Revise danet.py (#1507)

The dropout layer should be placed before the classification layer.

gluoncv/model_zoo/danet.py

@@ -90,16 +90,16 @@ def __init__(self, in_channels, out_channels, norm_layer=nn.BatchNorm, norm_kwar
         self.conv52.add(nn.Activation('relu'))
 
         self.conv6 = nn.HybridSequential()
-        self.conv6.add(nn.Conv2D(in_channels=512, channels=out_channels, kernel_size=1))
         self.conv6.add(nn.Dropout(0.1))
+        self.conv6.add(nn.Conv2D(in_channels=512, channels=out_channels, kernel_size=1))
 
         self.conv7 = nn.HybridSequential()
-        self.conv7.add(nn.Conv2D(in_channels=512, channels=out_channels, kernel_size=1))
         self.conv7.add(nn.Dropout(0.1))
+        self.conv7.add(nn.Conv2D(in_channels=512, channels=out_channels, kernel_size=1))
 
         self.conv8 = nn.HybridSequential()
-        self.conv8.add(nn.Conv2D(in_channels=512, channels=out_channels, kernel_size=1))
         self.conv8.add(nn.Dropout(0.1))
+        self.conv8.add(nn.Conv2D(in_channels=512, channels=out_channels, kernel_size=1))
 
     def hybrid_forward(self, F, x):
         feat1 = self.conv5a(x)
@@ -191,4 +191,4 @@ def get_danet_resnet101_citys(**kwargs):
     >>> model = get_danet_resnet101_citys(pretrained=True)
     >>> print(model)
     """
-    return get_danet('citys', 'resnet101', **kwargs)
+    return get_danet('citys', 'resnet101', **kwargs)