Adding Differential Binarization model from PaddleOCR to Keras3

keras_hub/api/layers/__init__.py

@@ -41,6 +41,9 @@
 from keras_hub.src.models.densenet.densenet_image_converter import (
     DenseNetImageConverter,
 )
+from keras_hub.src.models.differential_binarization.differential_binarization_image_converter import (
+    DifferentialBinarizationImageConverter,
+)
 from keras_hub.src.models.efficientnet.efficientnet_image_converter import (
     EfficientNetImageConverter,
 )

keras_hub/api/models/__init__.py

@@ -104,6 +104,15 @@
 from keras_hub.src.models.densenet.densenet_image_classifier_preprocessor import (
     DenseNetImageClassifierPreprocessor,
 )
+from keras_hub.src.models.differential_binarization.differential_binarization import (
+    DifferentialBinarization,
+)
+from keras_hub.src.models.differential_binarization.differential_binarization_backbone import (
+    DifferentialBinarizationBackbone,
+)
+from keras_hub.src.models.differential_binarization.differential_binarization_preprocessor import (
+    DifferentialBinarizationPreprocessor,
+)
 from keras_hub.src.models.distil_bert.distil_bert_backbone import (
     DistilBertBackbone,
 )

keras_hub/src/models/differential_binarization/__init__.py

@@ -0,0 +1,9 @@
+from keras_hub.src.models.differential_binarization.differential_binarization_backbone import (
+    DifferentialBinarizationBackbone,
+)
+from keras_hub.src.models.differential_binarization.differential_binarization_presets import (
+    backbone_presets,
+)
+from keras_hub.src.utils.preset_utils import register_presets
+
+register_presets(backbone_presets, DifferentialBinarizationBackbone)

keras_hub/src/models/differential_binarization/differential_binarization.py

@@ -0,0 +1,189 @@
+import math
+
+import keras
+from keras import layers
+
+from keras_hub.src.api_export import keras_hub_export
+from keras_hub.src.models.differential_binarization.differential_binarization_backbone import (
+    DifferentialBinarizationBackbone,
+)
+from keras_hub.src.models.differential_binarization.differential_binarization_preprocessor import (
+    DifferentialBinarizationPreprocessor,
+)
+from keras_hub.src.models.differential_binarization.losses import DBLoss
+from keras_hub.src.models.image_segmenter import ImageSegmenter
+
+
+@keras_hub_export("keras_hub.models.DifferentialBinarization")
+class DifferentialBinarization(ImageSegmenter):
+    """
+    A Keras model implementing the Differential Binarization
+    architecture for scene text detection, described in
+    [Real-time Scene Text Detection with Differentiable Binarization](
+    https://arxiv.org/abs/1911.08947).
+
+    Args:
+        backbone: A `keras_hub.models.DifferentialBinarizationBackbone`
+            instance.
+        head_kernel_list: list of ints. The number of filters for probability
+            and threshold maps. Defaults to [3, 2, 2].
+        step_function_k: float. `k` parameter used within the differential
+            binarization step function.
+        preprocessor: `None`, a `keras_hub.models.Preprocessor` instance,
+            a `keras.Layer` instance, or a callable. If `None` no preprocessing
+            will be applied to the inputs.
+
+    Examples:
+    ```python
+    input_data = np.ones(shape=(8, 224, 224, 3))
+
+    image_encoder = keras_hub.models.ResNetBackbone.from_preset(
+        "resnet_vd_50_imagenet"
+    )
+    backbone = keras_hub.models.DifferentialBinarizationBackbone(image_encoder)
+    detector = keras_hub.models.DifferentialBinarization(
+        backbone=backbone
+    )
+
+    detector(input_data)
+    ```
+    """
+
+    backbone_cls = DifferentialBinarizationBackbone
+    preprocessor_cls = DifferentialBinarizationPreprocessor
+
+    def __init__(
+        self,
+        backbone,
+        head_kernel_list=[3, 2, 2],
+        step_function_k=50.0,
+        preprocessor=None,
+        **kwargs,
+    ):
+
+        inputs = backbone.input
+        x = backbone(inputs)
+        probability_maps = diffbin_head(
+            x,
+            in_channels=backbone.fpn_channels,
+            kernel_list=head_kernel_list,
+            name="head_prob",
+        )
+        threshold_maps = diffbin_head(
+            x,
+            in_channels=backbone.fpn_channels,
+            kernel_list=head_kernel_list,
+            name="head_thresh",
+        )
+        binary_maps = step_function(
+            probability_maps, threshold_maps, k=step_function_k
+        )
+        outputs = layers.Concatenate(axis=-1)(
+            [probability_maps, threshold_maps, binary_maps]
+        )
+
+        super().__init__(inputs=inputs, outputs=outputs, **kwargs)
+
+        self.backbone = backbone
+        self.head_kernel_list = head_kernel_list
+        self.step_function_k = step_function_k
+        self.preprocessor = preprocessor
+
+    def compile(
+        self,
+        optimizer="auto",
+        loss="auto",
+        **kwargs,
+    ):
+        """Configures the `DifferentialBinarization` task for training.
+
+        `DifferentialBinarization` extends the default compilation signature of
+        `keras.Model.compile` with defaults for `optimizer` and `loss`. To
+        override these defaults, pass any value to these arguments during
+        compilation.
+
+        Args:
+            optimizer: `"auto"`, an optimizer name, or a `keras.Optimizer`
+                instance. Defaults to `"auto"`, which uses the default optimizer
+                for `DifferentialBinarization`. See `keras.Model.compile` and
+                `keras.optimizers` for more info on possible `optimizer` values.
+            loss: `"auto"`, a loss name, or a `keras.losses.Loss` instance.
+                Defaults to `"auto"`, in which case the default loss
+                computation of `DifferentialBinarization` will be applied. See
+                `keras.Model.compile` and `keras.losses` for more info on
+                possible `loss` values.
+            **kwargs: See `keras.Model.compile` for a full list of arguments
+                supported by the compile method.
+        """
+        if optimizer == "auto":
+            optimizer = keras.optimizers.SGD(
+                learning_rate=0.007, weight_decay=0.0001, momentum=0.9
+            )
+        if loss == "auto":
+            loss = DBLoss()
+        super().compile(
+            optimizer=optimizer,
+            loss=loss,
+            **kwargs,
+        )
+
+    def get_config(self):
+        # Backbone serialized in `super`
+        config = super().get_config()
+        config.update(
+            {
+                "head_kernel_list": self.head_kernel_list,
+                "step_function_k": self.step_function_k,
+            }
+        )
+        return config
+
+
+def step_function(x, y, k):
+    return 1.0 / (1.0 + keras.ops.exp(-k * (x - y)))
+
+
+def diffbin_head(inputs, in_channels, kernel_list, name):
+    x = layers.Conv2D(
+        in_channels // 4,
+        kernel_size=kernel_list[0],
+        padding="same",
+        use_bias=False,
+        name=f"{name}_conv0_weights",
+    )(inputs)
+    x = layers.BatchNormalization(
+        beta_initializer=keras.initializers.Constant(1e-4),
+        gamma_initializer=keras.initializers.Constant(1.0),
+        name=f"{name}_conv0_bn",
+    )(x)
+    x = layers.ReLU(name=f"{name}_conv0_relu")(x)
+    x = layers.Conv2DTranspose(
+        in_channels // 4,
+        kernel_size=kernel_list[1],
+        strides=2,
+        padding="valid",
+        bias_initializer=keras.initializers.RandomUniform(
+            minval=-1.0 / math.sqrt(in_channels // 4 * 1.0),
+            maxval=1.0 / math.sqrt(in_channels // 4 * 1.0),
+        ),
+        name=f"{name}_conv1_weights",
+    )(x)
+    x = layers.BatchNormalization(
+        beta_initializer=keras.initializers.Constant(1e-4),
+        gamma_initializer=keras.initializers.Constant(1.0),
+        name=f"{name}_conv1_bn",
+    )(x)
+    x = layers.ReLU(name=f"{name}_conv1_relu")(x)
+    x = layers.Conv2DTranspose(
+        1,
+        kernel_size=kernel_list[2],
+        strides=2,
+        padding="valid",
+        activation="sigmoid",
+        bias_initializer=keras.initializers.RandomUniform(
+            minval=-1.0 / math.sqrt(in_channels // 4 * 1.0),
+            maxval=1.0 / math.sqrt(in_channels // 4 * 1.0),
+        ),
+        name=f"{name}_conv2_weights",
+    )(x)
+    return x

keras_hub/src/models/differential_binarization/differential_binarization_backbone.py

@@ -0,0 +1,103 @@
+import keras
+from keras import layers
+
+from keras_hub.src.api_export import keras_hub_export
+from keras_hub.src.models.backbone import Backbone
+
+
+@keras_hub_export("keras_hub.models.DifferentialBinarizationBackbone")
+class DifferentialBinarizationBackbone(Backbone):
+    """
+    A Keras model implementing the Differential Binarization
+    architecture for scene text detection, described in
+    [Real-time Scene Text Detection with Differentiable Binarization](
+    https://arxiv.org/abs/1911.08947).
+
+    This class contains the backbone architecture containing the feature
+    pyramid network.
+
+    Args:
+        image_encoder: A `keras_hub.models.ResNetBackbone` instance.
+        fpn_channels: int. The number of channels to output by the feature
+            pyramid network. Defaults to 256.
+    """
+
+    def __init__(
+        self,
+        image_encoder,
+        fpn_channels=256,
+        **kwargs,
+    ):
+        inputs = image_encoder.input
+        x = image_encoder.pyramid_outputs
+        x = diffbin_fpn_model(x, out_channels=fpn_channels)
+
+        super().__init__(inputs=inputs, outputs=x, **kwargs)
+
+        self.image_encoder = image_encoder
+        self.fpn_channels = fpn_channels
+
+    def get_config(self):
+        config = super().get_config()
+        config["fpn_channels"] = self.fpn_channels
+        config["image_encoder"] = keras.layers.serialize(self.image_encoder)
+        return config
+
+    @classmethod
+    def from_config(cls, config):
+        config["image_encoder"] = keras.layers.deserialize(
+            config["image_encoder"]
+        )
+        return cls(**config)
+
+
+def diffbin_fpn_model(inputs, out_channels):
+    in2 = layers.Conv2D(
+        out_channels, kernel_size=1, use_bias=False, name="neck_in2"
+    )(inputs["P2"])
+    in3 = layers.Conv2D(
+        out_channels, kernel_size=1, use_bias=False, name="neck_in3"
+    )(inputs["P3"])
+    in4 = layers.Conv2D(
+        out_channels, kernel_size=1, use_bias=False, name="neck_in4"
+    )(inputs["P4"])
+    in5 = layers.Conv2D(
+        out_channels, kernel_size=1, use_bias=False, name="neck_in5"
+    )(inputs["P5"])
+    out4 = layers.Add(name="add1")([layers.UpSampling2D()(in5), in4])
+    out3 = layers.Add(name="add2")([layers.UpSampling2D()(out4), in3])
+    out2 = layers.Add(name="add3")([layers.UpSampling2D()(out3), in2])
+    p5 = layers.Conv2D(
+        out_channels // 4,
+        kernel_size=3,
+        padding="same",
+        use_bias=False,
+        name="neck_p5",
+    )(in5)
+    p4 = layers.Conv2D(
+        out_channels // 4,
+        kernel_size=3,
+        padding="same",
+        use_bias=False,
+        name="neck_p4",
+    )(out4)
+    p3 = layers.Conv2D(
+        out_channels // 4,
+        kernel_size=3,
+        padding="same",
+        use_bias=False,
+        name="neck_p3",
+    )(out3)
+    p2 = layers.Conv2D(
+        out_channels // 4,
+        kernel_size=3,
+        padding="same",
+        use_bias=False,
+        name="neck_p2",
+    )(out2)
+    p5 = layers.UpSampling2D((8, 8))(p5)
+    p4 = layers.UpSampling2D((4, 4))(p4)
+    p3 = layers.UpSampling2D((2, 2))(p3)
+
+    fused = layers.Concatenate(axis=-1)([p5, p4, p3, p2])
+    return fused

keras_hub/src/models/differential_binarization/differential_binarization_backbone_test.py

@@ -0,0 +1,45 @@
+from keras import ops
+
+from keras_hub.src.models.differential_binarization.differential_binarization_backbone import (
+    DifferentialBinarizationBackbone,
+)
+from keras_hub.src.models.differential_binarization.differential_binarization_preprocessor import (
+    DifferentialBinarizationPreprocessor,
+)
+from keras_hub.src.models.resnet.resnet_backbone import ResNetBackbone
+from keras_hub.src.tests.test_case import TestCase
+
+
+class DifferentialBinarizationTest(TestCase):
+    def setUp(self):
+        self.batch_size = 2
+        self.image_size = 16
+        self.images = ops.ones((2, 224, 224, 3))
+        self.image_encoder = ResNetBackbone(
+            input_conv_filters=[64],
+            input_conv_kernel_sizes=[7],
+            stackwise_num_filters=[64, 128, 256, 512],
+            stackwise_num_blocks=[3, 4, 6, 3],
+            stackwise_num_strides=[1, 2, 2, 2],
+            block_type="bottleneck_block",
+            image_shape=(224, 224, 3),
+        )
+        self.preprocessor = DifferentialBinarizationPreprocessor()
+        self.init_kwargs = {
+            "image_encoder": self.image_encoder,
+        }
+
+    def test_backbone_basics(self):
+        self.run_backbone_test(
+            cls=DifferentialBinarizationBackbone,
+            init_kwargs=self.init_kwargs,
+            input_data=self.images,
+            expected_output_shape=(
+                2,
+                56,
+                56,
+                256,
+            ),
+            run_mixed_precision_check=False,
+            run_quantization_check=False,
+        )

keras_hub/src/models/differential_binarization/differential_binarization_image_converter.py

@@ -0,0 +1,10 @@
+from keras_hub.src.api_export import keras_hub_export
+from keras_hub.src.layers.preprocessing.image_converter import ImageConverter
+from keras_hub.src.models.differential_binarization.differential_binarization_backbone import (
+    DifferentialBinarizationBackbone,
+)
+
+
+@keras_hub_export("keras_hub.layers.DifferentialBinarizationImageConverter")
+class DifferentialBinarizationImageConverter(ImageConverter):
+    backbone_cls = DifferentialBinarizationBackbone