configuration_selectivevit.py

# configuration_my_model.py
from transformers import PretrainedConfig

class SMSelectiveViTConfig(PretrainedConfig):
    model_type = "softmasked_selective_vit"

    def __init__(
        self,
        image_size=224,
        patch_size=16,
        num_classes=1000,
        embed_dim=768,
        atten_dim=768,
        depth=12,
        num_heads=12,
        mlp_dim=3072,
        channels=3,
        dropout=0.0,
        drop_path=0.0,
        attention_scale=0.0,
        mask_threshold=0.0,
        patch_drop=0.0,
        use_distil_token=False,
        **kwargs,
    ):
        super().__init__(**kwargs)
        # store everything as attributes (HF will save them in config.json)
        self.image_size = image_size
        self.patch_size = patch_size
        self.num_classes = num_classes
        self.embed_dim = embed_dim
        self.atten_dim = atten_dim
        self.depth = depth
        self.num_heads = num_heads
        self.mlp_dim = mlp_dim
        self.channels = channels
        self.dropout = dropout
        self.drop_path = drop_path
        self.attention_scale = attention_scale
        self.mask_threshold = mask_threshold
        self.patch_drop = patch_drop
        self.use_distil_token = use_distil_token