add sdv3 code

benchmarks/vectorized_randomly_zoomed_crop.py

@@ -249,10 +249,10 @@ def from_config(cls, config):
                 config["zoom_factor"]
             )
         if isinstance(config["aspect_ratio_factor"], dict):
-            config["aspect_ratio_factor"] = (
-                keras.utils.deserialize_keras_object(
-                    config["aspect_ratio_factor"]
-                )
+            config[
+                "aspect_ratio_factor"
+            ] = keras.utils.deserialize_keras_object(
+                config["aspect_ratio_factor"]
             )
         return cls(**config)
 

keras_cv/__init__.py

@@ -4,6 +4,7 @@
 since your modifications would be overwritten.
 """
 
+
 import os
 
 # Import everything from /api/ into keras.

keras_cv/api/__init__.py

@@ -4,6 +4,7 @@
 since your modifications would be overwritten.
 """
 
+
 from keras_cv.api import bounding_box
 from keras_cv.api import callbacks
 from keras_cv.api import core

keras_cv/api/bounding_box/__init__.py

@@ -4,6 +4,7 @@
 since your modifications would be overwritten.
 """
 
+
 from keras_cv.src.bounding_box.converters import convert_format
 from keras_cv.src.bounding_box.ensure_tensor import ensure_tensor
 from keras_cv.src.bounding_box.formats import CENTER_XYWH

keras_cv/api/callbacks/__init__.py

@@ -4,6 +4,7 @@
 since your modifications would be overwritten.
 """
 
+
 from keras_cv.src.callbacks.pycoco_callback import PyCOCOCallback
 from keras_cv.src.callbacks.waymo_evaluation_callback import (
     WaymoEvaluationCallback,

keras_cv/api/core/__init__.py

@@ -4,6 +4,7 @@
 since your modifications would be overwritten.
 """
 
+
 from keras_cv.src.core.factor_sampler.constant_factor_sampler import (
     ConstantFactorSampler,
 )

keras_cv/api/datasets/__init__.py

@@ -4,6 +4,7 @@
 since your modifications would be overwritten.
 """
 
+
 from keras_cv.api.datasets import imagenet
 from keras_cv.api.datasets import pascal_voc
 from keras_cv.api.datasets import waymo

keras_cv/api/datasets/imagenet/__init__.py

@@ -4,4 +4,5 @@
 since your modifications would be overwritten.
 """
 
+
 from keras_cv.src.datasets.imagenet.load import load

keras_cv/api/datasets/pascal_voc/__init__.py

@@ -4,5 +4,6 @@
 since your modifications would be overwritten.
 """
 
+
 from keras_cv.api.datasets.pascal_voc import segmentation
 from keras_cv.src.datasets.pascal_voc.load import load

keras_cv/api/datasets/pascal_voc/segmentation/__init__.py

@@ -4,4 +4,5 @@
 since your modifications would be overwritten.
 """
 
+
 from keras_cv.src.datasets.pascal_voc.segmentation import load

keras_cv/api/datasets/waymo/__init__.py

@@ -4,6 +4,7 @@
 since your modifications would be overwritten.
 """
 
+
 from keras_cv.src.datasets.waymo.load import load
 from keras_cv.src.datasets.waymo.transformer import (
     build_tensors_for_augmentation,

keras_cv/api/keypoint/__init__.py

@@ -4,6 +4,7 @@
 since your modifications would be overwritten.
 """
 
+
 from keras_cv.src.keypoint.converters import convert_format
 from keras_cv.src.keypoint.formats import REL_XY
 from keras_cv.src.keypoint.formats import XY

keras_cv/api/layers/__init__.py

@@ -4,6 +4,7 @@
 since your modifications would be overwritten.
 """
 
+
 from keras_cv.src.layers.augmenter import Augmenter
 from keras_cv.src.layers.feature_pyramid import FeaturePyramid
 from keras_cv.src.layers.fusedmbconv import FusedMBConvBlock
@@ -152,6 +153,7 @@
 from keras_cv.src.layers.vit_det_layers import WindowedTransformerEncoder
 from keras_cv.src.layers.vit_det_layers import WindowPartitioning
 from keras_cv.src.layers.vit_layers import PatchingAndEmbedding
+from keras_cv.src.layers.vit_layers import Unpatching
 from keras_cv.src.models.segmentation.segment_anything.sam_layers import (
     MultiHeadAttentionWithDownsampling,
 )

keras_cv/api/losses/__init__.py

@@ -4,6 +4,7 @@
 since your modifications would be overwritten.
 """
 
+
 from keras_cv.src.losses.centernet_box_loss import CenterNetBoxLoss
 from keras_cv.src.losses.ciou_loss import CIoULoss
 from keras_cv.src.losses.focal import FocalLoss

keras_cv/api/metrics/__init__.py

@@ -4,6 +4,7 @@
 since your modifications would be overwritten.
 """
 
+
 from keras_cv.src.metrics.object_detection.box_coco_metrics import (
     BoxCOCOMetrics,
 )

keras_cv/api/models/__init__.py

@@ -4,6 +4,7 @@
 since your modifications would be overwritten.
 """
 
+
 from keras_cv.api.models import classification
 from keras_cv.api.models import feature_extractor
 from keras_cv.api.models import object_detection

keras_cv/api/models/classification/__init__.py

@@ -4,5 +4,6 @@
 since your modifications would be overwritten.
 """
 
+
 from keras_cv.src.models.classification.image_classifier import ImageClassifier
 from keras_cv.src.models.classification.video_classifier import VideoClassifier

keras_cv/api/models/feature_extractor/__init__.py

@@ -4,6 +4,7 @@
 since your modifications would be overwritten.
 """
 
+
 from keras_cv.src.models.feature_extractor.clip.clip_encoder import (
     CLIPAttention,
 )

keras_cv/api/models/object_detection/__init__.py

@@ -4,6 +4,7 @@
 since your modifications would be overwritten.
 """
 
+
 from keras_cv.src.models.object_detection.retinanet.retinanet import RetinaNet
 from keras_cv.src.models.object_detection.yolo_v8.yolo_v8_detector import (
     YOLOV8Detector,

keras_cv/api/models/retinanet/__init__.py

@@ -4,6 +4,7 @@
 since your modifications would be overwritten.
 """
 
+
 from keras_cv.src.models.object_detection.retinanet.feature_pyramid import (
     FeaturePyramid,
 )

keras_cv/api/models/segmentation/__init__.py

@@ -4,6 +4,7 @@
 since your modifications would be overwritten.
 """
 
+
 from keras_cv.src.models.segmentation.basnet.basnet import BASNet
 from keras_cv.src.models.segmentation.deeplab_v3_plus.deeplab_v3_plus import (
     DeepLabV3Plus,

keras_cv/api/models/stable_diffusion/__init__.py

@@ -4,6 +4,7 @@
 since your modifications would be overwritten.
 """
 
+
 from keras_cv.src.models.stable_diffusion.clip_tokenizer import SimpleTokenizer
 from keras_cv.src.models.stable_diffusion.decoder import Decoder
 from keras_cv.src.models.stable_diffusion.diffusion_model import DiffusionModel

keras_cv/api/models/yolov8/__init__.py

@@ -4,6 +4,7 @@
 since your modifications would be overwritten.
 """
 
+
 from keras_cv.src.models.object_detection.yolo_v8.yolo_v8_label_encoder import (
     YOLOV8LabelEncoder as LabelEncoder,
 )

keras_cv/api/visualization/__init__.py

@@ -4,6 +4,7 @@
 since your modifications would be overwritten.
 """
 
+
 from keras_cv.src.visualization.draw_bounding_boxes import draw_bounding_boxes
 from keras_cv.src.visualization.plot_bounding_box_gallery import (
     plot_bounding_box_gallery,

keras_cv/src/layers/preprocessing/base_image_augmentation_layer.py

@@ -236,15 +236,15 @@ def _compute_output_signature(self, inputs):
         bounding_boxes = inputs.get(BOUNDING_BOXES, None)
 
         if bounding_boxes is not None:
-            fn_output_signature[BOUNDING_BOXES] = (
-                self._compute_bounding_box_signature(bounding_boxes)
-            )
+            fn_output_signature[
+                BOUNDING_BOXES
+            ] = self._compute_bounding_box_signature(bounding_boxes)
 
         segmentation_masks = inputs.get(SEGMENTATION_MASKS, None)
         if segmentation_masks is not None:
-            fn_output_signature[SEGMENTATION_MASKS] = (
-                self.compute_image_signature(segmentation_masks)
-            )
+            fn_output_signature[
+                SEGMENTATION_MASKS
+            ] = self.compute_image_signature(segmentation_masks)
 
         keypoints = inputs.get(KEYPOINTS, None)
         if keypoints is not None:

keras_cv/src/layers/preprocessing/random_crop_and_resize.py

@@ -272,10 +272,10 @@ def from_config(cls, config):
                 config["crop_area_factor"]
             )
         if isinstance(config["aspect_ratio_factor"], dict):
-            config["aspect_ratio_factor"] = (
-                keras.utils.deserialize_keras_object(
-                    config["aspect_ratio_factor"]
-                )
+            config[
+                "aspect_ratio_factor"
+            ] = keras.utils.deserialize_keras_object(
+                config["aspect_ratio_factor"]
             )
         return cls(**config)
 

keras_cv/src/layers/regularization/squeeze_excite.py

@@ -118,10 +118,10 @@ def get_config(self):
     @classmethod
     def from_config(cls, config):
         if isinstance(config["squeeze_activation"], dict):
-            config["squeeze_activation"] = (
-                keras.saving.deserialize_keras_object(
-                    config["squeeze_activation"]
-                )
+            config[
+                "squeeze_activation"
+            ] = keras.saving.deserialize_keras_object(
+                config["squeeze_activation"]
             )
         if isinstance(config["excite_activation"], dict):
             config["excite_activation"] = keras.saving.deserialize_keras_object(

keras_cv/src/layers/vit_layers.py

@@ -15,13 +15,14 @@
 import math
 
 import tensorflow as tf
-from tensorflow.keras import layers
 
 from keras_cv.src.api_export import keras_cv_export
+from keras_cv.src.backend import keras
+from keras_cv.src.backend import ops
 
 
 @keras_cv_export("keras_cv.layers.PatchingAndEmbedding")
-class PatchingAndEmbedding(layers.Layer):
+class PatchingAndEmbedding(keras.layers.Layer):
     """
     Layer to patchify images, prepend a class token, positionally embed and
     create a projection of patches for Vision Transformers
@@ -44,7 +45,7 @@ class PatchingAndEmbedding(layers.Layer):
         Patchified and linearly projected input images, including a prepended
         learnable class token with shape (batch, num_patches+1, project_dim)
 
-    Example:
+    Basic usage:
 
     ```
     images = #... batch of images
@@ -70,7 +71,7 @@ def __init__(self, project_dim, patch_size, padding="VALID", **kwargs):
                 f"Padding must be either 'SAME' or 'VALID', but {padding} was "
                 "passed."
             )
-        self.projection = layers.Conv2D(
+        self.projection = keras.layers.Conv2D(
             filters=self.project_dim,
             kernel_size=self.patch_size,
             strides=self.patch_size,
@@ -87,7 +88,7 @@ def build(self, input_shape):
             * input_shape[2]
             // self.patch_size
         )
-        self.position_embedding = layers.Embedding(
+        self.position_embedding = keras.layers.Embedding(
             input_dim=self.num_patches + 1, output_dim=self.project_dim
         )
 
@@ -221,3 +222,78 @@ def get_config(self):
         }
         base_config = super().get_config()
         return dict(list(base_config.items()) + list(config.items()))
+
+
+@keras_cv_export("keras_cv.layers.Unpatching")
+class Unpatching(keras.layers.Layer):
+    """
+    Layer to unpatchify image data.
+
+    This layer expects patches sorted by column and reorganizes the patches
+    such that they will each be positioned as a 2D shape with some
+    number of channels.
+
+    Any necessary padding or truncation will be applied to reach the target
+    shape.
+
+    Args:
+        target_shape: The target image shape after unpatching,
+         of form [height, width]
+    """
+
+    def __init__(self, target_shape):
+        self.target_shape = target_shape
+
+    def call(self, patches):
+        """
+        Reconstructs an unpatched image from the sequence of column sequence
+        patches.
+
+        If there are insufficient patches to construct the image of requested
+        dimensions, additional zero-patches will
+        be appended. If excessive patches are provided, unnecessary patches
+        will be truncated from the end.
+
+        Args:
+            patches: Patches of images in column sequence (i.e. each patch
+            is vertically oriented relative to the
+                previous patch). Expected shape of [batch_size, patch_num,
+                 patch_height, patch_width, channels].
+
+        Returns:
+            Unpatched image: Image reconstructed from the patches,
+                of expected shape [batch_size, height, width, channels].
+        """
+        (
+            batch_size,
+            patch_count,
+            patch_height,
+            patch_width,
+            channels,
+        ) = ops.shape(patches)
+
+        target_height, target_width = self.target_shape
+
+        patches_per_column = target_height // patch_height
+        patches_per_row = target_width // patch_width
+        required_patches = patches_per_column * patches_per_row
+
+        if required_patches > patch_count:
+            corrected_patches = ops.concatenate(
+                [
+                    patches,
+                    ops.zeros(
+                        (
+                            required_patches - patch_count,
+                            patch_height,
+                            patch_height,
+                            channels,
+                        )
+                    ),
+                ],
+                axis=1,
+            )
+        else:
+            corrected_patches = patches[:, :required_patches]
+
+        return ops.split(corrected_patches, patches_per_column, axis=1)

keras_cv/src/metrics/object_detection/box_coco_metrics.py

@@ -212,9 +212,9 @@ def result_fn(self, force=False):
             )
             result = {}
             for i, key in enumerate(METRIC_NAMES):
-                result[self.name_prefix() + METRIC_MAPPING[key]] = (
-                    py_func_result[i]
-                )
+                result[
+                    self.name_prefix() + METRIC_MAPPING[key]
+                ] = py_func_result[i]
             return result
 
         obj.result = types.MethodType(result_fn, obj)