ffn + Dropout

mmpose/models/backbones/vit_sam.py

@@ -259,7 +259,7 @@ class CustomAttentionFFN(nn.Module):
             nn.Linear(dim, dim * 4),
             nn.GELU(),
             nn.Linear(dim * 4, dim),
-            nn.DropPath(proj_drop)
+            nn.Dropout(proj_drop)
         )
         self.norm1 = nn.LayerNorm(dim)
         self.norm2 = nn.LayerNorm(dim)
@@ -330,8 +330,8 @@ class ViTSam(BaseBackbone):
             param.requires_grad = False
 
         # 交叉注意力
-        # self.cross_attn = Cross_Attention(embed_dim, num_heads=num_heads, qkv_bias=qkv_bias, \
+        self.cross_attn = Cross_Attention(embed_dim, num_heads=num_heads, qkv_bias=qkv_bias, \
-        #     qk_scale=qk_scale, attn_drop=attn_drop_rate, proj_drop=drop_rate)
+            qk_scale=qk_scale, attn_drop=attn_drop_rate, proj_drop=drop_rate)
 
         # vit_token做自注意力后，再和sam_token做交叉注意力，得到的结果再经过FFN
         # self.custom_attn_ffn = CustomAttentionFFN(embed_dim, num_heads=num_heads, qkv_bias=qkv_bias, \
@@ -342,7 +342,7 @@ class ViTSam(BaseBackbone):
             nn.Linear(embed_dim, embed_dim * 4),
             nn.GELU(),
             nn.Linear(embed_dim * 4, embed_dim),
-            nn.DropPath(drop_rate)
+            nn.Dropout(drop_rate)
         )
         self.sam_norm = norm_layer(embed_dim)