Anpassungen am Head, damit bias akzeptiert

modeling_my_segformer.py

@@ -32,7 +32,6 @@ class MySegformerForSemanticSegmentation(PreTrainedModel):
         )
 
         # Head initialization
-        # Use config.decode_head.in_channels directly, as it's defined in the config.
         in_channels = config.decode_head["in_channels"]
 
         self.segmentation_head = SegformerHead(
@@ -40,11 +39,9 @@ class MySegformerForSemanticSegmentation(PreTrainedModel):
             in_index=list(config.out_indices),
             channels=config.decode_head["channels"],
             dropout_ratio=config.decode_head["dropout_ratio"],
-            # 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
+            interpolate_mode=config.decode_head["interpolate_mode"],
+            use_conv_bias_in_convmodules=False
         )
 
         self.post_init()

segformer_plusplus/model/head/segformer_head.py

@@ -18,6 +18,10 @@ class SegformerHead(BaseModule):
     Args:
         interpolate_mode: The interpolate mode of MLP head upsample operation.
             Default: 'bilinear'.
+        use_conv_bias_in_convmodules (bool | str): If True, ConvModules will use bias.
+            If False, they won't. If 'auto', they follow ConvModule's default.
+            This is added for compatibility with models trained with no conv bias
+            when followed by BatchNorm, while keeping default local behavior.
     """
 
     def __init__(self,
@@ -28,7 +32,9 @@ class SegformerHead(BaseModule):
                  out_channels=19,
                  norm_cfg=None,
                  align_corners=False,
-                 interpolate_mode='bilinear'):
+                 interpolate_mode='bilinear',
+                 use_conv_bias_in_convmodules: bool | str = 'auto'
+                 ):
         super().__init__()
 
         self.in_channels = in_channels
@@ -39,6 +45,7 @@ class SegformerHead(BaseModule):
         self.norm_cfg = norm_cfg
         self.align_corners = align_corners
         self.interpolate_mode = interpolate_mode
+        self.use_conv_bias_in_convmodules = use_conv_bias_in_convmodules # Speichern des neuen Parameters
 
         self.act_cfg = dict(type='ReLU')
         self.conv_seg = nn.Conv2d(channels, self.out_channels, kernel_size=1)
@@ -49,9 +56,15 @@ class SegformerHead(BaseModule):
 
         num_inputs = len(self.in_channels)
 
-        print(num_inputs)
-        print(in_index)
-        #assert num_inputs == len(self.in_index)
+        conv_module_bias_setting = use_conv_bias_in_convmodules
+        if use_conv_bias_in_convmodules == 'auto':
+            pass
+        elif isinstance(use_conv_bias_in_convmodules, bool):
+            # Wenn True/False explizit übergeben wird, verwenden wir das
+            conv_module_bias_setting = use_conv_bias_in_convmodules
+        else:
+            raise ValueError("use_conv_bias_in_convmodules must be 'auto', True, or False")
+
 
         self.convs = nn.ModuleList()
         for i in range(num_inputs):
@@ -62,13 +75,17 @@ class SegformerHead(BaseModule):
                     kernel_size=1,
                     stride=1,
                     norm_cfg=self.norm_cfg,
-                    act_cfg=self.act_cfg))
+                    act_cfg=self.act_cfg,
+                    bias=conv_module_bias_setting # Verwende den bestimmten Bias-Wert
+                ))
 
         self.fusion_conv = ConvModule(
             in_channels=self.channels * num_inputs,
             out_channels=self.channels,
             kernel_size=1,
-            norm_cfg=self.norm_cfg)
+            norm_cfg=self.norm_cfg,
+            bias=conv_module_bias_setting # Verwende den bestimmten Bias-Wert
+        )
 
     def cls_seg(self, feat):
         """Classify each pixel."""
@@ -94,4 +111,4 @@ class SegformerHead(BaseModule):
 
         out = self.cls_seg(out)
 
-        return out
+        return out