[ViT] Vision Transformer (ViT) backbone, layers, and image classifier (#1989)

* vit base

* Add vit backbone, classifier and preprocessor layers

* update args

* add default args

* correct build method

* fix build issues

* fix bugs

* Update backbone args and configs

* correct position ids dtype

* build token layer

* token layer build

* assign correct dtype to TokenLayer

* fix build shape of token layer

* correct mlp dens var names

* use default norm mean and std as per hugging face config

* correct position_ids

* remove separate token layer

* correct position ids

* Checkpoint conversion script and minor changes

* correct flag type

* correct key name

* use flat list later we can extract in between layers if needed

* Add test cases and correct dtype polciy for model

* add proper docstrings

* correct test cases

* use numpy for test data

* nit

* nit

* add presets

* load vit preset from hugging face directly

* nit

* handle num classes case for ViT

* replace toke with first

* convert all vit checkpoints using tools

* Add custom ImageClassifier for ViT

* remove token pooling and rename representation_size to intermediate_dim

Author: sineeli

keras_hub/api/layers/__init__.py

@@ -63,6 +63,7 @@
     SegFormerImageConverter,
 )
 from keras_hub.src.models.vgg.vgg_image_converter import VGGImageConverter
+from keras_hub.src.models.vit.vit_image_converter import ViTImageConverter
 from keras_hub.src.models.whisper.whisper_audio_converter import (
     WhisperAudioConverter,
 )

keras_hub/api/models/__init__.py

@@ -330,6 +330,11 @@
 from keras_hub.src.models.vgg.vgg_image_classifier_preprocessor import (
     VGGImageClassifierPreprocessor,
 )
+from keras_hub.src.models.vit.vit_backbone import ViTBackbone
+from keras_hub.src.models.vit.vit_image_classifier import ViTImageClassifier
+from keras_hub.src.models.vit.vit_image_classifier_preprocessor import (
+    ViTImageClassifierPreprocessor,
+)
 from keras_hub.src.models.vit_det.vit_det_backbone import ViTDetBackbone
 from keras_hub.src.models.whisper.whisper_backbone import WhisperBackbone
 from keras_hub.src.models.whisper.whisper_tokenizer import WhisperTokenizer

keras_hub/src/models/vit/__init__.py

@@ -0,0 +1,5 @@
+from keras_hub.src.models.vit.vit_backbone import ViTBackbone
+from keras_hub.src.models.vit.vit_presets import backbone_presets
+from keras_hub.src.utils.preset_utils import register_presets
+
+register_presets(backbone_presets, ViTBackbone)

keras_hub/src/models/vit/vit_backbone.py

@@ -0,0 +1,152 @@
+import keras
+
+from keras_hub.src.api_export import keras_hub_export
+from keras_hub.src.models.backbone import Backbone
+from keras_hub.src.models.vit.vit_layers import ViTEncoder
+from keras_hub.src.models.vit.vit_layers import ViTPatchingAndEmbedding
+from keras_hub.src.utils.keras_utils import standardize_data_format
+
+
+@keras_hub_export("keras_hub.models.ViTBackbone")
+class ViTBackbone(Backbone):
+    """Vision Transformer (ViT) backbone.
+
+    This backbone implements the Vision Transformer architecture as described in
+    [An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale](https://arxiv.org/abs/2010.11929).
+    It transforms the input image into a sequence of patches, embeds them, and
+    then processes them through a series of Transformer encoder layers.
+
+    Args:
+        image_shape: A tuple or list of 3 integers representing the shape of the
+            input image `(height, width, channels)`, `height` and `width` must
+            be equal.
+        patch_size: int. The size of each image patch, the input image will be
+            divided into patches of shape `(patch_size, patch_size)`.
+        num_layers: int. The number of transformer encoder layers.
+        num_heads: int. specifying the number of attention heads in each
+            Transformer encoder layer.
+        hidden_dim: int. The dimensionality of the hidden representations.
+        mlp_dim: int. The dimensionality of the intermediate MLP layer in
+            each Transformer encoder layer.
+        dropout_rate: float. The dropout rate for the Transformer encoder
+            layers.
+        attention_dropout: float. The dropout rate for the attention mechanism
+            in each Transformer encoder layer.
+        layer_norm_epsilon: float. Value used for numerical stability in
+            layer normalization.
+        use_mha_bias: bool. Whether to use bias in the multi-head
+            attention layers.
+        use_mlp_bias: bool. Whether to use bias in the MLP layers.
+        data_format: str.  `"channels_last"` or `"channels_first"`, specifying
+            the data format for the input image. If `None`, defaults to
+            `"channels_last"`.
+        dtype: The dtype of the layer weights. Defaults to None.
+        **kwargs: Additional keyword arguments to be passed to the parent
+            `Backbone` class.
+    """
+
+    def __init__(
+        self,
+        image_shape,
+        patch_size,
+        num_layers,
+        num_heads,
+        hidden_dim,
+        mlp_dim,
+        dropout_rate=0.0,
+        attention_dropout=0.0,
+        layer_norm_epsilon=1e-6,
+        use_mha_bias=True,
+        use_mlp_bias=True,
+        data_format=None,
+        dtype=None,
+        **kwargs,
+    ):
+        # === Laters ===
+        data_format = standardize_data_format(data_format)
+        h_axis, w_axis, channels_axis = (
+            (-3, -2, -1) if data_format == "channels_last" else (-2, -1, -3)
+        )
+        # Check that the input image is well specified.
+        if image_shape[h_axis] is None or image_shape[w_axis] is None:
+            raise ValueError(
+                f"Image shape must have defined height and width. Found `None` "
+                f"at index {h_axis} (height) or {w_axis} (width). "
+                f"Image shape: {image_shape}"
+            )
+        if image_shape[h_axis] != image_shape[w_axis]:
+            raise ValueError(
+                f"Image height and width must be equal. Found height: "
+                f"{image_shape[h_axis]}, width: {image_shape[w_axis]} at "
+                f"indices {h_axis} and {w_axis} respectively. Image shape: "
+                f"{image_shape}"
+            )
+
+        num_channels = image_shape[channels_axis]
+
+        # === Functional Model ===
+        inputs = keras.layers.Input(shape=image_shape)
+
+        x = ViTPatchingAndEmbedding(
+            image_size=image_shape[h_axis],
+            patch_size=patch_size,
+            hidden_dim=hidden_dim,
+            num_channels=num_channels,
+            data_format=data_format,
+            dtype=dtype,
+            name="vit_patching_and_embedding",
+        )(inputs)
+
+        output = ViTEncoder(
+            num_layers=num_layers,
+            num_heads=num_heads,
+            hidden_dim=hidden_dim,
+            mlp_dim=mlp_dim,
+            dropout_rate=dropout_rate,
+            attention_dropout=attention_dropout,
+            layer_norm_epsilon=layer_norm_epsilon,
+            use_mha_bias=use_mha_bias,
+            use_mlp_bias=use_mlp_bias,
+            dtype=dtype,
+            name="vit_encoder",
+        )(x)
+
+        super().__init__(
+            inputs=inputs,
+            outputs=output,
+            dtype=dtype,
+            **kwargs,
+        )
+
+        # === Config ===
+        self.image_shape = image_shape
+        self.patch_size = patch_size
+        self.num_layers = num_layers
+        self.num_heads = num_heads
+        self.hidden_dim = hidden_dim
+        self.mlp_dim = mlp_dim
+        self.dropout_rate = dropout_rate
+        self.attention_dropout = attention_dropout
+        self.layer_norm_epsilon = layer_norm_epsilon
+        self.use_mha_bias = use_mha_bias
+        self.use_mlp_bias = use_mlp_bias
+        self.data_format = data_format
+
+    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,
+                "num_heads": self.num_heads,
+                "hidden_dim": self.hidden_dim,
+                "mlp_dim": self.mlp_dim,
+                "dropout_rate": self.dropout_rate,
+                "attention_dropout": self.attention_dropout,
+                "layer_norm_epsilon": self.layer_norm_epsilon,
+                "use_mha_bias": self.use_mha_bias,
+                "use_mlp_bias": self.use_mlp_bias,
+            }
+        )
+        return config

keras_hub/src/models/vit/vit_backbone_test.py

@@ -0,0 +1,37 @@
+import pytest
+from keras import ops
+
+from keras_hub.src.models.vit.vit_backbone import ViTBackbone
+from keras_hub.src.tests.test_case import TestCase
+
+
+class ViTBackboneTest(TestCase):
+    def setUp(self):
+        self.init_kwargs = {
+            "image_shape": (28, 28, 3),
+            "patch_size": 4,
+            "num_layers": 3,
+            "hidden_dim": 48,
+            "num_heads": 6,
+            "mlp_dim": 48 * 4,
+            "use_mha_bias": True,
+        }
+        self.input_size = 28
+        self.input_data = ops.ones((2, self.input_size, self.input_size, 3))
+
+    def test_backbone_basics(self):
+        self.run_backbone_test(
+            cls=ViTBackbone,
+            init_kwargs={**self.init_kwargs},
+            input_data=self.input_data,
+            expected_output_shape=(2, 50, 48),
+            run_quantization_check=False,
+        )
+
+    @pytest.mark.large
+    def test_saved_model(self):
+        self.run_model_saving_test(
+            cls=ViTBackbone,
+            init_kwargs=self.init_kwargs,
+            input_data=self.input_data,
+        )