增加keras-cv模型及训练代码

Keras/keras-cv/keras_cv/core/factor_sampler/__init__.py

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

Keras/keras-cv/keras_cv/core/factor_sampler/constant_factor_sampler.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.core.factor_sampler.factor_sampler import FactorSampler
+
+
+@tf.keras.utils.register_keras_serializable(package="keras_cv")
+class ConstantFactorSampler(FactorSampler):
+    """ConstantFactorSampler samples the same factor for every call to `__call__()`.
+
+    This is useful in cases where a user wants to always ensure that an augmentation
+    layer performs augmentations of the same strength.
+
+    Args:
+        value: the value to return from `__call__()`.
+
+    Usage:
+    ```python
+    constant_factor = keras_cv.ConstantFactorSampler(0.5)
+    random_sharpness = keras_cv.layers.RandomSharpness(factor=constant_factor)
+    # random_sharpness will now always use a factor of 0.5
+    ```
+    """
+
+    def __init__(self, value):
+        self.value = value
+
+    def __call__(self, shape=(), dtype="float32"):
+        return tf.ones(shape=shape, dtype=dtype) * self.value
+
+    def get_config(self):
+        return {"value": self.value}

Keras/keras-cv/keras_cv/core/factor_sampler/constant_factor_sampler_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
+
+
+class ConstantFactorSamplerTest(tf.test.TestCase):
+    def test_sample(self):
+        factor = keras_cv.ConstantFactorSampler(0.3)
+        self.assertEqual(factor(), 0.3)
+
+    def test_config(self):
+        factor = keras_cv.ConstantFactorSampler(0.3)
+        config = factor.get_config()
+        self.assertEqual(config["value"], 0.3)

Keras/keras-cv/keras_cv/core/factor_sampler/factor_sampler.py

+# Copyright 2022 The KerasCV Authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     https://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import tensorflow as tf
+
+
+@tf.keras.utils.register_keras_serializable(package="keras_cv")
+class FactorSampler:
+    """FactorSampler represents a strength factor for use in an augmentation layer.
+
+    FactorSampler should be subclassed and implement a `__call__()` method that returns
+    a tf.float32, or a float.  This method will be used by preprocessing layers to
+    determine the strength of their augmentation.  The specific range of values
+    supported may vary by layer, but for most layers is the range [0, 1].
+    """
+
+    def __call__(self, shape=None, dtype="float32"):
+        raise NotImplementedError(
+            "FactorSampler subclasses must implement a `__call__()` method."
+        )
+
+    def get_config(self):
+        return {}

Keras/keras-cv/keras_cv/core/factor_sampler/normal_factor_sampler.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.core.factor_sampler.factor_sampler import FactorSampler
+
+
+@tf.keras.utils.register_keras_serializable(package="keras_cv")
+class NormalFactorSampler(FactorSampler):
+    """NormalFactorSampler samples factors from a normal distribution.
+
+    This is useful in cases where a user wants to always ensure that an augmentation
+    layer performs augmentations of the same strength.
+
+    Args:
+        mean: mean value for the distribution.
+        stddev: standard deviation of the distribution.
+        min_value: values below min_value are clipped to min_value.
+        max_value: values above max_value are clipped to max_value.
+
+    Usage:
+    ```python
+    factor = keras_cv.core.NormalFactor(
+        mean=0.5,
+        stddev=0.1,
+        lower=0,
+        upper=1
+    )
+    random_sharpness = keras_cv.layers.RandomSharpness(factor=factor)
+    # random_sharpness will now sample normally around 0.5, with a lower of 0 and upper
+    # bound of 1.
+    ```
+    """
+
+    def __init__(self, mean, stddev, min_value, max_value, seed=None):
+        self.mean = mean
+        self.stddev = stddev
+        self.min_value = min_value
+        self.max_value = max_value
+        self.seed = seed
+
+    def __call__(self, shape=(), dtype="float32"):
+        return tf.clip_by_value(
+            tf.random.normal(
+                shape=shape,
+                mean=self.mean,
+                stddev=self.stddev,
+                seed=self.seed,
+                dtype=dtype,
+            ),
+            self.min_value,
+            self.max_value,
+        )
+
+    def get_config(self):
+        return {
+            "mean": self.mean,
+            "stddev": self.stddev,
+            "min_value": self.min_value,
+            "max_value": self.max_value,
+            "seed": self.seed,
+        }

Keras/keras-cv/keras_cv/core/factor_sampler/normal_factor_sampler_test_.py

+# Copyright 2022 The KerasCV Authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     https://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import tensorflow as tf
+
+from keras_cv import core
+
+
+class NormalFactorTest(tf.test.TestCase):
+    def test_sample(self):
+        factor = core.NormalFactor(mean=0.5, stddev=0.2, min_value=0, max_value=1)
+        self.assertTrue(0 <= factor() <= 1)
+
+    def test_config(self):
+        factor = core.NormalFactor(mean=0.5, stddev=0.2, min_value=0, max_value=1)
+        config = factor.get_config()
+        self.assertEqual(config["mean"], 0.5)
+        self.assertEqual(config["stddev"], 0.2)

Keras/keras-cv/keras_cv/core/factor_sampler/uniform_factor_sampler.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.core.factor_sampler.factor_sampler import FactorSampler
+
+
+@tf.keras.utils.register_keras_serializable(package="keras_cv")
+class UniformFactorSampler(FactorSampler):
+    """UniformFactorSampler samples factors uniformly from a range.
+
+    This is useful in cases where a user wants to always ensure that an augmentation
+    layer performs augmentations of the same strength.
+
+    Args:
+        lower: the lower bound of values returned from `__call__()`.
+        upper: the upper bound of values returned from `__call__()`.
+        seed: A shape int or Tensor, the seed to the random number generator. Must have
+            dtype int32 or int64. (When using XLA, only int32 is allowed.)
+    Usage:
+    ```python
+    uniform_factor = keras_cv.UniformFactorSampler(0, 0.5)
+    random_sharpness = keras_cv.layers.RandomSharpness(factor=uniform_factor)
+    # random_sharpness will now sample factors between 0, and 0.5
+    ```
+    """
+
+    def __init__(self, lower, upper, seed=None):
+        self.lower = lower
+        self.upper = upper
+        self.seed = seed
+
+    def __call__(self, shape=(), dtype="float32"):
+        return tf.random.uniform(
+            shape, seed=self.seed, minval=self.lower, maxval=self.upper, dtype=dtype
+        )
+
+    def get_config(self):
+        return {
+            "lower": self.lower,
+            "upper": self.upper,
+            "seed": self.seed,
+        }

Keras/keras-cv/keras_cv/core/factor_sampler/uniform_factor_sampler_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
+
+
+class UniformFactorSamplerTest(tf.test.TestCase):
+    def test_sample(self):
+        factor = keras_cv.UniformFactorSampler(0.3, 0.6)
+        self.assertTrue(0.3 <= factor() <= 0.6)
+
+    def test_config(self):
+        factor = keras_cv.UniformFactorSampler(0.3, 0.6)
+        config = factor.get_config()
+        self.assertEqual(config["lower"], 0.3)
+        self.assertEqual(config["upper"], 0.6)

Keras/keras-cv/keras_cv/custom_ops/BUILD

+licenses(["notice"])  # Apache 2.0
+
+package(default_visibility = ["//visibility:public"])
+
+config_setting(
+    name = "windows",
+    constraint_values = ["@bazel_tools//platforms:windows"],
+)
+
+cc_library(
+    name = "box_util",
+    srcs = ["box_util.cc"],
+    hdrs = ["box_util.h"],
+    deps = [
+        "@local_config_tf//:libtensorflow_framework",
+        "@local_config_tf//:tf_header_lib",
+    ],
+    copts = select({
+        ":windows": ["/DEIGEN_STRONG_INLINE=inline", "-DTENSORFLOW_MONOLITHIC_BUILD", "/DPLATFORM_WINDOWS", "/DEIGEN_HAS_C99_MATH", "/DTENSORFLOW_USE_EIGEN_THREADPOOL", "/DEIGEN_AVOID_STL_ARRAY", "/Iexternal/gemmlowp", "/wd4018", "/wd4577", "/DNOGDI", "/UTF_COMPILE_LIBRARY"],
+        "//conditions:default": ["-pthread", "-std=c++17"],
+    }),
+)
+
+cc_binary(
+    name = '_keras_cv_custom_ops.so',
+    srcs = [
+        "kernels/pairwise_iou_kernel.cc",
+        "ops/pairwise_iou_op.cc"
+    ],
+    linkshared = 1,
+    deps = [
+        "@local_config_tf//:libtensorflow_framework",
+        "@local_config_tf//:tf_header_lib",
+        ":box_util",
+    ],
+    features = select({
+        ":windows": ["windows_export_all_symbols"],
+        "//conditions:default": [],
+    }),
+    copts = select({
+        ":windows": ["/DEIGEN_STRONG_INLINE=inline", "-DTENSORFLOW_MONOLITHIC_BUILD", "/DPLATFORM_WINDOWS", "/DEIGEN_HAS_C99_MATH", "/DTENSORFLOW_USE_EIGEN_THREADPOOL", "/DEIGEN_AVOID_STL_ARRAY", "/Iexternal/gemmlowp", "/wd4018", "/wd4577", "/DNOGDI", "/UTF_COMPILE_LIBRARY"],
+        "//conditions:default": ["-pthread", "-std=c++17"],
+    }),
+)

Keras/keras-cv/keras_cv/custom_ops/__init__.py

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

Keras/keras-cv/keras_cv/custom_ops/box_util.cc

+/* Copyright 2022 The KerasCV Authors. All Rights Reserved.
+
+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
+
+    http://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.
+==============================================================================*/
+
+#include "keras_cv/custom_ops/box_util.h"
+
+#include <algorithm>
+#include <cmath>
+
+namespace tensorflow {
+namespace kerascv {
+namespace box {
+
+const double kEPS = 1e-8;
+
+// Min,max box dimensions (length, width, height). Boxes with dimensions that
+// exceed these values will have box intersections of 0.
+constexpr double kMinBoxDim = 1e-3;
+constexpr double kMaxBoxDim = 1e6;
+
+// A line with the representation a*x + b*y + c = 0.
+struct Line {
+  double a = 0;
+  double b = 0;
+  double c = 0;
+
+  Line(const Vertex& v1, const Vertex& v2)
+      : a(v2.y - v1.y), b(v1.x - v2.x), c(v2.x * v1.y - v2.y * v1.x) {}
+
+  // Computes the line value for a vertex v as a * v.x + b * v.y + c
+  double LineValue(const Vertex& v) const { return a * v.x + b * v.y + c; }
+
+  // Computes the intersection point with the other line.
+  Vertex IntersectionPoint(const Line& other) const {
+    const double w = a * other.b - b * other.a;
+    CHECK_GT(std::fabs(w), kEPS) << "No intersection between the two lines.";
+    return Vertex((b * other.c - c * other.b) / w,
+                  (c * other.a - a * other.c) / w);
+  }
+};
+
+// Computes the coordinates of its four vertices given a 2D rotated box,
+std::vector<Vertex> ComputeBoxVertices(const double cx, const double cy,
+                                       const double w, const double h,
+                                       const double heading) {
+  const double dxcos = (w / 2.) * std::cos(heading);
+  const double dxsin = (w / 2.) * std::sin(heading);
+  const double dycos = (h / 2.) * std::cos(heading);
+  const double dysin = (h / 2.) * std::sin(heading);
+  return {Vertex(cx - dxcos + dysin, cy - dxsin - dycos),
+          Vertex(cx + dxcos + dysin, cy + dxsin - dycos),
+          Vertex(cx + dxcos - dysin, cy + dxsin + dycos),
+          Vertex(cx - dxcos - dysin, cy - dxsin + dycos)};
+}
+
+// Computes the intersection points between two rotated boxes, by following:
+//
+// 1. Initiazlizes the current intersection points with the vertices of one box,
+//    and the other box is taken as the cutting box;
+//
+// 2. For each cutting line in the cutting box (four cutting lines in total):
+//      For each point in the current intersection points:
+//        If the point is inside of the cutting line:
+//          Adds it to the new intersection points;
+//        if current point and its next point are in the opposite side of the
+//          cutting line:
+//          Computes the line of current points and its next point as tmp_line;
+//          Computes the intersection point between the cutting line and
+//            tmp_line;
+//          Adds the intersection point to the new intersection points;
+//      After checking each cutting line, sets current intersection points as
+//      new intersection points;
+//
+// 3. Returns the final intersection points.
+std::vector<Vertex> ComputeIntersectionPoints(
+    const std::vector<Vertex>& rbox_1, const std::vector<Vertex>& rbox_2) {
+  std::vector<Vertex> intersection = rbox_1;
+  const int vertices_len = rbox_2.size();
+  for (int i = 0; i < rbox_2.size(); ++i) {
+    const int len = intersection.size();
+    if (len <= 2) {
+      break;
+    }
+    const Vertex& p = rbox_2[i];
+    const Vertex& q = rbox_2[(i + 1) % vertices_len];
+    Line cutting_line(p, q);
+    // Computes line value.
+    std::vector<double> line_values;
+    line_values.reserve(len);
+    for (int j = 0; j < len; ++j) {
+      line_values.push_back(cutting_line.LineValue(intersection[j]));
+    }
+    // Updates current intersection points.
+    std::vector<Vertex> new_intersection;
+    for (int j = 0; j < len; ++j) {
+      const double s_val = line_values[j];
+      const Vertex& s = intersection[j];
+      // Adds the current vertex.
+      if (s_val <= 0 || std::fabs(s_val) <= kEPS) {
+        new_intersection.push_back(s);
+      }
+      const double t_val = line_values[(j + 1) % len];
+      // Skips the checking of intersection point if the next vertex is on the
+      // line.
+      if (std::fabs(t_val) <= kEPS) {
+        continue;
+      }
+      // Adds the intersection point.
+      if ((s_val > 0 && t_val < 0) || (s_val < 0 && t_val > 0)) {
+        Line s_t_line(s, intersection[(j + 1) % len]);
+        new_intersection.push_back(cutting_line.IntersectionPoint(s_t_line));
+      }
+    }
+    intersection = new_intersection;
+  }
+  return intersection;
+}
+
+// Computes the area of a convex polygon,
+double ComputePolygonArea(const std::vector<Vertex>& convex_polygon) {
+  const int len = convex_polygon.size();
+  if (len <= 2) {
+    return 0;
+  }
+  double area = 0;
+  for (int i = 0; i < len; ++i) {
+    const Vertex& p = convex_polygon[i];
+    const Vertex& q = convex_polygon[(i + 1) % len];
+    area += p.x * q.y - p.y * q.x;
+  }
+  return std::fabs(0.5 * area);
+}
+
+RotatedBox2D::RotatedBox2D(const double cx, const double cy, const double w,
+                           const double h, const double heading)
+    : cx_(cx), cy_(cy), w_(w), h_(h), heading_(heading) {
+  // Compute loose bounds on dimensions of box that doesn't require computing
+  // full intersection.  We can do this by trying to compute the largest circle
+  // swept by rotating the box around its center.  The radius of that circle
+  // is the length of the ray from the center to the box corner.  The upper
+  // bound for this value is the length of the longer dimension divided by two
+  // and then multiplied by root(2) (worst-case being a square box); we choose
+  // 1.5 as slightly higher than root(2), and then use these extrema to do
+  // simple extrema box checks without having to compute the true cos/sin value.
+  double max_dim = std::max(w_, h_) / 2. * 1.5;
+  loose_min_x_ = cx_ - max_dim;
+  loose_max_x_ = cx_ + max_dim;
+  loose_min_y_ = cy_ - max_dim;
+  loose_max_y_ = cy_ + max_dim;
+
+  extreme_box_dim_ = (w_ <= kMinBoxDim || h_ <= kMinBoxDim);
+  extreme_box_dim_ |= (w_ >= kMaxBoxDim || h_ >= kMaxBoxDim);
+}
+
+double RotatedBox2D::Area() const {
+  if (area_ < 0) {
+    const double area = ComputePolygonArea(box_vertices());
+    area_ = std::fabs(area) <= kEPS ? 0 : area;
+  }
+  return area_;
+}
+
+const std::vector<Vertex>& RotatedBox2D::box_vertices() const {
+  if (box_vertices_.empty()) {
+    box_vertices_ = ComputeBoxVertices(cx_, cy_, w_, h_, heading_);
+  }
+
+  return box_vertices_;
+}
+
+bool RotatedBox2D::NonZeroAndValid() const { return !extreme_box_dim_; }
+
+bool RotatedBox2D::MaybeIntersects(const RotatedBox2D& other) const {
+  // If the box dimensions of either box are too small / large,
+  // assume they are not well-formed boxes (otherwise we are
+  // subject to issues due to catastrophic cancellation).
+  if (extreme_box_dim_ || other.extreme_box_dim_) {
+    return false;
+  }
+
+  // Check whether the loose extrema overlap -- if not, then there is
+  // no chance that the two boxes overlap even when computing the true,
+  // more expensive overlap.
+  if ((loose_min_x_ > other.loose_max_x_) ||
+      (loose_max_x_ < other.loose_min_x_) ||
+      (loose_min_y_ > other.loose_max_y_) ||
+      (loose_max_y_ < other.loose_min_y_)) {
+    return false;
+  }
+
+  return true;
+}
+
+double RotatedBox2D::Intersection(const RotatedBox2D& other) const {
+  // Do a fast intersection check - if the boxes are not near each other
+  // then we can return early.  If they are close enough to maybe overlap,
+  // we do the full check.
+  if (!MaybeIntersects(other)) {
+    return 0.0;
+  }
+
+  // Computes the intersection polygon.
+  const std::vector<Vertex> intersection_polygon =
+      ComputeIntersectionPoints(box_vertices(), other.box_vertices());
+  // Computes the intersection area.
+  const double intersection_area = ComputePolygonArea(intersection_polygon);
+
+  return std::fabs(intersection_area) <= kEPS ? 0 : intersection_area;
+}
+
+double RotatedBox2D::IoU(const RotatedBox2D& other) const {
+  // Computes the intersection area.
+  const double intersection_area = Intersection(other);
+  if (intersection_area == 0) {
+    return 0;
+  }
+  // Computes the union area.
+  const double union_area = Area() + other.Area() - intersection_area;
+  if (std::fabs(union_area) <= kEPS) {
+    return 0;
+  }
+  return intersection_area / union_area;
+}
+
+std::vector<Upright3DBox> ParseBoxesFromTensor(const Tensor& boxes_tensor) {
+  int num_boxes = boxes_tensor.dim_size(0);
+
+  const auto t_boxes_tensor = boxes_tensor.matrix<float>();
+
+  std::vector<Upright3DBox> bboxes3d;
+  bboxes3d.reserve(num_boxes);
+  for (int i = 0; i < num_boxes; ++i) {
+    const double center_x = t_boxes_tensor(i, 0);
+    const double center_y = t_boxes_tensor(i, 1);
+    const double center_z = t_boxes_tensor(i, 2);
+    const double dimension_x = t_boxes_tensor(i, 3);
+    const double dimension_y = t_boxes_tensor(i, 4);
+    const double dimension_z = t_boxes_tensor(i, 5);
+    const double heading = t_boxes_tensor(i, 6);
+    const double z_min = center_z - dimension_z / 2;
+    const double z_max = center_z + dimension_z / 2;
+    RotatedBox2D box2d(center_x, center_y, dimension_x, dimension_y, heading);
+    if (dimension_x <= 0 || dimension_y <= 0) {
+      bboxes3d.emplace_back(RotatedBox2D(), z_min, z_max);
+    } else {
+      bboxes3d.emplace_back(box2d, z_min, z_max);
+    }
+  }
+  return bboxes3d;
+}
+
+bool Upright3DBox::NonZeroAndValid() const {
+  // If min is larger than max, the upright box is invalid.
+  //
+  // If the min and max are equal, the height of the box is 0. and thus the box
+  // is zero.
+  if (z_min - z_max >= 0.) {
+    return false;
+  }
+
+  return rbox.NonZeroAndValid();
+}
+
+double Upright3DBox::IoU(const Upright3DBox& other) const {
+  // Check that both boxes are non-zero and valid.  Otherwise,
+  // return 0.
+  if (!NonZeroAndValid() || !other.NonZeroAndValid()) {
+    return 0;
+  }
+
+  // Quickly check whether z's overlap; if they don't, we can return 0.
+  const double z_inter =
+      std::max(.0, std::min(z_max, other.z_max) - std::max(z_min, other.z_min));
+  if (z_inter == 0) {
+    return 0;
+  }
+
+  const double base_inter = rbox.Intersection(other.rbox);
+  if (base_inter == 0) {
+    return 0;
+  }
+
+  const double volume_1 = rbox.Area() * (z_max - z_min);
+  const double volume_2 = other.rbox.Area() * (other.z_max - other.z_min);
+  const double volume_inter = base_inter * z_inter;
+  const double volume_union = volume_1 + volume_2 - volume_inter;
+  return volume_inter > 0 ? volume_inter / volume_union : 0;
+}
+
+double Upright3DBox::Overlap(const Upright3DBox& other) const {
+  // Check that both boxes are non-zero and valid.  Otherwise,
+  // return 0.
+  if (!NonZeroAndValid() || !other.NonZeroAndValid()) {
+    return 0;
+  }
+
+  const double z_inter =
+      std::max(.0, std::min(z_max, other.z_max) - std::max(z_min, other.z_min));
+  if (z_inter == 0) {
+    return 0;
+  }
+
+  const double base_inter = rbox.Intersection(other.rbox);
+  if (base_inter == 0) {
+    return 0;
+  }
+
+  const double volume_1 = rbox.Area() * (z_max - z_min);
+  const double volume_inter = base_inter * z_inter;
+  // Normalizes intersection of volume by the volume of this box.
+  return volume_inter > 0 ? volume_inter / volume_1 : 0;
+}
+
+}  // namespace box
+}  // namespace kerascv
+}  // namespace tensorflow

Keras/keras-cv/keras_cv/custom_ops/box_util.h

+/* Copyright 2022 The Keras CV Authors. All Rights Reserved.
+
+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
+
+    http://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.
+==============================================================================*/
+
+#ifndef THIRD_PARTY_PY_KERAS_CV_OPS_BOX_UTIL_H_
+#define THIRD_PARTY_PY_KERAS_CV_OPS_BOX_UTIL_H_
+
+#include <string>
+#include <vector>
+
+#include "tensorflow/core/framework/tensor.h"
+
+namespace tensorflow {
+namespace kerascv {
+namespace box {
+
+// A vertex with (x, y) coordinate.
+//
+// This is an internal implementation detail of RotatedBox2D.
+struct Vertex {
+  // Creates an empty Vertex.
+  Vertex() = default;
+
+  Vertex(const double x, const double y) : x(x), y(y) {}
+
+  double x = 0;
+  double y = 0;
+};
+
+// A rotated 2D bounding box represented as (cx, cy, w, h, r). cx, cy are the
+// box center coordinates; w, h are the box width and height; heading is the
+// rotation angle in radian relative to the 'positive x' direction.
+class RotatedBox2D {
+ public:
+  // Creates an empty rotated 2D box.
+  RotatedBox2D() : RotatedBox2D(0, 0, 0, 0, 0) {}
+
+  RotatedBox2D(const double cx, const double cy, const double w, const double h,
+               const double heading);
+
+  // Returns the area of the box.
+  double Area() const;
+
+  // Returns the intersection area between this box and the given box.
+  double Intersection(const RotatedBox2D& other) const;
+
+  // Returns the IoU between this box and the given box.
+  double IoU(const RotatedBox2D& other) const;
+
+  // Returns true if the box is valid (width and height are not extremely
+  // large or small).
+  bool NonZeroAndValid() const;
+
+ private:
+  // Computes / caches box_vertices_ calculation.
+  const std::vector<Vertex>& box_vertices() const;
+
+  // Returns true if this box and 'other' might intersect.
+  //
+  // If this returns false, the two boxes definitely do not intersect.  If this
+  // returns true, it is still possible that the two boxes do not intersect, and
+  // the more expensive intersection code will be called.
+  bool MaybeIntersects(const RotatedBox2D& other) const;
+
+  double cx_ = 0;
+  double cy_ = 0;
+  double w_ = 0;
+  double h_ = 0;
+  double heading_ = 0;
+
+  // Loose boundaries for fast intersection test.
+  double loose_min_x_ = -1;
+  double loose_max_x_ = -1;
+  double loose_min_y_ = -1;
+  double loose_max_y_ = -1;
+
+  // True if the dimensions of the box are very small or very large in any
+  // dimension.
+  bool extreme_box_dim_ = false;
+
+  // The following fields are computed on demand.  They are logically
+  // const.
+
+  // Cached area.  Access via Area() public API.
+  mutable double area_ = -1;
+
+  // Stores the vertices of the box.  Access via box_vertices().
+  mutable std::vector<Vertex> box_vertices_;
+};
+
+// A 3D box of 7-DOFs: only allows rotation around the z-axis.
+struct Upright3DBox {
+  RotatedBox2D rbox = RotatedBox2D();
+  double z_min = 0;
+  double z_max = 0;
+
+  // Creates an empty rotated 3D box.
+  Upright3DBox() = default;
+
+  // Creates a 3D box from the raw input data with size 7. The data format is
+  // (center_x, center_y, center_z, dimension_x, dimension_y, dimension_z,
+  // heading)
+  Upright3DBox(const std::vector<double>& raw)
+      : rbox(raw[0], raw[1], raw[3], raw[4], raw[6]),
+        z_min(raw[2] - raw[5] / 2.0),
+        z_max(raw[2] + raw[5] / 2.0) {}
+
+  Upright3DBox(const RotatedBox2D& rb, const double z_min, const double z_max)
+      : rbox(rb), z_min(z_min), z_max(z_max) {}
+
+  // Computes intersection over union (of the volume).
+  double IoU(const Upright3DBox& other) const;
+
+  // Computes overlap: intersection of this box and the given box normalized
+  // over the volume of this box.
+  double Overlap(const Upright3DBox& other) const;
+
+  // Returns true if the box is valid (width and height are not extremely
+  // large or small, and zmin < zmax).
+  bool NonZeroAndValid() const;
+};
+
+// Converts a [N, 7] tensor to a vector of N Upright3DBox objects.
+std::vector<Upright3DBox> ParseBoxesFromTensor(const Tensor& boxes_tensor);
+
+}  // namespace box
+}  // namespace kerascv
+}  // namespace tensorflow
+
+#endif  // THIRD_PARTY_PY_KERAS_CV_OPS_BOX_UTIL_H_

Keras/keras-cv/keras_cv/custom_ops/kernels/pairwise_iou_kernel.cc

+/* Copyright 2022 The KerasCV Authors. All Rights Reserved.
+
+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
+
+    http://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.
+==============================================================================*/
+
+#include <vector>
+
+#include "keras_cv/custom_ops/box_util.h"
+#include "tensorflow/core/framework/op_kernel.h"
+#include "tensorflow/core/framework/tensor.h"
+#include "tensorflow/core/framework/tensor_shape.h"
+#include "tensorflow/core/lib/core/errors.h"
+
+namespace tensorflow {
+namespace kerascv {
+namespace {
+
+class PairwiseIoUOp : public OpKernel {
+ public:
+  explicit PairwiseIoUOp(OpKernelConstruction* ctx) : OpKernel(ctx) {}
+
+  void Compute(OpKernelContext* ctx) override {
+    const Tensor& a = ctx->input(0);
+    const Tensor& b = ctx->input(1);
+    OP_REQUIRES(ctx, TensorShapeUtils::IsMatrix(a.shape()),
+                errors::InvalidArgument("In[0] must be a matrix, but get ",
+                                        a.shape().DebugString()));
+    OP_REQUIRES(ctx, TensorShapeUtils::IsMatrix(b.shape()),
+                errors::InvalidArgument("In[0] must be a matrix, but get ",
+                                        b.shape().DebugString()));
+    OP_REQUIRES(ctx, 7 == a.dim_size(1),
+                errors::InvalidArgument("Matrix size-incompatible: In[0]: ",
+                                        a.shape().DebugString()));
+    OP_REQUIRES(ctx, 7 == b.dim_size(1),
+                errors::InvalidArgument("Matrix size-incompatible: In[1]: ",
+                                        b.shape().DebugString()));
+
+    const int n_a = a.dim_size(0);
+    const int n_b = b.dim_size(0);
+
+    Tensor* iou_a_b = nullptr;
+    OP_REQUIRES_OK(
+        ctx, ctx->allocate_output("iou", TensorShape({n_a, n_b}), &iou_a_b));
+
+    auto t_iou_a_b = iou_a_b->matrix<float>();
+
+    std::vector<box::Upright3DBox> box_a = box::ParseBoxesFromTensor(a);
+    std::vector<box::Upright3DBox> box_b = box::ParseBoxesFromTensor(b);
+    for (int i_a = 0; i_a < n_a; ++i_a) {
+      for (int i_b = 0; i_b < n_b; ++i_b) {
+        t_iou_a_b(i_a, i_b) = box_a[i_a].IoU(box_b[i_b]);
+      }
+    }
+  }
+};
+
+REGISTER_KERNEL_BUILDER(Name("PairwiseIou3D").Device(DEVICE_CPU),
+                        PairwiseIoUOp);
+
+}  // namespace
+}  // namespace kerascv
+}  // namespace tensorflow

Keras/keras-cv/keras_cv/custom_ops/ops/pairwise_iou_op.cc

+/* Copyright 2022 The KerasCV Authors. All Rights Reserved.
+
+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
+
+    http://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.
+==============================================================================*/
+
+#include "tensorflow/core/framework/op.h"
+#include "tensorflow/core/framework/shape_inference.h"
+
+using namespace tensorflow;
+
+REGISTER_OP("PairwiseIou3D")
+    .Input("boxes_a: float")
+    .Input("boxes_b: float")
+    .Output("iou: float")
+    .SetShapeFn([](tensorflow::shape_inference::InferenceContext* c) {
+      c->set_output(
+          0, c->MakeShape({c->Dim(c->input(0), 0), c->Dim(c->input(1), 0)}));
+      return tensorflow::Status();
+    })
+    .Doc(R"doc(
+Calculate pairwise IoUs between two set of 3D bboxes. Every bbox is represented
+as [center_x, center_y, center_z, dim_x, dim_y, dim_z, heading].
+boxes_a: A tensor of shape [num_boxes_a, 7]
+boxes_b: A tensor of shape [num_boxes_b, 7]
+)doc");

Keras/keras-cv/keras_cv/datasets/__init__.py

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

Keras/keras-cv/keras_cv/datasets/imagenet/README.md

+### The ImageNet Dataset in keras_cv
+
+In order to load ImageNet with KerasCV, you'll need to download the [original ImageNet dataset](https://image-net.org) and parse the images into TFRecords.
+
+Tensorflow provides a [script](https://github.com/tensorflow/tpu/blob/master/tools/datasets/imagenet_to_gcs.py) to perform this parsing and upload images to Google Cloud Storage (or optionally to local storage).
+
+Please reference that script's instructions on producing ImageNet TFRecords, and then use the KerasCV loader to load records from wherever you choose to store them.

Keras/keras-cv/keras_cv/datasets/imagenet/__init__.py

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

Keras/keras-cv/keras_cv/datasets/imagenet/load.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
+
+
+def parse_imagenet_example(img_size, crop_to_aspect_ratio):
+    """Function to parse a TFRecord example into an image and label"""
+
+    resizing = None
+    if img_size:
+        resizing = layers.Resizing(
+            width=img_size[0],
+            height=img_size[1],
+            crop_to_aspect_ratio=crop_to_aspect_ratio,
+        )
+
+    def apply(example):
+        # Read example
+        image_key = "image/encoded"
+        label_key = "image/class/label"
+        keys_to_features = {
+            image_key: tf.io.FixedLenFeature((), tf.string, ""),
+            label_key: tf.io.FixedLenFeature([], tf.int64, -1),
+        }
+        parsed = tf.io.parse_single_example(example, keys_to_features)
+
+        # Decode and resize image
+        image_bytes = tf.reshape(parsed[image_key], shape=[])
+        image = tf.io.decode_jpeg(image_bytes, channels=3)
+        if resizing:
+            image = resizing(image)
+
+        # Decode label
+        label = tf.cast(tf.reshape(parsed[label_key], shape=()), dtype=tf.int32) - 1
+        label = tf.one_hot(label, 1000)
+
+        return image, label
+
+    return apply
+
+
+def load(
+    split,
+    tfrecord_path,
+    batch_size=None,
+    shuffle=True,
+    shuffle_buffer=None,
+    reshuffle_each_iteration=False,
+    img_size=None,
+    crop_to_aspect_ratio=True,
+):
+    """Loads the ImageNet dataset from TFRecords
+
+    Usage:
+    ```python
+    dataset, ds_info = keras_cv.datasets.imagenet.load(
+        split="train", tfrecord_path="gs://my-bucket/imagenet-tfrecords"
+    )
+    ```
+
+    Args:
+        split: the split to load.  Should be one of "train" or "validation."
+        tfrecord_path: the path to your preprocessed ImageNet TFRecords.
+            See keras_cv/datasets/imagenet/README.md for preprocessing instructions.
+        batch_size: how many instances to include in batches after loading.
+            Should only be specified if img_size is specified (so that images
+            can be resized to the same size before batching).
+        shuffle: whether or not to shuffle the dataset.  Defaults to True.
+        shuffle_buffer: the size of the buffer to use in shuffling.
+        reshuffle_each_iteration: whether to reshuffle the dataset on every epoch.
+            Defaults to False.
+        img_size: the size to resize the images to. Defaults to None, indicating
+            that images should not be resized.
+
+    Returns:
+        tf.data.Dataset containing ImageNet.  Each entry is a dictionary containing
+        keys {"image": image, "label": label} where images is a Tensor of shape
+        [H, W, 3] and label is a Tensor of shape [1000].
+    """
+
+    if batch_size is not None and img_size is None:
+        raise ValueError("Batching can only be performed if images are resized.")
+
+    num_splits = 1024 if split == "train" else 128
+    filenames = [
+        f"{tfrecord_path}/{split}-{i:05d}-of-{num_splits:05d}"
+        for i in range(0, num_splits)
+    ]
+
+    dataset = tf.data.TFRecordDataset(
+        filenames=filenames, num_parallel_reads=tf.data.AUTOTUNE
+    )
+
+    dataset = dataset.map(
+        parse_imagenet_example(img_size, crop_to_aspect_ratio),
+        num_parallel_calls=tf.data.AUTOTUNE,
+    )
+
+    if shuffle:
+        if not batch_size and not shuffle_buffer:
+            raise ValueError(
+                "If `shuffle=True`, either a `batch_size` or `shuffle_buffer` must be "
+                "provided to `keras_cv.datasets.imagenet.load().`"
+            )
+        shuffle_buffer = shuffle_buffer or 8 * batch_size
+        dataset = dataset.shuffle(
+            shuffle_buffer, reshuffle_each_iteration=reshuffle_each_iteration
+        )
+
+    if batch_size is not None:
+        dataset = dataset.batch(batch_size)
+
+    return dataset.prefetch(tf.data.AUTOTUNE)

Keras/keras-cv/keras_cv/datasets/pascal_voc/__init__.py

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

Keras/keras-cv/keras_cv/datasets/pascal_voc/load.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 tensorflow_datasets as tfds
+from tensorflow import keras
+
+from keras_cv import bounding_box
+
+
+def curry_map_function(bounding_box_format, img_size):
+    """Mapping function to create batched image and bbox coordinates"""
+
+    if img_size is not None:
+        resizing = keras.layers.Resizing(
+            height=img_size[0], width=img_size[1], crop_to_aspect_ratio=False
+        )
+
+    # TODO(lukewood): update `keras.layers.Resizing` to support bounding boxes.
+    def apply(inputs):
+        # Support image size none.
+        if img_size is not None:
+            inputs["image"] = resizing(inputs["image"])
+
+        inputs["objects"]["bbox"] = bounding_box.convert_format(
+            inputs["objects"]["bbox"],
+            images=inputs["image"],
+            source="rel_yxyx",
+            target=bounding_box_format,
+        )
+
+        bounding_boxes = inputs["objects"]["bbox"]
+        labels = tf.cast(inputs["objects"]["label"], tf.float32)
+        labels = tf.expand_dims(labels, axis=-1)
+        bounding_boxes = tf.concat([bounding_boxes, labels], axis=-1)
+        return {"images": inputs["image"], "bounding_boxes": bounding_boxes}
+
+    return apply
+
+
+def load(
+    split,
+    bounding_box_format,
+    batch_size=None,
+    shuffle_buffer=None,
+    shuffle_files=True,
+    img_size=None,
+):
+    """Loads the PascalVOC 2007 dataset.
+
+    Usage:
+    ```python
+    dataset, ds_info = keras_cv.datasets.pascal_voc.load(
+        split="train", bounding_box_format="xywh", batch_size=9
+    )
+    ```
+
+    Args:
+        split: the split string passed to the `tensorflow_datasets.load()` call.  Should
+            be one of "train", "test", or "validation."
+        bounding_box_format: the keras_cv bounding box format to load the boxes into.
+            For a list of supported formats, please  Refer
+            [to the keras.io docs](https://keras.io/api/keras_cv/bounding_box/formats/)
+            for more details on supported bounding box formats.
+        batch_size: (Optional) how many instances to include in batches after loading. If
+            not provided, no batching will occur.
+        shuffle_buffer: (Optional) the size of the buffer to use in shuffling.
+        shuffle_files: (Optional) whether or not to shuffle files, defaults to True.
+        img_size: (Optional) size to resize the images to.  By default, images are not
+            resized `tf.RaggedTensor` batches are produced if batching occurs.
+
+    Returns:
+        tf.data.Dataset containing PascalVOC.  Each entry is a dictionary containing
+        keys {"images": images, "bounding_boxes": bounding_boxes} where images is a
+        Tensor of shape [batch, H, W, 3] and bounding_boxes is a `tf.RaggedTensor` of
+        shape [batch, None, 5].
+    """
+    dataset, dataset_info = tfds.load(
+        "voc/2007", split=split, shuffle_files=shuffle_files, with_info=True
+    )
+    dataset = dataset.map(
+        curry_map_function(bounding_box_format=bounding_box_format, img_size=img_size),
+        num_parallel_calls=tf.data.AUTOTUNE,
+    )
+
+    if shuffle_buffer:
+        dataset = dataset.shuffle(shuffle_buffer, reshuffle_each_iteration=True)
+
+    if batch_size is not None:
+        dataset = dataset.apply(
+            tf.data.experimental.dense_to_ragged_batch(batch_size=batch_size)
+        )
+    return dataset, dataset_info