diff --git a/dynet/lstm.cc b/dynet/lstm.cc
index 55aef93c0..ec0f51919 100644
--- a/dynet/lstm.cc
+++ b/dynet/lstm.cc
@@ -93,17 +93,17 @@ void LSTMBuilder::start_new_sequence_impl(const vector<Expression>& hinit) {
       float retention_rate = 1.f - dropout_rate;
       float scale = 1.f / retention_rate;
       // in
-      masks_i.push_back(random_bernoulli(*_cg,{ idim}, retention_rate) * scale);
-      masks_i.push_back(random_bernoulli(*_cg,{ idim}, retention_rate) * scale);
-      masks_i.push_back(random_bernoulli(*_cg,{ idim}, retention_rate) * scale);
+      masks_i.push_back(random_bernoulli(*_cg,{ idim}, retention_rate, scale));
+      masks_i.push_back(random_bernoulli(*_cg,{ idim}, retention_rate, scale));
+      masks_i.push_back(random_bernoulli(*_cg,{ idim}, retention_rate, scale));
       // h
-      masks_i.push_back(random_bernoulli(*_cg,{ hidden_dim}, retention_rate) * scale);
-      masks_i.push_back(random_bernoulli(*_cg,{ hidden_dim}, retention_rate) * scale);
-      masks_i.push_back(random_bernoulli(*_cg,{ hidden_dim}, retention_rate) * scale);
+      masks_i.push_back(random_bernoulli(*_cg,{ hidden_dim}, retention_rate, scale));
+      masks_i.push_back(random_bernoulli(*_cg,{ hidden_dim}, retention_rate, scale));
+      masks_i.push_back(random_bernoulli(*_cg,{ hidden_dim}, retention_rate, scale));
       // c
-      masks_i.push_back(random_bernoulli(*_cg,{ hidden_dim}, retention_rate) * scale);
-      masks_i.push_back(random_bernoulli(*_cg,{ hidden_dim}, retention_rate) * scale);
-      masks_i.push_back(random_bernoulli(*_cg,{ hidden_dim}, retention_rate) * scale);
+      masks_i.push_back(random_bernoulli(*_cg,{ hidden_dim}, retention_rate, scale));
+      masks_i.push_back(random_bernoulli(*_cg,{ hidden_dim}, retention_rate, scale));
+      masks_i.push_back(random_bernoulli(*_cg,{ hidden_dim}, retention_rate, scale));
       masks.push_back(masks_i);
     }
   }