Fix self.reset_classifier num_classes update

timm/models/davit.py

@@ -633,6 +633,7 @@ class DaVit(nn.Module):
         return self.head.fc
 
     def reset_classifier(self, num_classes: int, global_pool: Optional[str] = None):
+        self.num_classes = num_classes
         self.head.reset(num_classes, global_pool)
 
     def forward_features(self, x):

timm/models/focalnet.py

@@ -455,6 +455,7 @@ class FocalNet(nn.Module):
         return self.head.fc
 
     def reset_classifier(self, num_classes: int, global_pool: Optional[str] = None):
+        self.num_classes = num_classes
         self.head.reset(num_classes, pool_type=global_pool)
 
     def forward_features(self, x):

timm/models/metaformer.py

@@ -584,6 +584,7 @@ class MetaFormer(nn.Module):
         return self.head.fc
 
     def reset_classifier(self, num_classes: int, global_pool: Optional[str] = None):
+        self.num_classes = num_classes
         if global_pool is not None:
             self.head.global_pool = SelectAdaptivePool2d(pool_type=global_pool)
             self.head.flatten = nn.Flatten(1) if global_pool else nn.Identity()

timm/models/nextvit.py

@@ -557,6 +557,7 @@ class NextViT(nn.Module):
         return self.head.fc
 
     def reset_classifier(self, num_classes: int, global_pool: Optional[str] = None):
+        self.num_classes = num_classes
         self.head.reset(num_classes, pool_type=global_pool)
 
     def forward_features(self, x):

timm/models/nfnet.py

@@ -434,6 +434,7 @@ class NormFreeNet(nn.Module):
         return self.head.fc
 
     def reset_classifier(self, num_classes: int, global_pool: Optional[str] = None):
+        self.num_classes = num_classes
         self.head.reset(num_classes, global_pool)
 
     def forward_features(self, x):

timm/models/pvt_v2.py

@@ -384,7 +384,7 @@ class PyramidVisionTransformerV2(nn.Module):
         if global_pool is not None:
             assert global_pool in ('avg', '')
             self.global_pool = global_pool
-        self.head = nn.Linear(self.embed_dim, num_classes) if num_classes > 0 else nn.Identity()
+        self.head = nn.Linear(self.num_features, num_classes) if num_classes > 0 else nn.Identity()
 
     def forward_features(self, x):
         x = self.patch_embed(x)

timm/models/rdnet.py

@@ -349,6 +349,7 @@ class RDNet(nn.Module):
         return self.head.fc
 
     def reset_classifier(self, num_classes: int, global_pool: Optional[str] = None):
+        self.num_classes = num_classes
         self.head.reset(num_classes, global_pool)
 
     def forward_features(self, x):

timm/models/regnet.py

@@ -515,6 +515,7 @@ class RegNet(nn.Module):
         return self.head.fc
 
     def reset_classifier(self, num_classes: int, global_pool: Optional[str] = None):
+        self.num_classes = num_classes
         self.head.reset(num_classes, pool_type=global_pool)
 
     def forward_intermediates(

timm/models/tresnet.py

@@ -225,6 +225,7 @@ class TResNet(nn.Module):
         return self.head.fc
 
     def reset_classifier(self, num_classes: int, global_pool: Optional[str] = None):
+        self.num_classes = num_classes
         self.head.reset(num_classes, pool_type=global_pool)
 
     def forward_features(self, x):

timm/models/vision_transformer_sam.py

@@ -537,6 +537,7 @@ class VisionTransformerSAM(nn.Module):
         return self.head
 
     def reset_classifier(self, num_classes: int, global_pool: Optional[str] = None):
+        self.num_classes = num_classes
         self.head.reset(num_classes, global_pool)
 
     def forward_intermediates(

timm/models/xception_aligned.py

@@ -275,6 +275,7 @@ class XceptionAligned(nn.Module):
         return self.head.fc
 
     def reset_classifier(self, num_classes: int, global_pool: Optional[str] = None):
+        self.num_classes = num_classes
         self.head.reset(num_classes, pool_type=global_pool)
 
     def forward_features(self, x):