Anpassungen Config

mix_vision_transformer_config.py

@@ -26,7 +26,6 @@ class MySegformerConfig(PretrainedConfig):
         super().__init__(**kwargs)
         self.embed_dims = embed_dims  # einfach int behalten
         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
@@ -41,5 +40,16 @@ class MySegformerConfig(PretrainedConfig):
         self.out_indices = out_indices
         self.num_classes = num_classes
 
-        # optionaler Block für Head-Config (falls gebraucht)
+        # optionaler Block für Head-Config (falls decode_head nicht übergeben wurde)
+        if decode_head is None:
+            decode_head = {
+                "in_channels": self.embed_dims,          # [32, 64, 160, 256]
+                "in_index": list(range(self.num_stages)),# [0, 1, 2, 3]
+                "channels": 256,
+                "dropout_ratio": 0.1,
+                "out_channels": self.num_classes,
+                "align_corners": False,
+                "interpolate_mode": "bilinear"
+            }
+
         self.decode_head = decode_head