keras-hub-nightly 0.15.0.dev20240823171555__py3-none-any.whl

keras_hub/src/models/opt/opt_preprocessor.py

@@ -0,0 +1,188 @@
+# Copyright 2024 The KerasHub Authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     https://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+import keras
+
+from keras_hub.src.api_export import keras_hub_export
+from keras_hub.src.layers.preprocessing.start_end_packer import StartEndPacker
+from keras_hub.src.models.opt.opt_tokenizer import OPTTokenizer
+from keras_hub.src.models.preprocessor import Preprocessor
+from keras_hub.src.utils.keras_utils import (
+    convert_inputs_to_list_of_tensor_segments,
+)
+
+
+@keras_hub_export("keras_hub.models.OPTPreprocessor")
+class OPTPreprocessor(Preprocessor):
+    """OPT preprocessing layer which tokenizes and packs inputs.
+
+    This preprocessing layer will do 2 things:
+
+    - Tokenize the input using the `tokenizer`.
+    - Construct a dictionary with keys `"token_ids"`, `"padding_mask"`, that can
+        be passed directly to a `keras_hub.models.OPTBackbone`.
+
+    This layer can be used directly with `tf.data.Dataset.map` to preprocess
+    string data in the `(x, y, sample_weight)` format used by
+    `keras.Model.fit`.
+
+    The call method of this layer accepts three arguments, `x`, `y`, and
+    `sample_weight`. `x` can be a python string or tensor representing a single
+    segment, a list of python strings representing a batch of single segments,
+    or a list of tensors representing multiple segments to be packed together.
+    `y` and `sample_weight` are both optional, can have any format, and will be
+    passed through unaltered.
+
+    `OPTPreprocessor` forces the input to have only one segment, as OPT is
+    mainly used for generation tasks. For tasks having multi-segment inputs
+    like "glue/mnli", please use a model designed for classification purposes
+    such as BERT or RoBERTa.
+
+    Args:
+        tokenizer: A `keras_hub.models.OPTTokenizer` instance.
+        sequence_length: The length of the packed inputs.
+        add_start_token: If `True`, the preprocessor will append the tokenizer
+            start token to each input sequence.
+        add_end_token: If `True`, the preprocessor will append the tokenizer
+            end token to each input sequence.
+
+    Call arguments:
+        x: A string, `tf.Tensor` or list of python strings.
+        y: Any label data. Will be passed through unaltered.
+        sample_weight: Any label weight data. Will be passed through unaltered.
+        sequence_length: Pass to override the configured `sequence_length` of
+            the layer.
+
+    Examples:
+
+    Directly calling the layer on data.
+    ```python
+    preprocessor = keras_hub.models.OPTPreprocessor.from_preset("opt_125m_en")
+
+    # Tokenize and pack a single sentence.
+    preprocessor("The quick brown fox jumped.")
+
+    # Tokenize a batch of single sentences.
+    preprocessor(["The quick brown fox jumped.", "Call me Ishmael."])
+
+    # Custom vocabulary.
+    features = ["a quick fox.", "a fox quick."]
+    vocab = {"<|endoftext|>": 0, "a": 4, "Ġquick": 5, "Ġfox": 6}
+    merges = ["Ġ q", "u i", "c k", "ui ck", "Ġq uick"]
+    merges += ["Ġ f", "o x", "Ġf ox"]
+    tokenizer = keras_hub.models.OPTTokenizer(
+        vocabulary=vocab,
+        merges=merges,
+    )
+    preprocessor = keras_hub.models.OPTPreprocessor(tokenizer=tokenizer)
+    preprocessor("The quick brown fox jumped.")
+    ```
+
+    Mapping with `tf.data.Dataset`.
+    ```python
+    preprocessor = keras_hub.models.OPTPreprocessor.from_preset("opt_125m_en")
+
+    text = tf.constant(["The quick brown fox jumped.", "Call me Ishmael."])
+    label = tf.constant([1, 1])
+
+    # Map labeled single sentences.
+    ds = tf.data.Dataset.from_tensor_slices((text, label))
+    ds = ds.map(preprocessor, num_parallel_calls=tf.data.AUTOTUNE)
+
+    # Map unlabeled single sentences.
+    ds = tf.data.Dataset.from_tensor_slices(text)
+    ds = ds.map(preprocessor, num_parallel_calls=tf.data.AUTOTUNE)
+    ```
+    """
+
+    tokenizer_cls = OPTTokenizer
+
+    def __init__(
+        self,
+        tokenizer,
+        sequence_length=1024,
+        add_start_token=True,
+        add_end_token=True,
+        **kwargs,
+    ):
+        super().__init__(**kwargs)
+
+        self.tokenizer = tokenizer
+        self.packer = None
+        self.sequence_length = sequence_length
+        self.add_start_token = add_start_token
+        self.add_end_token = add_end_token
+
+    def build(self, input_shape):
+        # Defer packer creation to `build()` so that we can be sure tokenizer
+        # assets have loaded when restoring a saved model.
+        self.packer = StartEndPacker(
+            start_value=self.tokenizer.start_token_id,
+            end_value=self.tokenizer.end_token_id,
+            pad_value=self.tokenizer.pad_token_id,
+            sequence_length=self.sequence_length,
+            return_padding_mask=True,
+        )
+        self.built = True
+
+    def get_config(self):
+        config = super().get_config()
+        config.update(
+            {
+                "sequence_length": self.sequence_length,
+                "add_start_token": self.add_start_token,
+                "add_end_token": self.add_end_token,
+            }
+        )
+        return config
+
+    def call(
+        self,
+        x,
+        y=None,
+        sample_weight=None,
+        sequence_length=None,
+    ):
+        x = convert_inputs_to_list_of_tensor_segments(x)
+        if len(x) != 1:
+            raise ValueError(
+                "OPT requires each input feature to contain only "
+                f"one segment, but received {len(x)}. If you are using OPT "
+                "for a multi-segment classification task, please refer to "
+                "classification models like BERT or RoBERTa."
+            )
+        sequence_length = sequence_length or self.sequence_length
+        token_ids, padding_mask = self.packer(
+            self.tokenizer(x[0]),
+            sequence_length=sequence_length,
+            add_start_value=self.add_start_token,
+            add_end_value=self.add_end_token,
+        )
+        x = {
+            "token_ids": token_ids,
+            "padding_mask": padding_mask,
+        }
+        return keras.utils.pack_x_y_sample_weight(x, y, sample_weight)
+
+    @property
+    def sequence_length(self):
+        """The padded length of model input sequences."""
+        return self._sequence_length
+
+    @sequence_length.setter
+    def sequence_length(self, value):
+        self._sequence_length = value
+        if self.packer is not None:
+            self.packer.sequence_length = value

keras_hub/src/models/opt/opt_presets.py

@@ -0,0 +1,72 @@
+# Copyright 2024 The KerasHub Authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     https://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""OPT model preset configurations."""
+
+# Metadata for loading pretrained model weights.
+backbone_presets = {
+    "opt_125m_en": {
+        "metadata": {
+            "description": (
+                "12-layer OPT model where case in maintained. Trained on "
+                "BookCorpus, CommonCrawl, Pile, and PushShift.io corpora."
+            ),
+            "params": 125237760,
+            "official_name": "OPT",
+            "path": "opt",
+            "model_card": "https://github.com/facebookresearch/metaseq/blob/main/projects/OPT/model_card.md",
+        },
+        "kaggle_handle": "kaggle://keras/opt/keras/opt_125m_en/2",
+    },
+    # We skip the 350m checkpoint because it does not match the structure of
+    # other checkpoints.
+    "opt_1.3b_en": {
+        "metadata": {
+            "description": (
+                "24-layer OPT model where case in maintained. Trained on "
+                "BookCorpus, CommonCrawl, Pile, and PushShift.io corpora."
+            ),
+            "params": 1315753984,
+            "official_name": "OPT",
+            "path": "opt",
+            "model_card": "https://github.com/facebookresearch/metaseq/blob/main/projects/OPT/model_card.md",
+        },
+        "kaggle_handle": "kaggle://keras/opt/keras/opt_1.3b_en/2",
+    },
+    "opt_2.7b_en": {
+        "metadata": {
+            "description": (
+                "32-layer OPT model where case in maintained. Trained on "
+                "BookCorpus, CommonCrawl, Pile, and PushShift.io corpora."
+            ),
+            "params": 2700000000,
+            "official_name": "OPT",
+            "path": "opt",
+            "model_card": "https://github.com/facebookresearch/metaseq/blob/main/projects/OPT/model_card.md",
+        },
+        "kaggle_handle": "kaggle://keras/opt/keras/opt_2.7b_en/2",
+    },
+    "opt_6.7b_en": {
+        "metadata": {
+            "description": (
+                "32-layer OPT model where case in maintained. Trained on "
+                "BookCorpus, CommonCrawl, Pile, and PushShift.io corpora."
+            ),
+            "params": 6700000000,
+            "official_name": "OPT",
+            "path": "opt",
+            "model_card": "https://github.com/facebookresearch/metaseq/blob/main/projects/OPT/model_card.md",
+        },
+        "kaggle_handle": "kaggle://keras/opt/keras/opt_6.7b_en/2",
+    },
+}

keras_hub/src/models/opt/opt_tokenizer.py

@@ -0,0 +1,116 @@
+# Copyright 2024 The KerasHub Authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     https://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+from keras_hub.src.api_export import keras_hub_export
+from keras_hub.src.tokenizers.byte_pair_tokenizer import BytePairTokenizer
+
+
+@keras_hub_export("keras_hub.models.OPTTokenizer")
+class OPTTokenizer(BytePairTokenizer):
+    """An OPT tokenizer using Byte-Pair Encoding subword segmentation.
+
+    This tokenizer class will tokenize raw strings into integer sequences and
+    is based on `keras_hub.tokenizers.BytePairTokenizer`. Unlike the
+    underlying tokenizer, it will check for all special tokens needed by OPT
+    models and provides a `from_preset()` method to automatically download
+    a matching vocabulary for a OPT preset.
+
+    This tokenizer does not provide truncation or padding of inputs.
+
+    If input is a batch of strings (rank > 0), the layer will output a
+    `tf.RaggedTensor` where the last dimension of the output is ragged.
+    If input is a scalar string (rank == 0), the layer will output a dense
+    `tf.Tensor` with static shape `[None]`.
+
+    Args:
+        vocabulary: string or dict, maps token to integer ids. If it is a
+            string, it should be the file path to a json file.
+        merges: string or list, contains the merge rule. If it is a string,
+            it should be the file path to merge rules. The merge rule file
+            should have one merge rule per line. Every merge rule contains
+            merge entities separated by a space.
+
+    Examples:
+    ```python
+    # Unbatched input.
+    tokenizer = keras_hub.models.OPTTokenizer.from_preset(
+        "opt_125m_en",
+    )
+    tokenizer("The quick brown fox jumped.")
+
+    # Batched input.
+    tokenizer(["The quick brown fox jumped.", "The fox slept."])
+
+    # Detokenization.
+    tokenizer.detokenize(tokenizer("The quick brown fox jumped."))
+
+    # Custom vocabulary.
+    vocab = {"<pad>": 1, "</s>": 2, "Ġquick": 4, "Ġfox": 5}
+    merges = ["Ġ q", "u i", "c k", "ui ck", "Ġq uick"]
+    merges += ["Ġ f", "o x", "Ġf ox"]
+    tokenizer = keras_hub.models.OPTTokenizer(vocabulary=vocab, merges=merges)
+    tokenizer("The quick brown fox jumped.")
+    ```
+    """
+
+    def __init__(
+        self,
+        vocabulary=None,
+        merges=None,
+        **kwargs,
+    ):
+        self.start_token = "</s>"
+        self.pad_token = "<pad>"
+        self.end_token = "</s>"
+
+        super().__init__(
+            vocabulary=vocabulary,
+            merges=merges,
+            unsplittable_tokens=[
+                self.start_token,
+                self.pad_token,
+                self.end_token,
+            ],
+            **kwargs,
+        )
+
+    def set_vocabulary_and_merges(self, vocabulary, merges):
+        super().set_vocabulary_and_merges(vocabulary, merges)
+
+        if vocabulary is not None:
+            # Check for necessary special tokens.
+            for token in [self.start_token, self.pad_token, self.end_token]:
+                if token not in self.vocabulary:
+                    raise ValueError(
+                        f"Cannot find token `'{token}'` in the provided "
+                        f"`vocabulary`. Please provide `'{token}'` in your "
+                        "`vocabulary` or use a pretrained `vocabulary` name."
+                    )
+
+            self.start_token_id = self.token_to_id(self.start_token)
+            self.pad_token_id = self.token_to_id(self.pad_token)
+            self.end_token_id = self.token_to_id(self.end_token)
+        else:
+            self.start_token_id = None
+            self.pad_token_id = None
+            self.end_token_id = None
+
+    def get_config(self):
+        config = super().get_config()
+        # In the constructor, we pass the list of special tokens to the
+        # `unsplittable_tokens` arg of the superclass' constructor. Hence, we
+        # delete it from the config here.
+        del config["unsplittable_tokens"]
+        return config

keras_hub/src/models/pali_gemma/__init__.py

@@ -0,0 +1,23 @@
+# Copyright 2024 The KerasHub Authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     https://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+from keras_hub.src.models.pali_gemma.pali_gemma_backbone import (
+    PaliGemmaBackbone,
+)
+from keras_hub.src.models.pali_gemma.pali_gemma_presets import backbone_presets
+from keras_hub.src.models.pali_gemma.pali_gemma_tokenizer import (
+    PaliGemmaTokenizer,
+)
+from keras_hub.src.utils.preset_utils import register_presets
+
+register_presets(backbone_presets, (PaliGemmaBackbone, PaliGemmaTokenizer))

keras_hub/src/models/pali_gemma/pali_gemma_backbone.py

@@ -0,0 +1,277 @@
+# Copyright 2024 The KerasHub Authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     https://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+import keras
+from keras import ops
+
+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.gemma.rms_normalization import RMSNormalization
+from keras_hub.src.models.pali_gemma.pali_gemma_decoder_block import (
+    PaliGemmaDecoderBlock,
+)
+from keras_hub.src.models.pali_gemma.pali_gemma_vit import PaliGemmaVit
+
+
+@keras_hub_export("keras_hub.models.PaliGemmaBackbone")
+class PaliGemmaBackbone(Backbone):
+    """PaliGemma core network with hyperparameters.
+
+    This backbone implements the mixed-modality PaliGemma architecture. It
+    contains a Visual Transformer network, as well as text token embedding
+    layer, followed by a backend-agnostic concatenation operation to
+    construct a sequence of representations of mixed type embeddings (visual
+    and textual). Then, the concatenated sequence is passed through a series
+    of Mixed Modality Decoder Blocks. The returned value from calling this model
+    represents probabilistic values for output tokens.
+
+    For a higher-level object for text-generation,
+    see `keras_hub.models.PaliGemmaCausalLM`.
+
+    The default constructor gives a fully customizable, randomly initialized
+    PaliGemma model with any number of vit layers, heads, embedding
+    dimensions, and equivalent configuration for Paligemma Decoder layers. To
+    load preset architectures and weights, use the `from_preset` constructor.
+
+    Args:
+        vocabulary_size: int. The size of the token vocabulary.
+        image_size: int. The resolution of the image in both width and height.
+            Note: input images must be square.
+        num_layers: int. The number of transformer mixed decoder layers.
+        num_query_heads: int. The number of heads for the query projections in
+            the mixed decoder attention layer.
+        num_key_value_heads: int. The number of heads for the key and value
+            projections in the mixed decoder attention layers.
+        hidden_dim: int. The size of the transformer hidden state at the end
+            of each mixed transformer layer.
+        intermediate_dim: int. The output dimension of the first Dense layer in
+            a two-layer feedforward network for each transformer decoder block.
+        head_dim: int. The size of each attention head in the mixed decoder.
+        vit_patch_size: int. The size of each square patch in the input image.
+        vit_num_heads: int. The number of attention heads for the vision(image)
+            transformer encoder.
+        vit_hidden_dim: int. The size of the transformer hidden state at the end
+            of each vision transformer layer.
+        vit_num_layers: int. The number of vision transformer layers.
+        vit_intermediate_dim: int. The output dimension of the first Dense layer
+            in a two-layer feedforward network for vision transformer.
+        vit_pooling: string. The encoded vision embeddings are pooled using the
+            specified polling setting. The accepted values are `"map"`, `"gap"`,
+            `"0"` or `"none"`. Defaults to `"none"`.
+        vit_classifier_activation: activation function. The activation that
+            is used for final output classification in the vision transformer.
+        vit_name: string. The name used for vision transformer layers.
+        include_rescaling: bool. If true, the image input will be rescaled from
+            the range `[0, 255]`, to the range `[0, 1]`.
+        layer_norm_epsilon: float. The epsilon value user for every layer norm
+            in all transformer blocks.
+        dropout: float. Dropout probability for the Transformer decoder blocks.
+        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
+            be done a float32 precision regardless of dtype.
+
+    Example:
+    ```python
+    input_data = {
+        "token_ids": np.ones(shape=(1, 12), dtype="int32"),
+        "images": np.random.uniform(size=(1, 224, 224, 3)),
+        "padding_mask": np.array([[1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0]]),
+    }
+
+    # Pretrained PaliGemma decoder.
+    model = keras_hub.models.PaliGemmaBackbone.from_preset("pali_gemma_mix_224")
+    model(input_data)
+
+    # Randomly initialized PaliGemma decoder with custom config.
+    model = keras_hub.models.PaliGemmaBackbone(
+        vocabulary_size=50257,
+        images_size=224,
+        num_layers=12,
+        num_query_heads=12,
+        num_key_value_heads=1,
+        hidden_dim=768,
+        intermediate_dim=3072,
+        head_dim=64,
+        vit_patch_size=14,
+        vit_num_heads=8,
+        vit_hidden_dim=768,
+        vit_intermediate_dim=3072,
+        vit_num_layers=2,
+    )
+    model(input_data)
+    ```
+    """
+
+    def __init__(
+        self,
+        vocabulary_size,
+        image_size,
+        num_layers,
+        num_query_heads,
+        num_key_value_heads,
+        hidden_dim,
+        intermediate_dim,
+        head_dim,
+        vit_patch_size,
+        vit_num_heads,
+        vit_hidden_dim,
+        vit_num_layers,
+        vit_intermediate_dim=None,  # TODO remove default
+        vit_pooling=None,
+        vit_classifier_activation=None,
+        vit_name=None,
+        include_rescaling=True,
+        layer_norm_epsilon=1e-6,
+        dropout=0,
+        dtype=None,
+        **kwargs,
+    ):
+        # === Layers ===
+        self.token_embedding = ReversibleEmbedding(
+            input_dim=vocabulary_size,
+            output_dim=hidden_dim,
+            tie_weights=True,
+            embeddings_initializer=keras.initializers.VarianceScaling(
+                scale=1.0,
+                mode="fan_in",
+                distribution="untruncated_normal",
+                seed=None,
+            ),
+            dtype=dtype,
+            name="token_embedding",
+        )
+        # TODO Remove this. Work around for previous serialization bug.
+        vit_intermediate_dim = vit_intermediate_dim or 4304
+        self.vit_encoder = PaliGemmaVit(
+            image_size=image_size,
+            include_rescaling=include_rescaling,
+            patch_size=vit_patch_size,
+            num_heads=vit_num_heads,
+            hidden_dim=vit_hidden_dim,
+            num_layers=vit_num_layers,
+            intermediate_dim=vit_intermediate_dim,
+            pooling=vit_pooling,
+            num_classes=hidden_dim,
+            classifier_activation=vit_classifier_activation,
+            dtype=dtype,
+            name=vit_name,
+        )
+        self.transformer_layers = []
+        for i in range(num_layers):
+            layer = PaliGemmaDecoderBlock(
+                hidden_dim=hidden_dim,
+                intermediate_dim=intermediate_dim,
+                num_query_heads=num_query_heads,
+                head_dim=head_dim,
+                num_key_value_heads=num_key_value_heads,
+                dropout=dropout,
+                dtype=dtype,
+                name=f"decoder_block_{i}",
+            )
+            self.transformer_layers.append(layer)
+        self.layer_norm = RMSNormalization(
+            epsilon=layer_norm_epsilon,
+            dtype=dtype,
+            name="final_normalization",
+        )
+
+        # === Functional Model ===
+        image_input = self.vit_encoder.inputs[0]
+        token_id_input = keras.Input(
+            shape=(None,), dtype="int32", name="token_ids"
+        )
+        padding_mask_input = keras.Input(
+            shape=(None,), dtype="int32", name="padding_mask"
+        )
+        response_mask_input = keras.Input(
+            shape=(None,), dtype="int32", name="response_mask"
+        )
+        img_embeddings = self.vit_encoder(image_input)
+        text_embeddings = self.token_embedding(token_id_input)
+        text_embeddings = text_embeddings * ops.cast(
+            ops.sqrt(hidden_dim), text_embeddings.dtype
+        )
+        x = ops.concatenate((img_embeddings, text_embeddings), axis=1)
+        for transformer_layer in self.transformer_layers:
+            x = transformer_layer(
+                x,
+                padding_mask=padding_mask_input,
+                response_mask=response_mask_input,
+            )
+        sequence_output = self.layer_norm(x)
+        super().__init__(
+            inputs={
+                "images": image_input,
+                "token_ids": token_id_input,
+                "padding_mask": padding_mask_input,
+                "response_mask": response_mask_input,
+            },
+            outputs=sequence_output,
+            dtype=dtype,
+            **kwargs,
+        )
+
+        # === Config ===
+        self.vocabulary_size = vocabulary_size
+        self.image_size = image_size
+        self.include_rescaling = include_rescaling
+        self.num_layers = num_layers
+        self.num_query_heads = num_query_heads
+        self.num_key_value_heads = num_key_value_heads
+        self.hidden_dim = hidden_dim
+        self.intermediate_dim = intermediate_dim
+        self.head_dim = head_dim
+        self.layer_norm_epsilon = layer_norm_epsilon
+        self.dropout = dropout
+        # VIT Params
+        self.vit_patch_size = vit_patch_size
+        self.vit_num_heads = vit_num_heads
+        self.vit_hidden_dim = vit_hidden_dim
+        self.vit_num_layers = vit_num_layers
+        self.vit_intermediate_dim = vit_intermediate_dim
+        self.vit_pooling = vit_pooling
+        self.vit_classifier_activation = vit_classifier_activation
+        self.vit_name = vit_name
+        # Keep the image_sequence_length as a backbone property for easy access.
+        self.image_sequence_length = self.vit_encoder.image_sequence_length
+
+    def get_config(self):
+        config = super().get_config()
+        config.update(
+            {
+                "vocabulary_size": self.vocabulary_size,
+                "image_size": self.image_size,
+                "include_rescaling": self.include_rescaling,
+                "num_layers": self.num_layers,
+                "num_query_heads": self.num_query_heads,
+                "num_key_value_heads": self.num_key_value_heads,
+                "hidden_dim": self.hidden_dim,
+                "intermediate_dim": self.intermediate_dim,
+                "head_dim": self.head_dim,
+                "layer_norm_epsilon": self.layer_norm_epsilon,
+                "dropout": self.dropout,
+                "vit_patch_size": self.vit_patch_size,
+                "vit_num_heads": self.vit_num_heads,
+                "vit_hidden_dim": self.vit_hidden_dim,
+                "vit_num_layers": self.vit_num_layers,
+                "vit_intermediate_dim": self.vit_intermediate_dim,
+                "vit_pooling": self.vit_pooling,
+                "vit_classifier_activation": self.vit_classifier_activation,
+                "vit_name": self.vit_name,
+            }
+        )
+        return config