Update davit.py

timm/models/davit.py

@@ -493,6 +493,14 @@ class DaViT(nn.Module):
         # non-normalized pyramid features + corresponding sizes
         return features, sizes
     
+    def forward_pyramid_features(self, x) -> List[Tensor]:
+        x, sizes = self.forward_network(x)
+        outs = []
+        for i, out in enumerate(x):
+            H, W = sizes[i]
+            outs.append(out.view(-1, H, W, self.embed_dims[i]).permute(0, 3, 1, 2).contiguous())
+        
+        return outs
 
     def forward_features(self, x):
         x, sizes = self.forward_network(x)
@@ -505,11 +513,14 @@ class DaViT(nn.Module):
     def forward_head(self, x, pre_logits: bool = False):
         return self.head(x, pre_logits=pre_logits)
         
-    def forward(self, x):
+    def forward_classifier(self, x):
         x = self.forward_features(x)
         x = self.forward_head(x)
         return x
         
+    def forward(self, x):
+        return self.forward_classifier(self, x)
+
 class DaViTFeatures(DaViT):
 
     def __init__(*args):
@@ -517,13 +528,8 @@ class DaViTFeatures(DaViT):
         self.feature_info = FeatureInfo(self.feature_info, kwargs.get('out_inices', (1, 2, 3, 4)))
     
     def forward(self, x) -> List[Tensor]:
-        x, sizes = self.forward_network(x)
+        return self.forward_pyramid_features(self, x)
-        outs = []
-        for i, out in enumerate(x):
-            H, W = sizes[i]
-            outs.append(out.view(-1, H, W, self.embed_dims[i]).permute(0, 3, 1, 2).contiguous())
     
-        return outs
     
 
 def checkpoint_filter_fn(state_dict, model):