Anpassungen an den COnfigs

config.json

@@ -1,17 +1,20 @@
 {
+  "model_type": "my_segformer",
   "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],
-  "out_indices": [0, 1, 2, 3],
   "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": {
     "in_channels": [32, 64, 160, 256],
     "in_index": [0, 1, 2, 3],

mix_vision_transformer_config.py

@@ -5,10 +5,11 @@ class MySegformerConfig(PretrainedConfig):
 
     def __init__(
         self,
-        embed_dims=[64, 128, 320, 512],
+        in_channels=3,
+        embed_dims=[32, 64, 160, 256],
         num_stages=4,
-        num_layers=[3, 4, 6, 3],
-        num_heads=[1, 2, 4, 8],
+        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],
@@ -16,16 +17,19 @@ class MySegformerConfig(PretrainedConfig):
         qkv_bias=True,
         drop_rate=0.0,
         attn_drop_rate=0.0,
-        drop_path_rate=0.0,
+        drop_path_rate=0.1,
         out_indices=(0, 1, 2, 3),
+        num_classes=19,
+        decode_head=None,
         **kwargs
     ):
         super().__init__(**kwargs)
 
-        # Absicherung, falls embed_dims als int übergeben wird
+        # 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
@@ -39,4 +43,7 @@ class MySegformerConfig(PretrainedConfig):
         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