|
| 1 | +from keras import layers |
| 2 | + |
| 3 | +from keras_hub.src.api_export import keras_hub_export |
| 4 | +from keras_hub.src.models.dinov3.dinov3_layers import DINOV3Embedding |
| 5 | +from keras_hub.src.models.dinov3.dinov3_layers import DINOV3Encoder |
| 6 | +from keras_hub.src.models.dinov3.dinov3_layers import ( |
| 7 | + DINOV3RopePositionEmbedding, |
| 8 | +) |
| 9 | +from keras_hub.src.models.feature_pyramid_backbone import FeaturePyramidBackbone |
| 10 | +from keras_hub.src.utils.keras_utils import standardize_data_format |
| 11 | + |
| 12 | + |
| 13 | +@keras_hub_export("keras_hub.models.DINOV3Backbone") |
| 14 | +class DINOV3Backbone(FeaturePyramidBackbone): |
| 15 | + """DINOV3 core network with hyperparameters. |
| 16 | +
|
| 17 | + Args: |
| 18 | + patch_size: int. The size of each square patch in the input image. |
| 19 | + num_layers: int. The number of transformer layers. |
| 20 | + hidden_dim: int. The size of the transformer hidden state at the end |
| 21 | + of each transformer layer. |
| 22 | + num_heads: int. The number of attention heads for each transformer. |
| 23 | + intermediate_dim: int. The output dimension of the first Dense layer in |
| 24 | + a two-layer feedforward network for each transformer. |
| 25 | + layer_scale_init_value: float. The initial value for the layer scale in |
| 26 | + the transformer layers. Defaults to `1.0`. |
| 27 | + num_register_tokens: int. The number of register tokens to use in the |
| 28 | + embedding layer. Defaults to `0`. |
| 29 | + use_mask_token: bool. Whether to use a mask token in the embedding |
| 30 | + layer. Defaults to `True`. |
| 31 | + hidden_activation: str or callable. Activation to use in the MLP. |
| 32 | + Defaults to `"gelu"`. |
| 33 | + use_gated_mlp: bool. Whether to use Gated MLP layers. Defaults to |
| 34 | + `False`. |
| 35 | + use_query_bias: bool. Whether to use a bias for the query projection. |
| 36 | + Defaults to `True`. |
| 37 | + use_key_bias: bool. Whether to use a bias for the key projection. |
| 38 | + Defaults to `True`. |
| 39 | + use_value_bias: bool. Whether to use a bias for the value projection. |
| 40 | + Defaults to `True`. |
| 41 | + use_proj_bias: bool. Whether to use a bias for the output projection. |
| 42 | + Defaults to `True`. |
| 43 | + use_mlp_bias: bool. Whether to use a bias for the dense layers in MLP. |
| 44 | + Defaults to `True`. |
| 45 | + attention_dropout: float. The dropout rate for the attention |
| 46 | + probabilities. Defaults to `0.0`. |
| 47 | + drop_path_rate: float. The drop path rate to use. Defaults to `0.0`. |
| 48 | + image_shape: tuple. The input shape without the batch size. Defaults to |
| 49 | + `(518, 518, 3)`. |
| 50 | + rope_theta: float. The base period of the rotary position embeddings. |
| 51 | + Defaults to `100.0`. |
| 52 | + apply_layernorm: bool. Whether to apply layer normalization to the |
| 53 | + outputs of each stage in the feature pyramid. Defaults to `False`. |
| 54 | + data_format: `None` or str. If specified, either `"channels_last"` or |
| 55 | + `"channels_first"`. The ordering of the dimensions in the |
| 56 | + inputs. `"channels_last"` corresponds to inputs with shape |
| 57 | + `(batch_size, height, width, channels)` |
| 58 | + while `"channels_first"` corresponds to inputs with shape |
| 59 | + `(batch_size, channels, height, width)`. It defaults to the |
| 60 | + `image_data_format` value found in your Keras config file at |
| 61 | + `~/.keras/keras.json`. If you never set it, then it will be |
| 62 | + `"channels_last"`. |
| 63 | + dtype: string or `keras.mixed_precision.DTypePolicy`. The dtype to use |
| 64 | + for the models computations and weights. Note that some |
| 65 | + computations, such as softmax and layer normalization will always |
| 66 | + be done a float32 precision regardless of dtype. |
| 67 | +
|
| 68 | + Example: |
| 69 | + ```python |
| 70 | + # Pretrained DINOV3 model. |
| 71 | + input_data = { |
| 72 | + "images": np.ones(shape=(1, 518, 518, 3), dtype="float32"), |
| 73 | + } |
| 74 | + model = keras_hub.models.DINOV3Backbone.from_preset( |
| 75 | + "dinov3_vit_small_lvd1689m" |
| 76 | + ) |
| 77 | + model(input_data) |
| 78 | +
|
| 79 | + # Pretrained DINOV3 model with custom image shape. |
| 80 | + input_data = { |
| 81 | + "images": np.ones(shape=(1, 224, 224, 3), dtype="float32"), |
| 82 | + } |
| 83 | + model = keras_hub.models.DINOV3Backbone.from_preset( |
| 84 | + "dinov3_vit_small_lvd1689m", image_shape=(224, 224, 3) |
| 85 | + ) |
| 86 | + model(input_data) |
| 87 | +
|
| 88 | + # Randomly initialized DINOV3 model with custom config. |
| 89 | + model = keras_hub.models.DINOV3Backbone( |
| 90 | + patch_size=14, |
| 91 | + num_layers=2, |
| 92 | + hidden_dim=32, |
| 93 | + num_heads=2, |
| 94 | + intermediate_dim=128, |
| 95 | + image_shape=(224, 224, 3), |
| 96 | + ) |
| 97 | + model(input_data) |
| 98 | +
|
| 99 | + # Accessing feature pyramid outputs. |
| 100 | + backbone = keras_hub.models.DINOV3Backbone.from_preset( |
| 101 | + "dinov3_vit_small_lvd1689m", image_shape=(224, 224, 3) |
| 102 | + ) |
| 103 | + model = keras.Model( |
| 104 | + inputs=backbone.inputs, |
| 105 | + outputs=backbone.pyramid_outputs, |
| 106 | + ) |
| 107 | + features = model(input_data) |
| 108 | + ``` |
| 109 | + """ |
| 110 | + |
| 111 | + def __init__( |
| 112 | + self, |
| 113 | + patch_size, |
| 114 | + num_layers, |
| 115 | + hidden_dim, |
| 116 | + num_heads, |
| 117 | + intermediate_dim, |
| 118 | + layer_scale_init_value=1.0, |
| 119 | + num_register_tokens=4, |
| 120 | + use_mask_token=True, |
| 121 | + hidden_activation="gelu", |
| 122 | + use_gated_mlp=False, |
| 123 | + use_query_bias=True, |
| 124 | + use_key_bias=True, |
| 125 | + use_value_bias=True, |
| 126 | + use_proj_bias=True, |
| 127 | + use_mlp_bias=True, |
| 128 | + attention_dropout=0.0, |
| 129 | + drop_path_rate=0.0, |
| 130 | + layer_norm_eps=1e-5, |
| 131 | + image_shape=(518, 518, 3), |
| 132 | + rope_theta=100.0, |
| 133 | + apply_layernorm=False, |
| 134 | + data_format=None, |
| 135 | + dtype=None, |
| 136 | + name=None, |
| 137 | + **kwargs, |
| 138 | + ): |
| 139 | + data_format = standardize_data_format(data_format) |
| 140 | + |
| 141 | + prefix = str(name) + "_" if name is not None else "" |
| 142 | + |
| 143 | + # === Layers === |
| 144 | + self.embeddings = DINOV3Embedding( |
| 145 | + hidden_dim=hidden_dim, |
| 146 | + patch_size=patch_size, |
| 147 | + num_register_tokens=num_register_tokens, |
| 148 | + use_mask_token=use_mask_token, |
| 149 | + data_format=data_format, |
| 150 | + dtype=dtype, |
| 151 | + name=f"{prefix}embeddings", |
| 152 | + ) |
| 153 | + self.rope_embedding = DINOV3RopePositionEmbedding( |
| 154 | + hidden_dim=hidden_dim, |
| 155 | + num_heads=num_heads, |
| 156 | + rope_theta=rope_theta, |
| 157 | + patch_size=patch_size, |
| 158 | + dtype=dtype, |
| 159 | + name=f"{prefix}rope_embedding", |
| 160 | + ) |
| 161 | + self.encoder = DINOV3Encoder( |
| 162 | + num_layers=num_layers, |
| 163 | + hidden_dim=hidden_dim, |
| 164 | + num_heads=num_heads, |
| 165 | + intermediate_dim=intermediate_dim, |
| 166 | + layer_scale_init_value=layer_scale_init_value, |
| 167 | + hidden_activation=hidden_activation, |
| 168 | + use_gated_mlp=use_gated_mlp, |
| 169 | + use_query_bias=use_query_bias, |
| 170 | + use_key_bias=use_key_bias, |
| 171 | + use_value_bias=use_value_bias, |
| 172 | + use_proj_bias=use_proj_bias, |
| 173 | + use_mlp_bias=use_mlp_bias, |
| 174 | + attention_dropout=attention_dropout, |
| 175 | + drop_path_rate=drop_path_rate, |
| 176 | + layer_norm_eps=layer_norm_eps, |
| 177 | + dtype=dtype, |
| 178 | + name=f"{prefix}encoder", |
| 179 | + ) |
| 180 | + self.layernorm = layers.LayerNormalization( |
| 181 | + epsilon=layer_norm_eps, dtype=dtype, name=f"{prefix}layernorm" |
| 182 | + ) |
| 183 | + |
| 184 | + # === Functional Model === |
| 185 | + pyramid_outputs = {} |
| 186 | + image_input = layers.Input(shape=image_shape, name="pixel_values") |
| 187 | + x = self.embeddings(image_input) |
| 188 | + pyramid_outputs["stem"] = x |
| 189 | + |
| 190 | + position_embeddings = self.rope_embedding(image_input) |
| 191 | + num_prefix_tokens = 1 + num_register_tokens |
| 192 | + |
| 193 | + x, encoder_pyramid_outputs = self.encoder( |
| 194 | + x, |
| 195 | + position_embeddings=position_embeddings, |
| 196 | + num_prefix_tokens=num_prefix_tokens, |
| 197 | + ) |
| 198 | + pyramid_outputs.update(encoder_pyramid_outputs) |
| 199 | + x = self.layernorm(x) |
| 200 | + if apply_layernorm: |
| 201 | + for key in pyramid_outputs: |
| 202 | + pyramid_outputs[key] = self.layernorm(pyramid_outputs[key]) |
| 203 | + outputs = x |
| 204 | + super().__init__( |
| 205 | + inputs={"pixel_values": image_input}, |
| 206 | + outputs=outputs, |
| 207 | + dtype=dtype, |
| 208 | + name=name, |
| 209 | + **kwargs, |
| 210 | + ) |
| 211 | + |
| 212 | + # === Config === |
| 213 | + self.patch_size = int(patch_size) |
| 214 | + self.num_layers = int(num_layers) |
| 215 | + self.hidden_dim = int(hidden_dim) |
| 216 | + self.num_heads = int(num_heads) |
| 217 | + self.intermediate_dim = int(intermediate_dim) |
| 218 | + self.layer_scale_init_value = float(layer_scale_init_value) |
| 219 | + self.num_register_tokens = int(num_register_tokens) |
| 220 | + self.use_mask_token = bool(use_mask_token) |
| 221 | + self.hidden_activation = hidden_activation |
| 222 | + self.use_gated_mlp = bool(use_gated_mlp) |
| 223 | + self.use_query_bias = bool(use_query_bias) |
| 224 | + self.use_key_bias = bool(use_key_bias) |
| 225 | + self.use_value_bias = bool(use_value_bias) |
| 226 | + self.use_proj_bias = bool(use_proj_bias) |
| 227 | + self.use_mlp_bias = bool(use_mlp_bias) |
| 228 | + self.attention_dropout = float(attention_dropout) |
| 229 | + self.drop_path_rate = float(drop_path_rate) |
| 230 | + self.layer_norm_eps = float(layer_norm_eps) |
| 231 | + self.image_shape = image_shape |
| 232 | + self.rope_theta = rope_theta |
| 233 | + self.apply_layernorm = apply_layernorm |
| 234 | + self.pyramid_outputs = pyramid_outputs |
| 235 | + |
| 236 | + def get_config(self): |
| 237 | + config = super().get_config() |
| 238 | + config.update( |
| 239 | + { |
| 240 | + "patch_size": self.patch_size, |
| 241 | + "num_layers": self.num_layers, |
| 242 | + "hidden_dim": self.hidden_dim, |
| 243 | + "num_heads": self.num_heads, |
| 244 | + "intermediate_dim": self.intermediate_dim, |
| 245 | + "num_register_tokens": self.num_register_tokens, |
| 246 | + "use_mask_token": self.use_mask_token, |
| 247 | + "layer_scale_init_value": self.layer_scale_init_value, |
| 248 | + "hidden_activation": self.hidden_activation, |
| 249 | + "use_gated_mlp": self.use_gated_mlp, |
| 250 | + "use_query_bias": self.use_query_bias, |
| 251 | + "use_key_bias": self.use_key_bias, |
| 252 | + "use_value_bias": self.use_value_bias, |
| 253 | + "use_proj_bias": self.use_proj_bias, |
| 254 | + "use_mlp_bias": self.use_mlp_bias, |
| 255 | + "attention_dropout": self.attention_dropout, |
| 256 | + "drop_path_rate": self.drop_path_rate, |
| 257 | + "layer_norm_eps": self.layer_norm_eps, |
| 258 | + "image_shape": self.image_shape, |
| 259 | + "rope_theta": self.rope_theta, |
| 260 | + "apply_layernorm": self.apply_layernorm, |
| 261 | + } |
| 262 | + ) |
| 263 | + return config |
0 commit comments