Merge branch 'dtk22.04' into 'main'

Dtk22.04

See merge request dcutoolkit/deeplearing/dlexamples_new!49

Keras/keras-cv/keras_cv/layers/preprocessing/base_image_augmentation_layer_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.layers.preprocessing.base_image_augmentation_layer import (
+    BaseImageAugmentationLayer,
+)
+
+
+class RandomAddLayer(BaseImageAugmentationLayer):
+    def __init__(self, value_range=(0.0, 1.0), fixed_value=None, **kwargs):
+        super().__init__(**kwargs)
+        self.value_range = value_range
+        self.fixed_value = fixed_value
+
+    def get_random_transformation(self, **kwargs):
+        if self.fixed_value:
+            return self.fixed_value
+        return self._random_generator.random_uniform(
+            [], minval=self.value_range[0], maxval=self.value_range[1]
+        )
+
+    def augment_image(self, image, transformation, **kwargs):
+        return image + transformation
+
+    def augment_label(self, label, transformation, **kwargs):
+        return label + transformation
+
+    def augment_bounding_boxes(self, bounding_boxes, transformation, **kwargs):
+        return bounding_boxes + transformation
+
+    def augment_keypoints(self, keypoints, transformation, **kwargs):
+        return keypoints + transformation
+
+    def augment_segmentation_mask(self, segmentation_mask, transformation, **kwargs):
+        return segmentation_mask + transformation
+
+
+class VectorizeDisabledLayer(BaseImageAugmentationLayer):
+    def __init__(self, **kwargs):
+        self.auto_vectorize = False
+        super().__init__(**kwargs)
+
+
+class BaseImageAugmentationLayerTest(tf.test.TestCase):
+    def test_augment_single_image(self):
+        add_layer = RandomAddLayer(fixed_value=2.0)
+        image = np.random.random(size=(8, 8, 3)).astype("float32")
+        output = add_layer(image)
+
+        self.assertAllClose(image + 2.0, output)
+
+    def test_augment_dict_return_type(self):
+        add_layer = RandomAddLayer(fixed_value=2.0)
+        image = np.random.random(size=(8, 8, 3)).astype("float32")
+        output = add_layer({"images": image})
+
+        self.assertIsInstance(output, dict)
+
+    def test_auto_vectorize_disabled(self):
+        vectorize_disabled_layer = VectorizeDisabledLayer()
+        self.assertFalse(vectorize_disabled_layer.auto_vectorize)
+        self.assertEqual(vectorize_disabled_layer._map_fn, tf.map_fn)
+
+    def test_augment_casts_dtypes(self):
+        add_layer = RandomAddLayer(fixed_value=2.0)
+        images = tf.ones((2, 8, 8, 3), dtype="uint8")
+        output = add_layer(images)
+
+        self.assertAllClose(tf.ones((2, 8, 8, 3), dtype="float32") * 3.0, output)
+
+    def test_augment_batch_images(self):
+        add_layer = RandomAddLayer()
+        images = np.random.random(size=(2, 8, 8, 3)).astype("float32")
+        output = add_layer(images)
+
+        diff = output - images
+        # Make sure the first image and second image get different augmentation
+        self.assertNotAllClose(diff[0], diff[1])
+
+    def test_augment_image_and_label(self):
+        add_layer = RandomAddLayer(fixed_value=2.0)
+        image = np.random.random(size=(8, 8, 3)).astype("float32")
+        label = np.random.random(size=(1,)).astype("float32")
+
+        output = add_layer({"images": image, "labels": label})
+        expected_output = {"images": image + 2.0, "labels": label + 2.0}
+        self.assertAllClose(output, expected_output)
+
+    def test_augment_image_and_target(self):
+        add_layer = RandomAddLayer(fixed_value=2.0)
+        image = np.random.random(size=(8, 8, 3)).astype("float32")
+        label = np.random.random(size=(1,)).astype("float32")
+
+        output = add_layer({"images": image, "targets": label})
+        expected_output = {"images": image + 2.0, "targets": label + 2.0}
+        self.assertAllClose(output, expected_output)
+
+    def test_augment_batch_images_and_labels(self):
+        add_layer = RandomAddLayer()
+        images = np.random.random(size=(2, 8, 8, 3)).astype("float32")
+        labels = np.random.random(size=(2, 1)).astype("float32")
+        output = add_layer({"images": images, "labels": labels})
+
+        image_diff = output["images"] - images
+        label_diff = output["labels"] - labels
+        # Make sure the first image and second image get different augmentation
+        self.assertNotAllClose(image_diff[0], image_diff[1])
+        self.assertNotAllClose(label_diff[0], label_diff[1])
+
+    def test_augment_leaves_extra_dict_entries_unmodified(self):
+        add_layer = RandomAddLayer(fixed_value=0.5)
+        images = np.random.random(size=(8, 8, 3)).astype("float32")
+        filenames = tf.constant("/path/to/first.jpg")
+        inputs = {"images": images, "filenames": filenames}
+
+        outputs = add_layer(inputs)
+
+        self.assertListEqual(list(inputs.keys()), list(outputs.keys()))
+        self.assertAllEqual(inputs["filenames"], outputs["filenames"])
+        self.assertNotAllClose(inputs["images"], outputs["images"])
+        self.assertAllEqual(inputs["images"], images)  # Assert original unchanged
+
+    def test_augment_leaves_batched_extra_dict_entries_unmodified(self):
+        add_layer = RandomAddLayer(fixed_value=0.5)
+        images = np.random.random(size=(2, 8, 8, 3)).astype("float32")
+        filenames = tf.constant(["/path/to/first.jpg", "/path/to/second.jpg"])
+        inputs = {"images": images, "filenames": filenames}
+
+        outputs = add_layer(inputs)
+
+        self.assertListEqual(list(inputs.keys()), list(outputs.keys()))
+        self.assertAllEqual(inputs["filenames"], outputs["filenames"])
+        self.assertNotAllClose(inputs["images"], outputs["images"])
+        self.assertAllEqual(inputs["images"], images)  # Assert original unchanged
+
+    def test_augment_image_and_localization_data(self):
+        add_layer = RandomAddLayer(fixed_value=2.0)
+        images = np.random.random(size=(8, 8, 3)).astype("float32")
+        bounding_boxes = np.random.random(size=(3, 5)).astype("float32")
+        keypoints = np.random.random(size=(3, 5, 2)).astype("float32")
+        segmentation_mask = np.random.random(size=(8, 8, 1)).astype("float32")
+
+        output = add_layer(
+            {
+                "images": images,
+                "bounding_boxes": bounding_boxes,
+                "keypoints": keypoints,
+                "segmentation_masks": segmentation_mask,
+            }
+        )
+
+        expected_output = {
+            "images": images + 2.0,
+            "bounding_boxes": bounding_boxes + 2.0,
+            "keypoints": keypoints + 2.0,
+            "segmentation_masks": segmentation_mask + 2.0,
+        }
+        self.assertAllClose(output, expected_output)
+
+    def test_augment_batch_image_and_localization_data(self):
+        add_layer = RandomAddLayer()
+        images = np.random.random(size=(2, 8, 8, 3)).astype("float32")
+        bounding_boxes = np.random.random(size=(2, 3, 5)).astype("float32")
+        keypoints = np.random.random(size=(2, 3, 5, 2)).astype("float32")
+        segmentation_masks = np.random.random(size=(2, 8, 8, 1)).astype("float32")
+
+        output = add_layer(
+            {
+                "images": images,
+                "bounding_boxes": bounding_boxes,
+                "keypoints": keypoints,
+                "segmentation_masks": segmentation_masks,
+            }
+        )
+
+        bounding_boxes_diff = output["bounding_boxes"] - bounding_boxes
+        keypoints_diff = output["keypoints"] - keypoints
+        segmentation_mask_diff = output["segmentation_masks"] - segmentation_masks
+        self.assertNotAllClose(bounding_boxes_diff[0], bounding_boxes_diff[1])
+        self.assertNotAllClose(keypoints_diff[0], keypoints_diff[1])
+        self.assertNotAllClose(segmentation_mask_diff[0], segmentation_mask_diff[1])
+
+        @tf.function
+        def in_tf_function(inputs):
+            return add_layer(inputs)
+
+        output = in_tf_function(
+            {
+                "images": images,
+                "bounding_boxes": bounding_boxes,
+                "keypoints": keypoints,
+                "segmentation_masks": segmentation_masks,
+            }
+        )
+
+        bounding_boxes_diff = output["bounding_boxes"] - bounding_boxes
+        keypoints_diff = output["keypoints"] - keypoints
+        segmentation_mask_diff = output["segmentation_masks"] - segmentation_masks
+        self.assertNotAllClose(bounding_boxes_diff[0], bounding_boxes_diff[1])
+        self.assertNotAllClose(keypoints_diff[0], keypoints_diff[1])
+        self.assertNotAllClose(segmentation_mask_diff[0], segmentation_mask_diff[1])
+
+    def test_augment_all_data_in_tf_function(self):
+        add_layer = RandomAddLayer()
+        images = np.random.random(size=(2, 8, 8, 3)).astype("float32")
+        bounding_boxes = np.random.random(size=(2, 3, 5)).astype("float32")
+        keypoints = np.random.random(size=(2, 3, 5, 2)).astype("float32")
+        segmentation_masks = np.random.random(size=(2, 8, 8, 1)).astype("float32")
+
+        @tf.function
+        def in_tf_function(inputs):
+            return add_layer(inputs)
+
+        output = in_tf_function(
+            {
+                "images": images,
+                "bounding_boxes": bounding_boxes,
+                "keypoints": keypoints,
+                "segmentation_masks": segmentation_masks,
+            }
+        )
+
+        bounding_boxes_diff = output["bounding_boxes"] - bounding_boxes
+        keypoints_diff = output["keypoints"] - keypoints
+        segmentation_mask_diff = output["segmentation_masks"] - segmentation_masks
+        self.assertNotAllClose(bounding_boxes_diff[0], bounding_boxes_diff[1])
+        self.assertNotAllClose(keypoints_diff[0], keypoints_diff[1])
+        self.assertNotAllClose(segmentation_mask_diff[0], segmentation_mask_diff[1])
+
+    def test_raise_error_missing_class_id(self):
+        add_layer = RandomAddLayer()
+        images = np.random.random(size=(2, 8, 8, 3)).astype("float32")
+        bounding_boxes = np.random.random(size=(2, 3, 4)).astype("float32")
+        keypoints = np.random.random(size=(2, 3, 5, 2)).astype("float32")
+        segmentation_masks = np.random.random(size=(2, 8, 8, 1)).astype("float32")
+
+        with self.assertRaisesRegex(
+            ValueError,
+            "Bounding boxes are missing class_id. If you would like to pad the "
+            "bounding boxes with class_id, use `keras_cv.bounding_box.add_"
+            "class_id`",
+        ):
+            add_layer(
+                {
+                    "images": images,
+                    "bounding_boxes": bounding_boxes,
+                    "keypoints": keypoints,
+                    "segmentation_masks": segmentation_masks,
+                }
+            )

Keras/keras-cv/keras_cv/layers/preprocessing/channel_shuffle.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,
+)
+
+
+@tf.keras.utils.register_keras_serializable(package="keras_cv")
+class ChannelShuffle(BaseImageAugmentationLayer):
+    """Shuffle channels of an input image.
+
+    Input shape:
+        The expected images should be [0-255] pixel ranges.
+        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:
+        groups: Number of groups to divide the input channels. Default 3.
+        seed: Integer. Used to create a random seed.
+
+    Call arguments:
+        inputs: Tensor representing images of shape
+            `(batch_size, width, height, channels)`, with dtype tf.float32 / tf.uint8,
+            ` or (width, height, channels)`, with dtype tf.float32 / tf.uint8
+        training: A boolean argument that determines whether the call should be run
+            in inference mode or training mode. Default: True.
+
+    Usage:
+    ```python
+    (images, labels), _ = tf.keras.datasets.cifar10.load_data()
+    channel_shuffle = keras_cv.layers.ChannelShuffle()
+    augmented_images = channel_shuffle(images)
+    ```
+    """
+
+    def __init__(self, groups=3, seed=None, **kwargs):
+        super().__init__(seed=seed, **kwargs)
+        self.groups = groups
+        self.seed = seed
+
+    def augment_image(self, image, transformation=None, **kwargs):
+        shape = tf.shape(image)
+        height, width = shape[0], shape[1]
+        num_channels = image.shape[2]
+
+        if not num_channels % self.groups == 0:
+            raise ValueError(
+                "The number of input channels should be "
+                "divisible by the number of groups."
+                f"Received: channels={num_channels}, groups={self.groups}"
+            )
+
+        channels_per_group = num_channels // self.groups
+        image = tf.reshape(image, [height, width, self.groups, channels_per_group])
+        image = tf.transpose(image, perm=[2, 0, 1, 3])
+        image = tf.random.shuffle(image, seed=self.seed)
+        image = tf.transpose(image, perm=[1, 2, 3, 0])
+        image = tf.reshape(image, [height, width, num_channels])
+
+        return image
+
+    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({"groups": self.groups, "seed": self.seed})
+        return config
+
+    def compute_output_shape(self, input_shape):
+        return input_shape

Keras/keras-cv/keras_cv/layers/preprocessing/channel_shuffle_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.preprocessing.channel_shuffle import ChannelShuffle
+
+
+class ChannelShuffleTest(tf.test.TestCase):
+    def test_return_shapes(self):
+        xs = tf.ones((2, 512, 512, 3))
+
+        layer = ChannelShuffle(groups=3)
+        xs = layer(xs, training=True)
+        self.assertEqual(xs.shape, [2, 512, 512, 3])
+
+    def test_channel_shuffle_call_results_one_channel(self):
+        xs = tf.cast(
+            tf.stack(
+                [3 * tf.ones((40, 40, 1)), 2 * tf.ones((40, 40, 1))],
+                axis=0,
+            ),
+            dtype=tf.float32,
+        )
+
+        layer = ChannelShuffle(groups=1)
+        xs = layer(xs, training=True)
+        self.assertTrue(tf.math.reduce_any(xs[0] == 3.0))
+        self.assertTrue(tf.math.reduce_any(xs[1] == 2.0))
+
+    def test_channel_shuffle_call_results_multi_channel(self):
+        xs = tf.cast(
+            tf.stack(
+                [3 * tf.ones((40, 40, 20)), 2 * tf.ones((40, 40, 20))],
+                axis=0,
+            ),
+            dtype=tf.float32,
+        )
+
+        layer = ChannelShuffle(groups=5)
+        xs = layer(xs, training=True)
+        self.assertTrue(tf.math.reduce_any(xs[0] == 3.0))
+        self.assertTrue(tf.math.reduce_any(xs[1] == 2.0))
+
+    def test_non_square_image(self):
+        xs = tf.cast(
+            tf.stack(
+                [2 * tf.ones((1024, 512, 1)), tf.ones((1024, 512, 1))],
+                axis=0,
+            ),
+            dtype=tf.float32,
+        )
+
+        layer = ChannelShuffle(groups=1)
+        xs = layer(xs, training=True)
+        self.assertTrue(tf.math.reduce_any(xs[0] == 2.0))
+        self.assertTrue(tf.math.reduce_any(xs[1] == 1.0))
+
+    def test_in_tf_function(self):
+        xs = tf.cast(
+            tf.stack([2 * tf.ones((100, 100, 1)), tf.ones((100, 100, 1))], axis=0),
+            dtype=tf.float32,
+        )
+
+        layer = ChannelShuffle(groups=1)
+
+        @tf.function
+        def augment(x):
+            return layer(x, training=True)
+
+        xs = augment(xs)
+        self.assertTrue(tf.math.reduce_any(xs[0] == 2.0))
+        self.assertTrue(tf.math.reduce_any(xs[1] == 1.0))
+
+    def test_in_single_image(self):
+        xs = tf.cast(
+            tf.ones((512, 512, 1)),
+            dtype=tf.float32,
+        )
+
+        layer = ChannelShuffle(groups=1)
+        xs = layer(xs, training=True)
+        self.assertTrue(tf.math.reduce_any(xs == 1.0))

Keras/keras-cv/keras_cv/layers/preprocessing/cut_mix.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 fill_utils
+
+
+@tf.keras.utils.register_keras_serializable(package="keras_cv")
+class CutMix(BaseImageAugmentationLayer):
+    """CutMix implements the CutMix data augmentation technique.
+
+    Args:
+        alpha: Float between 0 and 1.  Inverse scale parameter for the gamma
+            distribution.  This controls the shape of the distribution from which the
+            smoothing values are sampled.  Defaults 1.0, which is a recommended value
+            when training an imagenet1k classification model.
+        seed: Integer. Used to create a random seed.
+    References:
+       - [CutMix paper]( https://arxiv.org/abs/1905.04899).
+
+    Sample usage:
+    ```python
+    (images, labels), _ = tf.keras.datasets.cifar10.load_data()
+    labels = tf.one_hot(labels.squeeze(), 10)
+
+    cutmix = keras_cv.layers.preprocessing.cut_mix.CutMix(10)
+    output = cutmix({"images": images[:32], "labels": labels[:32]})
+    # output == {'images': updated_images, 'labels': updated_labels}
+    ```
+    """
+
+    def __init__(self, alpha=1.0, seed=None, **kwargs):
+        super().__init__(seed=seed, **kwargs)
+        self.alpha = alpha
+        self.seed = seed
+
+    def _sample_from_beta(self, alpha, beta, shape):
+        sample_alpha = tf.random.gamma(
+            shape, 1.0, beta=alpha, seed=self._random_generator.make_legacy_seed()
+        )
+        sample_beta = tf.random.gamma(
+            shape, 1.0, beta=beta, seed=self._random_generator.make_legacy_seed()
+        )
+        return sample_alpha / (sample_alpha + sample_beta)
+
+    def _batch_augment(self, inputs):
+        self._validate_inputs(inputs)
+        images = inputs.get("images", None)
+        labels = inputs.get("labels", None)
+        if images is None or labels is None:
+            raise ValueError(
+                "CutMix expects inputs in a dictionary with format "
+                '{"images": images, "labels": labels}.'
+                f"Got: inputs = {inputs}"
+            )
+        images, labels = self._update_labels(*self._cutmix(images, labels))
+        inputs["images"] = images
+        inputs["labels"] = labels
+        return inputs
+
+    def _augment(self, inputs):
+        raise ValueError(
+            "CutMix received a single image to `call`.  The layer relies on "
+            "combining multiple examples, and as such will not behave as "
+            "expected.  Please call the layer with 2 or more samples."
+        )
+
+    def _cutmix(self, images, labels):
+        """Apply cutmix."""
+        input_shape = tf.shape(images)
+        batch_size, image_height, image_width = (
+            input_shape[0],
+            input_shape[1],
+            input_shape[2],
+        )
+
+        permutation_order = tf.random.shuffle(tf.range(0, batch_size), seed=self.seed)
+        lambda_sample = self._sample_from_beta(self.alpha, self.alpha, (batch_size,))
+
+        ratio = tf.math.sqrt(1 - lambda_sample)
+
+        cut_height = tf.cast(
+            ratio * tf.cast(image_height, dtype=tf.float32), dtype=tf.int32
+        )
+        cut_width = tf.cast(
+            ratio * tf.cast(image_height, dtype=tf.float32), dtype=tf.int32
+        )
+
+        random_center_height = tf.random.uniform(
+            shape=[batch_size], minval=0, maxval=image_height, dtype=tf.int32
+        )
+        random_center_width = tf.random.uniform(
+            shape=[batch_size], minval=0, maxval=image_width, dtype=tf.int32
+        )
+
+        bounding_box_area = cut_height * cut_width
+        lambda_sample = 1.0 - bounding_box_area / (image_height * image_width)
+        lambda_sample = tf.cast(lambda_sample, dtype=tf.float32)
+
+        images = fill_utils.fill_rectangle(
+            images,
+            random_center_width,
+            random_center_height,
+            cut_width,
+            cut_height,
+            tf.gather(images, permutation_order),
+        )
+
+        return images, labels, lambda_sample, permutation_order
+
+    def _update_labels(self, images, labels, lambda_sample, permutation_order):
+        cutout_labels = tf.gather(labels, permutation_order)
+
+        lambda_sample = tf.reshape(lambda_sample, [-1, 1])
+        labels = lambda_sample * labels + (1.0 - lambda_sample) * cutout_labels
+        return images, labels
+
+    def _validate_inputs(self, inputs):
+        labels = inputs.get("labels", None)
+        if labels is None:
+            raise ValueError(
+                "CutMix expects 'labels' to be present in its inputs. "
+                "CutMix relies on both images an labels. "
+                "Please pass a dictionary with keys 'images' "
+                "containing the image Tensor, and 'labels' containing "
+                "the classification labels. "
+                "For example, `cut_mix({'images': images, 'labels': labels})`."
+            )
+        if not labels.dtype.is_floating:
+            raise ValueError(
+                f"CutMix received labels with type {labels.dtype}. "
+                "Labels must be of type float."
+            )
+
+    def get_config(self):
+        config = {
+            "alpha": self.alpha,
+            "seed": self.seed,
+        }
+        base_config = super().get_config()
+        return dict(list(base_config.items()) + list(config.items()))

Keras/keras-cv/keras_cv/layers/preprocessing/cut_mix_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.preprocessing.cut_mix import CutMix
+
+classes = 10
+
+
+class CutMixTest(tf.test.TestCase):
+    def test_return_shapes(self):
+        xs = tf.ones((2, 512, 512, 3))
+        # randomly sample labels
+        ys = tf.random.categorical(tf.math.log([[0.5, 0.5]]), 2)
+        ys = tf.squeeze(ys)
+        ys = tf.one_hot(ys, classes)
+
+        layer = CutMix(seed=1)
+        outputs = layer({"images": xs, "labels": ys})
+        xs, ys = outputs["images"], outputs["labels"]
+
+        self.assertEqual(xs.shape, [2, 512, 512, 3])
+        self.assertEqual(ys.shape, [2, 10])
+
+    def test_cut_mix_call_results(self):
+        xs = tf.cast(
+            tf.stack(
+                [2 * tf.ones((4, 4, 3)), tf.ones((4, 4, 3))],
+                axis=0,
+            ),
+            tf.float32,
+        )
+        ys = tf.one_hot(tf.constant([0, 1]), 2)
+
+        layer = CutMix(seed=1)
+        outputs = layer({"images": xs, "labels": ys})
+        xs, ys = outputs["images"], outputs["labels"]
+
+        # At least some pixels should be replaced in the CutMix operation
+        self.assertTrue(tf.math.reduce_any(xs[0] == 1.0))
+        self.assertTrue(tf.math.reduce_any(xs[0] == 2.0))
+        self.assertTrue(tf.math.reduce_any(xs[1] == 1.0))
+        self.assertTrue(tf.math.reduce_any(xs[1] == 2.0))
+        # No labels should still be close to their original values
+        self.assertNotAllClose(ys, 1.0)
+        self.assertNotAllClose(ys, 0.0)
+
+    def test_cut_mix_call_results_one_channel(self):
+        xs = tf.cast(
+            tf.stack(
+                [2 * tf.ones((4, 4, 1)), tf.ones((4, 4, 1))],
+                axis=0,
+            ),
+            tf.float32,
+        )
+        ys = tf.one_hot(tf.constant([0, 1]), 2)
+
+        layer = CutMix(seed=1)
+        outputs = layer({"images": xs, "labels": ys})
+        xs, ys = outputs["images"], outputs["labels"]
+
+        # At least some pixels should be replaced in the CutMix operation
+        self.assertTrue(tf.math.reduce_any(xs[0] == 1.0))
+        self.assertTrue(tf.math.reduce_any(xs[0] == 2.0))
+        self.assertTrue(tf.math.reduce_any(xs[1] == 1.0))
+        self.assertTrue(tf.math.reduce_any(xs[1] == 2.0))
+        # No labels should still be close to their original values
+        self.assertNotAllClose(ys, 1.0)
+        self.assertNotAllClose(ys, 0.0)
+
+    def test_in_tf_function(self):
+        xs = tf.cast(
+            tf.stack([2 * tf.ones((100, 100, 1)), tf.ones((100, 100, 1))], axis=0),
+            tf.float32,
+        )
+        ys = tf.one_hot(tf.constant([0, 1]), 2)
+
+        layer = CutMix(seed=1)
+
+        @tf.function
+        def augment(x, y):
+            return layer({"images": x, "labels": y})
+
+        outputs = augment(xs, ys)
+        xs, ys = outputs["images"], outputs["labels"]
+
+        # At least some pixels should be replaced in the CutMix operation
+        self.assertTrue(tf.math.reduce_any(xs[0] == 1.0))
+        self.assertTrue(tf.math.reduce_any(xs[0] == 2.0))
+        self.assertTrue(tf.math.reduce_any(xs[1] == 1.0))
+        self.assertTrue(tf.math.reduce_any(xs[1] == 2.0))
+        # No labels should still be close to their original values
+        self.assertNotAllClose(ys, 1.0)
+        self.assertNotAllClose(ys, 0.0)
+
+    def test_single_image_input(self):
+        xs = tf.ones((512, 512, 3))
+        ys = tf.one_hot(tf.constant([1]), 2)
+        inputs = {"images": xs, "labels": ys}
+        layer = CutMix()
+        with self.assertRaisesRegexp(
+            ValueError, "CutMix received a single image to `call`"
+        ):
+            _ = layer(inputs)
+
+    def test_missing_labels(self):
+        xs = tf.ones((2, 512, 512, 3))
+        inputs = {"images": xs}
+        layer = CutMix()
+        with self.assertRaisesRegexp(ValueError, "CutMix expects 'labels'"):
+            _ = layer(inputs)
+
+    def test_int_labels(self):
+        xs = tf.ones((2, 512, 512, 3))
+        ys = tf.one_hot(tf.constant([1, 0]), 2, dtype=tf.int32)
+        inputs = {"images": xs, "labels": ys}
+        layer = CutMix()
+        with self.assertRaisesRegexp(ValueError, "CutMix received labels with type"):
+            _ = layer(inputs)
+
+    def test_image_input(self):
+        xs = tf.ones((2, 512, 512, 3))
+        layer = CutMix()
+        with self.assertRaisesRegexp(
+            ValueError, "CutMix expects 'labels' to be present in its inputs"
+        ):
+            _ = layer(xs)

Keras/keras-cv/keras_cv/layers/preprocessing/equalization.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 Equalization(BaseImageAugmentationLayer):
+    """Equalization performs histogram equalization on a channel-wise basis.
+
+    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`.
+        bins: Integer indicating the number of bins to use in histogram equalization.
+            Should be in the range [0, 256].
+
+    Usage:
+    ```python
+    equalize = Equalization()
+
+    (images, labels), _ = tf.keras.datasets.cifar10.load_data()
+    # Note that images are an int8 Tensor with values in the range [0, 255]
+    images = equalize(images)
+    ```
+
+    Call arguments:
+        images: Tensor of pixels in range [0, 255], in RGB format.  Can be
+            of type float or int.  Should be in NHWC format.
+    """
+
+    def __init__(self, value_range, bins=256, **kwargs):
+        super().__init__(**kwargs)
+        self.bins = bins
+        self.value_range = value_range
+
+    def equalize_channel(self, image, channel_index):
+        """equalize_channel performs histogram equalization on a single channel.
+
+        Args:
+            image: int Tensor with pixels in range [0, 255], RGB format,
+                with channels last
+            channel_index: channel to equalize
+        """
+        image = image[..., channel_index]
+        # Compute the histogram of the image channel.
+        histogram = tf.histogram_fixed_width(image, [0, 255], nbins=self.bins)
+
+        # For the purposes of computing the step, filter out the nonzeros.
+        # Zeroes are replaced by a big number while calculating min to keep shape
+        # constant across input sizes for compatibility with vectorized_map
+
+        big_number = 1410065408
+        histogram_without_zeroes = tf.where(
+            tf.equal(histogram, 0),
+            big_number,
+            histogram,
+        )
+
+        step = (tf.reduce_sum(histogram) - tf.reduce_min(histogram_without_zeroes)) // (
+            self.bins - 1
+        )
+
+        def build_mapping(histogram, step):
+            # Compute the cumulative sum, shifting by step // 2
+            # and then normalization by step.
+            lookup_table = (tf.cumsum(histogram) + (step // 2)) // step
+            # Shift lookup_table, prepending with 0.
+            lookup_table = tf.concat([[0], lookup_table[:-1]], 0)
+            # Clip the counts to be in range.  This is done
+            # in the C code for image.point.
+            return tf.clip_by_value(lookup_table, 0, 255)
+
+        # If step is zero, return the original image.  Otherwise, build
+        # lookup table from the full histogram and step and then index from it.
+        result = tf.cond(
+            tf.equal(step, 0),
+            lambda: image,
+            lambda: tf.gather(build_mapping(histogram, step), image),
+        )
+
+        return result
+
+    def augment_image(self, image, **kwargs):
+        image = preprocessing.transform_value_range(
+            image, self.value_range, (0, 255), dtype=image.dtype
+        )
+        image = tf.cast(image, tf.int32)
+        image = tf.map_fn(
+            lambda channel: self.equalize_channel(image, channel),
+            tf.range(tf.shape(image)[-1]),
+        )
+
+        image = tf.transpose(image, [1, 2, 0])
+        image = tf.cast(image, tf.float32)
+        image = preprocessing.transform_value_range(image, (0, 255), self.value_range)
+        return image
+
+    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({"bins": self.bins, "value_range": self.value_range})
+        return config

Keras/keras-cv/keras_cv/layers/preprocessing/equalization_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.equalization import Equalization
+
+
+class EqualizationTest(tf.test.TestCase, parameterized.TestCase):
+    def test_return_shapes(self):
+        xs = 255 * tf.ones((2, 512, 512, 3), dtype=tf.int32)
+        layer = Equalization(value_range=(0, 255))
+        xs = layer(xs)
+
+        self.assertEqual(xs.shape, [2, 512, 512, 3])
+        self.assertAllEqual(xs, 255 * tf.ones((2, 512, 512, 3)))
+
+    def test_return_shapes_inside_model(self):
+        layer = Equalization(value_range=(0, 255))
+        inp = tf.keras.layers.Input(shape=[512, 512, 5])
+        out = layer(inp)
+        model = tf.keras.models.Model(inp, out)
+
+        self.assertEqual(model.layers[-1].output_shape, (None, 512, 512, 5))
+
+    def test_equalizes_to_all_bins(self):
+        xs = tf.random.uniform((2, 512, 512, 3), 0, 255, dtype=tf.float32)
+        layer = Equalization(value_range=(0, 255))
+        xs = layer(xs)
+
+        for i in range(0, 256):
+            self.assertTrue(tf.math.reduce_any(xs == i))
+
+    @parameterized.named_parameters(
+        ("float32", tf.float32), ("int32", tf.int32), ("int64", tf.int64)
+    )
+    def test_input_dtypes(self, dtype):
+        xs = tf.random.uniform((2, 512, 512, 3), 0, 255, dtype=dtype)
+        layer = Equalization(value_range=(0, 255))
+        xs = layer(xs)
+
+        for i in range(0, 256):
+            self.assertTrue(tf.math.reduce_any(xs == i))
+        self.assertAllInRange(xs, 0, 255)
+
+    @parameterized.named_parameters(("0_255", 0, 255), ("0_1", 0, 1))
+    def test_output_range(self, lower, upper):
+        xs = tf.random.uniform((2, 512, 512, 3), lower, upper, dtype=tf.float32)
+        layer = Equalization(value_range=(lower, upper))
+        xs = layer(xs)
+        self.assertAllInRange(xs, lower, upper)

Keras/keras-cv/keras_cv/layers/preprocessing/fourier_mix.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,
+)
+
+
+@tf.keras.utils.register_keras_serializable(package="keras_cv")
+class FourierMix(BaseImageAugmentationLayer):
+    """FourierMix implements the FMix data augmentation technique.
+
+    Args:
+        alpha: Float value for beta distribution.  Inverse scale parameter for the gamma
+            distribution.  This controls the shape of the distribution from which the
+            smoothing values are sampled.  Defaults to 0.5, which is a recommended value
+            in the paper.
+        decay_power: A float value representing the decay power.  Defaults to 3, as
+            recommended in the paper.
+        seed: Integer. Used to create a random seed.
+    References:
+        - [FMix paper](https://arxiv.org/abs/2002.12047).
+
+    Sample usage:
+    ```python
+    (images, labels), _ = tf.keras.datasets.cifar10.load_data()
+    fourier_mix = keras_cv.layers.preprocessing.FourierMix(0.5)
+    augmented_images, updated_labels = fourier_mix({'images': images, 'labels': labels})
+    # output == {'images': updated_images, 'labels': updated_labels}
+    ```
+    """
+
+    def __init__(self, alpha=0.5, decay_power=3, seed=None, **kwargs):
+        super().__init__(seed=seed, **kwargs)
+        self.alpha = alpha
+        self.decay_power = decay_power
+        self.seed = seed
+
+    def _sample_from_beta(self, alpha, beta, shape):
+        sample_alpha = tf.random.gamma(
+            shape, 1.0, beta=alpha, seed=self._random_generator.make_legacy_seed()
+        )
+        sample_beta = tf.random.gamma(
+            shape, 1.0, beta=beta, seed=self._random_generator.make_legacy_seed()
+        )
+        return sample_alpha / (sample_alpha + sample_beta)
+
+    @staticmethod
+    def _fftfreq(signal_size, sample_spacing=1):
+        """This function returns the sample frequencies of a discrete fourier transform.
+        The result array contains the frequency bin centers starting at 0 using the
+        sample spacing.
+        """
+
+        results = tf.concat(
+            [
+                tf.range((signal_size - 1) / 2 + 1, dtype=tf.int32),
+                tf.range(-(signal_size // 2), 0, dtype=tf.int32),
+            ],
+            0,
+        )
+
+        return results / (signal_size * sample_spacing)
+
+    def _apply_fftfreq(self, h, w):
+        # Applying the fourier transform across 2 dimensions (height and width).
+        fx = FourierMix._fftfreq(w)[: w // 2 + 1 + w % 2]
+        fy = FourierMix._fftfreq(h)
+        fy = tf.expand_dims(fy, -1)
+
+        return tf.math.sqrt(fx * fx + fy * fy)
+
+    def _get_spectrum(self, freqs, decay_power, channel, h, w):
+        # Function to apply a low pass filter by decaying its high frequency components.
+        scale = tf.ones(1) / tf.cast(
+            tf.math.maximum(freqs, tf.convert_to_tensor([1 / tf.reduce_max([w, h])]))
+            ** decay_power,
+            tf.float32,
+        )
+
+        param_size = tf.concat(
+            [tf.constant([channel]), tf.shape(freqs), tf.constant([2])], 0
+        )
+        param = self._random_generator.random_normal(param_size)
+
+        scale = tf.expand_dims(scale, -1)[None, :]
+
+        return scale * param
+
+    def _sample_mask_from_transform(self, decay, shape, ch=1):
+        # Sampling low frequency map from fourier transform.
+        freqs = self._apply_fftfreq(shape[0], shape[1])
+        spectrum = self._get_spectrum(freqs, decay, ch, shape[0], shape[1])
+        spectrum = tf.complex(spectrum[:, 0], spectrum[:, 1])
+
+        mask = tf.math.real(tf.signal.irfft2d(spectrum, shape))
+        mask = mask[:1, : shape[0], : shape[1]]
+
+        mask = mask - tf.reduce_min(mask)
+        mask = mask / tf.reduce_max(mask)
+        return mask
+
+    def _binarise_mask(self, mask, lam, in_shape):
+        # Create the final mask from the sampled values.
+        idx = tf.argsort(tf.reshape(mask, [-1]), direction="DESCENDING")
+        mask = tf.reshape(mask, [-1])
+        num = tf.cast(tf.math.round(lam * tf.cast(tf.size(mask), tf.float32)), tf.int32)
+
+        updates = tf.concat(
+            [
+                tf.ones((num,), tf.float32),
+                tf.zeros((tf.size(mask) - num,), tf.float32),
+            ],
+            0,
+        )
+
+        mask = tf.scatter_nd(
+            tf.expand_dims(idx, -1), updates, tf.expand_dims(tf.size(mask), -1)
+        )
+
+        mask = tf.reshape(mask, in_shape)
+        return mask
+
+    def _batch_augment(self, inputs):
+        images = inputs.get("images", None)
+        labels = inputs.get("labels", None)
+        if images is None or labels is None:
+            raise ValueError(
+                "FourierMix expects inputs in a dictionary with format "
+                '{"images": images, "labels": labels}.'
+                f"Got: inputs = {inputs}"
+            )
+        images, lambda_sample, permutation_order = self._fourier_mix(images)
+        if labels is not None:
+            labels = self._update_labels(labels, lambda_sample, permutation_order)
+            inputs["labels"] = labels
+        inputs["images"] = images
+        return inputs
+
+    def _augment(self, inputs):
+        raise ValueError(
+            "FourierMix received a single image to `call`.  The layer relies on "
+            "combining multiple examples, and as such will not behave as "
+            "expected.  Please call the layer with 2 or more samples."
+        )
+
+    def _fourier_mix(self, images):
+        shape = tf.shape(images)
+        permutation_order = tf.random.shuffle(tf.range(0, shape[0]), seed=self.seed)
+
+        lambda_sample = self._sample_from_beta(self.alpha, self.alpha, (shape[0],))
+
+        # generate masks utilizing mapped calls
+        masks = tf.map_fn(
+            lambda x: self._sample_mask_from_transform(self.decay_power, shape[1:-1]),
+            tf.range(shape[0], dtype=tf.float32),
+        )
+
+        # binarise masks utilizing mapped calls
+        masks = tf.map_fn(
+            lambda i: self._binarise_mask(masks[i], lambda_sample[i], shape[1:-1]),
+            tf.range(shape[0], dtype=tf.int32),
+            fn_output_signature=tf.float32,
+        )
+        masks = tf.expand_dims(masks, -1)
+
+        fmix_images = tf.gather(images, permutation_order)
+        images = masks * images + (1.0 - masks) * fmix_images
+
+        return images, lambda_sample, permutation_order
+
+    def _update_labels(self, labels, lambda_sample, permutation_order):
+        labels_for_fmix = tf.gather(labels, permutation_order)
+
+        # for broadcasting
+        batch_size = tf.expand_dims(tf.shape(labels)[0], -1)
+        labels_rank = tf.rank(labels)
+        broadcast_shape = tf.concat([batch_size, tf.ones(labels_rank - 1, tf.int32)], 0)
+        lambda_sample = tf.reshape(lambda_sample, broadcast_shape)
+
+        labels = lambda_sample * labels + (1.0 - lambda_sample) * labels_for_fmix
+        return labels
+
+    def get_config(self):
+        config = {
+            "alpha": self.alpha,
+            "decay_power": self.decay_power,
+            "seed": self.seed,
+        }
+        base_config = super().get_config()
+        return dict(list(base_config.items()) + list(config.items()))

Keras/keras-cv/keras_cv/layers/preprocessing/fourier_mix_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.preprocessing.fourier_mix import FourierMix
+
+classes = 10
+
+
+class FourierMixTest(tf.test.TestCase):
+    def test_return_shapes(self):
+        xs = tf.ones((2, 512, 512, 3))
+        # randomly sample labels
+        ys = tf.random.categorical(tf.math.log([[0.5, 0.5]]), 2)
+        ys = tf.squeeze(ys)
+        ys = tf.one_hot(ys, classes)
+
+        layer = FourierMix()
+        outputs = layer({"images": xs, "labels": ys})
+        xs, ys = (
+            outputs["images"],
+            outputs["labels"],
+        )
+
+        self.assertEqual(xs.shape, [2, 512, 512, 3])
+        self.assertEqual(ys.shape, [2, 10])
+
+    def test_fourier_mix_call_results(self):
+        xs = tf.cast(
+            tf.stack(
+                [2 * tf.ones((4, 4, 3)), tf.ones((4, 4, 3))],
+                axis=0,
+            ),
+            tf.float32,
+        )
+        ys = tf.one_hot(tf.constant([0, 1]), 2)
+
+        layer = FourierMix()
+        outputs = layer({"images": xs, "labels": ys})
+        xs, ys = outputs["images"], outputs["labels"]
+
+        # 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_in_tf_function(self):
+        xs = tf.cast(
+            tf.stack(
+                [2 * tf.ones((4, 4, 3)), tf.ones((4, 4, 3))],
+                axis=0,
+            ),
+            tf.float32,
+        )
+        ys = tf.one_hot(tf.constant([0, 1]), 2)
+
+        layer = FourierMix()
+
+        @tf.function
+        def augment(x, y):
+            return layer({"images": x, "labels": y})
+
+        outputs = augment(xs, ys)
+        xs, ys = outputs["images"], outputs["labels"]
+
+        # 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_image_input_only(self):
+        xs = tf.cast(
+            tf.stack([2 * tf.ones((100, 100, 1)), tf.ones((100, 100, 1))], axis=0),
+            tf.float32,
+        )
+        layer = FourierMix()
+        with self.assertRaisesRegexp(ValueError, "expects inputs in a dictionary"):
+            _ = layer(xs)
+
+    def test_single_image_input(self):
+        xs = tf.ones((512, 512, 3))
+        ys = tf.one_hot(tf.constant([1]), 2)
+        inputs = {"images": xs, "labels": ys}
+        layer = FourierMix()
+        with self.assertRaisesRegexp(
+            ValueError, "FourierMix received a single image to `call`"
+        ):
+            _ = layer(inputs)

Keras/keras-cv/keras_cv/layers/preprocessing/grayscale.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,
+)
+
+
+@tf.keras.utils.register_keras_serializable(package="keras_cv")
+class Grayscale(BaseImageAugmentationLayer):
+    """Grayscale is a preprocessing layer that transforms RGB images to Grayscale images.
+    Input images should have values in the range of [0, 255].
+
+    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:
+        output_channels.
+            Number color channels present in the output image.
+            The output_channels can be 1 or 3. RGB image with shape
+            (..., height, width, 3) will have the following shapes
+            after the `Grayscale` operation:
+                 a. (..., height, width, 1) if output_channels = 1
+                 b. (..., height, width, 3) if output_channels = 3.
+
+    Usage:
+    ```python
+    (images, labels), _ = tf.keras.datasets.cifar10.load_data()
+    to_grayscale = keras_cv.layers.preprocessing.Grayscale()
+    augmented_images = to_grayscale(images)
+    ```
+    """
+
+    def __init__(self, output_channels=1, **kwargs):
+        super().__init__(**kwargs)
+        self.output_channels = output_channels
+        # This layer may raise an error when running on GPU using auto_vectorize
+        self.auto_vectorize = False
+
+    def _check_input_params(self, output_channels):
+        if output_channels not in [1, 3]:
+            raise ValueError(
+                "Received invalid argument output_channels. "
+                f"output_channels must be in 1 or 3. Got {output_channels}"
+            )
+        self.output_channels = output_channels
+
+    def augment_image(self, image, transformation=None, **kwargs):
+        grayscale = tf.image.rgb_to_grayscale(image)
+        if self.output_channels == 1:
+            return grayscale
+        elif self.output_channels == 3:
+            return tf.image.grayscale_to_rgb(grayscale)
+        else:
+            raise ValueError("Unsupported value for `output_channels`.")
+
+    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 = {
+            "output_channels": self.output_channels,
+        }
+        base_config = super().get_config()
+        return dict(list(base_config.items()) + list(config.items()))

Keras/keras-cv/keras_cv/layers/preprocessing/grayscale_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 GrayscaleTest(tf.test.TestCase):
+    def test_return_shapes(self):
+        xs = tf.ones((2, 512, 512, 3))
+
+        layer = preprocessing.Grayscale(
+            output_channels=1,
+        )
+        xs1 = layer(xs, training=True)
+
+        layer = preprocessing.Grayscale(
+            output_channels=3,
+        )
+        xs2 = layer(xs, training=True)
+
+        self.assertEqual(xs1.shape, [2, 512, 512, 1])
+        self.assertEqual(xs2.shape, [2, 512, 512, 3])
+
+    def test_in_tf_function(self):
+        xs = tf.cast(
+            tf.stack([2 * tf.ones((100, 100, 3)), tf.ones((100, 100, 3))], axis=0),
+            tf.float32,
+        )
+
+        # test 1
+        layer = preprocessing.Grayscale(
+            output_channels=1,
+        )
+
+        @tf.function
+        def augment(x):
+            return layer(x, training=True)
+
+        xs1 = augment(xs)
+
+        # test 2
+        layer = preprocessing.Grayscale(
+            output_channels=3,
+        )
+
+        @tf.function
+        def augment(x):
+            return layer(x, training=True)
+
+        xs2 = augment(xs)
+
+        self.assertEqual(xs1.shape, [2, 100, 100, 1])
+        self.assertEqual(xs2.shape, [2, 100, 100, 3])
+
+    def test_non_square_image(self):
+        xs = tf.cast(
+            tf.stack([2 * tf.ones((512, 1024, 3)), tf.ones((512, 1024, 3))], axis=0),
+            tf.float32,
+        )
+
+        layer = preprocessing.Grayscale(
+            output_channels=1,
+        )
+        xs1 = layer(xs, training=True)
+
+        layer = preprocessing.Grayscale(
+            output_channels=3,
+        )
+        xs2 = layer(xs, training=True)
+
+        self.assertEqual(xs1.shape, [2, 512, 1024, 1])
+        self.assertEqual(xs2.shape, [2, 512, 1024, 3])
+
+    def test_in_single_image(self):
+        xs = tf.cast(
+            tf.ones((512, 512, 3)),
+            dtype=tf.float32,
+        )
+
+        layer = preprocessing.Grayscale(
+            output_channels=1,
+        )
+        xs1 = layer(xs, training=True)
+
+        layer = preprocessing.Grayscale(
+            output_channels=3,
+        )
+        xs2 = layer(xs, training=True)
+
+        self.assertEqual(xs1.shape, [512, 512, 1])
+        self.assertEqual(xs2.shape, [512, 512, 3])

Keras/keras-cv/keras_cv/layers/preprocessing/grid_mask.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 tensorflow.keras import layers
+
+from keras_cv import core
+from keras_cv.layers.preprocessing.base_image_augmentation_layer import (
+    BaseImageAugmentationLayer,
+)
+from keras_cv.utils import fill_utils
+from keras_cv.utils import preprocessing
+
+
+def _center_crop(mask, width, height):
+    masks_shape = tf.shape(mask)
+    h_diff = masks_shape[0] - height
+    w_diff = masks_shape[1] - width
+
+    h_start = tf.cast(h_diff / 2, tf.int32)
+    w_start = tf.cast(w_diff / 2, tf.int32)
+    return tf.image.crop_to_bounding_box(mask, h_start, w_start, height, width)
+
+
+@tf.keras.utils.register_keras_serializable(package="keras_cv")
+class GridMask(BaseImageAugmentationLayer):
+    """GridMask class for grid-mask augmentation.
+
+
+    Input shape:
+        Int or float tensor with values in the range [0, 255].
+        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:
+        ratio_factor: A float, tuple of two floats, or `keras_cv.FactorSampler`.
+            Ratio determines the ratio from spacings to grid masks.
+            Lower values make the grid
+            size smaller, and higher values make the grid mask large.
+            Floats should be in the range [0, 1].  0.5 indicates that grid and
+            spacing will be of equal size. To always use the same value, pass a
+            `keras_cv.ConstantFactorSampler()`.
+
+            Defaults to `(0, 0.5)`.
+        rotation_factor:
+            The rotation_factor will be used to randomly rotate the grid_mask during
+            training. Default to 0.1, which results in an output rotating by a
+            random amount in the range [-10% * 2pi, 10% * 2pi].
+
+            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: Pixels inside the gridblock are filled according to the given
+            mode (one of `{"constant", "gaussian_noise"}`). Default: "constant".
+            - *constant*: Pixels are filled with the same constant value.
+            - *gaussian_noise*: Pixels are filled with random gaussian noise.
+        fill_value: an integer represents of value to be filled inside the gridblock
+            when `fill_mode="constant"`. Valid integer range [0 to 255]
+        seed: Integer. Used to create a random seed.
+
+    Usage:
+    ```python
+    (images, labels), _ = tf.keras.datasets.cifar10.load_data()
+    random_gridmask = keras_cv.layers.preprocessing.GridMask()
+    augmented_images = random_gridmask(images)
+    ```
+
+    References:
+        - [GridMask paper](https://arxiv.org/abs/2001.04086)
+    """
+
+    def __init__(
+        self,
+        ratio_factor=(0, 0.5),
+        rotation_factor=0.15,
+        fill_mode="constant",
+        fill_value=0.0,
+        seed=None,
+        **kwargs,
+    ):
+        super().__init__(seed=seed, **kwargs)
+        self.ratio_factor = preprocessing.parse_factor(
+            ratio_factor, param_name="ratio_factor"
+        )
+
+        if isinstance(rotation_factor, core.FactorSampler):
+            raise ValueError(
+                "Currently `GridMask.rotation_factor` does not support the "
+                "`FactorSampler` API.  This will be supported in the next Keras "
+                "release. For now, please pass a float for the "
+                "`rotation_factor` argument."
+            )
+
+        self.fill_mode = fill_mode
+        self.fill_value = fill_value
+        self.rotation_factor = rotation_factor
+        self.random_rotate = layers.RandomRotation(
+            factor=rotation_factor,
+            fill_mode="constant",
+            fill_value=0.0,
+            seed=seed,
+        )
+        self.auto_vectorize = False
+        self._check_parameter_values()
+        self.seed = seed
+
+    def _check_parameter_values(self):
+        fill_mode, fill_value = self.fill_mode, self.fill_value
+
+        if fill_value not in range(0, 256):
+            raise ValueError(
+                f"fill_value should be in the range [0, 255]. Got {fill_value}"
+            )
+
+        if fill_mode not in ["constant", "gaussian_noise", "random"]:
+            raise ValueError(
+                '`fill_mode` should be "constant", '
+                f'"gaussian_noise", or "random".  Got `fill_mode`={fill_mode}'
+            )
+
+    def get_random_transformation(
+        self, image=None, label=None, bounding_boxes=None, **kwargs
+    ):
+        ratio = self.ratio_factor()
+
+        # compute grid mask
+        input_shape = tf.shape(image)
+        mask = self._compute_grid_mask(input_shape, ratio=ratio)
+
+        # convert mask to single-channel image
+        mask = tf.cast(mask, tf.float32)
+        mask = tf.expand_dims(mask, axis=-1)
+
+        # randomly rotate mask
+        mask = self.random_rotate(mask)
+
+        # compute fill
+        if self.fill_mode == "constant":
+            fill_value = tf.fill(input_shape, self.fill_value)
+        else:
+            # gaussian noise
+            fill_value = self._random_generator.random_normal(
+                shape=input_shape, dtype=image.dtype
+            )
+
+        return mask, fill_value
+
+    def _compute_grid_mask(self, input_shape, ratio):
+        height = tf.cast(input_shape[0], tf.float32)
+        width = tf.cast(input_shape[1], tf.float32)
+
+        # mask side length
+        input_diagonal_len = tf.sqrt(tf.square(width) + tf.square(height))
+        mask_side_len = tf.math.ceil(input_diagonal_len)
+
+        # grid unit size
+        unit_size = self._random_generator.random_uniform(
+            shape=(),
+            minval=tf.math.minimum(height * 0.5, width * 0.3),
+            maxval=tf.math.maximum(height * 0.5, width * 0.3) + 1,
+            dtype=tf.float32,
+        )
+        rectangle_side_len = tf.cast((ratio) * unit_size, tf.float32)
+
+        # sample x and y offset for grid units randomly between 0 and unit_size
+        delta_x = self._random_generator.random_uniform(
+            shape=(), minval=0.0, maxval=unit_size, dtype=tf.float32
+        )
+        delta_y = self._random_generator.random_uniform(
+            shape=(), minval=0.0, maxval=unit_size, dtype=tf.float32
+        )
+
+        # grid size (number of diagonal units in grid)
+        grid_size = mask_side_len // unit_size + 1
+        grid_size_range = tf.range(1, grid_size + 1)
+
+        # diagonal corner coordinates
+        unit_size_range = grid_size_range * unit_size
+        x1 = unit_size_range - delta_x
+        x0 = x1 - rectangle_side_len
+        y1 = unit_size_range - delta_y
+        y0 = y1 - rectangle_side_len
+
+        # compute grid coordinates
+        x0, y0 = tf.meshgrid(x0, y0)
+        x1, y1 = tf.meshgrid(x1, y1)
+
+        # flatten mesh grid
+        x0 = tf.reshape(x0, [-1])
+        y0 = tf.reshape(y0, [-1])
+        x1 = tf.reshape(x1, [-1])
+        y1 = tf.reshape(y1, [-1])
+
+        # convert coordinates to mask
+        corners = tf.stack([x0, y0, x1, y1], axis=-1)
+        mask_side_len = tf.cast(mask_side_len, tf.int32)
+        rectangle_masks = fill_utils.corners_to_mask(
+            corners, mask_shape=(mask_side_len, mask_side_len)
+        )
+        grid_mask = tf.reduce_any(rectangle_masks, axis=0)
+
+        return grid_mask
+
+    def augment_image(self, image, transformation=None, **kwargs):
+        mask, fill_value = transformation
+        input_shape = tf.shape(image)
+
+        # center crop mask
+        input_height = input_shape[0]
+        input_width = input_shape[1]
+        mask = _center_crop(mask, input_width, input_height)
+
+        # convert back to boolean mask
+        mask = tf.cast(mask, tf.bool)
+
+        return tf.where(mask, fill_value, image)
+
+    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 = {
+            "ratio_factor": self.ratio_factor,
+            "rotation_factor": self.rotation_factor,
+            "fill_mode": self.fill_mode,
+            "fill_value": self.fill_value,
+            "seed": self.seed,
+        }
+        base_config = super().get_config()
+        return dict(list(base_config.items()) + list(config.items()))

Keras/keras-cv/keras_cv/layers/preprocessing/grid_mask_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
+
+import keras_cv
+from keras_cv.layers.preprocessing.grid_mask import GridMask
+
+
+class GridMaskTest(tf.test.TestCase):
+    def test_return_shapes(self):
+        xs = tf.ones((2, 512, 512, 3))
+
+        layer = GridMask(ratio_factor=0.1, rotation_factor=(-0.2, 0.3))
+        xs = layer(xs, training=True)
+
+        self.assertEqual(xs.shape, [2, 512, 512, 3])
+
+    def test_gridmask_call_results_one_channel(self):
+        xs = tf.cast(
+            tf.stack(
+                [3 * tf.ones((40, 40, 1)), 2 * tf.ones((40, 40, 1))],
+                axis=0,
+            ),
+            dtype=tf.float32,
+        )
+
+        fill_value = 0.0
+        layer = GridMask(
+            ratio_factor=0.3,
+            rotation_factor=(0.2, 0.3),
+            fill_mode="constant",
+            fill_value=fill_value,
+        )
+        xs = layer(xs, training=True)
+
+        # Some pixels should be replaced with fill_value
+        self.assertTrue(tf.math.reduce_any(xs[0] == float(fill_value)))
+        self.assertTrue(tf.math.reduce_any(xs[0] == 3.0))
+        self.assertTrue(tf.math.reduce_any(xs[1] == float(fill_value)))
+        self.assertTrue(tf.math.reduce_any(xs[1] == 2.0))
+
+    def test_non_square_image(self):
+        xs = tf.cast(
+            tf.stack(
+                [2 * tf.ones((1024, 512, 1)), tf.ones((1024, 512, 1))],
+                axis=0,
+            ),
+            dtype=tf.float32,
+        )
+
+        fill_value = 100.0
+        layer = GridMask(
+            ratio_factor=0.6,
+            rotation_factor=0.3,
+            fill_mode="constant",
+            fill_value=fill_value,
+        )
+        xs = layer(xs, training=True)
+
+        # Some pixels should be replaced with fill_value
+        self.assertTrue(tf.math.reduce_any(xs[0] == float(fill_value)))
+        self.assertTrue(tf.math.reduce_any(xs[0] == 2.0))
+        self.assertTrue(tf.math.reduce_any(xs[1] == float(fill_value)))
+        self.assertTrue(tf.math.reduce_any(xs[1] == 1.0))
+
+    def test_in_tf_function(self):
+        xs = tf.cast(
+            tf.stack([2 * tf.ones((100, 100, 1)), tf.ones((100, 100, 1))], axis=0),
+            dtype=tf.float32,
+        )
+
+        fill_value = 255.0
+        layer = GridMask(
+            ratio_factor=keras_cv.ConstantFactorSampler(0.5),
+            rotation_factor=0.5,
+            fill_mode="constant",
+            fill_value=fill_value,
+        )
+
+        @tf.function
+        def augment(x):
+            return layer(x, training=True)
+
+        xs = augment(xs)
+
+        # Some pixels should be replaced with fill_value
+        self.assertTrue(tf.math.reduce_any(xs[0] == float(fill_value)))
+        self.assertTrue(tf.math.reduce_any(xs[0] == 2.0))
+        self.assertTrue(tf.math.reduce_any(xs[1] == float(fill_value)))
+        self.assertTrue(tf.math.reduce_any(xs[1] == 1.0))
+
+    def test_in_single_image(self):
+        xs = tf.cast(
+            tf.ones((512, 512, 1)),
+            dtype=tf.float32,
+        )
+
+        layer = GridMask(ratio_factor=(0.5, 0.5), fill_mode="constant", fill_value=0.0)
+        xs = layer(xs, training=True)
+        self.assertTrue(tf.math.reduce_any(xs == 0.0))
+        self.assertTrue(tf.math.reduce_any(xs == 1.0))

Keras/keras-cv/keras_cv/layers/preprocessing/maybe_apply.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,
+)
+
+
+@tf.keras.utils.register_keras_serializable(package="keras_cv")
+class MaybeApply(BaseImageAugmentationLayer):
+    """Apply provided layer to random elements in a batch.
+
+    Args:
+        layer: a keras `Layer` or `BaseImageAugmentationLayer`. This layer will be
+            applied to randomly chosen samples in a batch. Layer should not modify the
+            size of provided inputs.
+        rate: controls the frequency of applying the layer. 1.0 means all elements in
+            a batch will be modified. 0.0 means no elements will be modified.
+            Defaults to 0.5.
+        auto_vectorize: bool, whether to use tf.vectorized_map or tf.map_fn for
+            batched input. Setting this to True might give better performance but
+            currently doesn't work with XLA. Defaults to False.
+        seed: integer, controls random behaviour.
+
+    Example usage:
+    ```
+    # Let's declare an example layer that will set all image pixels to zero.
+    zero_out = tf.keras.layers.Lambda(lambda x: {"images": 0 * x["images"]})
+
+    # Create a small batch of random, single-channel, 2x2 images:
+    images = tf.random.stateless_uniform(shape=(5, 2, 2, 1), seed=[0, 1])
+    print(images[..., 0])
+    # <tf.Tensor: shape=(5, 2, 2), dtype=float32, numpy=
+    # array([[[0.08216608, 0.40928006],
+    #         [0.39318466, 0.3162533 ]],
+    #
+    #        [[0.34717774, 0.73199546],
+    #         [0.56369007, 0.9769211 ]],
+    #
+    #        [[0.55243933, 0.13101244],
+    #         [0.2941643 , 0.5130266 ]],
+    #
+    #        [[0.38977218, 0.80855536],
+    #         [0.6040567 , 0.10502195]],
+    #
+    #        [[0.51828027, 0.12730157],
+    #         [0.288486  , 0.252975  ]]], dtype=float32)>
+
+    # Apply the layer with 50% probability:
+    maybe_apply = MaybeApply(layer=zero_out, rate=0.5, seed=1234)
+    outputs = maybe_apply(images)
+    print(outputs[..., 0])
+    # <tf.Tensor: shape=(5, 2, 2), dtype=float32, numpy=
+    # array([[[0.        , 0.        ],
+    #         [0.        , 0.        ]],
+    #
+    #        [[0.34717774, 0.73199546],
+    #         [0.56369007, 0.9769211 ]],
+    #
+    #        [[0.55243933, 0.13101244],
+    #         [0.2941643 , 0.5130266 ]],
+    #
+    #        [[0.38977218, 0.80855536],
+    #         [0.6040567 , 0.10502195]],
+    #
+    #        [[0.        , 0.        ],
+    #         [0.        , 0.        ]]], dtype=float32)>
+
+    # We can observe that the layer has been randomly applied to 2 out of 5 samples.
+    ```
+    """
+
+    def __init__(self, layer, rate=0.5, auto_vectorize=False, seed=None, **kwargs):
+        super().__init__(seed=seed, **kwargs)
+
+        if not (0 <= rate <= 1.0):
+            raise ValueError(f"rate must be in range [0, 1]. Received rate: {rate}")
+
+        self._layer = layer
+        self._rate = rate
+        self.auto_vectorize = auto_vectorize
+        self.seed = seed
+
+    def _augment(self, inputs):
+        if self._random_generator.random_uniform(shape=()) > 1.0 - self._rate:
+            return self._layer(inputs)
+        else:
+            return inputs
+
+    def get_config(self):
+        config = super().get_config()
+        config.update(
+            {
+                "rate": self._rate,
+                "layer": self._layer,
+                "seed": self.seed,
+                "auto_vectorize": self.auto_vectorize,
+            }
+        )
+        return config

Keras/keras-cv/keras_cv/layers/preprocessing/maybe_apply_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.base_image_augmentation_layer import (
+    BaseImageAugmentationLayer,
+)
+from keras_cv.layers.preprocessing.maybe_apply import MaybeApply
+
+
+class ZeroOut(BaseImageAugmentationLayer):
+    """Zero out all entries, for testing purposes."""
+
+    def __init__(self):
+        super(ZeroOut, self).__init__()
+
+    def augment_image(self, image, transformation=None, **kwargs):
+        return 0 * image
+
+    def augment_label(self, label, transformation=None, **kwargs):
+        return 0 * label
+
+
+class MaybeApplyTest(tf.test.TestCase, parameterized.TestCase):
+    rng = tf.random.Generator.from_seed(seed=1234)
+
+    @parameterized.parameters([-0.5, 1.7])
+    def test_raises_error_on_invalid_rate_parameter(self, invalid_rate):
+        with self.assertRaises(ValueError):
+            MaybeApply(rate=invalid_rate, layer=ZeroOut())
+
+    def test_works_with_batched_input(self):
+        batch_size = 32
+        dummy_inputs = self.rng.uniform(shape=(batch_size, 224, 224, 3))
+        layer = MaybeApply(rate=0.5, layer=ZeroOut(), seed=1234)
+
+        outputs = layer(dummy_inputs)
+        num_zero_inputs = self._num_zero_batches(dummy_inputs)
+        num_zero_outputs = self._num_zero_batches(outputs)
+
+        self.assertEqual(num_zero_inputs, 0)
+        self.assertLess(num_zero_outputs, batch_size)
+        self.assertGreater(num_zero_outputs, 0)
+
+    @staticmethod
+    def _num_zero_batches(images):
+        num_batches = tf.shape(images)[0]
+        num_non_zero_batches = tf.math.count_nonzero(
+            tf.math.count_nonzero(images, axis=[1, 2, 3]), dtype=tf.int32
+        )
+        return num_batches - num_non_zero_batches
+
+    def test_inputs_unchanged_with_zero_rate(self):
+        dummy_inputs = self.rng.uniform(shape=(32, 224, 224, 3))
+        layer = MaybeApply(rate=0.0, layer=ZeroOut())
+
+        outputs = layer(dummy_inputs)
+
+        self.assertAllClose(outputs, dummy_inputs)
+
+    def test_all_inputs_changed_with_rate_equal_to_one(self):
+        dummy_inputs = self.rng.uniform(shape=(32, 224, 224, 3))
+        layer = MaybeApply(rate=1.0, layer=ZeroOut())
+
+        outputs = layer(dummy_inputs)
+
+        self.assertAllEqual(outputs, tf.zeros_like(dummy_inputs))
+
+    def test_works_with_single_image(self):
+        dummy_inputs = self.rng.uniform(shape=(224, 224, 3))
+        layer = MaybeApply(rate=1.0, layer=ZeroOut())
+
+        outputs = layer(dummy_inputs)
+
+        self.assertAllEqual(outputs, tf.zeros_like(dummy_inputs))
+
+    def test_can_modify_label(self):
+        dummy_inputs = self.rng.uniform(shape=(32, 224, 224, 3))
+        dummy_labels = tf.ones(shape=(32, 2))
+        layer = MaybeApply(rate=1.0, layer=ZeroOut())
+
+        outputs = layer({"images": dummy_inputs, "labels": dummy_labels})
+
+        self.assertAllEqual(outputs["labels"], tf.zeros_like(dummy_labels))
+
+    def test_works_with_native_keras_layers(self):
+        dummy_inputs = self.rng.uniform(shape=(32, 224, 224, 3))
+        zero_out = tf.keras.layers.Lambda(lambda x: {"images": 0 * x["images"]})
+        layer = MaybeApply(rate=1.0, layer=zero_out)
+
+        outputs = layer(dummy_inputs)
+
+        self.assertAllEqual(outputs, tf.zeros_like(dummy_inputs))
+
+    def test_works_with_xla(self):
+        dummy_inputs = self.rng.uniform(shape=(32, 224, 224, 3))
+        # auto_vectorize=True will crash XLA
+        layer = MaybeApply(rate=0.5, layer=ZeroOut(), auto_vectorize=False)
+
+        @tf.function(jit_compile=True)
+        def apply(x):
+            return layer(x)
+
+        apply(dummy_inputs)

Keras/keras-cv/keras_cv/layers/preprocessing/mix_up.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,
+)
+
+
+@tf.keras.utils.register_keras_serializable(package="keras_cv")
+class MixUp(BaseImageAugmentationLayer):
+    """MixUp implements the MixUp data augmentation technique.
+
+    Args:
+        alpha: Float between 0 and 1.  Inverse scale parameter for the gamma
+            distribution.  This controls the shape of the distribution from which the
+            smoothing values are sampled.  Defaults 0.2, which is a recommended value
+            when training an imagenet1k classification model.
+        seed: Integer. Used to create a random seed.
+
+    References:
+        - [MixUp paper](https://arxiv.org/abs/1710.09412).
+        - [MixUp for Object Detection paper](https://arxiv.org/pdf/1902.04103).
+
+    Sample usage:
+    ```python
+    (images, labels), _ = tf.keras.datasets.cifar10.load_data()
+    mixup = keras_cv.layers.preprocessing.MixUp(10)
+    augmented_images, updated_labels = mixup({'images': images, 'labels': labels})
+    # output == {'images': updated_images, 'labels': updated_labels}
+    ```
+    """
+
+    def __init__(self, alpha=0.2, seed=None, **kwargs):
+        super().__init__(seed=seed, **kwargs)
+        self.alpha = alpha
+        self.seed = seed
+
+    def _sample_from_beta(self, alpha, beta, shape):
+        sample_alpha = tf.random.gamma(
+            shape, 1.0, beta=alpha, seed=self._random_generator.make_legacy_seed()
+        )
+        sample_beta = tf.random.gamma(
+            shape, 1.0, beta=beta, seed=self._random_generator.make_legacy_seed()
+        )
+        return sample_alpha / (sample_alpha + sample_beta)
+
+    def _batch_augment(self, inputs):
+        self._validate_inputs(inputs)
+        images = inputs.get("images", None)
+        labels = inputs.get("labels", None)
+        bounding_boxes = inputs.get("bounding_boxes", None)
+        images, lambda_sample, permutation_order = self._mixup(images)
+        if labels is not None:
+            labels = self._update_labels(labels, lambda_sample, permutation_order)
+            inputs["labels"] = labels
+        if bounding_boxes is not None:
+            bounding_boxes = self._update_bounding_boxes(
+                bounding_boxes, permutation_order
+            )
+            inputs["bounding_boxes"] = bounding_boxes
+        inputs["images"] = images
+        return inputs
+
+    def _augment(self, inputs):
+        raise ValueError(
+            "MixUp received a single image to `call`.  The layer relies on "
+            "combining multiple examples, and as such will not behave as "
+            "expected.  Please call the layer with 2 or more samples."
+        )
+
+    def _mixup(self, images):
+        batch_size = tf.shape(images)[0]
+        permutation_order = tf.random.shuffle(tf.range(0, batch_size), seed=self.seed)
+
+        lambda_sample = self._sample_from_beta(self.alpha, self.alpha, (batch_size,))
+        lambda_sample = tf.reshape(lambda_sample, [-1, 1, 1, 1])
+
+        mixup_images = tf.gather(images, permutation_order)
+        images = lambda_sample * images + (1.0 - lambda_sample) * mixup_images
+
+        return images, tf.squeeze(lambda_sample), permutation_order
+
+    def _update_labels(self, labels, lambda_sample, permutation_order):
+        labels_for_mixup = tf.gather(labels, permutation_order)
+
+        lambda_sample = tf.reshape(lambda_sample, [-1, 1])
+        labels = lambda_sample * labels + (1.0 - lambda_sample) * labels_for_mixup
+
+        return labels
+
+    def _update_bounding_boxes(self, bounding_boxes, permutation_order):
+        boxes_for_mixup = tf.gather(bounding_boxes, permutation_order)
+        bounding_boxes = tf.concat([bounding_boxes, boxes_for_mixup], axis=1)
+
+        return bounding_boxes
+
+    def _validate_inputs(self, inputs):
+        images = inputs.get("images", None)
+        labels = inputs.get("labels", None)
+        bounding_boxes = inputs.get("bounding_boxes", None)
+        if images is None or (labels is None and bounding_boxes is None):
+            raise ValueError(
+                "MixUp expects inputs in a dictionary with format "
+                '{"images": images, "labels": labels}. or'
+                '{"images": images, "bounding_boxes": bounding_boxes}'
+                f"Got: inputs = {inputs}"
+            )
+        if labels is not None and not labels.dtype.is_floating:
+            raise ValueError(
+                f"MixUp received labels with type {labels.dtype}. "
+                "Labels must be of type float."
+            )
+
+    def get_config(self):
+        config = {
+            "alpha": self.alpha,
+            "seed": self.seed,
+        }
+        base_config = super().get_config()
+        return dict(list(base_config.items()) + list(config.items()))

Keras/keras-cv/keras_cv/layers/preprocessing/mix_up_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.preprocessing.mix_up import MixUp
+
+classes = 10
+
+
+class MixUpTest(tf.test.TestCase):
+    def test_return_shapes(self):
+        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, 5), 0, 1)
+
+        layer = MixUp()
+        # 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, 6, 5])
+
+    def test_mix_up_call_results(self):
+        xs = tf.cast(
+            tf.stack(
+                [2 * tf.ones((4, 4, 3)), tf.ones((4, 4, 3))],
+                axis=0,
+            ),
+            tf.float32,
+        )
+        ys = tf.one_hot(tf.constant([0, 1]), 2)
+
+        layer = MixUp()
+        outputs = layer({"images": xs, "labels": ys})
+        xs, ys = outputs["images"], outputs["labels"]
+
+        # 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_in_tf_function(self):
+        xs = tf.cast(
+            tf.stack(
+                [2 * tf.ones((4, 4, 3)), tf.ones((4, 4, 3))],
+                axis=0,
+            ),
+            tf.float32,
+        )
+        ys = tf.one_hot(tf.constant([0, 1]), 2)
+
+        layer = MixUp()
+
+        @tf.function
+        def augment(x, y):
+            return layer({"images": x, "labels": y})
+
+        outputs = augment(xs, ys)
+        xs, ys = outputs["images"], outputs["labels"]
+
+        # 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_image_input_only(self):
+        xs = tf.cast(
+            tf.stack([2 * tf.ones((100, 100, 1)), tf.ones((100, 100, 1))], axis=0),
+            tf.float32,
+        )
+        layer = MixUp()
+        with self.assertRaisesRegexp(ValueError, "expects inputs in a dictionary"):
+            _ = layer(xs)
+
+    def test_single_image_input(self):
+        xs = tf.ones((512, 512, 3))
+        ys = tf.one_hot(tf.constant([1]), 2)
+        inputs = {"images": xs, "labels": ys}
+        layer = MixUp()
+        with self.assertRaisesRegexp(
+            ValueError, "MixUp received a single image to `call`"
+        ):
+            _ = layer(inputs)
+
+    def test_int_labels(self):
+        xs = tf.ones((2, 512, 512, 3))
+        ys = tf.one_hot(tf.constant([1, 0]), 2, dtype=tf.int32)
+        inputs = {"images": xs, "labels": ys}
+        layer = MixUp()
+        with self.assertRaisesRegexp(ValueError, "MixUp received labels with type"):
+            _ = layer(inputs)
+
+    def test_image_input(self):
+        xs = tf.ones((2, 512, 512, 3))
+        layer = MixUp()
+        with self.assertRaisesRegexp(
+            ValueError, "MixUp expects inputs in a dictionary with format"
+        ):
+            _ = layer(xs)

Keras/keras-cv/keras_cv/layers/preprocessing/mosaic.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.preprocessing.base_image_augmentation_layer import (
+    BaseImageAugmentationLayer,
+)
+from keras_cv.utils import preprocessing
+
+
+@tf.keras.utils.register_keras_serializable(package="keras_cv")
+class Mosaic(BaseImageAugmentationLayer):
+    """Mosaic implements the mosaic data augmentation technique.
+
+    Mosaic data augmentation first takes 4 images from the batch and makes a grid.
+    After that based on the offset, a crop is taken to form the mosaic image. Labels
+    are in the same ratio as the the area of their images in the output image. Bounding
+    boxes are translated according to the position of the 4 images.
+
+    Args:
+        offset: A tuple of two floats, a single float or `keras_cv.FactorSampler`.
+            `offset` is used to determine the offset of the mosaic center from the
+            top-left corner of the mosaic. If a tuple is used, the x and y coordinates
+            of the mosaic center are 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)`. Defaults to
+            (0.25, 0.75).
+        bounding_box_format: a case-insensitive string (for example, "xyxy") to be
+            passed if bounding boxes are being augmented by this layer.
+            Each bounding box is defined by at least these 4 values. The inputs
+            may contain additional information such as classes and confidence after
+            these 4 values but these values will be ignored and returned as is. For
+            detailed information on the supported formats, see the
+            [KerasCV bounding box documentation](https://keras.io/api/keras_cv/bounding_box/formats/).
+            Defualts to None.
+        seed: Integer. Used to create a random seed.
+
+    References:
+        - [Yolov4 paper](https://arxiv.org/pdf/2004.10934).
+        - [Yolov5 implementation](https://github.com/ultralytics/yolov5).
+        - [YoloX implementation](https://github.com/Megvii-BaseDetection/YOLOX)
+
+    Sample usage:
+    ```python
+    (images, labels), _ = tf.keras.datasets.cifar10.load_data()
+    labels = tf.one_hot(labels,10)
+    labels = tf.cast(tf.squeeze(labels), tf.float32)
+    mosaic = keras_cv.layers.preprocessing.Mosaic()
+    output = mosaic({'images': images, 'labels': labels})
+    # output == {'images': updated_images, 'labels': updated_labels}
+    ```
+    """
+
+    def __init__(
+        self, offset=(0.25, 0.75), bounding_box_format=None, seed=None, **kwargs
+    ):
+        super().__init__(seed=seed, **kwargs)
+        self.offset = offset
+        self.bounding_box_format = bounding_box_format
+        self.center_sampler = preprocessing.parse_factor(offset)
+        self.seed = seed
+
+    def _batch_augment(self, inputs):
+        self._validate_inputs(inputs)
+        images = inputs.get("images", None)
+        labels = inputs.get("labels", None)
+        bounding_boxes = inputs.get("bounding_boxes", None)
+
+        batch_size = tf.shape(images)[0]
+        # pick 3 indices for every batch to create the mosaic output with.
+        permutation_order = tf.random.uniform(
+            (batch_size, 3),
+            minval=0,
+            maxval=batch_size,
+            dtype=tf.int32,
+            seed=self._random_generator.make_legacy_seed(),
+        )
+        # concatenate the batches with permutation order to get all 4 images of the mosaic
+        permutation_order = tf.concat(
+            [tf.expand_dims(tf.range(batch_size), axis=-1), permutation_order], axis=-1
+        )
+
+        input_height, input_width, _ = images.shape[1:]
+
+        mosaic_centers_x = (
+            self.center_sampler(tf.expand_dims(batch_size, axis=0)) * input_width
+        )
+        mosaic_centers_y = (
+            self.center_sampler(shape=tf.expand_dims(batch_size, axis=0)) * input_height
+        )
+        mosaic_centers = tf.stack((mosaic_centers_x, mosaic_centers_y), axis=-1)
+
+        # return the mosaics
+        images = tf.vectorized_map(
+            lambda index: self._update_image(
+                images, permutation_order, mosaic_centers, index
+            ),
+            tf.range(batch_size),
+        )
+
+        if labels is not None:
+            labels = tf.vectorized_map(
+                lambda index: self._update_label(
+                    images, labels, permutation_order, mosaic_centers, index
+                ),
+                tf.range(batch_size),
+            )
+            inputs["labels"] = labels
+
+        if bounding_boxes is not None:
+            # values to translate the boxes by in the mosaic image
+            translate_x = tf.stack(
+                [
+                    mosaic_centers_x - input_width,
+                    mosaic_centers_x,
+                    mosaic_centers_x - input_width,
+                    mosaic_centers_x,
+                ],
+                axis=-1,
+            )
+
+            translate_y = tf.stack(
+                [
+                    mosaic_centers_y - input_height,
+                    mosaic_centers_y - input_height,
+                    mosaic_centers_y,
+                    mosaic_centers_y,
+                ],
+                axis=-1,
+            )
+
+            bounding_boxes = tf.vectorized_map(
+                lambda index: self._update_bounding_box(
+                    images,
+                    bounding_boxes,
+                    permutation_order,
+                    translate_x,
+                    translate_y,
+                    index,
+                ),
+                tf.range(batch_size),
+            )
+            inputs["bounding_boxes"] = bounding_boxes
+        inputs["images"] = images
+        return inputs
+
+    def _augment(self, inputs):
+        raise ValueError(
+            "Mosaic received a single image to `call`.  The layer relies on "
+            "combining multiple examples, and as such will not behave as "
+            "expected.  Please call the layer with 4 or more samples."
+        )
+
+    def _update_image(self, images, permutation_order, mosaic_centers, index):
+        # forms mosaic for one image from the batch
+        input_height, input_width, _ = images.shape[1:]
+        mosaic_images = tf.gather(images, permutation_order[index])
+
+        top = tf.concat([mosaic_images[0], mosaic_images[1]], axis=1)
+        bottom = tf.concat([mosaic_images[2], mosaic_images[3]], axis=1)
+        output = tf.concat([top, bottom], axis=0)
+
+        # cropping coordinates for the mosaic
+        x1 = (input_width - mosaic_centers[index][0]) / (input_width * 2 - 1)
+        y1 = (input_height - mosaic_centers[index][1]) / (input_height * 2 - 1)
+        x2 = x1 + (input_width) / (input_width * 2 - 1)
+        y2 = y1 + (input_height) / (input_height * 2 - 1)
+
+        # helps avoid retracing caused by slicing, inspired by RRC implementation
+        output = tf.image.crop_and_resize(
+            tf.expand_dims(output, axis=0),
+            [[y1, x1, y2, x2]],
+            [0],
+            [input_height, input_width],
+        )
+        return tf.squeeze(output)
+
+    def _update_label(self, images, labels, permutation_order, mosaic_centers, index):
+        # updates labels for one output mosaic
+        input_height, input_width, _ = images.shape[1:]
+        labels_for_mosaic = tf.gather(labels, permutation_order[index])
+        center_x = mosaic_centers[index][0]
+        center_y = mosaic_centers[index][1]
+
+        area = input_height * input_width
+
+        # labels are in the same ratio as the area of the images
+        top_left_ratio = (center_x * center_y) / area
+        top_right_ratio = ((input_width - center_x) * center_y) / area
+        bottom_left_ratio = (center_x * (input_height - center_y)) / area
+        bottom_right_ratio = (
+            (input_width - center_x) * (input_height - center_y)
+        ) / area
+
+        label = (
+            labels_for_mosaic[0] * top_left_ratio
+            + labels_for_mosaic[1] * top_right_ratio
+            + labels_for_mosaic[2] * bottom_left_ratio
+            + labels_for_mosaic[3] * bottom_right_ratio
+        )
+        return label
+
+    def _update_bounding_box(
+        self, images, bounding_boxes, permutation_order, translate_x, translate_y, index
+    ):
+        # updates bboxes for one output mosaic
+        bounding_boxes = bounding_box.convert_format(
+            bounding_boxes,
+            source=self.bounding_box_format,
+            target="xyxy",
+            images=images,
+        )
+
+        boxes_for_mosaic = tf.gather(bounding_boxes, permutation_order[index])
+        boxes_for_mosaic, rest = tf.split(
+            boxes_for_mosaic, [4, bounding_boxes.shape[-1] - 4], axis=-1
+        )
+
+        # stacking translate values such that the shape is (4, 1, 4) or (num_images, broadcast dim, coordinates)
+        translate_values = tf.stack(
+            [
+                translate_x[index],
+                translate_y[index],
+                translate_x[index],
+                translate_y[index],
+            ],
+            axis=-1,
+        )
+        translate_values = tf.expand_dims(translate_values, axis=1)
+
+        # translating boxes
+        boxes_for_mosaic = boxes_for_mosaic + translate_values
+        boxes_for_mosaic = tf.concat([boxes_for_mosaic, rest], axis=-1)
+        boxes_for_mosaic = tf.reshape(boxes_for_mosaic, [-1, bounding_boxes.shape[-1]])
+
+        boxes_for_mosaic = bounding_box.clip_to_image(
+            boxes_for_mosaic,
+            bounding_box_format="xyxy",
+            images=images[index],
+        )
+        boxes_for_mosaic = bounding_box.convert_format(
+            boxes_for_mosaic,
+            source="xyxy",
+            target=self.bounding_box_format,
+            images=images[index],
+        )
+        return boxes_for_mosaic
+
+    def _validate_inputs(self, inputs):
+        images = inputs.get("images", None)
+        labels = inputs.get("labels", None)
+        bounding_boxes = inputs.get("bounding_boxes", None)
+        if images is None or (labels is None and bounding_boxes is None):
+            raise ValueError(
+                "Mosaic expects inputs in a dictionary with format "
+                '{"images": images, "labels": labels}. or'
+                '{"images": images, "bounding_boxes": bounding_boxes}'
+                f"Got: inputs = {inputs}"
+            )
+        if labels is not None and not labels.dtype.is_floating:
+            raise ValueError(
+                f"Mosaic received labels with type {labels.dtype}. "
+                "Labels must be of type float."
+            )
+        if bounding_boxes is not None and self.bounding_box_format is None:
+            raise ValueError(
+                "Mosaic received bounding boxes but no bounding_box_format. "
+                "Please pass a bounding_box_format from the supported list."
+            )
+
+    def get_config(self):
+        config = {
+            "offset": self.offset,
+            "bounding_box_format": self.bounding_box_format,
+            "seed": self.seed,
+        }
+        base_config = super().get_config()
+
+        return dict(list(base_config.items()) + list(config.items()))

Keras/keras-cv/keras_cv/layers/preprocessing/mosaic_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
+
+import keras_cv
+from keras_cv.layers.preprocessing.mosaic import Mosaic
+
+classes = 10
+
+
+class MosaicTest(tf.test.TestCase):
+    def DISABLED_test_integration_retina_net(self):
+        train_ds, train_dataset_info = keras_cv.datasets.pascal_voc.load(
+            bounding_box_format="xywh", split="train", batch_size=9
+        )
+        mosaic = keras_cv.layers.Mosaic(bounding_box_format="xywh")
+        train_ds = train_ds.map(mosaic, num_parallel_calls=tf.data.AUTOTUNE)
+        train_ds = train_ds.map(
+            lambda inputs: (inputs["images"], inputs["bounding_boxes"]),
+            num_parallel_calls=tf.data.AUTOTUNE,
+        )
+        model = keras_cv.models.RetinaNet(
+            classes=20,
+            bounding_box_format="xywh",
+            backbone="resnet50",
+            backbone_weights="imagenet",
+            include_rescaling=True,
+            evaluate_train_time_metrics=False,
+        )
+        model.backbone.trainable = False
+        optimizer = tf.optimizers.SGD(global_clipnorm=10.0)
+        model.compile(
+            run_eagerly=True,
+            classification_loss=keras_cv.losses.FocalLoss(
+                from_logits=True, reduction="none"
+            ),
+            box_loss=keras_cv.losses.SmoothL1Loss(l1_cutoff=1.0, reduction="none"),
+            optimizer=optimizer,
+        )
+        callbacks = [
+            tf.keras.callbacks.ReduceLROnPlateau(patience=5),
+            tf.keras.callbacks.TerminateOnNaN(),
+        ]
+        history = model.fit(train_ds, epochs=20, callbacks=callbacks)
+
+        for loss in history.history["loss"]:
+            self.assertFalse(tf.math.is_nan(loss))
+
+    def test_return_shapes(self):
+        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, 5), 0, 1)
+
+        layer = Mosaic(bounding_box_format="xywh")
+        # mosaic 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, 12, 5])
+
+    def test_in_tf_function(self):
+        xs = tf.cast(
+            tf.stack(
+                [2 * tf.ones((4, 4, 3)), tf.ones((4, 4, 3))],
+                axis=0,
+            ),
+            tf.float32,
+        )
+        ys = tf.one_hot(tf.constant([0, 1]), 2)
+
+        layer = Mosaic()
+
+        @tf.function
+        def augment(x, y):
+            return layer({"images": x, "labels": y})
+
+        outputs = augment(xs, ys)
+        xs, ys = outputs["images"], outputs["labels"]
+
+        self.assertEqual(xs.shape, [2, 4, 4, 3])
+        self.assertEqual(ys.shape, [2, 2])
+
+    def test_image_input_only(self):
+        xs = tf.cast(
+            tf.stack([2 * tf.ones((100, 100, 1)), tf.ones((100, 100, 1))], axis=0),
+            tf.float32,
+        )
+        layer = Mosaic()
+        with self.assertRaisesRegexp(ValueError, "expects inputs in a dictionary"):
+            _ = layer(xs)
+
+    def test_single_image_input(self):
+        xs = tf.ones((512, 512, 3))
+        ys = tf.one_hot(tf.constant([1]), 2)
+        inputs = {"images": xs, "labels": ys}
+        layer = Mosaic()
+        with self.assertRaisesRegexp(
+            ValueError, "Mosaic received a single image to `call`"
+        ):
+            _ = layer(inputs)
+
+    def test_int_labels(self):
+        xs = tf.ones((2, 512, 512, 3))
+        ys = tf.one_hot(tf.constant([1, 0]), 2, dtype=tf.int32)
+        inputs = {"images": xs, "labels": ys}
+        layer = Mosaic()
+        with self.assertRaisesRegexp(ValueError, "Mosaic received labels with type"):
+            _ = layer(inputs)
+
+    def test_image_input(self):
+        xs = tf.ones((2, 512, 512, 3))
+        layer = Mosaic()
+        with self.assertRaisesRegexp(
+            ValueError, "Mosaic expects inputs in a dictionary with format"
+        ):
+            _ = layer(xs)

Keras/keras-cv/keras_cv/layers/preprocessing/posterization.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.preprocessing import transform_value_range
+
+
+@tf.keras.utils.register_keras_serializable(package="keras_cv")
+class Posterization(BaseImageAugmentationLayer):
+    """Reduces the number of bits for each color channel.
+
+    References:
+    - [AutoAugment: Learning Augmentation Policies from Data](
+        https://arxiv.org/abs/1805.09501
+    )
+    - [RandAugment: Practical automated data augmentation with a reduced search space](
+        https://arxiv.org/abs/1909.13719
+    )
+
+    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`. Defaults to `(0, 255)`.
+        bits: integer. The number of bits to keep for each channel. Must be a value
+            between 1-8.
+
+     Usage:
+    ```python
+    (images, labels), _ = tf.keras.datasets.cifar10.load_data()
+    print(images[0, 0, 0])
+    # [59 62 63]
+    # Note that images are Tensors with values in the range [0, 255] and uint8 dtype
+    posterization = Posterization(bits=4, value_range=[0, 255])
+    images = posterization(images)
+    print(images[0, 0, 0])
+    # [48., 48., 48.]
+    # NOTE: the layer will output values in tf.float32, regardless of input dtype.
+    ```
+
+     Call arguments:
+        inputs: input tensor in two possible formats:
+            1. single 3D (HWC) image or 4D (NHWC) batch of images.
+            2. A dict of tensors where the images are under `"images"` key.
+    """
+
+    def __init__(self, value_range, bits, **kwargs):
+        super().__init__(**kwargs)
+
+        if not len(value_range) == 2:
+            raise ValueError(
+                "value_range must be a sequence of two elements. "
+                f"Received: {value_range}"
+            )
+
+        if not (0 < bits < 9):
+            raise ValueError(f"Bits value must be between 1-8. Received bits: {bits}.")
+
+        self._shift = 8 - bits
+        self._value_range = value_range
+
+    def augment_image(self, image, **kwargs):
+        image = transform_value_range(
+            images=image,
+            original_range=self._value_range,
+            target_range=[0, 255],
+        )
+        image = tf.cast(image, tf.uint8)
+
+        image = self._posterize(image)
+
+        image = tf.cast(image, self.compute_dtype)
+        return transform_value_range(
+            images=image,
+            original_range=[0, 255],
+            target_range=self._value_range,
+        )
+
+    def augment_bounding_boxes(self, bounding_boxes, **kwargs):
+        return bounding_boxes
+
+    def augment_segmentation_mask(self, segmentation_mask, transformation, **kwargs):
+        return segmentation_mask
+
+    def _batch_augment(self, inputs):
+        # Skip the use of vectorized_map or map_fn as the implementation is already
+        # vectorized
+        return self._augment(inputs)
+
+    def _posterize(self, image):
+        return tf.bitwise.left_shift(
+            tf.bitwise.right_shift(image, self._shift), self._shift
+        )
+
+    def augment_label(self, label, transformation=None, **kwargs):
+        return label
+
+    def get_config(self):
+        config = {"bits": 8 - self._shift, "value_range": self._value_range}
+        base_config = super().get_config()
+        return dict(list(base_config.items()) + list(config.items()))