ANpassung modeling

modeling_my_segformer.py

@@ -2,9 +2,9 @@ import torch
 import torch.nn as nn
 from transformers import PreTrainedModel
 
-from segformer_plusplus.model.backbone.mit import MixVisionTransformer   # Backbone
-from mix_vision_transformer_config import MySegformerConfig             # Config
-from segformer_plusplus.model.head.segformer_head import SegformerHead # <-- dein Head
+from segformer_plusplus.model.backbone.mit import MixVisionTransformer
+from mix_vision_transformer_config import MySegformerConfig
+from segformer_plusplus.model.head.segformer_head import SegformerHead
 
 
 class MySegformerForSemanticSegmentation(PreTrainedModel):
@@ -16,8 +16,7 @@ class MySegformerForSemanticSegmentation(PreTrainedModel):
 
         # Backbone (MixVisionTransformer)
         self.backbone = MixVisionTransformer(
-            # Pass only the first element of embed_dims for the initial patch embedding
-            embed_dims=config.embed_dims[0],   # <--- KORRIGIERTE ZEILE
+            embed_dims=config.embed_dims[0],
             num_stages=config.num_stages,
             num_layers=config.num_layers,
             num_heads=config.num_heads,
@@ -32,31 +31,31 @@ class MySegformerForSemanticSegmentation(PreTrainedModel):
             out_indices=config.out_indices
         )
 
-        # Head direkt importieren
+        # Head initialization
         # Use config.decode_head.in_channels directly, as it's defined in the config.
-        # This ensures consistency with the backbone's output channels for the head.
         in_channels = config.decode_head["in_channels"]
 
         self.segmentation_head = SegformerHead(
-            in_channels=in_channels,                  # Liste der Embeddings aus Backbone
-            in_index=list(config.out_indices),        # welche Feature Maps genutzt werden
-            channels=config.decode_head["channels"],   # channels parameter for SegformerHead itself
+            in_channels=in_channels,
+            in_index=list(config.out_indices),
+            channels=config.decode_head["channels"],
             dropout_ratio=config.decode_head["dropout_ratio"],
-            num_classes=getattr(config, "num_classes", 19), # Ensure num_classes is passed if not directly in decode_head config
-            align_corners=config.decode_head["align_corners"]
+            # REMOVED: num_classes=getattr(config, "num_classes", 19), <--- Entfernen Sie diese Zeile
+            align_corners=config.decode_head["align_corners"],
+            # Fügen Sie hier interpolate_mode hinzu, falls SegformerHead dies explizit erwartet
+            # und es in config.decode_head definiert ist (was es in Ihrer config.json ist)
+            interpolate_mode=config.decode_head["interpolate_mode"] # <-- Hinzugefügt
         )
 
         self.post_init()
 
     def forward(self, x):
-        # Backbone → Features (Liste von Tensors)
         features = self.backbone(x)
 
         # Debug: Ausgabe der Shapes der Backbone-Features
         for i, f in enumerate(features):
             print(f"Feature {i}: shape = {f.shape}")
 
-        # Head → logits
         logits = self.segmentation_head(features)
 
         return logits