keras-hub-nightly 0.19.0.dev202412120352__py3-none-any.whl → 0.19.0.dev202412140350__py3-none-any.whl

keras_hub/src/models/vit/vit_backbone.py

@@ -0,0 +1,152 @@
+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.vit.vit_layers import ViTEncoder
+from keras_hub.src.models.vit.vit_layers import ViTPatchingAndEmbedding
+from keras_hub.src.utils.keras_utils import standardize_data_format
+
+
+@keras_hub_export("keras_hub.models.ViTBackbone")
+class ViTBackbone(Backbone):
+    """Vision Transformer (ViT) backbone.
+
+    This backbone implements the Vision Transformer architecture as described in
+    [An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale](https://arxiv.org/abs/2010.11929).
+    It transforms the input image into a sequence of patches, embeds them, and
+    then processes them through a series of Transformer encoder layers.
+
+    Args:
+        image_shape: A tuple or list of 3 integers representing the shape of the
+            input image `(height, width, channels)`, `height` and `width` must
+            be equal.
+        patch_size: int. The size of each image patch, the input image will be
+            divided into patches of shape `(patch_size, patch_size)`.
+        num_layers: int. The number of transformer encoder layers.
+        num_heads: int. specifying the number of attention heads in each
+            Transformer encoder layer.
+        hidden_dim: int. The dimensionality of the hidden representations.
+        mlp_dim: int. The dimensionality of the intermediate MLP layer in
+            each Transformer encoder layer.
+        dropout_rate: float. The dropout rate for the Transformer encoder
+            layers.
+        attention_dropout: float. The dropout rate for the attention mechanism
+            in each Transformer encoder layer.
+        layer_norm_epsilon: float. Value used for numerical stability in
+            layer normalization.
+        use_mha_bias: bool. Whether to use bias in the multi-head
+            attention layers.
+        use_mlp_bias: bool. Whether to use bias in the MLP layers.
+        data_format: str.  `"channels_last"` or `"channels_first"`, specifying
+            the data format for the input image. If `None`, defaults to
+            `"channels_last"`.
+        dtype: The dtype of the layer weights. Defaults to None.
+        **kwargs: Additional keyword arguments to be passed to the parent
+            `Backbone` class.
+    """
+
+    def __init__(
+        self,
+        image_shape,
+        patch_size,
+        num_layers,
+        num_heads,
+        hidden_dim,
+        mlp_dim,
+        dropout_rate=0.0,
+        attention_dropout=0.0,
+        layer_norm_epsilon=1e-6,
+        use_mha_bias=True,
+        use_mlp_bias=True,
+        data_format=None,
+        dtype=None,
+        **kwargs,
+    ):
+        # === Laters ===
+        data_format = standardize_data_format(data_format)
+        h_axis, w_axis, channels_axis = (
+            (-3, -2, -1) if data_format == "channels_last" else (-2, -1, -3)
+        )
+        # Check that the input image is well specified.
+        if image_shape[h_axis] is None or image_shape[w_axis] is None:
+            raise ValueError(
+                f"Image shape must have defined height and width. Found `None` "
+                f"at index {h_axis} (height) or {w_axis} (width). "
+                f"Image shape: {image_shape}"
+            )
+        if image_shape[h_axis] != image_shape[w_axis]:
+            raise ValueError(
+                f"Image height and width must be equal. Found height: "
+                f"{image_shape[h_axis]}, width: {image_shape[w_axis]} at "
+                f"indices {h_axis} and {w_axis} respectively. Image shape: "
+                f"{image_shape}"
+            )
+
+        num_channels = image_shape[channels_axis]
+
+        # === Functional Model ===
+        inputs = keras.layers.Input(shape=image_shape)
+
+        x = ViTPatchingAndEmbedding(
+            image_size=image_shape[h_axis],
+            patch_size=patch_size,
+            hidden_dim=hidden_dim,
+            num_channels=num_channels,
+            data_format=data_format,
+            dtype=dtype,
+            name="vit_patching_and_embedding",
+        )(inputs)
+
+        output = ViTEncoder(
+            num_layers=num_layers,
+            num_heads=num_heads,
+            hidden_dim=hidden_dim,
+            mlp_dim=mlp_dim,
+            dropout_rate=dropout_rate,
+            attention_dropout=attention_dropout,
+            layer_norm_epsilon=layer_norm_epsilon,
+            use_mha_bias=use_mha_bias,
+            use_mlp_bias=use_mlp_bias,
+            dtype=dtype,
+            name="vit_encoder",
+        )(x)
+
+        super().__init__(
+            inputs=inputs,
+            outputs=output,
+            dtype=dtype,
+            **kwargs,
+        )
+
+        # === Config ===
+        self.image_shape = image_shape
+        self.patch_size = patch_size
+        self.num_layers = num_layers
+        self.num_heads = num_heads
+        self.hidden_dim = hidden_dim
+        self.mlp_dim = mlp_dim
+        self.dropout_rate = dropout_rate
+        self.attention_dropout = attention_dropout
+        self.layer_norm_epsilon = layer_norm_epsilon
+        self.use_mha_bias = use_mha_bias
+        self.use_mlp_bias = use_mlp_bias
+        self.data_format = data_format
+
+    def get_config(self):
+        config = super().get_config()
+        config.update(
+            {
+                "image_shape": self.image_shape,
+                "patch_size": self.patch_size,
+                "num_layers": self.num_layers,
+                "num_heads": self.num_heads,
+                "hidden_dim": self.hidden_dim,
+                "mlp_dim": self.mlp_dim,
+                "dropout_rate": self.dropout_rate,
+                "attention_dropout": self.attention_dropout,
+                "layer_norm_epsilon": self.layer_norm_epsilon,
+                "use_mha_bias": self.use_mha_bias,
+                "use_mlp_bias": self.use_mlp_bias,
+            }
+        )
+        return config

keras_hub/src/models/vit/vit_image_classifier.py

@@ -0,0 +1,187 @@
+import keras
+from keras import ops
+
+from keras_hub.src.api_export import keras_hub_export
+from keras_hub.src.models.image_classifier import ImageClassifier
+from keras_hub.src.models.task import Task
+from keras_hub.src.models.vit.vit_backbone import ViTBackbone
+from keras_hub.src.models.vit.vit_image_classifier_preprocessor import (
+    ViTImageClassifierPreprocessor,
+)
+
+
+@keras_hub_export("keras_hub.models.ViTImageClassifier")
+class ViTImageClassifier(ImageClassifier):
+    """ViT image classification task.
+
+    `ViTImageClassifier` tasks wrap a `keras_hub.models.ViTBackbone` and
+    a `keras_hub.models.Preprocessor` to create a model that can be used for
+    image classification. `ViTImageClassifier` tasks take an additional
+    `num_classes` argument, controlling the number of predicted output classes.
+
+    To fine-tune with `fit()`, pass a dataset containing tuples of `(x, y)`
+    labels where `x` is a string and `y` is a integer from `[0, num_classes)`.
+
+    Not that unlike `keras_hub.model.ImageClassifier`, the `ViTImageClassifier`
+    we pluck out `cls_token` which is first seqence from the backbone.
+
+    Args:
+        backbone: A `keras_hub.models.ViTBackbone` instance or a `keras.Model`.
+        num_classes: int. The number of classes to predict.
+        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.
+        pooling: String specifying the classification strategy. The choice
+            impacts the dimensionality and nature of the feature vector used for
+            classification.
+                `"token"`:  A single vector (class token) representing the
+                    overall image features.
+                `"gap"`: A single vector representing the average features
+                    across the spatial dimensions.
+        intermediate_dim: Optional dimensionality of the intermediate
+            representation layer before the final classification layer.
+            If `None`, the output of the transformer is directly used.
+            Defaults to `None`.
+        activation: `None`, str, or callable. The activation function to use on
+            the `Dense` layer. Set `activation=None` to return the output
+            logits. Defaults to `"softmax"`.
+        head_dtype: `None`, str, or `keras.mixed_precision.DTypePolicy`. The
+            dtype to use for the classification head's computations and weights.
+
+    Examples:
+
+    Call `predict()` to run inference.
+    ```python
+    # Load preset and train
+    images = np.random.randint(0, 256, size=(2, 224, 224, 3))
+    classifier = keras_hub.models.ViTImageClassifier.from_preset(
+        "vgg_16_imagenet"
+    )
+    classifier.predict(images)
+    ```
+
+    Call `fit()` on a single batch.
+    ```python
+    # Load preset and train
+    images = np.random.randint(0, 256, size=(2, 224, 224, 3))
+    labels = [0, 3]
+    classifier = keras_hub.models.VGGImageClassifier.from_preset(
+        "vit_base_patch16_224"
+    )
+    classifier.fit(x=images, y=labels, batch_size=2)
+    ```
+
+    Call `fit()` with custom loss, optimizer and backbone.
+    ```python
+    classifier = keras_hub.models.VGGImageClassifier.from_preset(
+        "vit_base_patch16_224"
+    )
+    classifier.compile(
+        loss=keras.losses.SparseCategoricalCrossentropy(from_logits=True),
+        optimizer=keras.optimizers.Adam(5e-5),
+    )
+    classifier.backbone.trainable = False
+    classifier.fit(x=images, y=labels, batch_size=2)
+    ```
+
+    Custom backbone.
+    ```python
+    images = np.random.randint(0, 256, size=(2, 224, 224, 3))
+    labels = [0, 3]
+    model = keras_hub.models.ViTBackbone(
+        image_shape = (224, 224, 3),
+        patch_size=16,
+        num_layers=6,
+        num_heads=3,
+        hidden_dim=768,
+        mlp_dim=2048
+    )
+    classifier = keras_hub.models.ViTImageClassifier(
+        backbone=backbone,
+        num_classes=4,
+    )
+    classifier.fit(x=images, y=labels, batch_size=2)
+    ```
+    """
+
+    backbone_cls = ViTBackbone
+    preprocessor_cls = ViTImageClassifierPreprocessor
+
+    def __init__(
+        self,
+        backbone,
+        num_classes,
+        preprocessor=None,
+        pooling="token",
+        intermediate_dim=None,
+        activation=None,
+        dropout=0.0,
+        head_dtype=None,
+        **kwargs,
+    ):
+        head_dtype = head_dtype or backbone.dtype_policy
+
+        # === Layers ===
+        self.backbone = backbone
+        self.preprocessor = preprocessor
+
+        if intermediate_dim is not None:
+            self.intermediate_layer = keras.layers.Dense(
+                intermediate_dim, activation="tanh", name="pre_logits"
+            )
+
+        self.dropout = keras.layers.Dropout(
+            rate=dropout,
+            dtype=head_dtype,
+            name="output_dropout",
+        )
+        self.output_dense = keras.layers.Dense(
+            num_classes,
+            activation=activation,
+            dtype=head_dtype,
+            name="predictions",
+        )
+
+        # === Functional Model ===
+        inputs = self.backbone.input
+        x = self.backbone(inputs)
+        if pooling == "token":
+            x = x[:, 0]
+        elif pooling == "gap":
+            ndim = len(ops.shape(x))
+            x = ops.mean(x, axis=list(range(1, ndim - 1)))  # (1,) or (1,2)
+
+        if intermediate_dim is not None:
+            x = self.intermediate_layer(x)
+
+        x = self.dropout(x)
+        outputs = self.output_dense(x)
+
+        # Skip the parent class functional model.
+        Task.__init__(
+            self,
+            inputs=inputs,
+            outputs=outputs,
+            **kwargs,
+        )
+
+        # === config ===
+        self.num_classes = num_classes
+        self.pooling = pooling
+        self.intermediate_dim = intermediate_dim
+        self.activation = activation
+        self.dropout = dropout
+
+    def get_config(self):
+        # Backbone serialized in `super`
+        config = super().get_config()
+        config.update(
+            {
+                "num_classes": self.num_classes,
+                "pooling": self.pooling,
+                "intermediate_dim": self.intermediate_dim,
+                "activation": self.activation,
+                "dropout": self.dropout,
+            }
+        )
+        return config

keras_hub/src/models/vit/vit_image_classifier_preprocessor.py

@@ -0,0 +1,12 @@
+from keras_hub.src.api_export import keras_hub_export
+from keras_hub.src.models.image_classifier_preprocessor import (
+    ImageClassifierPreprocessor,
+)
+from keras_hub.src.models.vit.vit_backbone import ViTBackbone
+from keras_hub.src.models.vit.vit_image_converter import ViTImageConverter
+
+
+@keras_hub_export("keras_hub.models.ViTImageClassifierPreprocessor")
+class ViTImageClassifierPreprocessor(ImageClassifierPreprocessor):
+    backbone_cls = ViTBackbone
+    image_converter_cls = ViTImageConverter

keras_hub/src/models/vit/vit_image_converter.py

@@ -0,0 +1,73 @@
+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.vit.vit_backbone import ViTBackbone
+from keras_hub.src.utils.tensor_utils import preprocessing_function
+
+
+@keras_hub_export("keras_hub.layers.ViTImageConverter")
+class ViTImageConverter(ImageConverter):
+    """Converts images to the format expected by a ViT model.
+
+    This layer performs image normalization using mean and standard deviation
+    values. By default, it uses the same normalization as the
+    "google/vit-large-patch16-224" model on Hugging Face:
+    `norm_mean=[0.5, 0.5, 0.5]` and `norm_std=[0.5, 0.5, 0.5]`
+    ([reference](https://huggingface.co/google/vit-large-patch16-224/blob/main/preprocessor_config.json)).
+    These defaults are suitable for models pretrained using this normalization.
+
+    Args:
+        norm_mean: list or tuple of floats. Mean values for image normalization.
+            Defaults to `[0.5, 0.5, 0.5]`.
+        norm_std: list or tuple of floats. Standard deviation values for
+            image normalization. Defaults to `[0.5, 0.5, 0.5]`.
+        **kwargs: Additional keyword arguments passed to
+            `keras_hub.layers.preprocessing.ImageConverter`.
+
+    Examples:
+    ```python
+    import keras
+    import numpy as np
+    from keras_hub.src.layers import ViTImageConverter
+
+    # Example image (replace with your actual image data)
+    image = np.random.rand(1, 224, 224, 3)  # Example: (B, H, W, C)
+
+    # Create a ViTImageConverter instance
+    converter = ViTImageConverter(
+        image_size=(28,28),
+        scale=1/255.
+    )
+    # Preprocess the image
+    preprocessed_image = converter(image)
+    ```
+    """
+
+    backbone_cls = ViTBackbone
+
+    def __init__(
+        self, norm_mean=[0.5, 0.5, 0.5], norm_std=[0.5, 0.5, 0.5], **kwargs
+    ):
+        super().__init__(**kwargs)
+        self.norm_mean = norm_mean
+        self.norm_std = norm_std
+
+    @preprocessing_function
+    def call(self, inputs):
+        x = super().call(inputs)
+        # By default normalize using imagenet mean and std
+        if self.norm_mean:
+            x = x - self._expand_non_channel_dims(self.norm_mean, x)
+        if self.norm_std:
+            x = x / self._expand_non_channel_dims(self.norm_std, x)
+
+        return x
+
+    def get_config(self):
+        config = super().get_config()
+        config.update(
+            {
+                "norm_mean": self.norm_mean,
+                "norm_std": self.norm_std,
+            }
+        )
+        return config