Fix PaliGemmaVitEncoder output shape

keras_hub/api/layers/__init__.py

@@ -70,6 +70,9 @@
 from keras_hub.src.models.segformer.segformer_image_converter import (
     SegFormerImageConverter,
 )
+from keras_hub.src.models.siglip.siglip_image_converter import (
+    SigLIPImageConverter,
+)
 from keras_hub.src.models.vgg.vgg_image_converter import VGGImageConverter
 from keras_hub.src.models.vit.vit_image_converter import ViTImageConverter
 from keras_hub.src.models.whisper.whisper_audio_converter import (

keras_hub/api/models/__init__.py

@@ -312,6 +312,13 @@
 )
 from keras_hub.src.models.seq_2_seq_lm import Seq2SeqLM
 from keras_hub.src.models.seq_2_seq_lm_preprocessor import Seq2SeqLMPreprocessor
+from keras_hub.src.models.siglip.siglip_backbone import SigLIPBackbone
+from keras_hub.src.models.siglip.siglip_preprocessor import SigLIPPreprocessor
+from keras_hub.src.models.siglip.siglip_text_encoder import SigLIPTextEncoder
+from keras_hub.src.models.siglip.siglip_tokenizer import SigLIPTokenizer
+from keras_hub.src.models.siglip.siglip_vision_encoder import (
+    SigLIPVisionEncoder,
+)
 from keras_hub.src.models.stable_diffusion_3.stable_diffusion_3_backbone import (
     StableDiffusion3Backbone,
 )

keras_hub/api/tokenizers/__init__.py

@@ -30,6 +30,7 @@
 )
 from keras_hub.src.models.phi3.phi3_tokenizer import Phi3Tokenizer
 from keras_hub.src.models.roberta.roberta_tokenizer import RobertaTokenizer
+from keras_hub.src.models.siglip.siglip_tokenizer import SigLIPTokenizer
 from keras_hub.src.models.t5.t5_tokenizer import T5Tokenizer
 from keras_hub.src.models.whisper.whisper_tokenizer import WhisperTokenizer
 from keras_hub.src.models.xlm_roberta.xlm_roberta_tokenizer import (

keras_hub/src/layers/preprocessing/image_converter.py

@@ -1,6 +1,7 @@
 import math
 
 import keras
+import ml_dtypes
 import numpy as np
 from keras import ops
 
@@ -18,6 +19,95 @@
 from keras_hub.src.utils.tensor_utils import preprocessing_function
 
 
+# TODO: Use `keras.layers.Resizing` once `antialias` is configurable.
+# https://github.com/keras-team/keras/pull/20972
+def _saturate_cast(x, dtype, backend_module):
+    def get_dtype_min_max(dtype):
+        if "bool" == dtype:
+            dtype_min = 0
+            dtype_max = 1
+        elif "int" in dtype:
+            dtype_min = ml_dtypes.iinfo(dtype).min
+            dtype_max = ml_dtypes.iinfo(dtype).max
+        else:
+            dtype_min = ml_dtypes.finfo(dtype).min
+            dtype_max = ml_dtypes.finfo(dtype).max
+        return dtype_min, dtype_max
+
+    dtype = keras.backend.standardize_dtype(dtype)
+    in_dtype = keras.backend.standardize_dtype(x.dtype)
+    in_min, in_max = get_dtype_min_max(in_dtype)
+    out_min, out_max = get_dtype_min_max(dtype)
+
+    min_limit = np.maximum(in_min, out_min).astype(in_dtype)
+    if min_limit < out_min:
+        min_limit = np.nextafter(min_limit, 0, dtype=in_dtype)
+    max_limit = np.minimum(in_max, out_max).astype(in_dtype)
+    if max_limit > out_max:
+        max_limit = np.nextafter(max_limit, 0, dtype=in_dtype)
+
+    x = backend_module.numpy.clip(x, min_limit, max_limit)
+    return backend_module.cast(x, dtype)
+
+
+class ResizingAntialiasConfigurable(keras.layers.Resizing):
+    """A preprocessing layer which resizes images.
+
+    This class is the same as `keras.layers.Resizing` but exposes `antialias` as
+    a configurable parameter.
+    """
+
+    def __init__(
+        self,
+        height,
+        width,
+        interpolation="bilinear",
+        antialias=False,
+        crop_to_aspect_ratio=False,
+        pad_to_aspect_ratio=False,
+        fill_mode="constant",
+        fill_value=0.0,
+        data_format=None,
+        **kwargs,
+    ):
+        super().__init__(
+            height=height,
+            width=width,
+            interpolation=interpolation,
+            crop_to_aspect_ratio=crop_to_aspect_ratio,
+            pad_to_aspect_ratio=pad_to_aspect_ratio,
+            fill_mode=fill_mode,
+            fill_value=fill_value,
+            data_format=data_format,
+            **kwargs,
+        )
+        self.antialias = bool(antialias)
+
+    def transform_images(self, images, transformation=None, training=True):
+        size = (self.height, self.width)
+        resized = self.backend.image.resize(
+            images,
+            size=size,
+            interpolation=self.interpolation,
+            antialias=self.antialias,  # Added.
+            data_format=self.data_format,
+            crop_to_aspect_ratio=self.crop_to_aspect_ratio,
+            pad_to_aspect_ratio=self.pad_to_aspect_ratio,
+            fill_mode=self.fill_mode,
+            fill_value=self.fill_value,
+        )
+        if resized.dtype == images.dtype:
+            return resized
+        if keras.backend.is_int_dtype(images.dtype):
+            resized = self.backend.numpy.round(resized)
+        return _saturate_cast(resized, images.dtype, self.backend)
+
+    def get_config(self):
+        config = super().get_config()
+        config.update({"antialias": self.antialias})
+        return config
+
+
 @keras_hub_export("keras_hub.layers.ImageConverter")
 class ImageConverter(PreprocessingLayer):
     """Preprocess raw images into model ready inputs.
@@ -65,6 +155,8 @@ class ImageConverter(PreprocessingLayer):
         interpolation: String, the interpolation method.
             Supports `"bilinear"`, `"nearest"`, `"bicubic"`,
             `"lanczos3"`, `"lanczos5"`. Defaults to `"bilinear"`.
+        antialias: Whether to use an antialiasing filter when downsampling an
+            image. Defaults to `False`.
         bounding_box_format: A string specifying the format of the bounding
             boxes, one of `"xyxy"`, `"rel_xyxy"`, `"xywh"`, `"center_xywh"`,
             `"yxyx"`, `"rel_yxyx"`. Specifies the format of the bounding boxes
@@ -107,6 +199,7 @@ def __init__(
         crop_to_aspect_ratio=True,
         pad_to_aspect_ratio=False,
         interpolation="bilinear",
+        antialias=False,
         bounding_box_format="yxyx",
         data_format=None,
         **kwargs,
@@ -132,12 +225,13 @@ def __init__(
         resizing_kwargs = {}
         if check_bounding_box_support():
             resizing_kwargs["bounding_box_format"] = bounding_box_format
-        self.resizing = keras.layers.Resizing(
+        self.resizing = ResizingAntialiasConfigurable(
             height=image_size[0] if image_size else None,
             width=image_size[1] if image_size else None,
             crop_to_aspect_ratio=crop_to_aspect_ratio,
             pad_to_aspect_ratio=pad_to_aspect_ratio,
             interpolation=interpolation,
+            antialias=antialias,
             data_format=data_format,
             dtype=self.dtype_policy,
             name="resizing",
@@ -148,6 +242,7 @@ def __init__(
         self.crop_to_aspect_ratio = crop_to_aspect_ratio
         self.pad_to_aspect_ratio = pad_to_aspect_ratio
         self.interpolation = interpolation
+        self.antialias = antialias
         self.bounding_box_format = bounding_box_format
         self.data_format = standardize_data_format(data_format)
 
@@ -211,6 +306,7 @@ def get_config(self):
                 "scale": self.scale,
                 "offset": self.offset,
                 "interpolation": self.interpolation,
+                "antialias": self.antialias,
                 "crop_to_aspect_ratio": self.crop_to_aspect_ratio,
                 "pad_to_aspect_ratio": self.pad_to_aspect_ratio,
                 "bounding_box_format": self.bounding_box_format,

keras_hub/src/models/pali_gemma/pali_gemma_vit.py

@@ -330,7 +330,11 @@ def compute_output_shape(self, inputs_shape):
             # `compute_output_spec` fails to propagate `inputs_shape`
             # correctly, causing it to be `None`.
             inputs_shape = [None, None, None]
-        return [inputs_shape[0], inputs_shape[1], self.hidden_dim]
+        return [
+            inputs_shape[0],
+            (inputs_shape[1] // self.patch_size) ** 2,
+            self.hidden_dim,
+        ]
 
     def get_config(self):
         config = super().get_config()

keras_hub/src/models/siglip/__init__.py

@@ -0,0 +1,5 @@
+from keras_hub.src.models.siglip.siglip_backbone import SigLIPBackbone
+from keras_hub.src.models.siglip.siglip_presets import backbone_presets
+from keras_hub.src.utils.preset_utils import register_presets
+
+register_presets(backbone_presets, SigLIPBackbone)