diff --git a/chainer/links/model/vision/googlenet.py b/chainer/links/model/vision/googlenet.py
index 32702f996907..7749968593dc 100644
--- a/chainer/links/model/vision/googlenet.py
+++ b/chainer/links/model/vision/googlenet.py
@@ -10,6 +10,7 @@
     available = False
     _import_error = e
 
+import chainer
 from chainer.dataset.convert import concat_examples
 from chainer.dataset import download
 from chainer import function
@@ -301,7 +302,7 @@ def predict(self, images, oversample=True):
         else:
             x = x[:, :, 16:240, 16:240]
         # Use no_backprop_mode to reduce memory consumption
-        with function.no_backprop_mode():
+        with function.no_backprop_mode(), chainer.using_config('train', False):
             x = Variable(self.xp.asarray(x))
             y = self(x, layers=['prob'])['prob']
             if oversample:
diff --git a/chainer/links/model/vision/resnet.py b/chainer/links/model/vision/resnet.py
index a8565dac051e..9283c868f80d 100644
--- a/chainer/links/model/vision/resnet.py
+++ b/chainer/links/model/vision/resnet.py
@@ -10,6 +10,7 @@
     available = False
     _import_error = e
 
+import chainer
 from chainer.dataset.convert import concat_examples
 from chainer.dataset import download
 from chainer import function
@@ -263,7 +264,7 @@ def predict(self, images, oversample=True):
         else:
             x = x[:, :, 16:240, 16:240]
         # Use no_backprop_mode to reduce memory consumption
-        with function.no_backprop_mode():
+        with function.no_backprop_mode(), chainer.using_config('train', False):
             x = Variable(self.xp.asarray(x))
             y = self(x, layers=['prob'])['prob']
             if oversample:
diff --git a/chainer/links/model/vision/vgg.py b/chainer/links/model/vision/vgg.py
index 4ed48b5d2c32..742af0c76854 100644
--- a/chainer/links/model/vision/vgg.py
+++ b/chainer/links/model/vision/vgg.py
@@ -10,6 +10,7 @@
     available = False
     _import_error = e
 
+import chainer
 from chainer.dataset.convert import concat_examples
 from chainer.dataset import download
 from chainer import function
@@ -266,7 +267,7 @@ def predict(self, images, oversample=True):
         else:
             x = x[:, :, 16:240, 16:240]
         # Use no_backprop_mode to reduce memory consumption
-        with function.no_backprop_mode():
+        with function.no_backprop_mode(), chainer.using_config('train', False):
             x = Variable(self.xp.asarray(x))
             y = self(x, layers=['prob'])['prob']
             if oversample: