keras-hub-nightly 0.19.0.dev202503060350__py3-none-any.whl → 0.20.0.dev202503150350__py3-none-any.whl

keras_hub/src/models/cspnet/cspnet_image_classifier.py

@@ -0,0 +1,12 @@
+from keras_hub.src.api_export import keras_hub_export
+from keras_hub.src.models.cspnet.cspnet_backbone import CSPNetBackbone
+from keras_hub.src.models.cspnet.cspnet_image_classifier_preprocessor import (
+    CSPNetImageClassifierPreprocessor,
+)
+from keras_hub.src.models.image_classifier import ImageClassifier
+
+
+@keras_hub_export("keras_hub.models.CSPNetImageClassifier")
+class CSPNetImageClassifier(ImageClassifier):
+    backbone_cls = CSPNetBackbone
+    preprocessor_cls = CSPNetImageClassifierPreprocessor

keras_hub/src/models/cspnet/cspnet_image_classifier_preprocessor.py

@@ -0,0 +1,14 @@
+from keras_hub.src.api_export import keras_hub_export
+from keras_hub.src.models.cspnet.cspnet_backbone import CSPNetBackbone
+from keras_hub.src.models.cspnet.cspnet_image_converter import (
+    CSPNetImageConverter,
+)
+from keras_hub.src.models.image_classifier_preprocessor import (
+    ImageClassifierPreprocessor,
+)
+
+
+@keras_hub_export("keras_hub.models.CSPNetImageClassifierPreprocessor")
+class CSPNetImageClassifierPreprocessor(ImageClassifierPreprocessor):
+    backbone_cls = CSPNetBackbone
+    image_converter_cls = CSPNetImageConverter

keras_hub/src/models/cspnet/cspnet_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.cspnet.cspnet_backbone import CSPNetBackbone
+
+
+@keras_hub_export("keras_hub.layers.CSPNetImageConverter")
+class CSPNetImageConverter(ImageConverter):
+    backbone_cls = CSPNetBackbone

keras_hub/src/models/cspnet/cspnet_presets.py

@@ -0,0 +1,16 @@
+"""CSPNet preset configurations."""
+
+backbone_presets = {
+    "csp_darknet_53_ra_imagenet": {
+        "metadata": {
+            "description": (
+                "A CSP-DarkNet (Cross-Stage-Partial) image classification model"
+                " pre-trained on the Randomly Augmented ImageNet 1k dataset at "
+                "a 224x224 resolution."
+            ),
+            "params": 26652512,
+            "path": "cspnet",
+        },
+        "kaggle_handle": "kaggle://keras/cspdarknet/keras/csp_darknet_53_ra_imagenet/1",
+    },
+}

keras_hub/src/models/gemma/gemma_attention.py

@@ -1,3 +1,5 @@
+import inspect
+
 import keras
 import numpy as np
 from keras import ops
@@ -5,6 +7,7 @@ from keras import ops
 from keras_hub.src.layers.modeling.rotary_embedding import RotaryEmbedding
 from keras_hub.src.utils.keras_utils import clone_initializer
 from keras_hub.src.utils.keras_utils import has_flash_attention_support
+from keras_hub.src.utils.keras_utils import running_on_tpu
 
 
 class CachedGemmaAttention(keras.layers.Layer):
@@ -103,6 +106,18 @@ class CachedGemmaAttention(keras.layers.Layer):
         )
         return x
 
+    def _can_use_flash_attention(self):
+        if not has_flash_attention_support():
+            return False
+        if self.dropout > 0.0:
+            return False
+        if self.logit_soft_cap is None:
+            return True
+        sig = inspect.signature(ops.dot_product_attention)
+        # We can currently only run soft capped attention for keras >= 3.10
+        # and only on TPU.
+        return running_on_tpu() and "attn_logits_soft_cap" in sig.parameters
+
     def _compute_attention(
         self,
         q,
@@ -118,27 +133,23 @@ class CachedGemmaAttention(keras.layers.Layer):
             query_normalization = 1 / np.sqrt(
                 self.hidden_dim // self.num_query_heads
             )
-        use_dot_product_attention = not (
-            self.dropout > 0.0 or (len(q.shape) != 4)
-        )
-        if has_flash_attention_support() and use_dot_product_attention:
-            if self.dropout > 0.0:
-                raise ValueError(
-                    "Flash attention does not support dropout. "
-                    "Please set `dropout` to 0.0."
-                )
+        if self._can_use_flash_attention():
             if attention_mask is not None:
                 attention_mask = ops.expand_dims(attention_mask, axis=1)
                 attention_mask = ops.cast(attention_mask, dtype="bool")
-
-            attention_output = ops.dot_product_attention(
+            # Only pass soft cap if needed as not all keras versions support.
+            if self.logit_soft_cap:
+                kwargs = {"attn_logits_soft_cap": self.logit_soft_cap}
+            else:
+                kwargs = {}
+            return ops.dot_product_attention(
                 query=q,
                 key=k,
                 value=v,
                 mask=attention_mask,
                 scale=query_normalization,
+                **kwargs,
             )
-            return attention_output
 
         q *= ops.cast(query_normalization, dtype=q.dtype)
         q_shape = ops.shape(q)

keras_hub/src/models/mobilenet/mobilenet_backbone.py

@@ -142,6 +142,8 @@ class DepthwiseConvBlock(keras.layers.Layer):
             signal into before reexciting back out. If (>1) technically, it's an
             excite & squeeze layer. If this doesn't exist there is no
             SqueezeExcite layer.
+        residual: bool, default False. True if we want a residual connection. If
+            False, there is no residual connection.
         name: str, name of the layer
         dtype: `None` or str or `keras.mixed_precision.DTypePolicy`. The dtype
             to use for the model's computations and weights.
@@ -161,6 +163,7 @@ class DepthwiseConvBlock(keras.layers.Layer):
         kernel_size=3,
         stride=2,
         squeeze_excite_ratio=None,
+        residual=False,
         name=None,
         dtype=None,
         **kwargs,
@@ -171,6 +174,7 @@ class DepthwiseConvBlock(keras.layers.Layer):
         self.kernel_size = kernel_size
         self.stride = stride
         self.squeeze_excite_ratio = squeeze_excite_ratio
+        self.residual = residual
         self.name = name
 
         channel_axis = (
@@ -256,11 +260,15 @@ class DepthwiseConvBlock(keras.layers.Layer):
         x = self.batch_normalization1(x)
         x = self.activation1(x)
 
-        if self.se_layer:
+        if self.squeeze_excite_ratio:
             x = self.se_layer(x)
 
         x = self.conv2(x)
         x = self.batch_normalization2(x)
+
+        if self.residual:
+            x = x + inputs
+
         return x
 
     def get_config(self):
@@ -272,6 +280,7 @@ class DepthwiseConvBlock(keras.layers.Layer):
                 "kernel_size": self.kernel_size,
                 "stride": self.stride,
                 "squeeze_excite_ratio": self.squeeze_excite_ratio,
+                "residual": self.residual,
                 "name": self.name,
             }
         )
@@ -675,6 +684,8 @@ class MobileNetBackbone(Backbone):
         stackwise_padding,
         output_num_filters,
         depthwise_filters,
+        depthwise_stride,
+        depthwise_residual,
         last_layer_filter,
         squeeze_and_excite=None,
         image_shape=(None, None, 3),
@@ -722,7 +733,9 @@ class MobileNetBackbone(Backbone):
         x = DepthwiseConvBlock(
             input_num_filters,
             depthwise_filters,
+            stride=depthwise_stride,
             squeeze_excite_ratio=squeeze_and_excite,
+            residual=depthwise_residual,
             name="block_0",
             dtype=dtype,
         )(x)
@@ -768,6 +781,8 @@ class MobileNetBackbone(Backbone):
         self.input_num_filters = input_num_filters
         self.output_num_filters = output_num_filters
         self.depthwise_filters = depthwise_filters
+        self.depthwise_stride = depthwise_stride
+        self.depthwise_residual = depthwise_residual
         self.last_layer_filter = last_layer_filter
         self.squeeze_and_excite = squeeze_and_excite
         self.input_activation = input_activation
@@ -790,6 +805,8 @@ class MobileNetBackbone(Backbone):
                 "input_num_filters": self.input_num_filters,
                 "output_num_filters": self.output_num_filters,
                 "depthwise_filters": self.depthwise_filters,
+                "depthwise_stride": self.depthwise_stride,
+                "depthwise_residual": self.depthwise_residual,
                 "last_layer_filter": self.last_layer_filter,
                 "squeeze_and_excite": self.squeeze_and_excite,
                 "input_activation": self.input_activation,

keras_hub/src/models/mobilenet/mobilenet_image_classifier.py

@@ -18,6 +18,7 @@ class MobileNetImageClassifier(ImageClassifier):
         self,
         backbone,
         num_classes,
+        num_features=1024,
         preprocessor=None,
         head_dtype=None,
         **kwargs,
@@ -33,7 +34,7 @@ class MobileNetImageClassifier(ImageClassifier):
         )
 
         self.output_conv = keras.layers.Conv2D(
-            filters=1024,
+            filters=num_features,
             kernel_size=(1, 1),
             strides=(1, 1),
             use_bias=True,
@@ -69,6 +70,7 @@ class MobileNetImageClassifier(ImageClassifier):
 
         # === Config ===
         self.num_classes = num_classes
+        self.num_features = num_features
 
     def get_config(self):
         # Skip ImageClassifier
@@ -76,6 +78,7 @@ class MobileNetImageClassifier(ImageClassifier):
         config.update(
             {
                 "num_classes": self.num_classes,
+                "num_features": self.num_features,
             }
         )
         return config

keras_hub/src/models/mobilenet/mobilenet_presets.py

@@ -4,12 +4,48 @@ backbone_presets = {
     "mobilenet_v3_small_050_imagenet": {
         "metadata": {
             "description": (
-                "Small MobileNet V3 model pre-trained on the ImageNet 1k "
-                "dataset at a 224x224 resolution."
+                "Small Mobilenet V3 model pre-trained on the ImageNet 1k "
+                "dataset at a 224x224 resolution. Has half channel multiplier."
             ),
             "params": 278784,
             "path": "mobilenetv3",
         },
         "kaggle_handle": "kaggle://keras/mobilenetv3/keras/mobilenet_v3_small_050_imagenet/1",
     },
+    "mobilenet_v3_small_100_imagenet": {
+        "metadata": {
+            "description": (
+                "Small Mobilenet V3 model pre-trained on the ImageNet 1k "
+                "dataset at a 224x224 resolution. Has baseline channel "
+                "multiplier."
+            ),
+            "params": 939120,
+            "path": "mobilenetv3",
+        },
+        "kaggle_handle": "kaggle://keras/mobilenetv3/keras/mobilenet_v3_small_100_imagenet/1",
+    },
+    "mobilenet_v3_large_100_imagenet": {
+        "metadata": {
+            "description": (
+                "Large Mobilenet V3 model pre-trained on the ImageNet 1k "
+                "dataset at a 224x224 resolution. Has baseline channel "
+                "multiplier."
+            ),
+            "params": 2996352,
+            "path": "mobilenetv3",
+        },
+        "kaggle_handle": "kaggle://keras/mobilenetv3/keras/mobilenet_v3_large_100_imagenet/1",
+    },
+    "mobilenet_v3_large_100_imagenet_21k": {
+        "metadata": {
+            "description": (
+                "Large Mobilenet V3 model pre-trained on the ImageNet 21k "
+                "dataset at a 224x224 resolution. Has baseline channel "
+                "multiplier."
+            ),
+            "params": 2996352,
+            "path": "mobilenetv3",
+        },
+        "kaggle_handle": "kaggle://keras/mobilenetv3/keras/mobilenet_v3_large_100_imagenet_21k/1",
+    },
 }

keras_hub/src/models/siglip/siglip_presets.py

@@ -10,7 +10,7 @@ backbone_presets = {
             "params": 203156230,
             "official_name": "SigLIP",
             "path": "siglip",
-            "model_card": "https://www.kaggle.com/models/kerashub/siglip",
+            "model_card": "https://huggingface.co/collections/google/siglip-659d5e62f0ae1a57ae0e83ba",
         },
         "kaggle_handle": "kaggle://kerashub/siglip/keras/siglip_base_patch16_224/2",
     },
@@ -22,7 +22,7 @@ backbone_presets = {
             "params": 203202370,
             "official_name": "SigLIP",
             "path": "siglip",
-            "model_card": "https://www.kaggle.com/models/kerashub/siglip",
+            "model_card": "https://huggingface.co/collections/google/siglip-659d5e62f0ae1a57ae0e83ba",
         },
         "kaggle_handle": "kaggle://kerashub/siglip/keras/siglip_base_patch16_256/1",
     },
@@ -34,7 +34,7 @@ backbone_presets = {
             "params": 203448450,
             "official_name": "SigLIP",
             "path": "siglip",
-            "model_card": "https://www.kaggle.com/models/kerashub/siglip",
+            "model_card": "https://huggingface.co/collections/google/siglip-659d5e62f0ae1a57ae0e83ba",
         },
         "kaggle_handle": "kaggle://kerashub/siglip/keras/siglip_base_patch16_384/1",
     },
@@ -46,7 +46,7 @@ backbone_presets = {
             "params": 203792962,
             "official_name": "SigLIP",
             "path": "siglip",
-            "model_card": "https://www.kaggle.com/models/kerashub/siglip",
+            "model_card": "https://huggingface.co/collections/google/siglip-659d5e62f0ae1a57ae0e83ba",
         },
         "kaggle_handle": "kaggle://kerashub/siglip/keras/siglip_base_patch16_512/1",
     },
@@ -58,7 +58,7 @@ backbone_presets = {
             "params": 652151106,
             "official_name": "SigLIP",
             "path": "siglip",
-            "model_card": "https://www.kaggle.com/models/kerashub/siglip",
+            "model_card": "https://huggingface.co/collections/google/siglip-659d5e62f0ae1a57ae0e83ba",
         },
         "kaggle_handle": "kaggle://kerashub/siglip/keras/siglip_large_patch16_256/1",
     },
@@ -70,7 +70,7 @@ backbone_presets = {
             "params": 652479106,
             "official_name": "SigLIP",
             "path": "siglip",
-            "model_card": "https://www.kaggle.com/models/kerashub/siglip",
+            "model_card": "https://huggingface.co/collections/google/siglip-659d5e62f0ae1a57ae0e83ba",
         },
         "kaggle_handle": "kaggle://kerashub/siglip/keras/siglip_large_patch16_384/1",
     },
@@ -83,7 +83,7 @@ backbone_presets = {
             "params": 877360578,
             "official_name": "SigLIP",
             "path": "siglip",
-            "model_card": "https://www.kaggle.com/models/kerashub/siglip",
+            "model_card": "https://huggingface.co/collections/google/siglip-659d5e62f0ae1a57ae0e83ba",
         },
         "kaggle_handle": "kaggle://kerashub/siglip/keras/siglip_so400m_patch14_224/2",
     },
@@ -96,7 +96,7 @@ backbone_presets = {
             "params": 877961291,
             "official_name": "SigLIP",
             "path": "siglip",
-            "model_card": "https://www.kaggle.com/models/kerashub/siglip",
+            "model_card": "https://huggingface.co/collections/google/siglip-659d5e62f0ae1a57ae0e83ba",
         },
         "kaggle_handle": "kaggle://kerashub/siglip/keras/siglip_so400m_patch14_384/1",
     },
@@ -109,7 +109,7 @@ backbone_presets = {
             "params": 1128759282,
             "official_name": "SigLIP",
             "path": "siglip",
-            "model_card": "https://www.kaggle.com/models/kerashub/siglip",
+            "model_card": "https://huggingface.co/collections/google/siglip-659d5e62f0ae1a57ae0e83ba",
         },
         "kaggle_handle": "kaggle://kerashub/siglip/keras/siglip_so400m_patch16_256_i18n/1",
     },
@@ -121,8 +121,204 @@ backbone_presets = {
             "params": 370626370,
             "official_name": "SigLIP",
             "path": "siglip",
-            "model_card": "https://www.kaggle.com/models/kerashub/siglip",
+            "model_card": "https://huggingface.co/collections/google/siglip-659d5e62f0ae1a57ae0e83ba",
         },
         "kaggle_handle": "kaggle://kerashub/siglip/keras/siglip_base_patch16_256_multilingual/1",
     },
+    # SigLIP2.
+    "siglip2_base_patch16_224": {
+        "metadata": {
+            "description": (
+                "375 million parameter, patch size 16, image size 224, "
+                "pre-trained on WebLi."
+            ),
+            "params": 375188230,
+            "official_name": "SigLIP2",
+            "path": "siglip",
+            "model_card": "https://huggingface.co/collections/google/siglip2-67b5dcef38c175486e240107",
+        },
+        "kaggle_handle": "kaggle://kerashub/siglip/keras/siglip2_base_patch16_224/1",
+    },
+    "siglip2_base_patch16_256": {
+        "metadata": {
+            "description": (
+                "375 million parameter, patch size 16, image size 256, "
+                "pre-trained on WebLi."
+            ),
+            "params": 375234370,
+            "official_name": "SigLIP2",
+            "path": "siglip",
+            "model_card": "https://huggingface.co/collections/google/siglip2-67b5dcef38c175486e240107",
+        },
+        "kaggle_handle": "kaggle://kerashub/siglip/keras/siglip2_base_patch16_256/1",
+    },
+    "siglip2_base_patch32_256": {
+        "metadata": {
+            "description": (
+                "376 million parameter, patch size 32, image size 256, "
+                "pre-trained on WebLi."
+            ),
+            "params": 376856194,
+            "official_name": "SigLIP2",
+            "path": "siglip",
+            "model_card": "https://huggingface.co/collections/google/siglip2-67b5dcef38c175486e240107",
+        },
+        "kaggle_handle": "kaggle://kerashub/siglip/keras/siglip2_base_patch32_256/1",
+    },
+    "siglip2_base_patch16_384": {
+        "metadata": {
+            "description": (
+                "376 million parameter, patch size 16, image size 384, "
+                "pre-trained on WebLi."
+            ),
+            "params": 376856194,
+            "official_name": "SigLIP2",
+            "path": "siglip",
+            "model_card": "https://huggingface.co/collections/google/siglip2-67b5dcef38c175486e240107",
+        },
+        "kaggle_handle": "kaggle://kerashub/siglip/keras/siglip2_base_patch16_384/1",
+    },
+    "siglip2_base_patch16_512": {
+        "metadata": {
+            "description": (
+                "375 million parameter, patch size 16, image size 512, "
+                "pre-trained on WebLi."
+            ),
+            "params": 375824962,
+            "official_name": "SigLIP2",
+            "path": "siglip",
+            "model_card": "https://huggingface.co/collections/google/siglip2-67b5dcef38c175486e240107",
+        },
+        "kaggle_handle": "kaggle://kerashub/siglip/keras/siglip2_base_patch16_512/1",
+    },
+    "siglip2_large_patch16_256": {
+        "metadata": {
+            "description": (
+                "881 million parameter, patch size 16, image size 256, "
+                "pre-trained on WebLi."
+            ),
+            "params": 881527106,
+            "official_name": "SigLIP2",
+            "path": "siglip",
+            "model_card": "https://huggingface.co/collections/google/siglip2-67b5dcef38c175486e240107",
+        },
+        "kaggle_handle": "kaggle://kerashub/siglip/keras/siglip2_large_patch16_256/1",
+    },
+    "siglip2_large_patch16_384": {
+        "metadata": {
+            "description": (
+                "881 million parameter, patch size 16, image size 384, "
+                "pre-trained on WebLi."
+            ),
+            "params": 881855106,
+            "official_name": "SigLIP2",
+            "path": "siglip",
+            "model_card": "https://huggingface.co/collections/google/siglip2-67b5dcef38c175486e240107",
+        },
+        "kaggle_handle": "kaggle://kerashub/siglip/keras/siglip2_large_patch16_384/1",
+    },
+    "siglip2_large_patch16_512": {
+        "metadata": {
+            "description": (
+                "882 million parameter, patch size 16, image size 512, "
+                "pre-trained on WebLi."
+            ),
+            "params": 882314306,
+            "official_name": "SigLIP2",
+            "path": "siglip",
+            "model_card": "https://huggingface.co/collections/google/siglip2-67b5dcef38c175486e240107",
+        },
+        "kaggle_handle": "kaggle://kerashub/siglip/keras/siglip2_large_patch16_512/1",
+    },
+    "siglip2_giant_opt_patch16_256": {
+        "metadata": {
+            "description": (
+                "1.8 billion parameter, patch size 16, image size 256, "
+                "pre-trained on WebLi."
+            ),
+            "params": 1871394226,
+            "official_name": "SigLIP2",
+            "path": "siglip",
+            "model_card": "https://huggingface.co/collections/google/siglip2-67b5dcef38c175486e240107",
+        },
+        "kaggle_handle": "kaggle://kerashub/siglip/keras/siglip2_giant_opt_patch16_256/1",
+    },
+    "siglip2_giant_opt_patch16_384": {
+        "metadata": {
+            "description": (
+                "1.8 billion parameter, patch size 16, image size 384, "
+                "pre-trained on WebLi."
+            ),
+            "params": 1871886066,
+            "official_name": "SigLIP2",
+            "path": "siglip",
+            "model_card": "https://huggingface.co/collections/google/siglip2-67b5dcef38c175486e240107",
+        },
+        "kaggle_handle": "kaggle://kerashub/siglip/keras/siglip2_giant_opt_patch16_384/1",
+    },
+    "siglip2_so400m_patch14_224": {
+        "metadata": {
+            "description": (
+                "1.1 billion parameter, patch size 14, image size 224, "
+                "shape-optimized version, pre-trained on WebLi."
+            ),
+            "params": 1135463922,
+            "official_name": "SigLIP2",
+            "path": "siglip",
+            "model_card": "https://huggingface.co/collections/google/siglip2-67b5dcef38c175486e240107",
+        },
+        "kaggle_handle": "kaggle://kerashub/siglip/keras/siglip2_so400m_patch14_224/1",
+    },
+    "siglip2_so400m_patch14_384": {
+        "metadata": {
+            "description": (
+                "1.1 billion parameter, patch size 14, image size 224, "
+                "shape-optimized version, pre-trained on WebLi."
+            ),
+            "params": 1136009291,
+            "official_name": "SigLIP2",
+            "path": "siglip",
+            "model_card": "https://huggingface.co/collections/google/siglip2-67b5dcef38c175486e240107",
+        },
+        "kaggle_handle": "kaggle://kerashub/siglip/keras/siglip2_so400m_patch14_384/1",
+    },
+    "siglip2_so400m_patch16_256": {
+        "metadata": {
+            "description": (
+                "1.1 billion parameter, patch size 16, image size 256, "
+                "shape-optimized version, pre-trained on WebLi."
+            ),
+            "params": 1135671282,
+            "official_name": "SigLIP2",
+            "path": "siglip",
+            "model_card": "https://huggingface.co/collections/google/siglip2-67b5dcef38c175486e240107",
+        },
+        "kaggle_handle": "kaggle://kerashub/siglip/keras/siglip2_so400m_patch16_256/1",
+    },
+    "siglip2_so400m_patch16_384": {
+        "metadata": {
+            "description": (
+                "1.1 billion parameter, patch size 16, image size 384, "
+                "shape-optimized version, pre-trained on WebLi."
+            ),
+            "params": 1136040242,
+            "official_name": "SigLIP2",
+            "path": "siglip",
+            "model_card": "https://huggingface.co/collections/google/siglip2-67b5dcef38c175486e240107",
+        },
+        "kaggle_handle": "kaggle://kerashub/siglip/keras/siglip2_so400m_patch16_384/1",
+    },
+    "siglip2_so400m_patch16_512": {
+        "metadata": {
+            "description": (
+                "1.1 billion parameter, patch size 16, image size 512, "
+                "shape-optimized version, pre-trained on WebLi."
+            ),
+            "params": 1136555698,
+            "official_name": "SigLIP2",
+            "path": "siglip",
+            "model_card": "https://huggingface.co/collections/google/siglip2-67b5dcef38c175486e240107",
+        },
+        "kaggle_handle": "kaggle://kerashub/siglip/keras/siglip2_so400m_patch16_512/1",
+    },
 }

keras_hub/src/models/siglip/siglip_text_encoder.py

@@ -27,6 +27,8 @@ class SigLIPTextEncoder(Backbone):
             Defaults to `1e-6`.
         max_sequence_length: int. The maximum sequence length that this encoder
             can consume. Defaults to `64`.
+        projection_dim: int. The size of the projection in the head. If not
+            specified, set to `hidden_dim`. Defaults to `None`.
         dtype: string or `keras.mixed_precision.DTypePolicy`. The dtype to use
             for the models computations and weights. Note that some
             computations, such as softmax and layer normalization will always
@@ -44,10 +46,12 @@ class SigLIPTextEncoder(Backbone):
         intermediate_activation="gelu_approximate",
         layer_norm_epsilon=1e-6,
         max_sequence_length=64,
+        projection_dim=None,
         dtype=None,
         name=None,
         **kwargs,
     ):
+        projection_dim = projection_dim or hidden_dim
         # `prefix` is used to prevent duplicate name when utilizing multiple
         # SigLIP encoders within a single model.
         prefix = str(name) + "_" if name is not None else ""
@@ -78,7 +82,7 @@ class SigLIPTextEncoder(Backbone):
             name=f"{prefix}post_layer_norm",
         )
         self.head = layers.Dense(
-            hidden_dim,
+            projection_dim,
             kernel_initializer=initializers.LecunNormal(),
             dtype=dtype,
             name=f"{prefix}head",
@@ -115,6 +119,7 @@ class SigLIPTextEncoder(Backbone):
         self.intermediate_activation = intermediate_activation
         self.layer_norm_epsilon = layer_norm_epsilon
         self.max_sequence_length = max_sequence_length
+        self.projection_dim = projection_dim
 
     def get_config(self):
         config = super().get_config()
@@ -129,6 +134,7 @@ class SigLIPTextEncoder(Backbone):
                 "intermediate_activation": self.intermediate_activation,
                 "layer_norm_epsilon": self.layer_norm_epsilon,
                 "max_sequence_length": self.max_sequence_length,
+                "projection_dim": self.projection_dim,
             }
         )
         return config

keras_hub/src/utils/keras_utils.py

@@ -72,3 +72,35 @@ def has_flash_attention_support():
         return True
     else:
         return False
+
+
+def running_on_tpu():
+    backend = keras.config.backend()
+    if backend == "jax":
+        import jax
+
+        devices = jax.devices()
+        return any(d.platform == "tpu" for d in devices)
+    elif backend == "tensorflow":
+        import tensorflow as tf
+
+        return bool(tf.config.list_logical_devices("TPU"))
+    elif backend == "torch":
+        return False
+
+
+def running_on_gpu():
+    backend = keras.config.backend()
+    if backend == "jax":
+        import jax
+
+        devices = jax.devices()
+        return any(d.platform == "gpu" for d in devices)
+    elif backend == "tensorflow":
+        import tensorflow as tf
+
+        return bool(tf.config.list_logical_devices("GPU"))
+    elif backend == "torch":
+        import torch
+
+        return torch.cuda.is_available()

keras_hub/src/utils/preset_utils.py

@@ -622,6 +622,7 @@ class PresetLoader:
             kwargs["preprocessor"] = self.load_preprocessor(
                 cls.preprocessor_cls,
             )
+
         return cls(**kwargs)
 
     def load_preprocessor(