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

keras_hub/src/models/pali_gemma/pali_gemma_presets.py

@@ -5,14 +5,12 @@ backbone_presets = {
     "pali_gemma_3b_mix_224": {
         "metadata": {
             "description": (
-                "image size 224, mix fine tuned, text sequence " "length is 256"
+                "image size 224, mix fine tuned, text sequence length is 256"
             ),
             "params": 2923335408,
-            "official_name": "PaliGemma",
             "path": "pali_gemma",
-            "model_card": "https://www.kaggle.com/models/google/paligemma",
         },
-        "kaggle_handle": "kaggle://keras/paligemma/keras/pali_gemma_3b_mix_224/2",
+        "kaggle_handle": "kaggle://keras/paligemma/keras/pali_gemma_3b_mix_224/4",
     },
     "pali_gemma_3b_mix_448": {
         "metadata": {
@@ -20,11 +18,9 @@ backbone_presets = {
                 "image size 448, mix fine tuned, text sequence length is 512"
             ),
             "params": 2924220144,
-            "official_name": "PaliGemma",
             "path": "pali_gemma",
-            "model_card": "https://www.kaggle.com/models/google/paligemma",
         },
-        "kaggle_handle": "kaggle://keras/paligemma/keras/pali_gemma_3b_mix_448/2",
+        "kaggle_handle": "kaggle://keras/paligemma/keras/pali_gemma_3b_mix_448/4",
     },
     "pali_gemma_3b_224": {
         "metadata": {
@@ -32,11 +28,9 @@ backbone_presets = {
                 "image size 224, pre trained, text sequence length is 128"
             ),
             "params": 2923335408,
-            "official_name": "PaliGemma",
             "path": "pali_gemma",
-            "model_card": "https://www.kaggle.com/models/google/paligemma",
         },
-        "kaggle_handle": "kaggle://keras/paligemma/keras/pali_gemma_3b_224/2",
+        "kaggle_handle": "kaggle://keras/paligemma/keras/pali_gemma_3b_224/4",
     },
     "pali_gemma_3b_448": {
         "metadata": {
@@ -44,22 +38,184 @@ backbone_presets = {
                 "image size 448, pre trained, text sequence length is 512"
             ),
             "params": 2924220144,
-            "official_name": "PaliGemma",
             "path": "pali_gemma",
-            "model_card": "https://www.kaggle.com/models/google/paligemma",
         },
-        "kaggle_handle": "kaggle://keras/paligemma/keras/pali_gemma_3b_448/2",
+        "kaggle_handle": "kaggle://keras/paligemma/keras/pali_gemma_3b_448/4",
     },
     "pali_gemma_3b_896": {
         "metadata": {
             "description": (
-                "image size 896, pre trained, text sequence length " "is 512"
+                "image size 896, pre trained, text sequence length is 512"
             ),
             "params": 2927759088,
-            "official_name": "PaliGemma",
             "path": "pali_gemma",
-            "model_card": "https://www.kaggle.com/models/google/paligemma",
         },
-        "kaggle_handle": "kaggle://keras/paligemma/keras/pali_gemma_3b_896/2",
+        "kaggle_handle": "kaggle://keras/paligemma/keras/pali_gemma_3b_896/4",
+    },
+    # PaliGemma2
+    "pali_gemma_2_ft_docci_3b_448": {
+        "metadata": {
+            "description": (
+                "3 billion parameter, image size 448, 27-layer for "
+                "SigLIP-So400m vision encoder and 26-layer Gemma2 2B lanuage "
+                "model. This model has been fine-tuned on the DOCCI dataset "
+                "for improved descriptions with fine-grained details."
+            ),
+            "params": 3032979696,
+            "official_name": "PaliGemma2",
+            "path": "pali_gemma2",
+            "model_card": "https://www.kaggle.com/models/google/paligemma-2",
+        },
+        "kaggle_handle": "kaggle://keras/paligemma2/keras/pali_gemma_2_ft_docci_3b_448/2",
+    },
+    "pali_gemma2_ft_docci_10b_448": {
+        "metadata": {
+            "description": (
+                "10 billion parameter, 27-layer for SigLIP-So400m vision "
+                "encoder and 42-layer Gemma2 9B lanuage model. This model has "
+                "been fine-tuned on the DOCCI dataset for improved "
+                "descriptions with fine-grained details."
+            ),
+            "params": 9663294192,
+            "official_name": "PaliGemma2",
+            "path": "pali_gemma2",
+            "model_card": "https://www.kaggle.com/models/google/paligemma-2",
+        },
+        "kaggle_handle": "kaggle://keras/paligemma2/keras/pali_gemma2_ft_docci_10b_448/2",
+    },
+    "pali_gemma2_pt_3b_224": {
+        "metadata": {
+            "description": (
+                "3 billion parameter, image size 224, 27-layer for "
+                "SigLIP-So400m vision encoder and 26-layer Gemma2 2B lanuage "
+                "model. This model has been pre-trained on a mixture of "
+                "datasets."
+            ),
+            "params": 3032094960,
+            "official_name": "PaliGemma2",
+            "path": "pali_gemma2",
+            "model_card": "https://www.kaggle.com/models/google/paligemma-2",
+        },
+        "kaggle_handle": "kaggle://keras/paligemma2/keras/pali_gemma2_pt_3b_224/2",
+    },
+    "pali_gemma2_pt_3b_448": {
+        "metadata": {
+            "description": (
+                "3 billion parameter, image size 448, 27-layer for "
+                "SigLIP-So400m vision encoder and 26-layer Gemma2 2B lanuage "
+                "model. This model has been pre-trained on a mixture of "
+                "datasets."
+            ),
+            "params": 3032979696,
+            "official_name": "PaliGemma2",
+            "path": "pali_gemma2",
+            "model_card": "https://www.kaggle.com/models/google/paligemma-2",
+        },
+        "kaggle_handle": "kaggle://keras/paligemma2/keras/pali_gemma2_pt_3b_448/2",
+    },
+    "pali_gemma2_pt_3b_896": {
+        "metadata": {
+            "description": (
+                "3 billion parameter, image size 896, 27-layer for "
+                "SigLIP-So400m vision encoder and 26-layer Gemma2 2B lanuage "
+                "model. This model has been pre-trained on a mixture of "
+                "datasets."
+            ),
+            "params": 3036518640,
+            "official_name": "PaliGemma2",
+            "path": "pali_gemma2",
+            "model_card": "https://www.kaggle.com/models/google/paligemma-2",
+        },
+        "kaggle_handle": "kaggle://keras/paligemma2/keras/pali_gemma2_pt_3b_896/2",
+    },
+    "pali_gemma2_pt_10b_224": {
+        "metadata": {
+            "description": (
+                "10 billion parameter, image size 224, 27-layer for "
+                "SigLIP-So400m vision encoder and 42-layer Gemma2 9B lanuage "
+                "model. This model has been pre-trained on a mixture of "
+                "datasets."
+            ),
+            "params": 9662409456,
+            "official_name": "PaliGemma2",
+            "path": "pali_gemma2",
+            "model_card": "https://www.kaggle.com/models/google/paligemma-2",
+        },
+        "kaggle_handle": "kaggle://keras/paligemma2/keras/pali_gemma2_pt_10b_224/2",
+    },
+    "pali_gemma2_pt_10b_448": {
+        "metadata": {
+            "description": (
+                "10 billion parameter, image size 448, 27-layer for "
+                "SigLIP-So400m vision encoder and 42-layer Gemma2 9B lanuage "
+                "model. This model has been pre-trained on a mixture of "
+                "datasets."
+            ),
+            "params": 9663294192,
+            "official_name": "PaliGemma2",
+            "path": "pali_gemma2",
+            "model_card": "https://www.kaggle.com/models/google/paligemma-2",
+        },
+        "kaggle_handle": "kaggle://keras/paligemma2/keras/pali_gemma2_pt_10b_448/2",
+    },
+    "pali_gemma2_pt_10b_896": {
+        "metadata": {
+            "description": (
+                "10 billion parameter, image size 896, 27-layer for "
+                "SigLIP-So400m vision encoder and 42-layer Gemma2 9B lanuage "
+                "model. This model has been pre-trained on a mixture of "
+                "datasets."
+            ),
+            "params": 9666833136,
+            "official_name": "PaliGemma2",
+            "path": "pali_gemma2",
+            "model_card": "https://www.kaggle.com/models/google/paligemma-2",
+        },
+        "kaggle_handle": "kaggle://keras/paligemma2/keras/pali_gemma2_pt_10b_896/2",
+    },
+    "pali_gemma2_pt_28b_224": {
+        "metadata": {
+            "description": (
+                "28 billion parameter, image size 224, 27-layer for "
+                "SigLIP-So400m vision encoder and 46-layer Gemma2 27B lanuage "
+                "model. This model has been pre-trained on a mixture of "
+                "datasets."
+            ),
+            "params": 9662409456,
+            "official_name": "PaliGemma2",
+            "path": "pali_gemma2",
+            "model_card": "https://www.kaggle.com/models/google/paligemma-2",
+        },
+        "kaggle_handle": "kaggle://keras/paligemma2/keras/pali_gemma2_pt_28b_224/3",
+    },
+    "pali_gemma2_pt_28b_448": {
+        "metadata": {
+            "description": (
+                "28 billion parameter, image size 448, 27-layer for "
+                "SigLIP-So400m vision encoder and 46-layer Gemma2 27B lanuage "
+                "model. This model has been pre-trained on a mixture of "
+                "datasets."
+            ),
+            "params": 9663294192,
+            "official_name": "PaliGemma2",
+            "path": "pali_gemma2",
+            "model_card": "https://www.kaggle.com/models/google/paligemma-2",
+        },
+        "kaggle_handle": "kaggle://keras/paligemma2/keras/pali_gemma2_pt_28b_448/2",
+    },
+    "pali_gemma2_pt_28b_896": {
+        "metadata": {
+            "description": (
+                "28 billion parameter, image size 896, 27-layer for "
+                "SigLIP-So400m vision encoder and 46-layer Gemma2 27B lanuage "
+                "model. This model has been pre-trained on a mixture of "
+                "datasets."
+            ),
+            "params": 9666833136,
+            "official_name": "PaliGemma2",
+            "path": "pali_gemma2",
+            "model_card": "https://www.kaggle.com/models/google/paligemma-2",
+        },
+        "kaggle_handle": "kaggle://keras/paligemma2/keras/pali_gemma2_pt_28b_896/2",
     },
 }

keras_hub/src/models/pali_gemma/pali_gemma_vit.py

@@ -12,7 +12,7 @@ class PaliGemmaVitEmbeddings(keras.layers.Layer):
         dtype=None,
         **kwargs,
     ):
-        super().__init__(**kwargs)
+        super().__init__(dtype=dtype, **kwargs)
         self.hidden_dim = hidden_dim
         self.image_size = image_size
         self.patch_size = patch_size
@@ -61,7 +61,7 @@ class PaliGemmaVitEmbeddings(keras.layers.Layer):
 
 class PaliGemmaVitAttention(keras.layers.Layer):
     """
-    Adapted from https://github.com/huggingface/transformers/blob/main/src/transformers/models/clip/modeling_clip.py # noqa: E501
+    Adapted from https://github.com/huggingface/transformers/blob/main/src/transformers/models/clip/modeling_clip.py
     """
 
     def __init__(
@@ -72,7 +72,7 @@ class PaliGemmaVitAttention(keras.layers.Layer):
         dtype=None,
         **kwargs,
     ):
-        super().__init__(**kwargs)
+        super().__init__(dtype=dtype, **kwargs)
 
         self.hidden_dim = hidden_dim
         self.num_heads = num_heads
@@ -120,7 +120,7 @@ class PaliGemmaVitAttention(keras.layers.Layer):
 
     def _transpose_for_scores(self, tensor, batch_size):
         """
-        Adapted from https://github.com/huggingface/transformers/blob/8e164c5400b7b413c7b8fb32e35132001effc970/src/transformers/models/bert/modeling_tf_bert.py#L252 # noqa: E501
+        Adapted from https://github.com/huggingface/transformers/blob/8e164c5400b7b413c7b8fb32e35132001effc970/src/transformers/models/bert/modeling_tf_bert.py#L252
         """
         # [batch_size, seq_len, all_head_dim] ->
         # [batch_size, seq_len, num_heads, head_dim]
@@ -282,7 +282,7 @@ class PaliGemmaVitEncoder(keras.layers.Layer):
         dtype=None,
         **kwargs,
     ):
-        super().__init__(**kwargs)
+        super().__init__(dtype=dtype, **kwargs)
         self.hidden_dim = hidden_dim
         self.num_layers = num_layers
         self.num_heads = num_heads
@@ -311,25 +311,26 @@ class PaliGemmaVitEncoder(keras.layers.Layer):
             for i in range(self.num_layers)
         ]
 
-    def build(self, input_shape):
-        self.vision_embeddings.build(input_shape)
+    def build(self, inputs_shape):
+        self.vision_embeddings.build(inputs_shape)
         for block in self.resblocks:
             block.build([None, None, self.hidden_dim])
         self.encoder_layer_norm.build([None, None, self.hidden_dim])
         self.built = True
 
-    def call(
-        self,
-        x,
-        mask=None,
-    ):
-        x = self.vision_embeddings(x)
+    def call(self, inputs, mask=None):
+        x = self.vision_embeddings(inputs)
         for block in self.resblocks:
             x = block(x, mask=mask)
         x = self.encoder_layer_norm(x)
         return x
 
     def compute_output_shape(self, inputs_shape):
+        if inputs_shape is None:
+            # Fix the compatibility issue with Keras 3.1 where
+            # `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]
 
     def get_config(self):
@@ -410,8 +411,6 @@ class PaliGemmaVit(keras.Model):
     Args:
         image_size: int. The height/width of the image. Both height and width is
             expected to be the same.
-        include_rescaling: bool. If true, the image input will be rescaled from
-            the range `[0, 255]`, to the range `[0, 1]`.
         patch_size: int. The size of each square patch in the input image.
         num_heads: int. The number of attention heads for the vision(image)
             transformer encoder.
@@ -452,7 +451,6 @@ class PaliGemmaVit(keras.Model):
         num_layers,
         intermediate_dim,
         num_classes,
-        include_rescaling=True,
         pooling=None,
         classifier_activation=None,
         dtype=None,
@@ -463,14 +461,6 @@ class PaliGemmaVit(keras.Model):
             shape=(image_size, image_size, 3), name="images"
         )
         x = image_input  # Intermediate result.
-        # TODO we have moved this rescaling to preprocessing layers for most
-        # models. We should consider removing it here, though it would break
-        # compatibility.
-        if include_rescaling:
-            rescaling = keras.layers.Rescaling(
-                scale=1.0 / 127.5, offset=-1.0, name="rescaling"
-            )
-            x = rescaling(image_input)
         x = PaliGemmaVitEncoder(
             hidden_dim=hidden_dim,
             num_layers=num_layers,
@@ -520,7 +510,6 @@ class PaliGemmaVit(keras.Model):
         self.pooling = pooling
         self.num_classes = num_classes
         self.image_size = image_size
-        self.include_rescaling = include_rescaling
         self.patch_size = patch_size
         self.classifier_activation = keras.activations.get(
             classifier_activation
@@ -549,7 +538,6 @@ class PaliGemmaVit(keras.Model):
                     self.classifier_activation
                 ),
                 "image_size": self.image_size,
-                "include_rescaling": self.include_rescaling,
                 "patch_size": self.patch_size,
             }
         )

keras_hub/src/models/phi3/phi3_causal_lm.py

@@ -41,7 +41,9 @@ class Phi3CausalLM(CausalLM):
         self.preprocessor = preprocessor
 
         # === Functional Model ===
-        inputs = backbone.inputs
+        # This must be "backbone.input" i.e. the full input structure,
+        # rather than "backbone.inputs" which is the flattened list of inputs.
+        inputs = backbone.input
         hidden_states = backbone(inputs)
         outputs = backbone.token_embedding(hidden_states, reverse=True)
         super().__init__(

keras_hub/src/models/phi3/phi3_decoder.py

@@ -53,7 +53,6 @@ class Phi3Decoder(keras.layers.Layer):
         self.kernel_initializer = keras.initializers.get(kernel_initializer)
 
     def build(self, decoder_sequence_shape):
-
         # Pre-attention layernorm.
         self.pre_attention_layernorm = Phi3LayerNorm(
             epsilon=self.layer_norm_epsilon,

keras_hub/src/models/phi3/phi3_presets.py

@@ -12,11 +12,9 @@ backbone_presets = {
                 "reasoning-dense properties."
             ),
             "params": 3821079552,
-            "official_name": "Phi-3",
             "path": "phi3",
-            "model_card": "https://huggingface.co/microsoft/Phi-3-mini-4k-instruct",
         },
-        "kaggle_handle": "kaggle://keras/phi3/keras/phi3_mini_4k_instruct_en",
+        "kaggle_handle": "kaggle://keras/phi3/keras/phi3_mini_4k_instruct_en/2",
     },
     "phi3_mini_128k_instruct_en": {
         "metadata": {
@@ -28,10 +26,8 @@ backbone_presets = {
                 "reasoning-dense properties."
             ),
             "params": 3821079552,
-            "official_name": "Phi-3",
             "path": "phi3",
-            "model_card": "https://huggingface.co/microsoft/Phi-3-mini-128k-instruct",
         },
-        "kaggle_handle": "kaggle://keras/phi3/keras/phi3_mini_128k_instruct_en",
+        "kaggle_handle": "kaggle://keras/phi3/keras/phi3_mini_128k_instruct_en/2",
     },
 }

keras_hub/src/models/phi3/phi3_rotary_embedding.py

@@ -43,7 +43,7 @@ class Phi3SuScaledRotaryEmbedding(RotaryEmbedding):
         max_sequence_length=4096,
         pretraining_sequence_length=4096,
         max_wavelength=10000,
-        **kwargs
+        **kwargs,
     ):
         super().__init__(max_wavelength=max_wavelength, **kwargs)
         self.max_sequence_length = max_sequence_length

keras_hub/src/models/preprocessor.py

@@ -8,7 +8,7 @@ from keras_hub.src.utils.preset_utils import PREPROCESSOR_CONFIG_FILE
 from keras_hub.src.utils.preset_utils import builtin_presets
 from keras_hub.src.utils.preset_utils import find_subclass
 from keras_hub.src.utils.preset_utils import get_preset_loader
-from keras_hub.src.utils.preset_utils import save_serialized_object
+from keras_hub.src.utils.preset_utils import get_preset_saver
 from keras_hub.src.utils.python_utils import classproperty
 
 
@@ -32,7 +32,7 @@ class Preprocessor(PreprocessingLayer):
     image_converter_cls = None
 
     def __init__(self, *args, **kwargs):
-        self.config_name = kwargs.pop("config_name", PREPROCESSOR_CONFIG_FILE)
+        self.config_file = kwargs.pop("config_file", PREPROCESSOR_CONFIG_FILE)
         super().__init__(*args, **kwargs)
         self._tokenizer = None
         self._image_converter = None
@@ -71,6 +71,22 @@ class Preprocessor(PreprocessingLayer):
     def image_converter(self, value):
         self._image_converter = value
 
+    @property
+    def image_size(self):
+        """Shortcut to get/set the image size of the image converter."""
+        if self.image_converter is None:
+            return None
+        return self.image_converter.image_size
+
+    @image_size.setter
+    def image_size(self, value):
+        if self.image_converter is None:
+            raise ValueError(
+                "Cannot set `image_size` on preprocessor if `image_converter` "
+                " is `None`."
+            )
+        self.image_converter.image_size = value
+
     def get_config(self):
         config = super().get_config()
         if self.tokenizer:
@@ -85,7 +101,7 @@ class Preprocessor(PreprocessingLayer):
             )
         config.update(
             {
-                "config_name": self.config_name,
+                "config_file": self.config_file,
             }
         )
         return config
@@ -117,7 +133,7 @@ class Preprocessor(PreprocessingLayer):
     def from_preset(
         cls,
         preset,
-        config_name=PREPROCESSOR_CONFIG_FILE,
+        config_file=PREPROCESSOR_CONFIG_FILE,
         **kwargs,
     ):
         """Instantiate a `keras_hub.models.Preprocessor` from a model preset.
@@ -145,12 +161,12 @@ class Preprocessor(PreprocessingLayer):
         Examples:
         ```python
         # Load a preprocessor for Gemma generation.
-        preprocessor = keras_hub.models.GemmaCausalLMPreprocessor.from_preset(
+        preprocessor = keras_hub.models.CausalLMPreprocessor.from_preset(
             "gemma_2b_en",
         )
 
         # Load a preprocessor for Bert classification.
-        preprocessor = keras_hub.models.BertTextClassifierPreprocessor.from_preset(
+        preprocessor = keras_hub.models.TextClassifierPreprocessor.from_preset(
             "bert_base_en",
         )
         ```
@@ -167,7 +183,7 @@ class Preprocessor(PreprocessingLayer):
         # Detect the correct subclass if we need to.
         if cls.backbone_cls != backbone_cls:
             cls = find_subclass(preset, cls, backbone_cls)
-        return loader.load_preprocessor(cls, config_name, **kwargs)
+        return loader.load_preprocessor(cls, config_file, **kwargs)
 
     @classmethod
     def _add_missing_kwargs(cls, loader, kwargs):
@@ -209,7 +225,5 @@ class Preprocessor(PreprocessingLayer):
         Args:
             preset_dir: The path to the local model preset directory.
         """
-        save_serialized_object(self, preset_dir, config_file=self.config_name)
-        for layer in self._flatten_layers(include_self=False):
-            if hasattr(layer, "save_to_preset"):
-                layer.save_to_preset(preset_dir)
+        saver = get_preset_saver(preset_dir)
+        saver.save_preprocessor(self)

keras_hub/src/models/resnet/resnet_backbone.py

@@ -51,16 +51,6 @@ class ResNetBackbone(FeaturePyramidBackbone):
             `True` for ResNetV2, `False` for ResNet.
         image_shape: tuple. The input shape without the batch size.
             Defaults to `(None, None, 3)`.
-        pooling: `None` or str. Pooling mode for feature extraction. Defaults
-            to `"avg"`.
-            - `None` means that the output of the model will be the 4D tensor
-                from the last convolutional block.
-            - `avg` means that global average pooling will be applied to the
-                output of the last convolutional block, resulting in a 2D
-                tensor.
-            - `max` means that global max pooling will be applied to the
-                output of the last convolutional block, resulting in a 2D
-                tensor.
         data_format: `None` or str. If specified, either `"channels_last"` or
             `"channels_first"`. The ordering of the dimensions in the
             inputs. `"channels_last"` corresponds to inputs with shape
@@ -75,10 +65,10 @@ class ResNetBackbone(FeaturePyramidBackbone):
 
     Examples:
     ```python
-    input_data = np.random.uniform(0, 255, size=(2, 224, 224, 3))
+    input_data = np.random.uniform(0, 1, size=(2, 224, 224, 3))
 
     # Pretrained ResNet backbone.
-    model = keras_hub.models.ResNetBackbone.from_preset("resnet50")
+    model = keras_hub.models.ResNetBackbone.from_preset("resnet_50_imagenet")
     model(input_data)
 
     # Randomly initialized ResNetV2 backbone with a custom config.
@@ -90,7 +80,6 @@ class ResNetBackbone(FeaturePyramidBackbone):
         stackwise_num_strides=[1, 2, 2],
         block_type="basic_block",
         use_pre_activation=True,
-        pooling="avg",
     )
     model(input_data)
     ```
@@ -188,7 +177,7 @@ class ResNetBackbone(FeaturePyramidBackbone):
                 use_bias=False,
                 padding="same",
                 dtype=dtype,
-                name=f"conv{conv_index+1}_conv",
+                name=f"conv{conv_index + 1}_conv",
             )(x)
 
         if not use_pre_activation: