),
+    ("RandomShear", preprocessing.RandomShear, {"x_factor": 0.3, "x_factor": 0.3}),
+    ("Solarization", preprocessing.Solarization, {"value_range": (0, 255)}),
+]
+
+
+class WithLabelsTest(tf.test.TestCase, parameterized.TestCase):
+    @parameterized.named_parameters(
+        *TEST_CONFIGURATIONS,
+        ("CutMix", preprocessing.CutMix, {}),
+        ("Mosaic", preprocessing.Mosaic, {}),
+    )
+    def test_can_run_with_labels(self, layer_cls, init_args):
+        layer = layer_cls(**init_args)
+
+        img = tf.random.uniform(
+            shape=(3, 512, 512, 3), minval=0, maxval=1, dtype=tf.float32
+        )
+        labels = tf.ones((3,), dtype=tf.float32)
+
+        inputs = {"images": img, "labels": labels}
+        outputs = layer(inputs)
+
+        self.assertIn("labels", outputs)
+
+    # this has to be a separate test case to exclude CutMix, MixUp, Mosaic etc.
+    @parameterized.named_parameters(*TEST_CONFIGURATIONS)
+    def test_can_run_with_labels_single_image(self, layer_cls, init_args):
+        layer = layer_cls(**init_args)
+        img = tf.random.uniform(
+            shape=(512, 512, 3), minval=0, maxval=1, dtype=tf.float32
+        )
+        labels = tf.ones((), dtype=tf.float32)
+
+        inputs = {"images": img, "labels": labels}
+        outputs = layer(inputs)
+
+        self.assertIn("labels", outputs)

Keras/keras-cv/keras_cv/layers/preprocessing/with_segmentation_masks_test.py

+# Copyright 2022 The KerasCV Authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     https://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+import tensorflow as tf
+from absl.testing import parameterized
+
+from keras_cv.layers import preprocessing
+
+TEST_CONFIGURATIONS = [
+    ("AutoContrast", preprocessing.AutoContrast, {"value_range": (0, 255)}),
+    ("ChannelShuffle", preprocessing.ChannelShuffle, {}),
+    ("Equalization", preprocessing.Equalization, {"value_range": (0, 255)}),
+    ("Grayscale", preprocessing.Grayscale, {}),
+    ("GridMask", preprocessing.GridMask, {}),
+    (
+        "Posterization",
+        preprocessing.Posterization,
+        {"bits": 3, "value_range": (0, 255)},
+    ),
+    (
+        "RandomColorDegeneration",
+        preprocessing.RandomColorDegeneration,
+        {"factor": 0.5},
+    ),
+    (
+        "RandomHue",
+        preprocessing.RandomHue,
+        {"factor": 0.5, "value_range": (0, 255)},
+    ),
+    (
+        "RandomChannelShift",
+        preprocessing.RandomChannelShift,
+        {"value_range": (0, 255), "factor": 0.5},
+    ),
+    (
+        "RandomColorJitter",
+        preprocessing.RandomColorJitter,
+        {
+            "value_range": (0, 255),
+            "brightness_factor": (-0.2, 0.5),
+            "contrast_factor": (0.5, 0.9),
+            "saturation_factor": (0.5, 0.9),
+            "hue_factor": (0.5, 0.9),
+            "seed": 1,
+        },
+    ),
+    (
+        "RandomGaussianBlur",
+        preprocessing.RandomGaussianBlur,
+        {"kernel_size": 3, "factor": (0.0, 3.0)},
+    ),
+    ("RandomJpegQuality", preprocessing.RandomJpegQuality, {"factor": (75, 100)}),
+    ("RandomSaturation", preprocessing.RandomSaturation, {"factor": 0.5}),
+    (
+        "RandomSharpness",
+        preprocessing.RandomSharpness,
+        {"factor": 0.5, "value_range": (0, 255)},
+    ),
+    ("Solarization", preprocessing.Solarization, {"value_range": (0, 255)}),
+]
+
+
+class WithSegmentationMasksTest(tf.test.TestCase, parameterized.TestCase):
+    @parameterized.named_parameters(*TEST_CONFIGURATIONS)
+    def test_can_run_with_segmentation_masks(self, layer_cls, init_args):
+        classes = 10
+        layer = layer_cls(**init_args)
+
+        img = tf.random.uniform(
+            shape=(3, 512, 512, 3), minval=0, maxval=1, dtype=tf.float32
+        )
+        segmentation_masks = tf.random.uniform(
+            shape=(3, 512, 512, 1), minval=0, maxval=classes, dtype=tf.int32
+        )
+
+        inputs = {"images": img, "segmentation_masks": segmentation_masks}
+        outputs = layer(inputs)
+
+        self.assertIn("segmentation_masks", outputs)
+        # This currently asserts that all layers are no-ops.
+        # When preprocessing layers are updated to mutate segmentation masks,
+        # this condition should only be asserted for no-op layers.
+        self.assertAllClose(inputs["segmentation_masks"], outputs["segmentation_masks"])
+
+    # This has to be a separate test case to exclude CutMix and MixUp
+    # (which are not yet supported for segmentation mask augmentation)
+    @parameterized.named_parameters(*TEST_CONFIGURATIONS)
+    def test_can_run_with_segmentation_mask_single_image(self, layer_cls, init_args):
+        classes = 10
+        layer = layer_cls(**init_args)
+        img = tf.random.uniform(
+            shape=(512, 512, 3), minval=0, maxval=1, dtype=tf.float32
+        )
+        segmentation_mask = tf.random.uniform(
+            shape=(512, 512, 1), minval=0, maxval=classes, dtype=tf.int32
+        )
+
+        inputs = {"images": img, "segmentation_masks": segmentation_mask}
+        outputs = layer(inputs)
+
+        self.assertIn("segmentation_masks", outputs)
+        # This currently asserts that all layers are no-ops.
+        # When preprocessing layers are updated to mutate segmentation masks,
+        # this condition should only be asserted for no-op layers.
+        self.assertAllClose(inputs["segmentation_masks"], outputs["segmentation_masks"])

Keras/keras-cv/keras_cv/layers/regularization/__init__.py

+# Copyright 2022 The KerasCV Authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     https://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from keras_cv.layers.regularization.drop_path import DropPath
+from keras_cv.layers.regularization.dropblock_2d import DropBlock2D
+from keras_cv.layers.regularization.squeeze_excite import SqueezeAndExcite2D
+from keras_cv.layers.regularization.stochastic_depth import StochasticDepth

Keras/keras-cv/keras_cv/layers/regularization/drop_path.py

+# Copyright 2022 The KerasCV Authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     https://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+import tensorflow as tf
+
+
+@tf.keras.utils.register_keras_serializable(package="keras_cv")
+class DropPath(tf.keras.__internal__.layers.BaseRandomLayer):
+    """
+    Implements the DropPath layer. DropPath randomly drops samples during training
+     with a probability of `rate`. Note that this layer drops individual samples
+    within a batch and not the entire batch. DropPath randomly drops some of the
+    individual samples from a batch, whereas StachasticDepth randomly drops the
+    entire batch.
+
+    References:
+        - [FractalNet](https://arxiv.org/abs/1605.07648v4).
+        - [rwightman/pytorch-image-models](https://github.com/rwightman/pytorch-image-models/blob/7c67d6aca992f039eece0af5f7c29a43d48c00e4/timm/models/layers/drop.py#L135)
+
+    Args:
+        rate: float, the probability of the residual branch being dropped.
+        seed: (Optional) Integer. Used to create a random seed.
+
+    Usage:
+    `DropPath` can be used in any network as follows:
+    ```python
+
+    # (...)
+    input = tf.ones((1, 3, 3, 1), dtype=tf.float32)
+    residual = tf.keras.layers.Conv2D(1, 1)(input)
+    output = keras_cv.layers.DropPath()(input)
+    # (...)
+    ```
+    """
+
+    def __init__(self, rate=0.5, seed=None, **kwargs):
+        super().__init__(seed=seed, **kwargs)
+        self.rate = rate
+        self.seed = seed
+
+    def call(self, x, training=None):
+        if self.rate == 0.0 or not training:
+            return x
+        else:
+            keep_prob = 1 - self.rate
+            drop_map_shape = (x.shape[0],) + (1,) * (len(x.shape) - 1)
+            drop_map = tf.keras.backend.random_bernoulli(
+                drop_map_shape, p=keep_prob, seed=self.seed
+            )
+            x = x / keep_prob
+            x = x * drop_map
+            return x
+
+    def get_config(self):
+        config = {"rate": self.rate, "seed": self.seed}
+        base_config = super().get_config()
+        return dict(list(base_config.items()) + list(config.items()))

Keras/keras-cv/keras_cv/layers/regularization/drop_path_test.py

+# Copyright 2022 The KerasCV Authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     https://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import tensorflow as tf
+
+from keras_cv.layers import DropPath
+
+
+class DropPathTest(tf.test.TestCase):
+    FEATURE_SHAPE = (16, 14, 14, 256)
+
+    def test_input_unchanged_in_eval_mode(self):
+        layer = DropPath(rate=0.5, seed=42)
+        inputs = tf.random.uniform(self.FEATURE_SHAPE)
+
+        outputs = layer(inputs, training=False)
+
+        self.assertAllClose(inputs, outputs)
+
+    def test_input_unchanged_with_rate_equal_to_zero(self):
+        layer = DropPath(rate=0, seed=42)
+        inputs = tf.random.uniform(self.FEATURE_SHAPE)
+
+        outputs = layer(inputs, training=True)
+
+        self.assertAllClose(inputs, outputs)
+
+    def test_input_gets_partially_zeroed_out_in_train_mode(self):
+        layer = DropPath(rate=0.2, seed=42)
+        inputs = tf.random.uniform(self.FEATURE_SHAPE)
+
+        outputs = layer(inputs, training=True)
+
+        non_zeros_inputs = tf.math.count_nonzero(inputs, dtype=tf.int32)
+        non_zeros_outputs = tf.math.count_nonzero(outputs, dtype=tf.int32)
+
+        self.assertGreaterEqual(non_zeros_inputs, non_zeros_outputs)
+
+    def test_strict_input_gets_partially_zeroed_out_in_train_mode(self):
+        layer = DropPath(rate=0.5, seed=42)
+        inputs = tf.random.uniform(self.FEATURE_SHAPE)
+
+        total_non_zero_inputs = 0
+        total_non_zero_outputs = 0
+
+        for _ in range(300):
+            outputs = layer(inputs, training=True)
+
+            non_zeros_inputs = tf.math.count_nonzero(inputs, dtype=tf.int32)
+            non_zeros_outputs = tf.math.count_nonzero(outputs, dtype=tf.int32)
+
+            total_non_zero_inputs += non_zeros_inputs
+            total_non_zero_outputs += non_zeros_outputs
+
+        self.assertAllInRange(
+            total_non_zero_outputs,
+            int(0.49 * tf.cast(total_non_zero_inputs, tf.float32)),
+            int(0.51 * tf.cast(total_non_zero_inputs, tf.float32)),
+        )