keras-hub-nightly 0.21.0.dev202505270408__tar.gz → 0.22.0.dev202505290412__tar.gz

{keras_hub_nightly-0.21.0.dev202505270408 → keras_hub_nightly-0.22.0.dev202505290412}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: keras-hub-nightly
-Version: 0.21.0.dev202505270408
+Version: 0.22.0.dev202505290412
 Summary: Pretrained models for Keras.
 Author-email: Keras team <keras-users@googlegroups.com>
 License-Expression: Apache-2.0

{keras_hub_nightly-0.21.0.dev202505270408 → keras_hub_nightly-0.22.0.dev202505290412}/keras_hub/src/layers/preprocessing/multi_segment_packer.py

@@ -3,6 +3,7 @@ from keras_hub.src.layers.preprocessing.preprocessing_layer import (
     PreprocessingLayer,
 )
 from keras_hub.src.utils.tensor_utils import convert_to_ragged_batch
+from keras_hub.src.utils.tensor_utils import pad
 from keras_hub.src.utils.tensor_utils import preprocessing_function
 
 try:
@@ -66,6 +67,8 @@ class MultiSegmentPacker(PreprocessingLayer):
                 "waterfall" algorithm that allocates quota in a
                 left-to-right manner and fills up the buckets until we run
                 out of budget. It support arbitrary number of segments.
+        padding_side: str. Whether to pad the input on the "left" or "right".
+            Defaults to "right".
 
     Returns:
         A tuple with two elements. The first is the dense, packed token
@@ -124,6 +127,7 @@ class MultiSegmentPacker(PreprocessingLayer):
         sep_value=None,
         pad_value=None,
         truncate="round_robin",
+        padding_side="right",
         **kwargs,
     ):
         super().__init__(**kwargs)
@@ -162,6 +166,7 @@ class MultiSegmentPacker(PreprocessingLayer):
         self.end_value = end_value
 
         self.pad_value = pad_value
+        self.padding_side = padding_side
 
     def get_config(self):
         config = super().get_config()
@@ -173,6 +178,7 @@ class MultiSegmentPacker(PreprocessingLayer):
                 "sep_value": self._sep_value,
                 "pad_value": self.pad_value,
                 "truncate": self.truncate,
+                "padding_side": self.padding_side,
             }
         )
         return config
@@ -287,10 +293,18 @@ class MultiSegmentPacker(PreprocessingLayer):
         # Pad to dense tensor output.
         sequence_length = sequence_length or self.sequence_length
         shape = tf.cast([-1, sequence_length], "int64")
-        token_ids = token_ids.to_tensor(
-            shape=shape, default_value=self.pad_value
+        token_ids = pad(
+            token_ids,
+            shape=shape,
+            padding_side=self.padding_side,
+            pad_value=self.pad_value,
+        )
+        segment_ids = pad(
+            segment_ids,
+            shape=shape,
+            padding_side=self.padding_side,
+            pad_value=0,
         )
-        segment_ids = segment_ids.to_tensor(shape=shape)
         # Remove the batch dim if added.
         if unbatched:
             token_ids = tf.squeeze(token_ids, 0)

{keras_hub_nightly-0.21.0.dev202505270408 → keras_hub_nightly-0.22.0.dev202505290412}/keras_hub/src/layers/preprocessing/start_end_packer.py

@@ -3,6 +3,7 @@ from keras_hub.src.layers.preprocessing.preprocessing_layer import (
     PreprocessingLayer,
 )
 from keras_hub.src.utils.tensor_utils import convert_to_ragged_batch
+from keras_hub.src.utils.tensor_utils import pad
 from keras_hub.src.utils.tensor_utils import preprocessing_function
 
 try:
@@ -39,6 +40,8 @@ class StartEndPacker(PreprocessingLayer):
             0 or "" will be added depending on the dtype of the input tensor.
         return_padding_mask: bool. Whether to return a boolean padding mask of
             all locations that are filled in with the `pad_value`.
+        padding_side: str. Whether to pad the input on the "left" or "right".
+            Defaults to "right".
 
     Call arguments:
         inputs: A `tf.Tensor`, `tf.RaggedTensor`, or list of python strings.
@@ -111,6 +114,7 @@ class StartEndPacker(PreprocessingLayer):
         pad_value=None,
         return_padding_mask=False,
         name=None,
+        padding_side="right",
         **kwargs,
     ):
         super().__init__(name=name, **kwargs)
@@ -139,6 +143,7 @@ class StartEndPacker(PreprocessingLayer):
 
         self.pad_value = pad_value
         self.return_padding_mask = return_padding_mask
+        self.padding_side = padding_side
 
     @preprocessing_function
     def call(
@@ -154,6 +159,13 @@ class StartEndPacker(PreprocessingLayer):
         batch_size = tf.shape(x)[0]
         sequence_length = sequence_length or self.sequence_length
         dtype = inputs.dtype
+        # Truncate.
+        truncation_length = sequence_length
+        if add_start_value and self.start_value is not None:
+            truncation_length -= len(self.start_value)
+        if add_end_value and self.end_value is not None:
+            truncation_length -= len(self.end_value)
+        x = x[..., :truncation_length]
 
         # Concatenate start and end tokens.
         if add_start_value and self.start_value is not None:
@@ -167,23 +179,28 @@ class StartEndPacker(PreprocessingLayer):
             end_token_id_tensor = tf.repeat(
                 end_value[tf.newaxis, :], repeats=batch_size, axis=0
             )
-            # Trim to leave room for end token.
-            x = x[..., : sequence_length - len(self.end_value)]
             x = tf.concat([x, end_token_id_tensor], axis=-1)
 
         # Pad to desired length.
-        outputs = x.to_tensor(
-            default_value=self.pad_value,
+        outputs = pad(
+            x,
+            pad_value=self.pad_value,
+            padding_side=self.padding_side,
             shape=(batch_size, sequence_length),
         )
         outputs = tf.squeeze(outputs, axis=0) if unbatched else outputs
 
         if self.return_padding_mask:
             mask = tf.ones_like(x, dtype="bool")
-            mask = mask.to_tensor(shape=(batch_size, sequence_length))
+
+            mask = pad(
+                mask,
+                pad_value=False,
+                padding_side=self.padding_side,
+                shape=(batch_size, sequence_length),
+            )
             mask = tf.squeeze(mask, axis=0) if unbatched else mask
             return outputs, mask
-
         return outputs
 
     def get_config(self):
@@ -195,6 +212,7 @@ class StartEndPacker(PreprocessingLayer):
                 "end_value": self._end_value,
                 "pad_value": self.pad_value,
                 "return_padding_mask": self.return_padding_mask,
+                "padding_side": self.padding_side,
             }
         )
         return config

{keras_hub_nightly-0.21.0.dev202505270408 → keras_hub_nightly-0.22.0.dev202505290412}/keras_hub/src/models/backbone.py

@@ -189,23 +189,26 @@ class Backbone(keras.Model):
         saver = get_preset_saver(preset_dir)
         saver.save_backbone(self, max_shard_size=max_shard_size)
 
-    def get_lora_target_names(self):
-        """Returns list of layer names which are to be LoRA-fied.
-
-        Subclasses can override this method if the names of layers to be
-        LoRa-fied are different.
-        """
+    def default_lora_layer_names(self):
+        """Returns list of layer names which are to be LoRA-fied."""
         return ["query_dense", "value_dense", "query", "value"]
 
-    def enable_lora(self, rank, target_names=None):
+    def enable_lora(self, rank, target_layer_names=None):
         """Enable Lora on the backbone.
 
         Calling this method will freeze all weights on the backbone,
         while enabling Lora on the query & value `EinsumDense` layers
         of the attention layers.
+
+        Args:
+            rank: The rank of the LoRA factorization.
+            target_layer_names: A list of strings, the names of the layers to
+                apply LoRA to. If `None`, this will be populated with the
+                default LoRA layer names as returned by
+                `backbone.default_lora_layer_names()`.
         """
-        if target_names is None:
-            target_names = self.get_lora_target_names()
+        if target_layer_names is None:
+            target_layer_names = self.default_lora_layer_names()
         self.trainable = True
         self._lora_enabled_layers = []
         self._lora_rank = rank
@@ -214,7 +217,7 @@ class Backbone(keras.Model):
         all_layers = self._flatten_layers(include_self=False)
         all_layers = [lyr for lyr in all_layers if lyr.weights]
         for i, layer in enumerate(all_layers):
-            for name in target_names:
+            for name in target_layer_names:
                 if layer.name == name:
                     if hasattr(layer, "enable_lora"):
                         layer.trainable = True

{keras_hub_nightly-0.21.0.dev202505270408 → keras_hub_nightly-0.22.0.dev202505290412}/keras_hub/src/models/gemma3/gemma3_backbone.py

@@ -402,8 +402,8 @@ class Gemma3Backbone(Backbone):
         )
         return config
 
-    def get_lora_target_names(self):
-        target_names = super().get_lora_target_names()
+    def default_lora_layer_names(self):
+        target_names = super().default_lora_layer_names()
 
         # Add these for `Gemma3VITAttention`.
         if not self.text_only_model:

{keras_hub_nightly-0.21.0.dev202505270408 → keras_hub_nightly-0.22.0.dev202505290412}/keras_hub/src/models/pali_gemma/pali_gemma_backbone.py

@@ -274,8 +274,8 @@ class PaliGemmaBackbone(Backbone):
         # Keep the image_sequence_length as a backbone property for easy access.
         self.image_sequence_length = self.vit_encoder.image_sequence_length
 
-    def get_lora_target_names(self):
-        target_names = super().get_lora_target_names()
+    def default_lora_layer_names(self):
+        target_names = super().default_lora_layer_names()
 
         # Add these for `PaliGemmaVITAttention`.
         target_names += ["query_proj", "value_proj"]

{keras_hub_nightly-0.21.0.dev202505270408 → keras_hub_nightly-0.22.0.dev202505290412}/keras_hub/src/models/vit/vit_backbone.py

@@ -18,10 +18,10 @@ class ViTBackbone(Backbone):
 
     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)`.
+            input image `(height, width, channels)`.
+        patch_size: int or (int, int). The size of each image patch, the input
+            image will be divided into patches of shape
+            `(patch_size_h, patch_size_w)`.
         num_layers: int. The number of transformer encoder layers.
         num_heads: int. specifying the number of attention heads in each
             Transformer encoder layer.
@@ -37,6 +37,10 @@ class ViTBackbone(Backbone):
         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.
+        use_class_token: bool. Whether to use class token to be part of
+            patch embedding. Defaults to `True`.
+        use_patch_bias: bool. Whether to use bias in Conv2d of patch embedding
+            layer. Defaults to `True`.
         data_format: str.  `"channels_last"` or `"channels_first"`, specifying
             the data format for the input image. If `None`, defaults to
             `"channels_last"`.
@@ -58,6 +62,8 @@ class ViTBackbone(Backbone):
         layer_norm_epsilon=1e-6,
         use_mha_bias=True,
         use_mlp_bias=True,
+        use_class_token=True,
+        use_patch_bias=True,
         data_format=None,
         dtype=None,
         **kwargs,
@@ -74,24 +80,34 @@ class ViTBackbone(Backbone):
                 f"at index {h_axis} (height) or {w_axis} (width). "
                 f"Image shape: {image_shape}"
             )
-        if image_shape[h_axis] != image_shape[w_axis]:
+
+        if isinstance(patch_size, int):
+            patch_size = (patch_size, patch_size)
+
+        if image_shape[h_axis] % patch_size[0] != 0:
+            raise ValueError(
+                f"Input height {image_shape[h_axis]} should be divisible by "
+                f"patch size {patch_size[0]}."
+            )
+
+        if image_shape[w_axis] % patch_size[1] != 0:
             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}"
+                f"Input width {image_shape[h_axis]} should be divisible by "
+                f"patch size {patch_size[1]}."
             )
 
         num_channels = image_shape[channels_axis]
 
         # === Functional Model ===
-        inputs = keras.layers.Input(shape=image_shape)
+        inputs = keras.layers.Input(shape=image_shape, name="images")
 
         x = ViTPatchingAndEmbedding(
-            image_size=image_shape[h_axis],
+            image_size=(image_shape[h_axis], image_shape[w_axis]),
             patch_size=patch_size,
             hidden_dim=hidden_dim,
             num_channels=num_channels,
+            use_class_token=use_class_token,
+            use_patch_bias=use_patch_bias,
             data_format=data_format,
             dtype=dtype,
             name="vit_patching_and_embedding",
@@ -130,6 +146,8 @@ class ViTBackbone(Backbone):
         self.layer_norm_epsilon = layer_norm_epsilon
         self.use_mha_bias = use_mha_bias
         self.use_mlp_bias = use_mlp_bias
+        self.use_class_token = use_class_token
+        self.use_patch_bias = use_patch_bias
         self.data_format = data_format
 
     def get_config(self):
@@ -147,6 +165,8 @@ class ViTBackbone(Backbone):
                 "layer_norm_epsilon": self.layer_norm_epsilon,
                 "use_mha_bias": self.use_mha_bias,
                 "use_mlp_bias": self.use_mlp_bias,
+                "use_class_token": self.use_class_token,
+                "use_patch_bias": self.use_patch_bias,
             }
         )
         return config

keras_hub_nightly-0.22.0.dev202505290412/keras_hub/src/models/vit/vit_image_converter.py

@@ -0,0 +1,8 @@
+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
+
+
+@keras_hub_export("keras_hub.layers.ViTImageConverter")
+class ViTImageConverter(ImageConverter):
+    backbone_cls = ViTBackbone

{keras_hub_nightly-0.21.0.dev202505270408 → keras_hub_nightly-0.22.0.dev202505290412}/keras_hub/src/models/vit/vit_layers.py

@@ -75,12 +75,13 @@ class ViTPatchingAndEmbedding(keras.layers.Layer):
     """Patches the image and embeds the patches.
 
     Args:
-        image_size: int. Size of the input image (height or width).
-            Assumed to be square.
-        patch_size: int. Size of each image patch.
+        image_size: (int, int). Size of the input image.
+        patch_size: (int, int). Size of each image patch.
         hidden_dim: int. Dimensionality of the patch embeddings.
         num_channels: int. Number of channels in the input image. Defaults to
             `3`.
+        use_class_token: bool. Whether to use class token to be part of
+            patch embedding. Defaults to `True`.
         data_format: str. `"channels_last"` or `"channels_first"`. Defaults to
             `None` (which uses `"channels_last"`).
         **kwargs: Additional keyword arguments passed to `keras.layers.Layer`
@@ -92,12 +93,15 @@ class ViTPatchingAndEmbedding(keras.layers.Layer):
         patch_size,
         hidden_dim,
         num_channels=3,
+        use_class_token=True,
+        use_patch_bias=True,
         data_format=None,
         **kwargs,
     ):
         super().__init__(**kwargs)
-        num_patches = (image_size // patch_size) ** 2
-        num_positions = num_patches + 1
+        grid_size = tuple([s // p for s, p in zip(image_size, patch_size)])
+        num_patches = grid_size[0] * grid_size[1]
+        num_positions = num_patches + 1 if use_class_token else num_patches
 
         # === Config ===
         self.image_size = image_size
@@ -106,19 +110,22 @@ class ViTPatchingAndEmbedding(keras.layers.Layer):
         self.num_channels = num_channels
         self.num_patches = num_patches
         self.num_positions = num_positions
+        self.use_class_token = use_class_token
+        self.use_patch_bias = use_patch_bias
         self.data_format = standardize_data_format(data_format)
 
     def build(self, input_shape):
-        self.class_token = self.add_weight(
-            shape=(
-                1,
-                1,
-                self.hidden_dim,
-            ),
-            initializer="random_normal",
-            dtype=self.variable_dtype,
-            name="class_token",
-        )
+        if self.use_class_token:
+            self.class_token = self.add_weight(
+                shape=(
+                    1,
+                    1,
+                    self.hidden_dim,
+                ),
+                initializer="random_normal",
+                dtype=self.variable_dtype,
+                name="class_token",
+            )
         self.patch_embedding = keras.layers.Conv2D(
             filters=self.hidden_dim,
             kernel_size=self.patch_size,
@@ -127,6 +134,7 @@ class ViTPatchingAndEmbedding(keras.layers.Layer):
             activation=None,
             dtype=self.dtype_policy,
             data_format=self.data_format,
+            use_bias=self.use_patch_bias,
             name="patch_embedding",
         )
         self.patch_embedding.build(input_shape)
@@ -153,10 +161,16 @@ class ViTPatchingAndEmbedding(keras.layers.Layer):
         patch_embeddings = ops.reshape(
             patch_embeddings, [embeddings_shape[0], -1, embeddings_shape[-1]]
         )
-        class_token = ops.tile(self.class_token, (embeddings_shape[0], 1, 1))
         position_embeddings = self.position_embedding(self.position_ids)
-        embeddings = ops.concatenate([class_token, patch_embeddings], axis=1)
-        return ops.add(embeddings, position_embeddings)
+
+        if self.use_class_token:
+            class_token = ops.tile(
+                self.class_token, (embeddings_shape[0], 1, 1)
+            )
+            patch_embeddings = ops.concatenate(
+                [class_token, patch_embeddings], axis=1
+            )
+        return ops.add(patch_embeddings, position_embeddings)
 
     def compute_output_shape(self, input_shape):
         return (
@@ -175,6 +189,7 @@ class ViTPatchingAndEmbedding(keras.layers.Layer):
                 "num_channels": self.num_channels,
                 "num_patches": self.num_patches,
                 "num_positions": self.num_positions,
+                "use_class_token": self.use_class_token,
             }
         )
         return config

{keras_hub_nightly-0.21.0.dev202505270408 → keras_hub_nightly-0.22.0.dev202505290412}/keras_hub/src/models/vit/vit_presets.py

@@ -11,7 +11,7 @@ backbone_presets = {
             "params": 85798656,
             "path": "vit",
         },
-        "kaggle_handle": "kaggle://keras/vit/keras/vit_base_patch16_224_imagenet/2",
+        "kaggle_handle": "kaggle://keras/vit/keras/vit_base_patch16_224_imagenet/3",
     },
     "vit_base_patch16_384_imagenet": {
         "metadata": {
@@ -22,7 +22,7 @@ backbone_presets = {
             "params": 86090496,
             "path": "vit",
         },
-        "kaggle_handle": "kaggle://keras/vit/keras/vit_base_patch16_384_imagenet/2",
+        "kaggle_handle": "kaggle://keras/vit/keras/vit_base_patch16_384_imagenet/3",
     },
     "vit_large_patch16_224_imagenet": {
         "metadata": {
@@ -33,7 +33,7 @@ backbone_presets = {
             "params": 303301632,
             "path": "vit",
         },
-        "kaggle_handle": "kaggle://keras/vit/keras/vit_large_patch16_224_imagenet/2",
+        "kaggle_handle": "kaggle://keras/vit/keras/vit_large_patch16_224_imagenet/3",
     },
     "vit_large_patch16_384_imagenet": {
         "metadata": {
@@ -44,7 +44,7 @@ backbone_presets = {
             "params": 303690752,
             "path": "vit",
         },
-        "kaggle_handle": "kaggle://keras/vit/keras/vit_large_patch16_384_imagenet/2",
+        "kaggle_handle": "kaggle://keras/vit/keras/vit_large_patch16_384_imagenet/3",
     },
     "vit_base_patch32_384_imagenet": {
         "metadata": {
@@ -55,7 +55,7 @@ backbone_presets = {
             "params": 87528192,
             "path": "vit",
         },
-        "kaggle_handle": "kaggle://keras/vit/keras/vit_base_patch32_384_imagenet/1",
+        "kaggle_handle": "kaggle://keras/vit/keras/vit_base_patch32_384_imagenet/2",
     },
     "vit_large_patch32_384_imagenet": {
         "metadata": {
@@ -66,7 +66,7 @@ backbone_presets = {
             "params": 305607680,
             "path": "vit",
         },
-        "kaggle_handle": "kaggle://keras/vit/keras/vit_large_patch32_384_imagenet/1",
+        "kaggle_handle": "kaggle://keras/vit/keras/vit_large_patch32_384_imagenet/2",
     },
     "vit_base_patch16_224_imagenet21k": {
         "metadata": {
@@ -77,7 +77,7 @@ backbone_presets = {
             "params": 85798656,
             "path": "vit",
         },
-        "kaggle_handle": "kaggle://keras/vit/keras/vit_base_patch16_224_imagenet21k/1",
+        "kaggle_handle": "kaggle://keras/vit/keras/vit_base_patch16_224_imagenet21k/2",
     },
     "vit_base_patch32_224_imagenet21k": {
         "metadata": {
@@ -88,7 +88,7 @@ backbone_presets = {
             "params": 87455232,
             "path": "vit",
         },
-        "kaggle_handle": "kaggle://keras/vit/keras/vit_base_patch32_224_imagenet21k/1",
+        "kaggle_handle": "kaggle://keras/vit/keras/vit_base_patch32_224_imagenet21k/2",
     },
     "vit_huge_patch14_224_imagenet21k": {
         "metadata": {
@@ -99,7 +99,7 @@ backbone_presets = {
             "params": 630764800,
             "path": "vit",
         },
-        "kaggle_handle": "kaggle://keras/vit/keras/vit_huge_patch14_224_imagenet21k/1",
+        "kaggle_handle": "kaggle://keras/vit/keras/vit_huge_patch14_224_imagenet21k/2",
     },
     "vit_large_patch16_224_imagenet21k": {
         "metadata": {
@@ -110,7 +110,7 @@ backbone_presets = {
             "params": 303301632,
             "path": "vit",
         },
-        "kaggle_handle": "kaggle://keras/vit/keras/vit_large_patch16_224_imagenet21k/1",
+        "kaggle_handle": "kaggle://keras/vit/keras/vit_large_patch16_224_imagenet21k/2",
     },
     "vit_large_patch32_224_imagenet21k": {
         "metadata": {
@@ -121,6 +121,6 @@ backbone_presets = {
             "params": 305510400,
             "path": "vit",
         },
-        "kaggle_handle": "kaggle://keras/vit/keras/vit_large_patch32_224_imagenet21k/1",
+        "kaggle_handle": "kaggle://keras/vit/keras/vit_large_patch32_224_imagenet21k/2",
     },
 }

{keras_hub_nightly-0.21.0.dev202505270408 → keras_hub_nightly-0.22.0.dev202505290412}/keras_hub/src/utils/tensor_utils.py

@@ -21,6 +21,20 @@ except ImportError:
 NO_CONVERT_COUNTER = threading.local()
 
 
+def pad(x, shape, padding_side, pad_value):
+    if padding_side == "left":
+        x = x[..., ::-1]
+
+    outputs = x.to_tensor(
+        default_value=pad_value,
+        shape=shape,
+    )
+
+    if padding_side == "left":
+        outputs = outputs[..., ::-1]
+    return outputs
+
+
 @contextlib.contextmanager
 def no_convert_scope():
     try:

{keras_hub_nightly-0.21.0.dev202505270408 → keras_hub_nightly-0.22.0.dev202505290412}/keras_hub/src/version.py

@@ -1,7 +1,7 @@
 from keras_hub.src.api_export import keras_hub_export
 
 # Unique source of truth for the version number.
-__version__ = "0.21.0.dev202505270408"
+__version__ = "0.22.0.dev202505290412"
 
 
 @keras_hub_export("keras_hub.version")

{keras_hub_nightly-0.21.0.dev202505270408 → keras_hub_nightly-0.22.0.dev202505290412}/keras_hub_nightly.egg-info/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: keras-hub-nightly
-Version: 0.21.0.dev202505270408
+Version: 0.22.0.dev202505290412
 Summary: Pretrained models for Keras.
 Author-email: Keras team <keras-users@googlegroups.com>
 License-Expression: Apache-2.0

keras_hub_nightly-0.21.0.dev202505270408/keras_hub/src/models/vit/vit_image_converter.py

@@ -1,78 +0,0 @@
-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):
-        # TODO: Remove this whole function. Why can just use scale and offset
-        # in the base class.
-        x = super().call(inputs)
-        if self.norm_mean:
-            norm_mean = self._expand_non_channel_dims(self.norm_mean, x)
-            x, norm_mean = self._convert_types(x, norm_mean, self.compute_dtype)
-            x = x - norm_mean
-        if self.norm_std:
-            norm_std = self._expand_non_channel_dims(self.norm_std, x)
-            x, norm_std = self._convert_types(x, norm_std, x.dtype)
-            x = x / norm_std
-
-        return x
-
-    def get_config(self):
-        config = super().get_config()
-        config.update(
-            {
-                "norm_mean": self.norm_mean,
-                "norm_std": self.norm_std,
-            }
-        )
-        return config