NaViT 的速度低于预期

作者: AmitMY创建于 2025年9月9日更新于 2025年12月7日

class NaViTConfig(PretrainedConfig): model_type = "navit" def init(self, image_size: int = 256, # only used to set a default; NaViT handles var-size patch_size: int = 16, hidden_size: int = 512, dim: int = 512, depth: int = 6, heads: int = 8, mlp_dim: int = 1024, dropout: float = 0.0, emb_dropout: float = 0.0, token_dropout_prob: float = 0.1, **kwargs, ): super().init(**kwargs) self.image_size = image_size self.patch_size = patch_size self.hidden_size = hidden_size self.dim = dim self.depth = depth self.heads = heads self.mlp_dim = mlp_dim self.dropout = dropout self.emb_dropout = emb_dropout self.token_dropout_prob = token_dropout_prob

内容来源: lucidrains/vit-pytorch