keras-hub 0.25.0.dev0__py3-none-any.whl → 0.26.0.dev0__py3-none-any.whl

keras_hub/src/models/sam3/sam3_vision_encoder.py

@@ -0,0 +1,738 @@
+import numpy as np
+from keras import layers
+from keras import ops
+
+from keras_hub.src.api_export import keras_hub_export
+from keras_hub.src.models.sam3.sam3_layers import SAM3MLP
+from keras_hub.src.models.sam3.sam3_layers import SAM3Embedding
+from keras_hub.src.models.sam3.sam3_layers import SAM3RoPEAttention
+from keras_hub.src.models.sam3.sam3_layers import SAM3SinePositionEmbedding
+from keras_hub.src.models.sam3.sam3_utils import window_partition
+from keras_hub.src.models.sam3.sam3_utils import window_unpartition
+
+
+class SAM3ViTRotaryEmbedding(layers.Layer):
+    def __init__(self, rope_theta, head_dim, end_x, end_y, scale=1.0, **kwargs):
+        super().__init__(**kwargs)
+        self.rope_theta = float(rope_theta)
+        self.head_dim = int(head_dim)
+        self.end_x = int(end_x)
+        self.end_y = int(end_y)
+        self.scale = float(scale)
+
+        # Ensure even dimension for proper axial splitting.
+        if self.head_dim % 4 != 0:
+            raise ValueError("Dimension must be divisible by 4 for axial RoPE")
+
+    def build(self, input_shape):
+        freqs = 1.0 / (
+            self.rope_theta
+            ** (
+                np.arange(0, self.head_dim, 4)[: (self.head_dim // 4)]
+                / self.head_dim
+            )
+        )
+        flattened_indices = np.arange(self.end_x * self.end_y, dtype=np.int64)
+        x_positions = (flattened_indices % self.end_x) * self.scale
+        y_positions = (
+            np.floor_divide(flattened_indices, self.end_x) * self.scale
+        )
+        freqs_x = np.outer(x_positions, freqs).astype(np.float32)
+        freqs_y = np.outer(y_positions, freqs).astype(np.float32)
+        inv_freq = np.concatenate([freqs_x, freqs_y], axis=-1)
+        inv_freq = np.repeat(inv_freq, repeats=2, axis=-1)
+        rope_embeddings_cos = np.cos(inv_freq)
+        rope_embeddings_sin = np.sin(inv_freq)
+        self.rope_embeddings_cos = self.add_weight(
+            name="rope_embeddings_cos",
+            shape=rope_embeddings_cos.shape,
+            dtype=self.variable_dtype,
+            trainable=False,
+            initializer=rope_embeddings_cos,
+        )
+        self.rope_embeddings_sin = self.add_weight(
+            name="rope_embeddings_sin",
+            shape=rope_embeddings_sin.shape,
+            dtype=self.variable_dtype,
+            trainable=False,
+            initializer=rope_embeddings_sin,
+        )
+
+    def call(self, inputs):
+        return self.rope_embeddings_cos, self.rope_embeddings_sin
+
+    def get_config(self):
+        config = super().get_config()
+        config.update(
+            {
+                "rope_theta": self.rope_theta,
+                "head_dim": self.head_dim,
+                "end_x": self.end_x,
+                "end_y": self.end_y,
+                "scale": self.scale,
+            }
+        )
+        return config
+
+    def compute_output_shape(self, input_shape):
+        embedding_shape = (self.end_x * self.end_y, self.head_dim)
+        return (embedding_shape, embedding_shape)
+
+    def load_own_variables(self, store):
+        try:
+            return super().load_own_variables(store)
+        except ValueError:
+            # `SAM3ViTRotaryEmbedding` has precomputed weights only. The issue
+            # of the loading logic could be ignored.
+            pass
+
+
+class SAM3ViTLayer(layers.Layer):
+    def __init__(
+        self,
+        image_shape,
+        patch_size,
+        hidden_dim,
+        intermediate_dim,
+        num_heads,
+        hidden_activation="gelu",
+        rope_theta=10000.0,
+        window_size=0,
+        rotary_scale=1.0,
+        attention_dropout_rate=0.0,
+        hidden_dropout_rate=0.0,
+        layer_norm_epsilon=1e-6,
+        **kwargs,
+    ):
+        super().__init__(**kwargs)
+        self.image_shape = (
+            int(image_shape[0]),
+            int(image_shape[1]),
+            int(image_shape[2]),
+        )
+        self.patch_size = int(patch_size)
+        self.hidden_dim = int(hidden_dim)
+        self.intermediate_dim = int(intermediate_dim)
+        self.num_heads = int(num_heads)
+        self.hidden_activation = hidden_activation
+        self.rope_theta = float(rope_theta)
+        self.window_size = int(window_size)
+        self.rotary_scale = float(rotary_scale)
+        self.hidden_dropout_rate = float(hidden_dropout_rate)
+        self.attention_dropout_rate = float(attention_dropout_rate)
+        self.layer_norm_epsilon = float(layer_norm_epsilon)
+        self.head_dim = self.hidden_dim // self.num_heads
+        input_size = (
+            self.image_shape[0] // self.patch_size,
+            self.image_shape[1] // self.patch_size,
+        )
+        if self.window_size > 0 and (
+            input_size[0] % self.window_size != 0
+            or input_size[1] % self.window_size != 0
+        ):
+            raise ValueError(
+                "Image size must be divisible by `patch_size` and "
+                "`window_size` for windowed attention. "
+                f"Received image size: {image_shape}, "
+                f"patch_size: {patch_size}, window_size: {window_size}"
+            )
+        rotary_input_size = (
+            input_size if window_size == 0 else (window_size, window_size)
+        )
+
+        self.layer_norm1 = layers.LayerNormalization(
+            epsilon=layer_norm_epsilon,
+            dtype=self.dtype_policy,
+            name="layer_norm1",
+        )
+        self.rotary_emb = SAM3ViTRotaryEmbedding(
+            rope_theta=rope_theta,
+            head_dim=self.head_dim,
+            end_x=rotary_input_size[0],
+            end_y=rotary_input_size[1],
+            scale=self.rotary_scale,
+            dtype=self.dtype_policy,
+            name="rotary_emb",
+        )
+        self.attention = SAM3RoPEAttention(
+            hidden_dim=self.hidden_dim,
+            num_heads=self.num_heads,
+            attention_dropout_rate=self.attention_dropout_rate,
+            dtype=self.dtype_policy,
+            name="attention",
+        )
+        self.layer_norm2 = layers.LayerNormalization(
+            epsilon=layer_norm_epsilon,
+            dtype=self.dtype_policy,
+            name="layer_norm2",
+        )
+        self.mlp = SAM3MLP(
+            hidden_dim=self.hidden_dim,
+            intermediate_dim=self.intermediate_dim,
+            activation=self.hidden_activation,
+            dropout_rate=self.hidden_dropout_rate,
+            dtype=self.dtype_policy,
+            name="mlp",
+        )
+        self.dropout = layers.Dropout(
+            self.hidden_dropout_rate, dtype=self.dtype_policy, name="dropout"
+        )
+
+    def build(self, input_shape):
+        self.input_hidden_dim = int(input_shape[-1])
+        self.layer_norm1.build(input_shape)
+        input_shape = self.layer_norm1.compute_output_shape(input_shape)
+        self.rotary_emb.build(input_shape)
+        input_shape_before_attention = input_shape
+        if self.window_size > 0:
+            input_shape = list(input_shape)
+            input_shape = (
+                None,
+                self.window_size,
+                self.window_size,
+                input_shape[-1],
+            )
+        self.attention.build(input_shape)
+        input_shape = self.attention.compute_output_shape(input_shape)
+        if self.window_size > 0:
+            input_shape = input_shape_before_attention
+        self.layer_norm2.build(input_shape)
+        self.mlp.build(input_shape)
+        self.dropout.build(input_shape)
+
+    def call(self, hidden_states, training=None):
+        residual = hidden_states
+        hidden_states = self.layer_norm1(hidden_states, training=training)
+        if self.window_size > 0:
+            height, width = (
+                self.image_shape[0] // self.patch_size,
+                self.image_shape[1] // self.patch_size,
+            )
+            # Partition into non-overlapping windows for efficient attention.
+            hidden_states = window_partition(
+                hidden_states,
+                height,
+                width,
+                self.window_size,
+                self.input_hidden_dim,
+            )
+
+        position_embeddings = self.rotary_emb(hidden_states, training=training)
+        hidden_states = self.attention(
+            hidden_states, position_embeddings, training=training
+        )
+        if self.window_size > 0:
+            # Reverse window partition to restore original spatial layout.
+            hidden_states = window_unpartition(
+                hidden_states, height, width, self.window_size, self.hidden_dim
+            )
+        hidden_states = ops.add(residual, hidden_states)
+        residual = hidden_states
+        hidden_states = self.layer_norm2(hidden_states, training=training)
+        hidden_states = self.mlp(hidden_states, training=training)
+        hidden_states = ops.add(
+            residual, self.dropout(hidden_states, training=training)
+        )
+        return hidden_states
+
+    def get_config(self):
+        config = super().get_config()
+        config.update(
+            {
+                "image_shape": self.image_shape,
+                "patch_size": self.patch_size,
+                "hidden_dim": self.hidden_dim,
+                "intermediate_dim": self.intermediate_dim,
+                "num_heads": self.num_heads,
+                "hidden_activation": self.hidden_activation,
+                "rope_theta": self.rope_theta,
+                "window_size": self.window_size,
+                "rotary_scale": self.rotary_scale,
+                "attention_dropout_rate": self.attention_dropout_rate,
+                "hidden_dropout_rate": self.hidden_dropout_rate,
+                "layer_norm_epsilon": self.layer_norm_epsilon,
+            }
+        )
+        return config
+
+    def compute_output_shape(self, input_shape):
+        return input_shape
+
+
+class SAM3ViTEncoder(layers.Layer):
+    def __init__(
+        self,
+        image_shape,
+        patch_size,
+        num_layers,
+        hidden_dim,
+        intermediate_dim,
+        num_heads,
+        pretrain_image_shape=(336, 336, 3),
+        hidden_activation="gelu",
+        rope_theta=100000.0,
+        window_size=0,
+        global_attn_indexes=None,
+        attention_dropout_rate=0.0,
+        hidden_dropout_rate=0.0,
+        layer_norm_epsilon=1e-6,
+        **kwargs,
+    ):
+        super().__init__(**kwargs)
+        self.image_shape = (
+            int(image_shape[0]),
+            int(image_shape[1]),
+            int(image_shape[2]),
+        )
+        self.patch_size = int(patch_size)
+        self.num_layers = int(num_layers)
+        self.hidden_dim = int(hidden_dim)
+        self.intermediate_dim = int(intermediate_dim)
+        self.num_heads = int(num_heads)
+        self.hidden_activation = hidden_activation
+        self.rope_theta = float(rope_theta)
+        self.window_size = int(window_size)
+        if global_attn_indexes is not None:
+            self.global_attn_indexes = [int(i) for i in global_attn_indexes]
+        else:
+            self.global_attn_indexes = None
+        self.pretrain_image_shape = (
+            int(pretrain_image_shape[0]),
+            int(pretrain_image_shape[1]),
+            int(pretrain_image_shape[2]),
+        )
+        self.hidden_dropout_rate = float(hidden_dropout_rate)
+        self.attention_dropout_rate = float(attention_dropout_rate)
+        self.layer_norm_epsilon = float(layer_norm_epsilon)
+        height = self.image_shape[0] // self.patch_size
+
+        self.embeddings = SAM3Embedding(
+            hidden_dim=self.hidden_dim,
+            patch_size=self.patch_size,
+            image_shape=self.image_shape,
+            dropout_rate=self.hidden_dropout_rate,
+            pretrain_image_shape=self.pretrain_image_shape,
+            dtype=self.dtype_policy,
+            name="embeddings",
+        )
+        self.layer_norm = layers.LayerNormalization(
+            epsilon=self.layer_norm_epsilon,
+            dtype=self.dtype_policy,
+            name="layer_norm",
+        )
+        self.layers = [
+            SAM3ViTLayer(
+                image_shape=self.image_shape,
+                patch_size=self.patch_size,
+                hidden_dim=self.hidden_dim,
+                intermediate_dim=self.intermediate_dim,
+                num_heads=self.num_heads,
+                hidden_activation=self.hidden_activation,
+                rope_theta=self.rope_theta,
+                window_size=(
+                    self.window_size if i not in self.global_attn_indexes else 0
+                ),
+                rotary_scale=(
+                    1.0
+                    if i not in self.global_attn_indexes
+                    else float(self.window_size) / height
+                ),
+                attention_dropout_rate=self.attention_dropout_rate,
+                hidden_dropout_rate=self.hidden_dropout_rate,
+                layer_norm_epsilon=self.layer_norm_epsilon,
+                dtype=self.dtype_policy,
+                name=f"layer_{i}",
+            )
+            for i in range(self.num_layers)
+        ]
+
+    def build(self, input_shape):
+        self.embeddings.build(input_shape)
+        input_shape = self.embeddings.compute_output_shape(input_shape)
+        input_shape = list(input_shape)
+        height = self.image_shape[0] // self.patch_size
+        width = self.image_shape[1] // self.patch_size
+        input_shape = [input_shape[0], height, width, self.hidden_dim]
+        self.layer_norm.build(input_shape)
+        for layer in self.layers:
+            layer.build(input_shape)
+
+    def call(self, pixel_values, training=None):
+        hidden_states = self.embeddings(pixel_values, training=training)
+        height = self.image_shape[0] // self.patch_size
+        width = self.image_shape[1] // self.patch_size
+        # Reshape to spatial format for windowed attention:
+        # [batch_size, height, width, hidden_size]
+        hidden_states = ops.reshape(
+            hidden_states, (-1, height, width, self.hidden_dim)
+        )
+        hidden_states = self.layer_norm(hidden_states, training=training)
+        for i, layer in enumerate(self.layers):
+            hidden_states = layer(hidden_states, training=training)
+
+        # Reshape back to sequence format:
+        # [batch_size, height*width, hidden_size]
+        return ops.reshape(hidden_states, (-1, height * width, self.hidden_dim))
+
+    def get_config(self):
+        config = super().get_config()
+        config.update(
+            {
+                "image_shape": self.image_shape,
+                "patch_size": self.patch_size,
+                "num_layers": self.num_layers,
+                "hidden_dim": self.hidden_dim,
+                "intermediate_dim": self.intermediate_dim,
+                "num_heads": self.num_heads,
+                "pretrain_image_shape": self.pretrain_image_shape,
+                "hidden_activation": self.hidden_activation,
+                "rope_theta": self.rope_theta,
+                "window_size": self.window_size,
+                "global_attn_indexes": self.global_attn_indexes,
+                "attention_dropout_rate": self.attention_dropout_rate,
+                "hidden_dropout_rate": self.hidden_dropout_rate,
+                "layer_norm_epsilon": self.layer_norm_epsilon,
+            }
+        )
+        return config
+
+    def compute_output_shape(self, input_shape):
+        input_shape = self.embeddings.compute_output_shape(input_shape)
+        return input_shape
+
+
+class SAM3FPNLayer(layers.Layer):
+    def __init__(self, input_dim, fpn_dim, scale_factor, **kwargs):
+        super().__init__(**kwargs)
+        self.input_dim = int(input_dim)
+        self.fpn_dim = int(fpn_dim)
+        self.scale_factor = float(scale_factor)
+
+        # Build the upsampling/downsampling layers based on scale factor.
+        if self.scale_factor == 4.0:
+            self.scale_layers = [
+                layers.Conv2DTranspose(
+                    self.input_dim // 2,
+                    kernel_size=2,
+                    strides=2,
+                    dtype=self.dtype_policy,
+                    name="scale_layers_0",
+                ),
+                layers.Activation(
+                    "gelu", dtype=self.dtype_policy, name="scale_layers_1"
+                ),
+                layers.Conv2DTranspose(
+                    self.input_dim // 4,
+                    kernel_size=2,
+                    strides=2,
+                    dtype=self.dtype_policy,
+                    name="scale_layers_2",
+                ),
+            ]
+        elif self.scale_factor == 2.0:
+            self.scale_layers = [
+                layers.Conv2DTranspose(
+                    self.input_dim // 2,
+                    kernel_size=2,
+                    strides=2,
+                    dtype=self.dtype_policy,
+                    name="scale_layers_0",
+                )
+            ]
+        elif self.scale_factor == 1.0:
+            self.scale_layers = []
+        elif self.scale_factor == 0.5:
+            self.scale_layers = [
+                layers.MaxPooling2D(
+                    pool_size=2,
+                    strides=2,
+                    dtype=self.dtype_policy,
+                    name="scale_layers_0",
+                )
+            ]
+        else:
+            raise ValueError(
+                f"Unsupported scale factor: {self.scale_factor}. "
+                "Supported scale factors are 4.0, 2.0, 1.0, and 0.5."
+            )
+        self.proj1 = layers.Conv2D(
+            self.fpn_dim, kernel_size=1, dtype=self.dtype_policy, name="proj1"
+        )
+        self.pad = layers.ZeroPadding2D(
+            padding=1, dtype=self.dtype_policy, name="pad"
+        )
+        self.proj2 = layers.Conv2D(
+            self.fpn_dim, kernel_size=3, dtype=self.dtype_policy, name="proj2"
+        )
+
+    def build(self, input_shape):
+        for layer in self.scale_layers:
+            layer.build(input_shape)
+            input_shape = layer.compute_output_shape(input_shape)
+        self.proj1.build(input_shape)
+        input_shape = self.proj1.compute_output_shape(input_shape)
+        self.pad.build(input_shape)
+        input_shape = self.pad.compute_output_shape(input_shape)
+        self.proj2.build(input_shape)
+
+    def call(self, inputs, training=None):
+        hidden_states = inputs
+        for layer in self.scale_layers:
+            hidden_states = layer(hidden_states, training=training)
+        hidden_states = self.proj1(hidden_states, training=training)
+        hidden_states = self.pad(hidden_states, training=training)
+        return self.proj2(hidden_states, training=training)
+
+    def get_config(self):
+        config = super().get_config()
+        config.update(
+            {
+                "input_dim": self.input_dim,
+                "fpn_dim": self.fpn_dim,
+                "scale_factor": self.scale_factor,
+            }
+        )
+        return config
+
+    def compute_output_shape(self, input_shape):
+        output_shape = input_shape
+        for layer in self.scale_layers:
+            output_shape = layer.compute_output_shape(output_shape)
+        output_shape = self.proj1.compute_output_shape(output_shape)
+        output_shape = self.pad.compute_output_shape(output_shape)
+        return self.proj2.compute_output_shape(output_shape)
+
+
+class SAM3VisionNeck(layers.Layer):
+    def __init__(self, hidden_dim, fpn_hidden_dim, scale_factors, **kwargs):
+        super().__init__(**kwargs)
+        self.hidden_dim = int(hidden_dim)
+        self.fpn_hidden_dim = int(fpn_hidden_dim)
+        self.scale_factors = scale_factors
+
+        self.position_encoding = SAM3SinePositionEmbedding(
+            num_pos_feats=self.fpn_hidden_dim // 2,
+            normalize=True,
+            dtype=self.dtype_policy,
+            name="position_encoding",
+        )
+        self.fpn_layers = [
+            SAM3FPNLayer(
+                input_dim=self.hidden_dim,
+                fpn_dim=self.fpn_hidden_dim,
+                scale_factor=scale,
+                dtype=self.dtype_policy,
+                name=f"fpn_layer_{i}",
+            )
+            for i, scale in enumerate(self.scale_factors)
+        ]
+
+    def build(self, input_shape):
+        self.position_encoding.build()
+        self.fpn_image_shapes = []
+        for layer in self.fpn_layers:
+            layer.build(input_shape)
+            fpn_shape = layer.compute_output_shape(input_shape)
+            self.fpn_image_shapes.append([int(fpn_shape[1]), int(fpn_shape[2])])
+
+    def call(self, hidden_states, training=None):
+        fpn_hidden_states = []
+        fpn_position_encodings = []
+        for i, layer in enumerate(self.fpn_layers):
+            fpn_output = layer(hidden_states, training=training)
+            fpn_hidden_states.append(fpn_output)
+            height, width = self.fpn_image_shapes[i]
+            pos_enc = self.position_encoding(
+                fpn_output, height=height, width=width, training=training
+            )
+            fpn_position_encodings.append(pos_enc)
+        return fpn_hidden_states, fpn_position_encodings
+
+    def get_config(self):
+        config = super().get_config()
+        config.update(
+            {
+                "hidden_dim": self.hidden_dim,
+                "fpn_hidden_dim": self.fpn_hidden_dim,
+                "scale_factors": self.scale_factors,
+            }
+        )
+        return config
+
+    def compute_output_shape(self, input_shape):
+        fpn_hidden_state_shapes = []
+        for layer in self.fpn_layers:
+            fpn_hidden_state_shapes.append(
+                layer.compute_output_shape(input_shape)
+            )
+        # fpn_hidden_states and fpn_position_encodings have the same shapes.
+        return fpn_hidden_state_shapes, fpn_hidden_state_shapes
+
+
+@keras_hub_export("keras_hub.layers.SAM3VisionEncoder")
+class SAM3VisionEncoder(layers.Layer):
+    """A vision encoder for the Segment Anything Model 3 (SAM3).
+
+    This layer implements a Vision Transformer (ViT) backbone followed by a
+    Feature Pyramid Network (FPN) neck. It processes input images and produces
+    multi-scale feature maps and their corresponding position encodings.
+
+    Args:
+        image_shape: tuple. The shape of the input image
+            (height, width, channels).
+        patch_size: int. The size of the patches to be extracted from the image.
+        num_layers: int. The number of transformer layers in the ViT backbone.
+        hidden_dim: int. The hidden dimension of the transformer layers.
+        intermediate_dim: int. The dimension of the intermediate layer in the
+            transformer's MLP.
+        num_heads: int. The number of attention heads.
+        fpn_hidden_dim: int. The hidden dimension of the FPN.
+        fpn_scale_factors: list of floats. The scale factors for each level of
+            the feature pyramid.
+        pretrain_image_shape: tuple. The shape of the image used during
+            pretraining, for position embedding interpolation. Defaults to
+            `(336, 336, 3)`.
+        hidden_activation: str. The activation function for the transformer
+            layers. Defaults to `"gelu"`.
+        rope_theta: float. The theta value for rotary position embeddings.
+            Defaults to `10000.0`.
+        window_size: int. The size of the window for windowed attention.
+            Defaults to `0`.
+        global_attn_indexes: list of ints. The indices of the layers that use
+            global attention instead of windowed attention.
+        attention_dropout_rate: float. The dropout rate for attention. Defaults
+            to `0`.
+        hidden_dropout_rate: float. The dropout rate for the MLP. Defaults to
+            `0.0`.
+        layer_norm_epsilon: float. The epsilon value for layer normalization.
+            Defaults to `1e-6`.
+    """
+
+    def __init__(
+        self,
+        image_shape,
+        patch_size,
+        num_layers,
+        hidden_dim,
+        intermediate_dim,
+        num_heads,
+        fpn_hidden_dim,
+        fpn_scale_factors,
+        pretrain_image_shape=(336, 336, 3),
+        hidden_activation="gelu",
+        rope_theta=10000.0,
+        window_size=0,
+        global_attn_indexes=None,
+        attention_dropout_rate=0.0,
+        hidden_dropout_rate=0.0,
+        layer_norm_epsilon=1e-6,
+        **kwargs,
+    ):
+        super().__init__(**kwargs)
+        self.image_shape = (
+            int(image_shape[0]),
+            int(image_shape[1]),
+            int(image_shape[2]),
+        )
+        self.patch_size = int(patch_size)
+        self.num_layers = int(num_layers)
+        self.hidden_dim = int(hidden_dim)
+        self.intermediate_dim = int(intermediate_dim)
+        self.num_heads = int(num_heads)
+        self.fpn_hidden_dim = int(fpn_hidden_dim)
+        self.fpn_scale_factors = fpn_scale_factors
+        self.hidden_activation = hidden_activation
+        self.rope_theta = float(rope_theta)
+        self.window_size = int(window_size)
+        if global_attn_indexes is not None:
+            self.global_attn_indexes = [int(i) for i in global_attn_indexes]
+        else:
+            self.global_attn_indexes = None
+        self.pretrain_image_shape = (
+            int(pretrain_image_shape[0]),
+            int(pretrain_image_shape[1]),
+            int(pretrain_image_shape[2]),
+        )
+        self.hidden_dropout_rate = float(hidden_dropout_rate)
+        self.attention_dropout_rate = float(attention_dropout_rate)
+        self.layer_norm_epsilon = float(layer_norm_epsilon)
+
+        self.backbone = SAM3ViTEncoder(
+            image_shape=self.image_shape,
+            patch_size=self.patch_size,
+            num_layers=self.num_layers,
+            hidden_dim=self.hidden_dim,
+            intermediate_dim=self.intermediate_dim,
+            num_heads=self.num_heads,
+            pretrain_image_shape=self.pretrain_image_shape,
+            hidden_activation=self.hidden_activation,
+            rope_theta=self.rope_theta,
+            window_size=self.window_size,
+            global_attn_indexes=self.global_attn_indexes,
+            attention_dropout_rate=self.attention_dropout_rate,
+            hidden_dropout_rate=self.hidden_dropout_rate,
+            layer_norm_epsilon=self.layer_norm_epsilon,
+            dtype=self.dtype_policy,
+            name="backbone",
+        )
+        self.vision_neck = SAM3VisionNeck(
+            hidden_dim=self.hidden_dim,
+            fpn_hidden_dim=self.fpn_hidden_dim,
+            scale_factors=self.fpn_scale_factors,
+            dtype=self.dtype_policy,
+            name="vision_neck",
+        )
+
+    def build(self, input_shape):
+        self.backbone.build(input_shape)
+        input_shape = self.backbone.compute_output_shape(input_shape)
+        height = self.image_shape[0] // self.patch_size
+        width = self.image_shape[1] // self.patch_size
+        input_shape = (input_shape[0], height, width, input_shape[-1])
+        self.vision_neck.build(input_shape)
+
+    def call(self, pixel_values, training=None):
+        hidden_states = self.backbone(pixel_values, training=training)
+        height = self.image_shape[0] // self.patch_size
+        width = self.image_shape[1] // self.patch_size
+        spatial_hidden_states = ops.reshape(
+            hidden_states, (-1, height, width, self.hidden_dim)
+        )
+        fpn_hidden_states, fpn_position_encodings = self.vision_neck(
+            spatial_hidden_states, training=training
+        )
+        return fpn_hidden_states, fpn_position_encodings
+
+    def get_config(self):
+        config = super().get_config()
+        config.update(
+            {
+                "image_shape": self.image_shape,
+                "patch_size": self.patch_size,
+                "num_layers": self.num_layers,
+                "hidden_dim": self.hidden_dim,
+                "intermediate_dim": self.intermediate_dim,
+                "num_heads": self.num_heads,
+                "fpn_hidden_dim": self.fpn_hidden_dim,
+                "fpn_scale_factors": self.fpn_scale_factors,
+                "pretrain_image_shape": self.pretrain_image_shape,
+                "hidden_activation": self.hidden_activation,
+                "rope_theta": self.rope_theta,
+                "window_size": self.window_size,
+                "global_attn_indexes": self.global_attn_indexes,
+                "attention_dropout_rate": self.attention_dropout_rate,
+                "hidden_dropout_rate": self.hidden_dropout_rate,
+                "layer_norm_epsilon": self.layer_norm_epsilon,
+            }
+        )
+        return config
+
+    def compute_output_shape(self, input_shape):
+        input_shape = self.backbone.compute_output_shape(input_shape)
+        height = self.image_shape[0] // self.patch_size
+        width = self.image_shape[1] // self.patch_size
+        input_shape = (input_shape[0], height, width, input_shape[-1])
+        fpn_hidden_state_shapes, fpn_position_encoding_shapes = (
+            self.vision_neck.compute_output_shape(input_shape)
+        )
+        return fpn_hidden_state_shapes, fpn_position_encoding_shapes