google/tipsv1-s14

1"""TIPSv2 model configuration."""2 3from transformers import PretrainedConfig4 5 6_VISION_FN_BY_GEOMETRY = {7    (384, 12): "vit_small",8    (768, 12): "vit_base",9    (1024, 24): "vit_large",10    (1152, 27): "vit_so400m",11    (1536, 40): "vit_giant2",12}13 14 15class TIPSv2Config(PretrainedConfig):16    """Configuration for TIPSv2 vision-language model."""17 18    model_type = "tipsv2"19 20    def __init__(21        self,22        vision_config=None,23        text_config=None,24        temperature_init_value=0.01,25        **kwargs,26    ):27        super().__init__(**kwargs)28        vision_config = vision_config or {}29        text_config = text_config or {}30        hidden_size = vision_config.get("hidden_size", 768)31        num_hidden_layers = vision_config.get("num_hidden_layers", 12)32        self.vision_fn = _VISION_FN_BY_GEOMETRY[(hidden_size, num_hidden_layers)]33        self.embed_dim = hidden_size34        self.patch_size = vision_config.get("patch_size", 14)35        self.img_size = vision_config.get("image_size", 448)36        self.ffn_layer = "swiglu" if vision_config.get("use_swiglu_ffn", False) else "mlp"37        self.init_values = vision_config.get("layerscale_value", 1.0)38        self.num_register_tokens = vision_config.get("num_register_tokens", 1)39        self.text_hidden_size = text_config.get("hidden_size", 768)40        self.text_mlp_dim = text_config.get("intermediate_size", 3072)41        self.text_num_heads = text_config.get("num_attention_heads", 12)42        self.text_num_layers = text_config.get("num_hidden_layers", 12)43        self.vocab_size = text_config.get("vocab_size", 32000)44        self.max_len = text_config.get("max_position_embeddings", 64)45        self.temperature = temperature_init_value46