Refactor vit pooling to add more reduction options, separately callable

Author: rwightman

timm/models/vision_transformer.py

@@ -386,6 +386,31 @@ def forward(self, x: torch.Tensor) -> torch.Tensor:
             return self._forward(x)
 
 
+def global_pool_nlc(
+        x: torch.Tensor,
+        pool_type: str = 'token',
+        num_prefix_tokens: int = 1,
+        reduce_include_prefix: bool = False,
+):
+    if not pool_type:
+        return x
+
+    if pool_type == 'token':
+        x = x[:, 0]  # class token
+    else:
+        x = x if reduce_include_prefix else x[:, num_prefix_tokens:]
+        if pool_type == 'avg':
+            x = x.mean(dim=1)
+        elif pool_type == 'avgmax':
+            x = 0.5 * (x.amax(dim=1) + x.mean(dim=1))
+        elif pool_type == 'max':
+            x = x.amax(dim=1)
+        else:
+            assert not pool_type, f'Unknown pool type {pool_type}'
+
+    return x
+
+
 class VisionTransformer(nn.Module):
     """ Vision Transformer
 
@@ -400,7 +425,7 @@ def __init__(
             patch_size: Union[int, Tuple[int, int]] = 16,
             in_chans: int = 3,
             num_classes: int = 1000,
-            global_pool: Literal['', 'avg', 'max', 'token', 'map'] = 'token',
+            global_pool: Literal['', 'avg', 'avgmax', 'max', 'token', 'map'] = 'token',
             embed_dim: int = 768,
             depth: int = 12,
             num_heads: int = 12,
@@ -459,10 +484,10 @@ def __init__(
             block_fn: Transformer block layer.
         """
         super().__init__()
-        assert global_pool in ('', 'avg', 'max', 'token', 'map')
+        assert global_pool in ('', 'avg', 'avgmax', 'max', 'token', 'map')
         assert class_token or global_pool != 'token'
         assert pos_embed in ('', 'none', 'learn')
-        use_fc_norm = global_pool in ['avg', 'max'] if fc_norm is None else fc_norm
+        use_fc_norm = global_pool in ('avg', 'avgmax', 'max') if fc_norm is None else fc_norm
         norm_layer = get_norm_layer(norm_layer) or partial(nn.LayerNorm, eps=1e-6)
         act_layer = get_act_layer(act_layer) or nn.GELU
 
@@ -596,10 +621,10 @@ def set_grad_checkpointing(self, enable: bool = True) -> None:
     def get_classifier(self) -> nn.Module:
         return self.head
 
-    def reset_classifier(self, num_classes: int, global_pool = None) -> None:
+    def reset_classifier(self, num_classes: int, global_pool: Optional[str] = None):
         self.num_classes = num_classes
         if global_pool is not None:
-            assert global_pool in ('', 'avg', 'token', 'map')
+            assert global_pool in ('', 'avg', 'avgmax', 'max', 'token', 'map')
             if global_pool == 'map' and self.attn_pool is None:
                 assert False, "Cannot currently add attention pooling in reset_classifier()."
             elif global_pool != 'map ' and self.attn_pool is not None:
@@ -756,15 +781,16 @@ def forward_features(self, x: torch.Tensor) -> torch.Tensor:
         x = self.norm(x)
         return x
 
-    def forward_head(self, x: torch.Tensor, pre_logits: bool = False) -> torch.Tensor:
+    def pool(self, x: torch.Tensor, pool_type: Optional[str] = None) -> torch.Tensor:
         if self.attn_pool is not None:
             x = self.attn_pool(x)
-        elif self.global_pool == 'avg':
-            x = x[:, self.num_prefix_tokens:].mean(dim=1)
-        elif self.global_pool == 'max':
-            x, _ = torch.max(x[:, self.num_prefix_tokens:], dim=1)
-        elif self.global_pool:
-            x = x[:, 0]  # class token
+            return x
+        pool_type = self.global_pool if pool_type is None else pool_type
+        x = global_pool_nlc(x, pool_type=pool_type, num_prefix_tokens=self.num_prefix_tokens)
+        return x
+
+    def forward_head(self, x: torch.Tensor, pre_logits: bool = False) -> torch.Tensor:
+        x = self.pool(x)
         x = self.fc_norm(x)
         x = self.head_drop(x)
         return x if pre_logits else self.head(x)