configuration_siglip2.py

from transformers.configuration_utils import PretrainedConfig
from transformers.utils import logging


logger = logging.get_logger(__name__)


class Siglip2TextConfig(PretrainedConfig):
    r"""
    Args:
        vocab_size (`int`, *optional*, defaults to 32000):
            Vocabulary size of the Siglip2 text model. Defines the number of different tokens that can be represented by
            the `inputs_ids` passed when calling [`Siglip2Model`].
        hidden_size (`int`, *optional*, defaults to 768):
            Dimensionality of the encoder layers and the pooler layer.
        intermediate_size (`int`, *optional*, defaults to 3072):
            Dimensionality of the "intermediate" (i.e., feed-forward) layer in the Transformer encoder.
        num_hidden_layers (`int`, *optional*, defaults to 12):
            Number of hidden layers in the Transformer encoder.
        num_attention_heads (`int`, *optional*, defaults to 12):
            Number of attention heads for each attention layer in the Transformer encoder.
        max_position_embeddings (`int`, *optional*, defaults to 64):
            The maximum sequence length that this model might ever be used with. Typically set this to something large
            just in case (e.g., 512 or 1024 or 2048).
        hidden_act (`str` or `function`, *optional*, defaults to `"gelu_pytorch_tanh"`):
            The non-linear activation function (function or string) in the encoder and pooler. If string, `"gelu"`,
            `"relu"`, `"selu"` and `"gelu_new"` `"quick_gelu"` are supported.
        layer_norm_eps (`float`, *optional*, defaults to 1e-06):
            The epsilon used by the layer normalization layers.
        attention_dropout (`float`, *optional*, defaults to 0.0):
            The dropout ratio for the attention probabilities.
        pad_token_id (`int`, *optional*, defaults to 1):
            The id of the padding token in the vocabulary.
        bos_token_id (`int`, *optional*, defaults to 49406):
            The id of the beginning-of-sequence token in the vocabulary.
        eos_token_id (`int`, *optional*, defaults to 49407):
            The id of the end-of-sequence token in the vocabulary.
        projection_size (`int`, *optional*, defaults to `hidden_size`):
            The size of the projection head.

    """

    model_type = "siglip2_text_model"
    base_config_key = "text_config"

    def __init__(
        self,
        vocab_size=32000,
        hidden_size=768,
        intermediate_size=3072,
        num_hidden_layers=12,
        num_attention_heads=12,
        max_position_embeddings=64,
        hidden_act="gelu_pytorch_tanh",
        layer_norm_eps=1e-6,
        attention_dropout=0.0,
        pad_token_id=1,
        bos_token_id=49406,
        eos_token_id=49407,
        projection_size=None,
        **kwargs,
    ):
        super().__init__(pad_token_id=pad_token_id, bos_token_id=bos_token_id, eos_token_id=eos_token_id, **kwargs)

        self.vocab_size = vocab_size
        self.hidden_size = hidden_size
        self.intermediate_size = intermediate_size
        self.num_hidden_layers = num_hidden_layers
        self.num_attention_heads = num_attention_heads
        self.max_position_embeddings = max_position_embeddings
        self.layer_norm_eps = layer_norm_eps
        self.hidden_act = hidden_act
        self.attention_dropout = attention_dropout
        self.projection_size = projection_size if projection_size is not None else hidden_size


class Siglip2VisionConfig(PretrainedConfig):
    r"""
    Args:
        hidden_size (`int`, *optional*, defaults to 768):
            Dimensionality of the encoder layers and the pooler layer.
        intermediate_size (`int`, *optional*, defaults to 3072):
            Dimensionality of the "intermediate" (i.e., feed-forward) layer in the Transformer encoder.
        num_hidden_layers (`int`, *optional*, defaults to 12):
            Number of hidden layers in the Transformer encoder.
        num_attention_heads (`int`, *optional*, defaults to 12):
            Number of attention heads for each attention layer in the Transformer encoder.
        num_channels (`int`, *optional*, defaults to 3):
            Number of channels in the input images.
        num_patches (`int`, *optional*, defaults to 256):
            The number of patches in the image with the size of (`patch_size`, `patch_size`).
            The image is resized to fill maximum of this number of patches, and to preserve
            the aspect ratio. In case the resulted number of patches is lower, the image is
            padded in "patch" dimension.
        patch_size (`int`, *optional*, defaults to 16):
            The size (resolution) of each patch.
        hidden_act (`str` or `function`, *optional*, defaults to `"gelu_pytorch_tanh"`):
            The non-linear activation function (function or string) in the encoder and pooler. If string, `"gelu"`,
            `"relu"`, `"selu"` and `"gelu_new"` `"quick_gelu"` are supported.
        layer_norm_eps (`float`, *optional*, defaults to 1e-06):
            The epsilon used by the layer normalization layers.
        attention_dropout (`float`, *optional*, defaults to 0.0):
            The dropout ratio for the attention probabilities.

    """

    model_type = "siglip2_vision_model"
    base_config_key = "vision_config"

    def __init__(
        self,
        hidden_size=768,
        out_hidden_size=2048,
        intermediate_size=3072,
        num_hidden_layers=12,
        num_attention_heads=12,
        num_channels=3,
        num_patches=256,
        patch_size=16,
        hidden_act="gelu_pytorch_tanh",
        layer_norm_eps=1e-6,
        attention_dropout=0.0,
        **kwargs,
    ):
        super().__init__(**kwargs)

        self.hidden_size = hidden_size
        self.out_hidden_size = out_hidden_size
        self.intermediate_size = intermediate_size
        self.num_hidden_layers = num_hidden_layers
        self.num_attention_heads = num_attention_heads
        self.num_channels = num_channels
        self.patch_size = patch_size
        self.attention_dropout = attention_dropout
        self.layer_norm_eps = layer_norm_eps
        self.hidden_act = hidden_act
        self.num_patches = num_patches
        self.in_features = -1


class Siglip2Config(PretrainedConfig):
    r"""
    Args:
        text_config (`dict`, *optional*):
            Dictionary of configuration options used to initialize [`Siglip2TextConfig`].
        vision_config (`dict`, *optional*):
            Dictionary of configuration options used to initialize [`Siglip2VisionConfig`].
        kwargs (*optional*):
            Dictionary of keyword arguments.

    """

    model_type = "siglip2"
    sub_configs = {"text_config": Siglip2TextConfig, "vision_config": Siglip2VisionConfig}

    def __init__(self, text_config=None, vision_config=None, **kwargs):
        super().__init__(**kwargs)

        if text_config is None:
            text_config = {}
            logger.info("`text_config` is `None`. Initializing the `Siglip2TextConfig` with default values.")

        if vision_config is None:
            vision_config = {}
            logger.info("`vision_config` is `None`. initializing the `Siglip2VisionConfig` with default values.")

        self.text_config = Siglip2TextConfig(**text_config)
        self.vision_config = Siglip2VisionConfig(**vision_config)

        self.initializer_factor = 1.0

    @classmethod
    def from_text_vision_configs(cls, text_config: Siglip2TextConfig, vision_config: Siglip2VisionConfig, **kwargs):

        return cls(text_config=text_config.to_dict(), vision_config=vision_config.to_dict(), **kwargs)


__all__ = ["Siglip2Config", "Siglip2TextConfig", "Siglip2VisionConfig"]