Update rw models, fix heads

timm/models/mambaout.py

@@ -8,10 +8,10 @@ InceptionNeXt (https://github.com/sail-sg/inceptionnext)
 from typing import Optional
 
 import torch
-import torch.nn as nn
+from torch import nn
 
 from timm.data import IMAGENET_DEFAULT_MEAN, IMAGENET_DEFAULT_STD
-from timm.layers import trunc_normal_, DropPath, LayerNorm, LayerScale
+from timm.layers import trunc_normal_, DropPath, LayerNorm, LayerScale, ClNormMlpClassifierHead
 from ._builder import build_model_with_cfg
 from ._manipulate import checkpoint_seq
 from ._registry import register_model
@@ -122,6 +122,7 @@ class MlpHead(nn.Module):
             self,
             dim,
             num_classes=1000,
+            pool_type='avg',
             act_layer=nn.GELU,
             mlp_ratio=4,
             norm_layer=LayerNorm,
@@ -130,17 +131,25 @@ class MlpHead(nn.Module):
     ):
         super().__init__()
         hidden_features = int(mlp_ratio * dim)
+        self.pool_type = pool_type
+
+        self.norm1 = norm_layer(dim)
         self.fc1 = nn.Linear(dim, hidden_features, bias=bias)
         self.act = act_layer()
-        self.norm = norm_layer(hidden_features)
+        self.norm2 = norm_layer(hidden_features)
         self.fc2 = nn.Linear(hidden_features, num_classes, bias=bias)
         self.head_dropout = nn.Dropout(drop_rate)
 
-    def forward(self, x):
+    def forward(self, x, pre_logits: bool = False):
+        if self.pool_type == 'avg':
+            x = x.mean((1, 2))
+        x = self.norm1(x)
         x = self.fc1(x)
         x = self.act(x)
-        x = self.norm(x)
+        x = self.norm2(x)
         x = self.head_dropout(x)
+        if pre_logits:
+            return x
         x = self.fc2(x)
         return x
 
@@ -208,7 +217,7 @@ class MambaOutStage(nn.Module):
             expansion_ratio=8 / 3,
             kernel_size=7,
             conv_ratio=1.0,
-            downsample: bool = False,
+            downsample: str = '',
             ls_init_value: Optional[float] = None,
             norm_layer=LayerNorm,
             act_layer=nn.GELU,
@@ -218,8 +227,10 @@ class MambaOutStage(nn.Module):
         dim_out = dim_out or dim
         self.grad_checkpointing = False
 
-        if downsample:
+        if downsample == 'conv':
             self.downsample = Downsample(dim, dim_out, norm_layer=norm_layer)
+        elif downsample == 'conv_nf':
+            self.downsample = DownsampleNormFirst(dim, dim_out, norm_layer=norm_layer)
         else:
             assert dim == dim_out
             self.downsample = nn.Identity()
@@ -276,10 +287,10 @@ class MambaOut(nn.Module):
             kernel_size=7,
             stem_mid_norm=True,
             ls_init_value=None,
+            downsample='conv',
             drop_path_rate=0.,
             drop_rate=0.,
-            output_norm=LayerNorm,
-            head_fn=MlpHead,
+            head_fn='default',
             **kwargs,
     ):
         super().__init__()
@@ -312,7 +323,7 @@ class MambaOut(nn.Module):
                 depth=depths[i],
                 kernel_size=kernel_size,
                 conv_ratio=conv_ratio,
-                downsample=i > 0,
+                downsample=downsample if i > 0 else '',
                 ls_init_value=ls_init_value,
                 norm_layer=norm_layer,
                 act_layer=act_layer,
@@ -322,9 +333,25 @@ class MambaOut(nn.Module):
             prev_dim = dim
             cur += depths[i]
 
-        self.norm = output_norm(prev_dim)
-
-        self.head = head_fn(prev_dim, num_classes, drop_rate=drop_rate)
+        if head_fn == 'default':
+            # specific to this model, unusual norm -> pool -> fc -> act -> norm -> fc combo
+            self.head = MlpHead(
+                prev_dim,
+                num_classes,
+                pool_type='avg',
+                drop_rate=drop_rate,
+                norm_layer=norm_layer,
+            )
+        else:
+            # more typical norm -> pool -> fc -> act -> fc
+            self.head = ClNormMlpClassifierHead(
+                prev_dim,
+                num_classes,
+                hidden_size=int(prev_dim * 4),
+                pool_type='avg',
+                norm_layer=norm_layer,
+                drop_rate=drop_rate,
+            )
 
         self.apply(self._init_weights)
 
@@ -336,7 +363,7 @@ class MambaOut(nn.Module):
 
     @torch.jit.ignore
     def no_weight_decay(self):
-        return {'norm'}
+        return {}
 
     def forward_features(self, x):
         x = self.stem(x)
@@ -345,9 +372,7 @@ class MambaOut(nn.Module):
         return x
 
     def forward_head(self, x, pre_logits: bool = False):
-        x = x.mean((1, 2))
-        x = self.norm(x)
-        x = self.head(x)
+        x = self.head(x, pre_logits=pre_logits) if pre_logits else self.head(x)
         return x
 
     def forward(self, x):
@@ -366,6 +391,10 @@ def checkpoint_filter_fn(state_dict, model):
         k = k.replace('downsample_layers.0.', 'stem.')
         k = re.sub(r'stages.([0-9]+).([0-9]+)', r'stages.\1.blocks.\2', k)
         k = re.sub(r'downsample_layers.([0-9]+)', r'stages.\1.downsample', k)
+        if k.startswith('norm.'):
+            k = k.replace('norm.', 'head.norm1.')
+        elif k.startswith('head.norm.'):
+            k = k.replace('head.norm.', 'head.norm2.')
         out_dict[k] = v
 
     return out_dict
@@ -443,7 +472,9 @@ def mambaout_small_rw(pretrained=False, **kwargs):
         depths=[3, 4, 27, 3],
         dims=[96, 192, 384, 576],
         stem_mid_norm=False,
+        downsample='conv_nf',
         ls_init_value=1e-6,
+        head_fn='norm_mlp',
     )
     return _create_mambaout('mambaout_small_rw', pretrained=pretrained, **dict(model_args, **kwargs))
 
@@ -455,5 +486,6 @@ def mambaout_base_rw(pretrained=False, **kwargs):
         dims=(128, 256, 512, 768),
         stem_mid_norm=False,
         ls_init_value=1e-6,
+        head_fn='norm_mlp',
     )
     return _create_mambaout('mambaout_base_rw', pretrained=pretrained, **dict(model_args, **kwargs))