keras-hub-nightly 0.23.0.dev202510150419__py3-none-any.whl → 0.24.0.dev202511080419__py3-none-any.whl

keras_hub/src/models/dinov3/dinov3_presets.py

@@ -0,0 +1,4 @@
+"""DINOV3 model preset configurations."""
+
+# Metadata for loading pretrained model weights.
+backbone_presets = {}

keras_hub/src/models/gemma/gemma_presets.py

@@ -206,4 +206,26 @@ backbone_presets = {
         },
         "kaggle_handle": "kaggle://keras/vaultgemma/keras/vault_gemma_1b_en/2",
     },
+    "c2s_scale_gemma_2_2b_en": {
+        "metadata": {
+            "description": (
+                "A 2 billion parameter, single-cell biology-aware model "
+                "built on the Gemma-2 architecture."
+            ),
+            "params": 2614341888,
+            "path": "gemma",
+        },
+        "kaggle_handle": "kaggle://keras/cell2sentence/keras/c2s_scale_gemma_2_2b_en/1",
+    },
+    "c2s_scale_gemma_2_27b_en": {
+        "metadata": {
+            "description": (
+                "A 27 billion parameter, single-cell biology-aware model "
+                "built on the Gemma-2 architecture."
+            ),
+            "params": 27227128320,
+            "path": "gemma",
+        },
+        "kaggle_handle": "kaggle://keras/cell2sentence/keras/c2s_scale_gemma_2_27b_en/1",
+    },
 }

keras_hub/src/models/gemma3/gemma3_presets.py

@@ -181,4 +181,43 @@ backbone_presets = {
         },
         "kaggle_handle": "kaggle://keras/gemma3/keras/gemma3_instruct_270m/4",
     },
+    "medgemma_instruct_4b": {
+        "metadata": {
+            "description": (
+                "A 4 billion parameter model based on Gemma 3. "
+                "This model is trained for performance on medical text"
+                "and image comprehension and is optimized for medical"
+                "applications that involve a text generation component."
+            ),
+            "params": 4300079472,
+            "path": "gemma3",
+        },
+        "kaggle_handle": "kaggle://keras/medgemma/keras/medgemma_instruct_4b/1",
+    },
+    "medgemma_instruct_27b": {
+        "metadata": {
+            "description": (
+                "A 27 billion parameter model based on Gemma 3. "
+                "This model trained for performance on medical text "
+                "and image comprehension and is optimized for medical "
+                "applications that involve a text generation component."
+            ),
+            "params": 27432406640,
+            "path": "gemma3",
+        },
+        "kaggle_handle": "kaggle://keras/medgemma/keras/medgemma_instruct_27b/1",
+    },
+    "medgemma_instruct_27b_text": {
+        "metadata": {
+            "description": (
+                "A 27 billion parameter text-only model based on Gemma 3. "
+                "This model is trained for performance on medical text "
+                "comprehension and is optimized for medical applications "
+                "that involve a text generation component."
+            ),
+            "params": 27009002240,
+            "path": "gemma3",
+        },
+        "kaggle_handle": "kaggle://keras/medgemma/keras/medgemma_instruct_27b_text/1",
+    },
 }

keras_hub/src/models/image_to_image.py

@@ -415,3 +415,8 @@ class ImageToImage(Task):
         # Image-to-image.
         outputs = [generate(*x) for x in inputs]
         return self._normalize_generate_outputs(outputs, input_is_scalar)
+
+    def _post_quantize(self, mode, **kwargs):
+        super()._post_quantize(mode, **kwargs)
+        # Reset the compiled generate function.
+        self.generate_function = None

keras_hub/src/models/inpaint.py

@@ -518,3 +518,8 @@ class Inpaint(Task):
         # Inpaint.
         outputs = [generate(*x) for x in inputs]
         return self._normalize_generate_outputs(outputs, input_is_scalar)
+
+    def _post_quantize(self, mode, **kwargs):
+        super()._post_quantize(mode, **kwargs)
+        # Reset the compiled generate function.
+        self.generate_function = None

keras_hub/src/models/mobilenetv5/__init__.py

@@ -0,0 +1,9 @@
+from keras_hub.src.models.mobilenetv5.mobilenetv5_backbone import (
+    MobileNetV5Backbone,
+)
+from keras_hub.src.models.mobilenetv5.mobilenetv5_presets import (
+    backbone_presets,
+)
+from keras_hub.src.utils.preset_utils import register_presets
+
+register_presets(backbone_presets, MobileNetV5Backbone)

keras_hub/src/models/mobilenetv5/mobilenetv5_presets.py

@@ -0,0 +1,15 @@
+"""MobileNetV5 preset configurations."""
+
+backbone_presets = {
+    "mobilenetv5_300m_enc_gemma3n": {
+        "metadata": {
+            "description": (
+                "Lightweight 300M-parameter convolutional vision encoder used "
+                "as the image backbone for Gemma 3n"
+            ),
+            "params": 294_284_096,
+            "path": "mobilenetv5",
+        },
+        "kaggle_handle": "kaggle://keras/mobilenetv5/keras/mobilenetv5_300m_enc_gemma3n/1",
+    }
+}

keras_hub/src/models/parseq/__init__.py

@@ -0,0 +1,5 @@
+from keras_hub.src.models.parseq.parseq_backbone import PARSeqBackbone
+from keras_hub.src.models.parseq.parseq_presets import backbone_presets
+from keras_hub.src.utils.preset_utils import register_presets
+
+register_presets(backbone_presets, PARSeqBackbone)

keras_hub/src/models/parseq/parseq_presets.py

@@ -0,0 +1,15 @@
+"""PARSeq preset configurations."""
+
+backbone_presets = {
+    "parseq": {
+        "metadata": {
+            "description": (
+                "Permuted autoregressive sequence (PARSeq) base "
+                "model for scene text recognition"
+            ),
+            "params": 23_832_671,
+            "path": "parseq",
+        },
+        "kaggle_handle": "kaggle://keras/parseq/keras/parseq/1",
+    }
+}

keras_hub/src/models/siglip/siglip_presets.py

@@ -321,4 +321,19 @@ backbone_presets = {
         },
         "kaggle_handle": "kaggle://keras/siglip/keras/siglip2_so400m_patch16_512/1",
     },
+    "medsiglip_900m_448": {
+        "metadata": {
+            "description": (
+                "A 900 million parameter variant of SigLIP trained to encode "
+                "medical images and text into a common embedding space. "
+                "MedSigLIP contains a vision encoder and a text encoder, and "
+                "supports 448x448 image resolution with up to 64 text tokens."
+            ),
+            "params": 878301426,
+            "official_name": "SigLIP2",
+            "path": "siglip",
+            "model_card": "https://huggingface.co/google/medsiglip-448#medsiglip-model-card",
+        },
+        "kaggle_handle": "kaggle://keras/medsiglip/keras/medsiglip_900m_448/1",
+    },
 }

keras_hub/src/models/smollm3/smollm3_backbone.py

@@ -0,0 +1,211 @@
+import keras
+
+from keras_hub.src.api_export import keras_hub_export
+from keras_hub.src.layers.modeling.reversible_embedding import (
+    ReversibleEmbedding,
+)
+from keras_hub.src.models.backbone import Backbone
+from keras_hub.src.models.smollm3.smollm3_layers import SmolLM3DecoderLayer
+
+
+@keras_hub_export(
+    [
+        "keras_hub.models.SmolLM3Backbone",
+        "keras_hub.models.SmolLMBackbone",
+    ]
+)
+class SmolLM3Backbone(Backbone):
+    """SmolLM3 core network with hyperparameters.
+
+    This network implements a Transformer-based decoder network,
+    SmolLM3, as described in the SmolLM3 model architecture.
+    It includes the embedding lookups and transformer layers.
+
+    The default constructor gives a fully customizable, randomly initialized
+    SmolLM3 model with any number of layers, heads, and embedding
+    dimensions. To load preset architectures and weights, use the `from_preset`
+    constructor.
+
+    Args:
+        vocabulary_size: int. The size of the token vocabulary.
+        hidden_dim: int. The size of the transformer hidden state at the end
+            of each transformer layer.
+        intermediate_dim: int. The output dimension of the first Dense layer in
+            the MLP network of each transformer layer.
+        num_layers: int. The number of transformer layers.
+        num_attention_heads: int. The number of attention heads for each
+            transformer layer.
+        num_key_value_heads: int. The number of key-value heads for grouped
+            query attention in each transformer layer.
+        attention_bias: bool. Whether to use bias in the query, key, value, and
+            output projection layers in the attention blocks.
+        attention_dropout: float. Dropout probability for the attention layers.
+        rope_layer_enabled_list: list of bool. List indicating whether RoPE
+            (Rotary Position Embedding) is enabled for each layer. Typically,
+            some layers may disable RoPE for architectural variations.
+        layer_types: list of str. List of layer types for each transformer
+            layer (e.g., "attention" or other custom types).
+        mlp_bias: bool. Whether to use bias in the MLP (feedforward) layers.
+        layer_norm_epsilon: float. Epsilon value for layer normalization layers
+            to prevent division by zero.
+        max_position_embeddings: int. The maximum sequence length that this
+            model might ever be used with.
+        rope_theta: float. The base period of the RoPE embeddings.
+        partial_rotary_factor: float. The percentage of hidden dimensions to
+            rotate in RoPE. A value of 1.0 rotates all dimensions, while values
+            less than 1.0 only rotate a subset.
+
+    Examples:
+
+    ```python
+    input_data = {
+        "token_ids": np.ones(shape=(1, 12), dtype="int32"),
+        "padding_mask": np.array([[1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0]]),
+    }
+
+    # Pretrained SmolLM3 decoder.
+    model = keras_hub.models.SmolLM3Backbone.from_preset(
+        "hf://HuggingFaceTB/SmolLM3-3B"
+    )
+    model(input_data)
+
+    # Randomly initialized SmolLM3 decoder with custom config.
+    model = keras_hub.models.SmolLM3Backbone(
+        vocabulary_size=49152,
+        hidden_dim=576,
+        intermediate_dim=1536,
+        num_layers=30,
+        num_attention_heads=9,
+        num_key_value_heads=3,
+        attention_bias=False,
+        attention_dropout=0.0,
+        rope_layer_enabled_list=[True] * 30,
+        layer_types=["attention"] * 30,
+        mlp_bias=False,
+        layer_norm_epsilon=1e-5,
+        max_position_embeddings=2048,
+        rope_theta=10000.0,
+        partial_rotary_factor=1.0,
+    )
+    model(input_data)
+    ```
+    """
+
+    def __init__(
+        self,
+        vocabulary_size,
+        hidden_dim,
+        intermediate_dim,
+        num_layers,
+        num_attention_heads,
+        num_key_value_heads,
+        attention_bias,
+        attention_dropout,
+        rope_layer_enabled_list,
+        layer_types,
+        mlp_bias,
+        layer_norm_epsilon,
+        max_position_embeddings,
+        rope_theta,
+        partial_rotary_factor,
+        **kwargs,
+    ):
+        # === Layers ===
+        self.token_embedding = ReversibleEmbedding(
+            input_dim=vocabulary_size,
+            output_dim=hidden_dim,
+            name="token_embedding",
+        )
+        self.transformer_layers = []
+        for i in range(num_layers):
+            layer = SmolLM3DecoderLayer(
+                hidden_size=hidden_dim,
+                num_attention_heads=num_attention_heads,
+                num_key_value_heads=num_key_value_heads,
+                attention_bias=attention_bias,
+                attention_dropout=attention_dropout,
+                rope_layer_enabled_list=rope_layer_enabled_list,
+                layer_types=layer_types,
+                layer_idx=i,
+                intermediate_size=intermediate_dim,
+                mlp_bias=mlp_bias,
+                layer_norm_epsilon=layer_norm_epsilon,
+                max_position_embeddings=max_position_embeddings,
+                rope_theta=rope_theta,
+                partial_rotary_factor=partial_rotary_factor,
+                name=f"transformer_layer_{i}",
+            )
+            self.transformer_layers.append(layer)
+
+        self.norm = keras.layers.RMSNormalization(
+            epsilon=layer_norm_epsilon,
+            name="sequence_output_layernorm",
+        )
+
+        # === Functional Model ===
+        token_id_input = keras.Input(
+            shape=(None,), dtype="int32", name="token_ids"
+        )
+
+        padding_mask_input = keras.Input(
+            shape=(None,), dtype="int32", name="padding_mask"
+        )
+
+        x = self.token_embedding(token_id_input)
+
+        for decoder_layer in self.transformer_layers:
+            x = decoder_layer(
+                x,
+                decoder_padding_mask=padding_mask_input,
+                **kwargs,
+            )
+
+        sequence_output = self.norm(x)
+        super().__init__(
+            inputs={
+                "token_ids": token_id_input,
+                "padding_mask": padding_mask_input,
+            },
+            outputs=sequence_output,
+            **kwargs,
+        )
+
+        # === Config ===
+        self.vocabulary_size = vocabulary_size
+        self.hidden_dim = hidden_dim
+        self.intermediate_dim = intermediate_dim
+        self.num_layers = num_layers
+        self.num_attention_heads = num_attention_heads
+        self.num_key_value_heads = num_key_value_heads
+        self.attention_bias = attention_bias
+        self.attention_dropout = attention_dropout
+        self.rope_layer_enabled_list = rope_layer_enabled_list
+        self.layer_types = layer_types
+        self.mlp_bias = mlp_bias
+        self.layer_norm_epsilon = layer_norm_epsilon
+        self.max_position_embeddings = max_position_embeddings
+        self.rope_theta = rope_theta
+        self.partial_rotary_factor = partial_rotary_factor
+
+    def get_config(self):
+        config = super().get_config()
+        config.update(
+            {
+                "vocabulary_size": self.vocabulary_size,
+                "hidden_dim": self.hidden_dim,
+                "intermediate_dim": self.intermediate_dim,
+                "num_layers": self.num_layers,
+                "num_attention_heads": self.num_attention_heads,
+                "num_key_value_heads": self.num_key_value_heads,
+                "attention_bias": self.attention_bias,
+                "attention_dropout": self.attention_dropout,
+                "rope_layer_enabled_list": self.rope_layer_enabled_list,
+                "layer_types": self.layer_types,
+                "mlp_bias": self.mlp_bias,
+                "layer_norm_epsilon": self.layer_norm_epsilon,
+                "max_position_embeddings": self.max_position_embeddings,
+                "rope_theta": self.rope_theta,
+                "partial_rotary_factor": self.partial_rotary_factor,
+            }
+        )
+        return config