diff --git a/src/opr/impl/dnn/batch_norm.cpp b/src/opr/impl/dnn/batch_norm.cpp
index d58c17f6fd47804fc434b77767932438030ae3eb..73fd6956bcbe6fecf41a1adb84b7b374b797800d 100644
--- a/src/opr/impl/dnn/batch_norm.cpp
+++ b/src/opr/impl/dnn/batch_norm.cpp
@@ -50,8 +50,9 @@ BatchNormForward::BatchNormForward(VarNode *x,
             mgb_throw_if(!(dest_opr->same_type<SharedDeviceTensor>() ||
                     dest_opr->same_type<VolatileSharedDeviceTensor>()),
                     GraphError,
-                    "mean and variance in BatchNorm must be SharedDeviceTensor "
-                    "or VolatileSharedDeviceTensor; got %s{%s} actually",
+                    "mean and variance in training mode BatchNorm must be",
+                    "SharedDeviceTensor or VolatileSharedDeviceTensor;",
+                    "got %s{%s} actually",
                     dest_opr->cname(), dest_opr->dyn_typeinfo()->name);
         };
         check_dest(mean);