lmms-lab-encoder/onevision-encoder-large

1from transformers.configuration_utils import PretrainedConfig2from transformers.utils import logging3 4 5logger = logging.get_logger(__name__)6 7 8class OneVisionEncoderConfig(PretrainedConfig):9    r"""10    This is the configuration class to store the configuration of a [`OneVisionEncoderModel`]. It is used to instantiate a11    OneVision Encoder model according to the specified arguments, defining the model architecture. Instantiating a configuration12    with the defaults will yield a similar configuration to that of the OneVision Encoder architecture.13 14    Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the15    documentation from [`PretrainedConfig`] for more information.16 17    Args:18        hidden_size (`int`, *optional*, defaults to 1024):19            Dimensionality of the encoder layers and the pooler layer.20        intermediate_size (`int`, *optional*, defaults to 4096):21            Dimensionality of the "intermediate" (i.e., feed-forward) layer in the Transformer encoder.22        num_hidden_layers (`int`, *optional*, defaults to 24):23            Number of hidden layers in the Transformer encoder.24        num_attention_heads (`int`, *optional*, defaults to 16):25            Number of attention heads for each attention layer in the Transformer encoder.26        num_channels (`int`, *optional*, defaults to 3):27            The number of input channels.28        image_size (`int`, *optional*, defaults to 224):29            The size (resolution) of each image.30        patch_size (`int`, *optional*, defaults to 14):31            The size (resolution) of each patch.32        hidden_act (`str` or `function`, *optional*, defaults to `"gelu"`):33            The non-linear activation function (function or string) in the encoder and pooler.34        layer_norm_eps (`float`, *optional*, defaults to 1e-6):35            The epsilon used by the layer normalization layers.36        layer_norm_type (`str`, *optional*, defaults to `"layer_norm"`):37            The type of layer normalization to use. Supported values: `"layer_norm"`, `"rms_norm"`.38        attention_dropout (`float`, *optional*, defaults to 0.0):39            The dropout ratio for the attention probabilities.40        initializer_range (`float`, *optional*, defaults to 0.02):41            The standard deviation of the truncated_normal_initializer for initializing all weight matrices.42        rope_theta (`float`, *optional*, defaults to 10000.0):43            The base period of the RoPE embeddings.44        use_head (`bool`, *optional*, defaults to `True`):45            Whether to use the pooling head.46 47    Example:48 49    ```python50    >>> from configuration_onevision_encoder import OneVisionEncoderConfig51    >>> from modeling_onevision_encoder import OneVisionEncoderModel52 53    >>> # Initializing a OneVisionEncoder configuration54    >>> configuration = OneVisionEncoderConfig()55 56    >>> # Initializing a model (with random weights) from the configuration57    >>> model = OneVisionEncoderModel(configuration)58 59    >>> # Accessing the model configuration60    >>> configuration = model.config61    ```62    """63 64    model_type = "onevision_encoder"65 66    def __init__(67        self,68        hidden_size=1024,69        intermediate_size=4096,70        num_hidden_layers=24,71        num_attention_heads=16,72        num_channels=3,73        image_size=448,74        patch_size=14,75        hidden_act="gelu",76        layer_norm_eps=1e-6,77        layer_norm_type="layer_norm",78        attention_dropout=0.0,79        initializer_range=0.02,80        rope_theta=10000.0,81        rope_temporal_size=64,82        use_head=True,83        **kwargs,84    ):85        super().__init__(**kwargs)86        self.hidden_size = hidden_size87        self.intermediate_size = intermediate_size88        self.num_hidden_layers = num_hidden_layers89        self.num_attention_heads = num_attention_heads90        self.num_channels = num_channels91        self.image_size = image_size92        self.patch_size = patch_size93        self.hidden_act = hidden_act94        self.layer_norm_eps = layer_norm_eps95        self.layer_norm_type = layer_norm_type96        self.attention_dropout = attention_dropout97        self.initializer_range = initializer_range98        self.rope_theta = rope_theta99        self.rope_temporal_size = rope_temporal_size  # None=use actual frames, int=fixed size (legacy: 64)100        self.use_head = use_head101