mix_vision_transformer_config.py

from transformers import PretrainedConfig

class MySegformerConfig(PretrainedConfig):
    model_type = "my_segformer"

    def __init__(
        self,
        in_channels=3,
        embed_dims=[32, 64, 160, 256],
        num_stages=4,
        num_layers=[2, 2, 2, 2],
        num_heads=[1, 2, 5, 8],
        patch_sizes=[7, 3, 3, 3],
        strides=[4, 2, 2, 2],
        sr_ratios=[8, 4, 2, 1],
        mlp_ratio=4,
        qkv_bias=True,
        drop_rate=0.0,
        attn_drop_rate=0.0,
        drop_path_rate=0.1,
        out_indices=(0, 1, 2, 3),
        num_classes=19,
        decode_head=None,
        **kwargs
    ):
        super().__init__(**kwargs)

        # Falls embed_dims als int angegeben ist, Liste erzwingen
        if isinstance(embed_dims, int):
            embed_dims = [embed_dims]

        self.in_channels = in_channels
        self.embed_dims = embed_dims
        self.num_stages = num_stages
        self.num_layers = num_layers
        self.num_heads = num_heads
        self.patch_sizes = patch_sizes
        self.strides = strides
        self.sr_ratios = sr_ratios
        self.mlp_ratio = mlp_ratio
        self.qkv_bias = qkv_bias
        self.drop_rate = drop_rate
        self.attn_drop_rate = attn_drop_rate
        self.drop_path_rate = drop_path_rate
        self.out_indices = out_indices
        self.num_classes = num_classes

        # optionaler Block für Head-Config (falls gebraucht)
        self.decode_head = decode_head