Update keras cv docs to the newest release

guides/ipynb/keras_cv/coco_metrics.ipynb

@@ -40,26 +40,26 @@
    "source": [
     "## Input format\n",
     "\n",
-    "KerasCV COCO metrics require a specific input format.\n",
+    "All KerasCV components that process bounding boxes, including COCO metrics, require a\n",
+    "`bounding_box_format` parameter.  This parameter is used to tell the components what\n",
+    "format your bounding boxes are in.  While this guide uses the `xyxy` format, a full\n",
+    "list of supported formats is available in\n",
+    "[the bounding_box API documentation](/api/keras_cv/bounding_box/formats).\n",
     "\n",
     "The metrics expect `y_true` and be a `float` Tensor with the shape `[batch,\n",
-    "num_images, num_boxes, 5]`. The final axis stores the locational and class\n",
-    "information for each specific bounding box. The dimensions in order are: `[left,\n",
-    "top, right, bottom, class]`.\n",
-    "\n",
-    "The metrics expect `y_pred` and be a `float` Tensor with the shape `[batch,\n",
-    "num_images, num_boxes, 56]`. The final axis stores the locational and class\n",
-    "information for each specific bounding box. The dimensions in order are: `[left,\n",
-    "top, right, bottom, class, confidence]`.\n",
+    "num_images, num_boxes, 5]`, with the ordering of last set of axes determined by the\n",
+    "provided format.  The same is true of `y_pred`, except that an additional `confidence`\n",
+    "axis must be provided.\n",
     "\n",
     "Due to the fact that each image may have a different number of bounding boxes,\n",
     "the `num_boxes` dimension may actually have a mismatching shape between images.\n",
     "KerasCV works around this by allowing you to either pass a `RaggedTensor` as an\n",
     "input to the KerasCV COCO metrics, or padding unused bounding boxes with `-1`.\n",
     "\n",
     "Utility functions to manipulate bounding boxes, transform between formats, and\n",
-    "pad bounding box Tensors with `-1s` are available at\n",
-    "[`keras_cv.bounding_box`](https://github.com/keras-team/keras-cv/blob/master/keras_cv/bounding_box)."
+    "pad bounding box Tensors with `-1s` are available from the\n",
+    "[`keras_cv.bounding_box`](https://github.com/keras-team/keras-cv/blob/master/keras_cv/bounding_box)\n",
+    "package."
    ]
   },
   {
@@ -102,7 +102,9 @@
     "from tensorflow import keras\n",
     "\n",
     "# only consider boxes with areas less than a 32x32 square.\n",
-    "metric = keras_cv.metrics.COCORecall(class_ids=[1, 2, 3], area_range=(0, 32**2))"
+    "metric = keras_cv.metrics.COCORecall(\n",
+    "    bounding_box_format=\"xyxy\", class_ids=[1, 2, 3], area_range=(0, 32**2)\n",
+    ")"
    ]
   },
   {
@@ -254,7 +256,11 @@
    "outputs": [],
    "source": [
     "recall = keras_cv.metrics.COCORecall(\n",
-    "    max_detections=100, class_ids=[1], area_range=(0, 64**2), name=\"coco_recall\"\n",
+    "    bounding_box_format=\"xyxy\",\n",
+    "    max_detections=100,\n",
+    "    class_ids=[1],\n",
+    "    area_range=(0, 64**2),\n",
+    "    name=\"coco_recall\",\n",
     ")\n",
     "model.compile(metrics=[recall])"
    ]

guides/ipynb/keras_cv/custom_image_augmentations.ipynb

@@ -658,6 +658,11 @@
     "colab_type": "text"
    },
    "source": [
+    "Additionally, be sure to accept `**kwargs` to your `augment_*` methods to ensure\n",
+    "forwards compatibility.  KerasCV will add additional label types in the future, and\n",
+    "if you do not include a `**kwargs` argument your augmentation layers will not be\n",
+    "forward compatible.\n",
+    "\n",
     "## Conclusion and next steps\n",
     "\n",
     "KerasCV offers a standard set of APIs to streamline the process of implementing your\n",

guides/ipynb/keras_cv/cut_mix_mix_up_and_rand_augment.ipynb

@@ -26,7 +26,8 @@
     "pipelines for image classification and object detection tasks. KerasCV offers a wide\n",
     "suite of preprocessing layers implementing common data augmentation techniques.\n",
     "\n",
-    "Perhaps three of the most useful layers are `CutMix`, `MixUp`, and `RandAugment`. These\n",
+    "Perhaps three of the most useful layers are `keras_cv.layers.CutMix`,\n",
+    "`keras_cv.layers.MixUp`, and `keras_cv.layers.RandAugment`. These\n",
     "layers are used in nearly all state-of-the-art image classification pipelines.\n",
     "\n",
     "This guide will show you how to compose these layers into your own data\n",
@@ -124,7 +125,7 @@
     "num_classes = dataset_info.features[\"label\"].num_classes\n",
     "\n",
     "\n",
-    "def prepare(image, label):\n",
+    "def to_dict(image, label):\n",
     "    image = tf.image.resize(image, IMAGE_SIZE)\n",
     "    image = tf.cast(image, tf.float32)\n",
     "    label = tf.one_hot(label, num_classes)\n",
@@ -135,11 +136,11 @@
     "    if split == \"train\":\n",
     "        return (\n",
     "            dataset.shuffle(10 * BATCH_SIZE)\n",
-    "            .map(prepare, num_parallel_calls=AUTOTUNE)\n",
+    "            .map(to_dict, num_parallel_calls=AUTOTUNE)\n",
     "            .batch(BATCH_SIZE)\n",
     "        )\n",
     "    if split == \"test\":\n",
-    "        return dataset.map(prepare, num_parallel_calls=AUTOTUNE).batch(BATCH_SIZE)\n",
+    "        return dataset.map(to_dict, num_parallel_calls=AUTOTUNE).batch(BATCH_SIZE)\n",
     "\n",
     "\n",
     "def load_dataset(split=\"train\"):\n",
@@ -305,11 +306,7 @@
     "\n",
     "In this example, we will use `CutMix` and `MixUp` independently in a manually created\n",
     "preprocessing pipeline. In most state of the art pipelines images are randomly\n",
-    "augmented by either `CutMix`, `MixUp`, or neither. The function below implements this\n",
-    "in an equal 1/3 split.\n",
-    "\n",
-    "Note that our `cut_mix_and_mix_up` function is annotated with a `tf.function` to ensure\n",
-    "optimal performance."
+    "augmented by either `CutMix`, `MixUp`, or neither. The function below implements both."
    ]
   },
   {
@@ -354,7 +351,8 @@
     "## Customizing your augmentation pipeline\n",
     "\n",
     "Perhaps you want to exclude an augmentation from `RandAugment`, or perhaps you want to\n",
-    "include the `GridMask()` as an option alongside the default `RandAugment` augmentations.\n",
+    "include the `keras_cv.layers.GridMask` as an option alongside the default `RandAugment`\n",
+    "augmentations.\n",
     "\n",
     "KerasCV allows you to construct production grade custom data augmentation pipelines using\n",
     "the `keras_cv.layers.RandomAugmentationPipeline` layer. This class operates similarly to\n",
@@ -419,7 +417,7 @@
     "colab_type": "text"
    },
    "source": [
-    "Next, let's add `GridMask` to our layers:"
+    "Next, let's add `keras_cv.layers.GridMask` to our layers:"
    ]
   },
   {
@@ -452,7 +450,13 @@
    "source": [
     "pipeline = keras_cv.layers.RandomAugmentationPipeline(\n",
     "    layers=layers, augmentations_per_image=3\n",
-    ")"
+    ")\n",
+    "\n",
+    "\n",
+    "def apply_pipeline(inputs):\n",
+    "    inputs[\"images\"] = pipeline(inputs[\"images\"])\n",
+    "    return inputs\n",
+    ""
    ]
   },
   {
@@ -472,12 +476,6 @@
    },
    "outputs": [],
    "source": [
-    "\n",
-    "def apply_pipeline(inputs):\n",
-    "    inputs[\"images\"] = pipeline(inputs[\"images\"])\n",
-    "    return inputs\n",
-    "\n",
-    "\n",
     "train_dataset = load_dataset().map(apply_pipeline, num_parallel_calls=AUTOTUNE)\n",
     "visualize_dataset(train_dataset, title=\"After custom pipeline\")"
    ]
@@ -523,11 +521,6 @@
    },
    "outputs": [],
    "source": [
-    "\n",
-    "def apply_pipeline(inputs):\n",
-    "    inputs[\"images\"] = pipeline(inputs[\"images\"])\n",
-    "    return inputs\n",
-    "\n",
     "\n",
     "train_dataset = load_dataset().map(apply_pipeline, num_parallel_calls=AUTOTUNE)\n",
     "visualize_dataset(train_dataset, title=\"After custom pipeline\")"
@@ -612,11 +605,9 @@
     "\n",
     "\n",
     "def get_model():\n",
-    "    inputs = keras.layers.Input(input_shape)\n",
-    "    x = applications.ResNet50V2(\n",
-    "        input_shape=input_shape, classes=num_classes, weights=None\n",
-    "    )(inputs)\n",
-    "    model = keras.Model(inputs, x)\n",
+    "    model = keras_cv.models.DenseNet121(\n",
+    "        include_rescaling=True, include_top=True, classes=num_classes\n",
+    "    )\n",
     "    model.compile(\n",
     "        loss=losses.CategoricalCrossentropy(label_smoothing=0.1),\n",
     "        optimizer=optimizers.SGD(momentum=0.9),\n",

guides/keras_cv/coco_metrics.py

@@ -22,26 +22,26 @@
 """
 ## Input format
 
-KerasCV COCO metrics require a specific input format.
+All KerasCV components that process bounding boxes, including COCO metrics, require a
+`bounding_box_format` parameter.  This parameter is used to tell the components what
+format your bounding boxes are in.  While this guide uses the `xyxy` format, a full
+list of supported formats is available in
+[the bounding_box API documentation](/api/keras_cv/bounding_box/formats).
 
 The metrics expect `y_true` and be a `float` Tensor with the shape `[batch,
-num_images, num_boxes, 5]`. The final axis stores the locational and class
-information for each specific bounding box. The dimensions in order are: `[left,
-top, right, bottom, class]`.
-
-The metrics expect `y_pred` and be a `float` Tensor with the shape `[batch,
-num_images, num_boxes, 56]`. The final axis stores the locational and class
-information for each specific bounding box. The dimensions in order are: `[left,
-top, right, bottom, class, confidence]`.
+num_images, num_boxes, 5]`, with the ordering of last set of axes determined by the
+provided format.  The same is true of `y_pred`, except that an additional `confidence`
+axis must be provided.
 
 Due to the fact that each image may have a different number of bounding boxes,
 the `num_boxes` dimension may actually have a mismatching shape between images.
 KerasCV works around this by allowing you to either pass a `RaggedTensor` as an
 input to the KerasCV COCO metrics, or padding unused bounding boxes with `-1`.
 
 Utility functions to manipulate bounding boxes, transform between formats, and
-pad bounding box Tensors with `-1s` are available at
-[`keras_cv.bounding_box`](https://github.com/keras-team/keras-cv/blob/master/keras_cv/bounding_box).
+pad bounding box Tensors with `-1s` are available from the
+[`keras_cv.bounding_box`](https://github.com/keras-team/keras-cv/blob/master/keras_cv/bounding_box)
+package.
 
 """
 
@@ -67,7 +67,9 @@
 from tensorflow import keras
 
 # only consider boxes with areas less than a 32x32 square.
-metric = keras_cv.metrics.COCORecall(class_ids=[1, 2, 3], area_range=(0, 32**2))
+metric = keras_cv.metrics.COCORecall(
+    bounding_box_format="xyxy", class_ids=[1, 2, 3], area_range=(0, 32**2)
+)
 
 """
 2.) Create Some Bounding Boxes:
@@ -130,7 +132,11 @@
 """
 
 recall = keras_cv.metrics.COCORecall(
-    max_detections=100, class_ids=[1], area_range=(0, 64**2), name="coco_recall"
+    bounding_box_format="xyxy",
+    max_detections=100,
+    class_ids=[1],
+    area_range=(0, 64**2),
+    name="coco_recall",
 )
 model.compile(metrics=[recall])
 

guides/keras_cv/custom_image_augmentations.py

@@ -430,6 +430,12 @@ def __init__(self, **kwargs):
 
 
 """
+
+Additionally, be sure to accept `**kwargs` to your `augment_*` methods to ensure
+forwards compatibility.  KerasCV will add additional label types in the future, and
+if you do not include a `**kwargs` argument your augmentation layers will not be
+forward compatible.
+
 ## Conclusion and next steps
 
 KerasCV offers a standard set of APIs to streamline the process of implementing your

guides/keras_cv/cut_mix_mix_up_and_rand_augment.py

@@ -13,7 +13,8 @@
 pipelines for image classification and object detection tasks. KerasCV offers a wide
 suite of preprocessing layers implementing common data augmentation techniques.
 
-Perhaps three of the most useful layers are `CutMix`, `MixUp`, and `RandAugment`. These
+Perhaps three of the most useful layers are `keras_cv.layers.CutMix`,
+`keras_cv.layers.MixUp`, and `keras_cv.layers.RandAugment`. These
 layers are used in nearly all state-of-the-art image classification pipelines.
 
 This guide will show you how to compose these layers into your own data
@@ -70,7 +71,7 @@
 num_classes = dataset_info.features["label"].num_classes
 
 
-def prepare(image, label):
+def to_dict(image, label):
     image = tf.image.resize(image, IMAGE_SIZE)
     image = tf.cast(image, tf.float32)
     label = tf.one_hot(label, num_classes)
@@ -81,11 +82,11 @@ def prepare_dataset(dataset, split):
     if split == "train":
         return (
             dataset.shuffle(10 * BATCH_SIZE)
-            .map(prepare, num_parallel_calls=AUTOTUNE)
+            .map(to_dict, num_parallel_calls=AUTOTUNE)
             .batch(BATCH_SIZE)
         )
     if split == "test":
-        return dataset.map(prepare, num_parallel_calls=AUTOTUNE).batch(BATCH_SIZE)
+        return dataset.map(to_dict, num_parallel_calls=AUTOTUNE).batch(BATCH_SIZE)
 
 
 def load_dataset(split="train"):
@@ -214,7 +215,8 @@ def cut_mix_and_mix_up(samples):
 ## Customizing your augmentation pipeline
 
 Perhaps you want to exclude an augmentation from `RandAugment`, or perhaps you want to
-include the `GridMask()` as an option alongside the default `RandAugment` augmentations.
+include the `keras_cv.layers.GridMask` as an option alongside the default `RandAugment`
+augmentations.
 
 KerasCV allows you to construct production grade custom data augmentation pipelines using
 the `keras_cv.layers.RandomAugmentationPipeline` layer. This class operates similarly to
@@ -246,7 +248,7 @@ def cut_mix_and_mix_up(samples):
 ]
 
 """
-Next, let's add `GridMask` to our layers:
+Next, let's add `keras_cv.layers.GridMask` to our layers:
 """
 
 layers = layers + [keras_cv.layers.GridMask()]
@@ -259,16 +261,16 @@ def cut_mix_and_mix_up(samples):
     layers=layers, augmentations_per_image=3
 )
 
-"""
-Let's check out the results!
-"""
-
 
 def apply_pipeline(inputs):
     inputs["images"] = pipeline(inputs["images"])
     return inputs
 
 
+"""
+Let's check out the results!
+"""
+
 train_dataset = load_dataset().map(apply_pipeline, num_parallel_calls=AUTOTUNE)
 visualize_dataset(train_dataset, title="After custom pipeline")
 
@@ -287,11 +289,6 @@ def apply_pipeline(inputs):
 """
 
 
-def apply_pipeline(inputs):
-    inputs["images"] = pipeline(inputs["images"])
-    return inputs
-
-
 train_dataset = load_dataset().map(apply_pipeline, num_parallel_calls=AUTOTUNE)
 visualize_dataset(train_dataset, title="After custom pipeline")
 
@@ -343,11 +340,9 @@ def preprocess_for_model(inputs):
 
 
 def get_model():
-    inputs = keras.layers.Input(input_shape)
-    x = applications.ResNet50V2(
-        input_shape=input_shape, classes=num_classes, weights=None
-    )(inputs)
-    model = keras.Model(inputs, x)
+    model = keras_cv.models.DenseNet121(
+        include_rescaling=True, include_top=True, classes=num_classes
+    )
     model.compile(
         loss=losses.CategoricalCrossentropy(label_smoothing=0.1),
         optimizer=optimizers.SGD(momentum=0.9),