Update davit.py

timm/models/davit.py

@@ -411,32 +411,35 @@ class DaViT(nn.Module):
         for stage_id, stage_param in enumerate(self.architecture):
             layer_offset_id = len(list(itertools.chain(*self.architecture[:stage_id])))
 
-            stage = nn.Sequential([
+            stage = nn.Sequential(
-                nn.Sequential([
+                nn.ModuleList([
-                    ChannelBlock(
+                    nn.Sequential(
-                        dim=self.embed_dims[item],
+                        nn.ModuleList([
-                        num_heads=self.num_heads[item],
+                            ChannelBlock(
-                        mlp_ratio=mlp_ratio,
+                                dim=self.embed_dims[item],
-                        qkv_bias=qkv_bias,
+                                num_heads=self.num_heads[item],
-                        drop_path=dpr[2 * (layer_id + layer_offset_id) + attention_id],
+                                mlp_ratio=mlp_ratio,
-                        norm_layer=nn.LayerNorm,
+                                qkv_bias=qkv_bias,
-                        ffn=ffn,
+                                drop_path=dpr[2 * (layer_id + layer_offset_id) + attention_id],
-                        cpe_act=cpe_act
+                                norm_layer=nn.LayerNorm,
-                    ) if attention_type == 'channel' else
+                                ffn=ffn,
-                    SpatialBlock(
+                                cpe_act=cpe_act
-                        dim=self.embed_dims[item],
+                            ) if attention_type == 'channel' else
-                        num_heads=self.num_heads[item],
+                            SpatialBlock(
-                        mlp_ratio=mlp_ratio,
+                                dim=self.embed_dims[item],
-                        qkv_bias=qkv_bias,
+                                num_heads=self.num_heads[item],
-                        drop_path=dpr[2 * (layer_id + layer_offset_id) + attention_id],
+                                mlp_ratio=mlp_ratio,
-                        norm_layer=nn.LayerNorm,
+                                qkv_bias=qkv_bias,
-                        ffn=ffn,
+                                drop_path=dpr[2 * (layer_id + layer_offset_id) + attention_id],
-                        cpe_act=cpe_act,
+                                norm_layer=nn.LayerNorm,
-                        window_size=window_size,
+                                ffn=ffn,
-                    ) if attention_type == 'spatial' else None
+                                cpe_act=cpe_act,
-                    for attention_id, attention_type in enumerate(attention_types)]
+                                window_size=window_size,
-                ) for layer_id, item in enumerate(stage_param)
+                            ) if attention_type == 'spatial' else None
-            ])
+                        for attention_id, attention_type in enumerate(attention_types)])
+                    ) for layer_id, item in enumerate(stage_param)
+                ])
+            )
             
             self.stages.add_module(f'stage_{stage_id}', stage)