Merge branch 'master' into mlpmixer

README.md

@@ -31,59 +31,73 @@ To learn more about the future project direction, please check the [roadmap](.gi
 - [Roadmap](.github/ROADMAP.md)
 - [API Design Guidelines](.github/API_DESIGN.md)
 
-## Quickstart
+## Installation
+
+To install the latest official release:
+
+```
+pip install keras-cv tensorflow --upgrade
+```
 
-Create a preprocessing pipeline:
+To install the latest unreleased changes to the library, we recommend using
+pip to install directly from the master branch on github:
+
+```
+pip install git+https://github.com/keras-team/keras-cv.git tensorflow --upgrade
+```
+
+## Quickstart
 
 ```python
 import keras_cv
 import tensorflow as tf
 from tensorflow import keras
 import tensorflow_datasets as tfds
 
+# Create a preprocessing pipeline
 augmenter = keras_cv.layers.Augmenter(
-  layers=[
-      keras_cv.layers.RandomFlip(),
-      keras_cv.layers.RandAugment(value_range=(0, 255)),
-      keras_cv.layers.CutMix(),
-      keras_cv.layers.MixUp()
+    layers=[
+        keras_cv.layers.RandomFlip(),
+        keras_cv.layers.RandAugment(value_range=(0, 255)),
+        keras_cv.layers.CutMix(),
+        keras_cv.layers.MixUp()
     ]
 )
 
-def augment_data(images, labels):
-  labels = tf.one_hot(labels, 3)
-  inputs = {"images": images, "labels": labels}
-  outputs = augmenter(inputs)
-  return outputs['images'], outputs['labels']
-```
-
-Augment a `tf.data.Dataset`:
-
-```python
-dataset = tfds.load('rock_paper_scissors', as_supervised=True, split='train')
-dataset = dataset.batch(64)
-dataset = dataset.map(augment_data, num_parallel_calls=tf.data.AUTOTUNE)
-```
-
-Create a model:
-
-```python
+def preprocess_data(images, labels, augment=False):
+    labels = tf.one_hot(labels, 3)
+    inputs = {"images": images, "labels": labels}
+    outputs = augmenter(inputs) if augment else inputs
+    return outputs['images'], outputs['labels']
+
+# Augment a `tf.data.Dataset`
+train_dataset, test_dataset = tfds.load(
+    'rock_paper_scissors',
+    as_supervised=True,
+    split=['train', 'test'],
+)
+train_dataset = train_dataset.batch(16).map(
+    lambda x, y: preprocess_data(x, y, augment=True),
+        num_parallel_calls=tf.data.AUTOTUNE).prefetch(
+            tf.data.AUTOTUNE)
+test_dataset = test_dataset.batch(16).map(
+    preprocess_data, num_parallel_calls=tf.data.AUTOTUNE).prefetch(
+        tf.data.AUTOTUNE)
+
+# Create a model
 densenet = keras_cv.models.DenseNet121(
-  include_rescaling=True,
-  include_top=True,
-  classes=3
+    include_rescaling=True,
+    include_top=True,
+    classes=3
 )
 densenet.compile(
-  loss='categorical_crossentropy',
-  optimizer='adam',
-  metrics=['accuracy']
+    loss='categorical_crossentropy',
+    optimizer='adam',
+    metrics=['accuracy']
 )
-```
 
-Train your model:
-
-```python
-densenet.fit(dataset)
+# Train your model
+densenet.fit(train_dataset, validation_data=test_dataset)
 ```
 
 ## Contributors

benchmarks/metrics/coco/mean_average_precision_bucket_performance.py

@@ -35,10 +35,7 @@ def produce_random_data(include_confidence=False, num_images=128, classes=20):
             )
         )
 
-    images = [
-        keras_cv.bounding_box.pad_batch_to_shape(x, [25, images[0].shape[1]])
-        for x in images
-    ]
+    images = [keras_cv.bounding_box.to_dense(x, max_boxes=25) for x in images]
     return tf.stack(images, axis=0)
 
 

benchmarks/metrics/coco/mean_average_precision_performance.py

@@ -35,10 +35,7 @@ def produce_random_data(include_confidence=False, num_images=128, classes=20):
             )
         )
 
-    images = [
-        keras_cv.bounding_box.pad_batch_to_shape(x, [25, images[0].shape[1]])
-        for x in images
-    ]
+    images = [keras_cv.bounding_box.to_dense(x, max_boxes=25) for x in images]
     return tf.stack(images, axis=0)
 
 

benchmarks/metrics/coco/recall_performance.py

@@ -35,10 +35,7 @@ def produce_random_data(include_confidence=False, num_images=128, classes=20):
             )
         )
 
-    images = [
-        keras_cv.bounding_box.pad_batch_to_shape(x, [25, images[0].shape[1]])
-        for x in images
-    ]
+    images = [keras_cv.bounding_box.to_dense(x, max_boxes=25) for x in images]
     return tf.stack(images, axis=0)
 
 
@@ -56,7 +53,7 @@ def produce_random_data(include_confidence=False, num_images=128, classes=20):
 for images in n_images:
     y_true = produce_random_data(num_images=images)
     y_pred = produce_random_data(num_images=images, include_confidence=True)
-    metric = coco._COCORecall(class_ids)
+    metric = coco._BoxRecall(class_ids)
     # warm up
     metric.update_state(y_true, y_pred)
     metric.result()

benchmarks/vectorization_strategy_benchmark.py

@@ -72,7 +72,9 @@ def single_rectangle_mask(corners, mask_shape):
     return masks
 
 
-def fill_single_rectangle(image, centers_x, centers_y, widths, heights, fill_values):
+def fill_single_rectangle(
+    image, centers_x, centers_y, widths, heights, fill_values
+):
     """Fill rectangles with fill value into images.
 
     Args:

benchmarks/vectorized_auto_contrast.py

@@ -0,0 +1,176 @@
+# Copyright 2023 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 time
+
+import matplotlib.pyplot as plt
+import tensorflow as tf
+import tensorflow.keras as keras
+
+from keras_cv.layers import AutoContrast
+from keras_cv.layers.preprocessing.base_image_augmentation_layer import (
+    BaseImageAugmentationLayer,
+)
+from keras_cv.utils import preprocessing
+
+
+class OldAutoContrast(BaseImageAugmentationLayer):
+    """Performs the AutoContrast operation on an image.
+
+    Auto contrast stretches the values of an image across the entire available
+    `value_range`.  This makes differences between pixels more obvious.  An example of
+    this is if an image only has values `[0, 1]` out of the range `[0, 255]`, auto
+    contrast will change the `1` values to be `255`.
+
+    Args:
+        value_range: the range of values the incoming images will have.
+            Represented as a two number tuple written [low, high].
+            This is typically either `[0, 1]` or `[0, 255]` depending
+            on how your preprocessing pipeline is setup.
+    """
+
+    def __init__(
+        self,
+        value_range,
+        **kwargs,
+    ):
+        super().__init__(**kwargs)
+        self.value_range = value_range
+
+    def augment_image(self, image, transformation=None, **kwargs):
+        original_image = image
+        image = preprocessing.transform_value_range(
+            image,
+            original_range=self.value_range,
+            target_range=(0, 255),
+            dtype=self.compute_dtype,
+        )
+
+        low = tf.reduce_min(tf.reduce_min(image, axis=0), axis=0)
+        high = tf.reduce_max(tf.reduce_max(image, axis=0), axis=0)
+        scale = 255.0 / (high - low)
+        offset = -low * scale
+
+        image = image * scale[None, None] + offset[None, None]
+        result = tf.clip_by_value(image, 0.0, 255.0)
+        result = preprocessing.transform_value_range(
+            result,
+            original_range=(0, 255),
+            target_range=self.value_range,
+            dtype=self.compute_dtype,
+        )
+        # don't process NaN channels
+        result = tf.where(tf.math.is_nan(result), original_image, result)
+        return result
+
+    def augment_bounding_boxes(self, bounding_boxes, **kwargs):
+        return bounding_boxes
+
+    def augment_label(self, label, transformation=None, **kwargs):
+        return label
+
+    def augment_segmentation_mask(
+        self, segmentation_mask, transformation, **kwargs
+    ):
+        return segmentation_mask
+
+    def get_config(self):
+        config = super().get_config()
+        config.update({"value_range": self.value_range})
+        return config
+
+
+class AutoContrastConsistencyTest(tf.test.TestCase):
+    def test_consistency_with_old_implementation(self):
+        images = tf.random.uniform(shape=(16, 32, 32, 3))
+
+        output = AutoContrast(value_range=(0, 1))(images)
+        old_output = OldAutoContrast(value_range=(0, 1))(images)
+
+        self.assertAllClose(old_output, output)
+
+
+if __name__ == "__main__":
+    (x_train, _), _ = keras.datasets.cifar10.load_data()
+    x_train = x_train.astype(float)
+
+    images = []
+    num_images = [1000, 2000, 5000, 10000]
+    results = {}
+
+    for aug in [AutoContrast, OldAutoContrast]:
+        c = aug.__name__
+
+        layer = aug(value_range=(0, 255))
+
+        runtimes = []
+        print(f"Timing {c}")
+
+        for n_images in num_images:
+            # warmup
+            layer(x_train[:n_images])
+
+            t0 = time.time()
+            r1 = layer(x_train[:n_images])
+            t1 = time.time()
+            runtimes.append(t1 - t0)
+            print(f"Runtime for {c}, n_images={n_images}: {t1 - t0}")
+
+        results[c] = runtimes
+
+        c = aug.__name__ + " Graph Mode"
+
+        layer = aug(value_range=(0, 255))
+
+        @tf.function()
+        def apply_aug(inputs):
+            return layer(inputs)
+
+        runtimes = []
+        print(f"Timing {c}")
+
+        for n_images in num_images:
+            # warmup
+            apply_aug(x_train[:n_images])
+
+            t0 = time.time()
+            r1 = apply_aug(x_train[:n_images])
+            t1 = time.time()
+            runtimes.append(t1 - t0)
+            print(f"Runtime for {c}, n_images={n_images}: {t1 - t0}")
+
+        results[c] = runtimes
+
+    plt.figure()
+    for key in results:
+        plt.plot(num_images, results[key], label=key)
+        plt.xlabel("Number images")
+
+    plt.ylabel("Runtime (seconds)")
+    plt.legend()
+    plt.show()
+
+    # So we can actually see more relevant margins
+    del results["OldAutoContrast"]
+
+    plt.figure()
+    for key in results:
+        plt.plot(num_images, results[key], label=key)
+        plt.xlabel("Number images")
+
+    plt.ylabel("Runtime (seconds)")
+    plt.legend()
+    plt.show()
+
+    # Compare two implementations
+    tf.test.main()