Fix pit implementation to be clsoer to deit/levit re distillation head handling

timm/models/deit.py

@@ -72,7 +72,7 @@ class VisionTransformerDistilled(VisionTransformer):
         self.dist_token = nn.Parameter(torch.zeros(1, 1, self.embed_dim))
         self.pos_embed = nn.Parameter(torch.zeros(1, self.patch_embed.num_patches + self.num_tokens, self.embed_dim))
         self.head_dist = nn.Linear(self.embed_dim, self.num_classes) if self.num_classes > 0 else nn.Identity()
-        self.distilled_training = False
+        self.distilled_training = False  # must set this True to train w/ distillation token
 
         self.init_weights(weight_init)
 

timm/models/levit.py

@@ -539,7 +539,7 @@ class LevitDistilled(Levit):
     def __init__(self, *args, **kwargs):
         super().__init__(*args, **kwargs)
         self.head_dist = NormLinear(self.num_features, self.num_classes) if self.num_classes > 0 else nn.Identity()
-        self.distilled_training = False
+        self.distilled_training = False  # must set this True to train w/ distillation token
 
     @torch.jit.ignore
     def get_classifier(self):

timm/models/pit.py

@@ -147,9 +147,10 @@ class PoolingVisionTransformer(nn.Module):
     A PyTorch implement of 'Rethinking Spatial Dimensions of Vision Transformers'
         - https://arxiv.org/abs/2103.16302
     """
-    def __init__(self, img_size, patch_size, stride, base_dims, depth, heads,
-                 mlp_ratio, num_classes=1000, in_chans=3, distilled=False, global_pool='token',
-                 attn_drop_rate=.0, drop_rate=.0, drop_path_rate=.0):
+    def __init__(
+            self, img_size, patch_size, stride, base_dims, depth, heads,
+            mlp_ratio, num_classes=1000, in_chans=3, global_pool='token',
+            distilled=False, attn_drop_rate=.0, drop_rate=.0, drop_path_rate=.0):
         super(PoolingVisionTransformer, self).__init__()
         assert global_pool in ('token',)
 
@@ -193,6 +194,7 @@ class PoolingVisionTransformer(nn.Module):
         self.head_dist = None
         if distilled:
             self.head_dist = nn.Linear(self.embed_dim, self.num_classes) if num_classes > 0 else nn.Identity()
+        self.distilled_training = False  # must set this True to train w/ distillation token
 
         trunc_normal_(self.pos_embed, std=.02)
         trunc_normal_(self.cls_token, std=.02)
@@ -207,6 +209,10 @@ class PoolingVisionTransformer(nn.Module):
     def no_weight_decay(self):
         return {'pos_embed', 'cls_token'}
 
+    @torch.jit.ignore
+    def set_distilled_training(self, enable=True):
+        self.distilled_training = enable
+
     @torch.jit.ignore
     def set_grad_checkpointing(self, enable=True):
         assert not enable, 'gradient checkpointing not supported'
@@ -231,16 +237,30 @@ class PoolingVisionTransformer(nn.Module):
         cls_tokens = self.norm(cls_tokens)
         return cls_tokens
 
-    def forward(self, x):
-        x = self.forward_features(x)
+    def forward_head(self, x, pre_logits: bool = False) -> torch.Tensor:
         if self.head_dist is not None:
-            x, x_dist = self.head(x[:, 0]), self.head_dist(x[:, 1])  # x must be a tuple
-            if self.training and not torch.jit.is_scripting():
+            assert self.global_pool == 'token'
+            x, x_dist = x[:, 0], x[:, 1]
+            if not pre_logits:
+                x = self.head(x)
+                x_dist = self.head_dist(x_dist)
+            if self.distilled_training and self.training and not torch.jit.is_scripting():
+                # only return separate classification predictions when training in distilled mode
                 return x, x_dist
             else:
+                # during standard train / finetune, inference average the classifier predictions
                 return (x + x_dist) / 2
         else:
-            return self.head(x[:, 0])
+            if self.global_pool == 'token':
+                x = x[:, 0]
+            if not pre_logits:
+                x = self.head(x)
+            return x
+
+    def forward(self, x):
+        x = self.forward_features(x)
+        x = self.forward_head(x)
+        return x
 
 
 def checkpoint_filter_fn(state_dict, model):