keras-hub-nightly 0.16.1.dev202410020340__py3-none-any.whl → 0.19.0.dev202501260345__py3-none-any.whl

keras_hub/src/models/deeplab_v3/deeplab_v3_backbone.py

@@ -0,0 +1,200 @@
+import keras
+
+from keras_hub.src.api_export import keras_hub_export
+from keras_hub.src.models.backbone import Backbone
+from keras_hub.src.models.deeplab_v3.deeplab_v3_layers import (
+    SpatialPyramidPooling,
+)
+
+
+@keras_hub_export("keras_hub.models.DeepLabV3Backbone")
+class DeepLabV3Backbone(Backbone):
+    """DeepLabV3 & DeepLabV3Plus architecture for semantic segmentation.
+
+    This class implements a DeepLabV3 & DeepLabV3Plus architecture as described
+    in [Encoder-Decoder with Atrous Separable Convolution for Semantic Image Segmentation](https://arxiv.org/abs/1802.02611)
+    (ECCV 2018)
+    and [Rethinking Atrous Convolution for Semantic Image Segmentation](https://arxiv.org/abs/1706.05587)
+    (CVPR 2017)
+
+    Args:
+        image_encoder: `keras.Model`. An instance that is used as a feature
+            extractor for the Encoder. Should either be a
+            `keras_hub.models.Backbone` or a `keras.Model` that implements the
+            `pyramid_outputs` property with keys "P2", "P3" etc as values.
+            A somewhat sensible backbone to use in many cases is
+            the `keras_hub.models.ResNetBackbone.from_preset("resnet_v2_50")`.
+        projection_filters: int. Number of filters in the convolution layer
+            projecting low-level features from the `image_encoder`.
+        spatial_pyramid_pooling_key: str. A layer level to extract and perform
+            `spatial_pyramid_pooling`, one of the key from the `image_encoder`
+            `pyramid_outputs` property such as  "P4", "P5" etc.
+        upsampling_size: int or tuple of 2 integers. The upsampling factors for
+            rows and columns of `spatial_pyramid_pooling` layer.
+            If `low_level_feature_key` is given then `spatial_pyramid_pooling`s
+            layer resolution should match with the `low_level_feature`s layer
+            resolution to concatenate both the layers for combined encoder
+            outputs.
+        dilation_rates: list. A `list` of integers for parallel dilated conv
+            applied to `SpatialPyramidPooling`. Usually a
+            sample choice of rates are `[6, 12, 18]`.
+        low_level_feature_key: str optional. A layer level to extract the
+            feature from one of the key from the `image_encoder`s
+            `pyramid_outputs` property such as  "P2", "P3" etc which will be the
+            Decoder block. Required only when the DeepLabV3Plus architecture
+            needs to be applied.
+        image_shape: tuple. The input shape without the batch size.
+            Defaults to `(None, None, 3)`.
+
+    Example:
+    ```python
+    # Load a trained backbone to extract features from it's `pyramid_outputs`.
+    image_encoder = keras_hub.models.ResNetBackbone.from_preset(
+        "resnet_50_imagenet"
+    )
+
+    model = keras_hub.models.DeepLabV3Backbone(
+        image_encoder=image_encoder,
+        projection_filters=48,
+        low_level_feature_key="P2",
+        spatial_pyramid_pooling_key="P5",
+        upsampling_size = 8,
+        dilation_rates = [6, 12, 18]
+    )
+    ```
+    """  # noqa: E501
+
+    def __init__(
+        self,
+        image_encoder,
+        spatial_pyramid_pooling_key,
+        upsampling_size,
+        dilation_rates,
+        low_level_feature_key=None,
+        projection_filters=48,
+        image_shape=(None, None, 3),
+        **kwargs,
+    ):
+        if not isinstance(image_encoder, keras.Model):
+            raise ValueError(
+                "Argument `image_encoder` must be a `keras.Model` instance. "
+                "Received instead "
+                f"{image_encoder} (of type {type(image_encoder)})."
+            )
+        data_format = keras.config.image_data_format()
+        channel_axis = -1 if data_format == "channels_last" else 1
+
+        # === Layers ===
+        inputs = keras.layers.Input(image_shape, name="inputs")
+
+        fpn_model = keras.Model(
+            image_encoder.inputs, image_encoder.pyramid_outputs
+        )
+
+        fpn_outputs = fpn_model(inputs)
+
+        spatial_pyramid_pooling = SpatialPyramidPooling(
+            dilation_rates=dilation_rates
+        )
+        spatial_backbone_features = fpn_outputs[spatial_pyramid_pooling_key]
+        spp_outputs = spatial_pyramid_pooling(spatial_backbone_features)
+
+        encoder_outputs = keras.layers.UpSampling2D(
+            size=upsampling_size,
+            interpolation="bilinear",
+            name="encoder_output_upsampling",
+            data_format=data_format,
+        )(spp_outputs)
+
+        if low_level_feature_key:
+            decoder_feature = fpn_outputs[low_level_feature_key]
+            low_level_projected_features = apply_low_level_feature_network(
+                decoder_feature, projection_filters, channel_axis
+            )
+
+            encoder_outputs = keras.layers.Concatenate(
+                axis=channel_axis, name="encoder_decoder_concat"
+            )([encoder_outputs, low_level_projected_features])
+        # upsampling to the original image size
+        upsampling = (2 ** int(spatial_pyramid_pooling_key[-1])) // (
+            int(upsampling_size[0])
+            if isinstance(upsampling_size, tuple)
+            else upsampling_size
+        )
+        # === Functional Model ===
+        x = keras.layers.Conv2D(
+            name="segmentation_head_conv",
+            filters=256,
+            kernel_size=1,
+            padding="same",
+            use_bias=False,
+            data_format=data_format,
+        )(encoder_outputs)
+        x = keras.layers.BatchNormalization(
+            name="segmentation_head_norm", axis=channel_axis
+        )(x)
+        x = keras.layers.ReLU(name="segmentation_head_relu")(x)
+        x = keras.layers.UpSampling2D(
+            size=upsampling,
+            interpolation="bilinear",
+            data_format=data_format,
+            name="backbone_output_upsampling",
+        )(x)
+
+        super().__init__(inputs=inputs, outputs=x, **kwargs)
+
+        # === Config ===
+        self.image_shape = image_shape
+        self.image_encoder = image_encoder
+        self.projection_filters = projection_filters
+        self.upsampling_size = upsampling_size
+        self.dilation_rates = dilation_rates
+        self.low_level_feature_key = low_level_feature_key
+        self.spatial_pyramid_pooling_key = spatial_pyramid_pooling_key
+
+    def get_config(self):
+        config = super().get_config()
+        config.update(
+            {
+                "image_encoder": keras.saving.serialize_keras_object(
+                    self.image_encoder
+                ),
+                "projection_filters": self.projection_filters,
+                "dilation_rates": self.dilation_rates,
+                "upsampling_size": self.upsampling_size,
+                "low_level_feature_key": self.low_level_feature_key,
+                "spatial_pyramid_pooling_key": self.spatial_pyramid_pooling_key,
+                "image_shape": self.image_shape,
+            }
+        )
+        return config
+
+    @classmethod
+    def from_config(cls, config):
+        if "image_encoder" in config and isinstance(
+            config["image_encoder"], dict
+        ):
+            config["image_encoder"] = keras.layers.deserialize(
+                config["image_encoder"]
+            )
+        return super().from_config(config)
+
+
+def apply_low_level_feature_network(
+    input_tensor, projection_filters, channel_axis
+):
+    data_format = keras.config.image_data_format()
+    x = keras.layers.Conv2D(
+        name="decoder_conv",
+        filters=projection_filters,
+        kernel_size=1,
+        padding="same",
+        use_bias=False,
+        data_format=data_format,
+    )(input_tensor)
+
+    x = keras.layers.BatchNormalization(name="decoder_norm", axis=channel_axis)(
+        x
+    )
+    x = keras.layers.ReLU(name="decoder_relu")(x)
+    return x

keras_hub/src/models/deeplab_v3/deeplab_v3_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.deeplab_v3.deeplab_v3_backbone import (
+    DeepLabV3Backbone,
+)
+
+
+@keras_hub_export("keras_hub.layers.DeepLabV3ImageConverter")
+class DeepLabV3ImageConverter(ImageConverter):
+    backbone_cls = DeepLabV3Backbone

keras_hub/src/models/deeplab_v3/deeplab_v3_image_segmeter_preprocessor.py

@@ -0,0 +1,16 @@
+from keras_hub.src.api_export import keras_hub_export
+from keras_hub.src.models.deeplab_v3.deeplab_v3_backbone import (
+    DeepLabV3Backbone,
+)
+from keras_hub.src.models.deeplab_v3.deeplab_v3_image_converter import (
+    DeepLabV3ImageConverter,
+)
+from keras_hub.src.models.image_segmenter_preprocessor import (
+    ImageSegmenterPreprocessor,
+)
+
+
+@keras_hub_export("keras_hub.models.DeepLabV3ImageSegmenterPreprocessor")
+class DeepLabV3ImageSegmenterPreprocessor(ImageSegmenterPreprocessor):
+    backbone_cls = DeepLabV3Backbone
+    image_converter_cls = DeepLabV3ImageConverter

keras_hub/src/models/deeplab_v3/deeplab_v3_layers.py

@@ -0,0 +1,215 @@
+import keras
+from keras import ops
+
+
+class SpatialPyramidPooling(keras.layers.Layer):
+    """Implements the Atrous Spatial Pyramid Pooling.
+
+    Reference for Atrous Spatial Pyramid Pooling [Rethinking Atrous Convolution
+    for Semantic Image Segmentation](https://arxiv.org/pdf/1706.05587.pdf) and
+    [Encoder-Decoder with Atrous Separable Convolution for Semantic Image
+    Segmentation](https://arxiv.org/pdf/1802.02611.pdf)
+
+    Args:
+    dilation_rates: list of ints. The dilation rate for parallel dilated conv.
+        Usually a sample choice of rates are `[6, 12, 18]`.
+    num_channels: int. The number of output channels, defaults to `256`.
+    activation: str. Activation to be used, defaults to `relu`.
+    dropout: float. The dropout rate of the final projection output after the
+        activations and batch norm, defaults to `0.0`, which means no dropout is
+        applied to the output.
+
+    Example:
+    ```python
+    inp = keras.layers.Input((384, 384, 3))
+    backbone = keras.applications.EfficientNetB0(
+        input_tensor=inp,
+        include_top=False)
+    output = backbone(inp)
+    output = SpatialPyramidPooling(
+        dilation_rates=[6, 12, 18])(output)
+    ```
+    """
+
+    def __init__(
+        self,
+        dilation_rates,
+        num_channels=256,
+        activation="relu",
+        dropout=0.0,
+        **kwargs,
+    ):
+        super().__init__(**kwargs)
+        self.dilation_rates = dilation_rates
+        self.num_channels = num_channels
+        self.activation = activation
+        self.dropout = dropout
+        self.data_format = keras.config.image_data_format()
+        self.channel_axis = -1 if self.data_format == "channels_last" else 1
+
+    def build(self, input_shape):
+        channels = input_shape[self.channel_axis]
+
+        # This is the parallel networks that process the input features with
+        # different dilation rates. The output from each channel will be merged
+        # together and feed to the output.
+        self.aspp_parallel_channels = []
+
+        # Channel1 with Conv2D and 1x1 kernel size.
+        conv_sequential = keras.Sequential(
+            [
+                keras.layers.Conv2D(
+                    filters=self.num_channels,
+                    kernel_size=(1, 1),
+                    use_bias=False,
+                    data_format=self.data_format,
+                    name="aspp_conv_1",
+                ),
+                keras.layers.BatchNormalization(
+                    axis=self.channel_axis, name="aspp_bn_1"
+                ),
+                keras.layers.Activation(
+                    self.activation, name="aspp_activation_1"
+                ),
+            ]
+        )
+        conv_sequential.build(input_shape)
+        self.aspp_parallel_channels.append(conv_sequential)
+
+        # Channel 2 and afterwards are based on self.dilation_rates, and each of
+        # them will have conv2D with 3x3 kernel size.
+        for i, dilation_rate in enumerate(self.dilation_rates):
+            conv_sequential = keras.Sequential(
+                [
+                    keras.layers.Conv2D(
+                        filters=self.num_channels,
+                        kernel_size=(3, 3),
+                        padding="same",
+                        dilation_rate=dilation_rate,
+                        use_bias=False,
+                        data_format=self.data_format,
+                        name=f"aspp_conv_{i + 2}",
+                    ),
+                    keras.layers.BatchNormalization(
+                        axis=self.channel_axis, name=f"aspp_bn_{i + 2}"
+                    ),
+                    keras.layers.Activation(
+                        self.activation, name=f"aspp_activation_{i + 2}"
+                    ),
+                ]
+            )
+            conv_sequential.build(input_shape)
+            self.aspp_parallel_channels.append(conv_sequential)
+
+        # Last channel is the global average pooling with conv2D 1x1 kernel.
+        if self.channel_axis == -1:
+            reshape = keras.layers.Reshape((1, 1, channels), name="reshape")
+        else:
+            reshape = keras.layers.Reshape((channels, 1, 1), name="reshape")
+        pool_sequential = keras.Sequential(
+            [
+                keras.layers.GlobalAveragePooling2D(
+                    data_format=self.data_format, name="average_pooling"
+                ),
+                reshape,
+                keras.layers.Conv2D(
+                    filters=self.num_channels,
+                    kernel_size=(1, 1),
+                    use_bias=False,
+                    data_format=self.data_format,
+                    name="conv_pooling",
+                ),
+                keras.layers.BatchNormalization(
+                    axis=self.channel_axis, name="bn_pooling"
+                ),
+                keras.layers.Activation(
+                    self.activation, name="activation_pooling"
+                ),
+            ]
+        )
+        pool_sequential.build(input_shape)
+        self.aspp_parallel_channels.append(pool_sequential)
+
+        # Final projection layers
+        projection = keras.Sequential(
+            [
+                keras.layers.Conv2D(
+                    filters=self.num_channels,
+                    kernel_size=(1, 1),
+                    use_bias=False,
+                    data_format=self.data_format,
+                    name="conv_projection",
+                ),
+                keras.layers.BatchNormalization(
+                    axis=self.channel_axis, name="bn_projection"
+                ),
+                keras.layers.Activation(
+                    self.activation, name="activation_projection"
+                ),
+                keras.layers.Dropout(rate=self.dropout, name="dropout"),
+            ],
+        )
+        projection_input_channels = (
+            2 + len(self.dilation_rates)
+        ) * self.num_channels
+        if self.data_format == "channels_first":
+            projection.build(
+                (input_shape[0],)
+                + (projection_input_channels,)
+                + (input_shape[2:])
+            )
+        else:
+            projection.build((input_shape[:-1]) + (projection_input_channels,))
+        self.projection = projection
+        self.built = True
+
+    def call(self, inputs):
+        """Calls the Atrous Spatial Pyramid Pooling layer on an input.
+
+        Args:
+            inputs: A tensor of shape [batch, height, width, channels]
+
+        Returns:
+            A tensor of shape [batch, height, width, num_channels]
+        """
+        result = []
+
+        for channel in self.aspp_parallel_channels:
+            temp = ops.cast(channel(inputs), inputs.dtype)
+            result.append(temp)
+
+        image_shape = ops.shape(inputs)
+        if self.channel_axis == -1:
+            height, width = image_shape[1], image_shape[2]
+        else:
+            height, width = image_shape[2], image_shape[3]
+        result[-1] = keras.layers.Resizing(
+            height,
+            width,
+            interpolation="bilinear",
+            data_format=self.data_format,
+            name="resizing",
+        )(result[-1])
+
+        result = ops.concatenate(result, axis=self.channel_axis)
+        return self.projection(result)
+
+    def compute_output_shape(self, inputs_shape):
+        if self.data_format == "channels_first":
+            return tuple(
+                (inputs_shape[0],) + (self.num_channels,) + (inputs_shape[2:])
+            )
+        else:
+            return tuple((inputs_shape[:-1]) + (self.num_channels,))
+
+    def get_config(self):
+        config = super().get_config()
+        config.update(
+            {
+                "dilation_rates": self.dilation_rates,
+                "num_channels": self.num_channels,
+                "activation": self.activation,
+                "dropout": self.dropout,
+            }
+        )
+        return config

keras_hub/src/models/deeplab_v3/deeplab_v3_presets.py

@@ -0,0 +1,17 @@
+"""DeepLabV3 preset configurations."""
+
+backbone_presets = {
+    "deeplab_v3_plus_resnet50_pascalvoc": {
+        "metadata": {
+            "description": (
+                "DeepLabV3+ model with ResNet50 as image encoder and trained "
+                "on augmented Pascal VOC dataset by Semantic Boundaries "
+                "Dataset(SBD) which is having categorical accuracy of 90.01 "
+                "and 0.63 Mean IoU."
+            ),
+            "params": 39190656,
+            "path": "deeplab_v3",
+        },
+        "kaggle_handle": "kaggle://keras/deeplabv3plus/keras/deeplab_v3_plus_resnet50_pascalvoc/4",
+    },
+}

keras_hub/src/models/deeplab_v3/deeplab_v3_segmenter.py

@@ -0,0 +1,111 @@
+import keras
+
+from keras_hub.src.api_export import keras_hub_export
+from keras_hub.src.models.deeplab_v3.deeplab_v3_backbone import (
+    DeepLabV3Backbone,
+)
+from keras_hub.src.models.deeplab_v3.deeplab_v3_image_segmeter_preprocessor import (  # noqa: E501
+    DeepLabV3ImageSegmenterPreprocessor,
+)
+from keras_hub.src.models.image_segmenter import ImageSegmenter
+
+
+@keras_hub_export("keras_hub.models.DeepLabV3ImageSegmenter")
+class DeepLabV3ImageSegmenter(ImageSegmenter):
+    """DeepLabV3 and DeeplabV3 and DeeplabV3Plus segmentation task.
+
+    Args:
+        backbone: A `keras_hub.models.DeepLabV3` instance.
+        num_classes: int. The number of classes for the detection model. Note
+            that the `num_classes` contains the background class, and the
+            classes from the data should be represented by integers with range
+            `[0, num_classes]`.
+        activation: str or callable. The activation function to use on
+            the `Dense` layer. Set `activation=None` to return the output
+            logits. Defaults to `None`.
+        preprocessor: A `keras_hub.models.DeepLabV3ImageSegmenterPreprocessor`
+            or `None`. If `None`, this model will not apply preprocessing, and
+            inputs should be preprocessed before calling the model.
+
+    Example:
+    Load a DeepLabV3 preset with all the 21 class, pretrained segmentation head.
+    ```python
+    images = np.ones(shape=(1, 96, 96, 3))
+    labels = np.zeros(shape=(1, 96, 96, 2))
+    segmenter = keras_hub.models.DeepLabV3ImageSegmenter.from_preset(
+        "deeplab_v3_plus_resnet50_pascalvoc",
+    )
+    segmenter.predict(images)
+    ```
+
+    Specify `num_classes` to load randomly initialized segmentation head.
+    ```python
+    segmenter = keras_hub.models.DeepLabV3ImageSegmenter.from_preset(
+        "deeplab_v3_plus_resnet50_pascalvoc",
+        num_classes=2,
+    )
+    segmenter.preprocessor.image_size = (96, 96)
+    segmenter.fit(images, labels, epochs=3)
+    segmenter.predict(images)  # Trained 2 class segmentation.
+    ```
+
+    Load DeepLabv3+ presets a extension of DeepLabv3 by adding a simple yet
+    effective decoder module to refine the segmentation results especially
+    along object boundaries.
+    ```python
+    segmenter = keras_hub.models.DeepLabV3ImageSegmenter.from_preset(
+        "deeplabv3_plus_resnet50_pascalvoc",
+    )
+    segmenter.predict(images)
+    ```
+    """
+
+    backbone_cls = DeepLabV3Backbone
+    preprocessor_cls = DeepLabV3ImageSegmenterPreprocessor
+
+    def __init__(
+        self,
+        backbone,
+        num_classes,
+        activation=None,
+        preprocessor=None,
+        **kwargs,
+    ):
+        data_format = keras.config.image_data_format()
+        # === Layers ===
+        self.output_conv = keras.layers.Conv2D(
+            name="segmentation_output",
+            filters=num_classes,
+            kernel_size=1,
+            use_bias=False,
+            padding="same",
+            activation=activation,
+            data_format=data_format,
+        )
+
+        # === Functional Model ===
+        inputs = backbone.input
+        x = backbone(inputs)
+        outputs = self.output_conv(x)
+        super().__init__(
+            inputs=inputs,
+            outputs=outputs,
+            **kwargs,
+        )
+
+        # === Config ===
+        self.backbone = backbone
+        self.num_classes = num_classes
+        self.activation = activation
+        self.preprocessor = preprocessor
+
+    def get_config(self):
+        # Backbone serialized in `super`
+        config = super().get_config()
+        config.update(
+            {
+                "num_classes": self.num_classes,
+                "activation": self.activation,
+            }
+        )
+        return config

keras_hub/src/models/densenet/densenet_backbone.py

@@ -29,7 +29,9 @@ class DenseNetBackbone(FeaturePyramidBackbone):
     input_data = np.ones(shape=(8, 224, 224, 3))
 
     # Pretrained backbone
-    model = keras_hub.models.DenseNetBackbone.from_preset("densenet121_imagenet")
+    model = keras_hub.models.DenseNetBackbone.from_preset(
+        "densenet_121_imagenet"
+    )
     model(input_data)
 
     # Randomly initialized backbone with a custom config
@@ -79,14 +81,14 @@ class DenseNetBackbone(FeaturePyramidBackbone):
                 channel_axis,
                 stackwise_num_repeats[stack_index],
                 growth_rate,
-                name=f"stack{stack_index+1}",
+                name=f"stack{stack_index + 1}",
             )
             pyramid_outputs[f"P{index}"] = x
             x = apply_transition_block(
                 x,
                 channel_axis,
                 compression_ratio,
-                name=f"transition{stack_index+1}",
+                name=f"transition{stack_index + 1}",
             )
 
         x = apply_dense_block(
@@ -138,7 +140,7 @@ def apply_dense_block(x, channel_axis, num_repeats, growth_rate, name=None):
 
     for i in range(num_repeats):
         x = apply_conv_block(
-            x, channel_axis, growth_rate, name=f"{name}_block{i+1}"
+            x, channel_axis, growth_rate, name=f"{name}_block{i + 1}"
         )
     return x