diff --git a/paddle/fluid/imperative/layer.cc b/paddle/fluid/imperative/layer.cc
index e65ac865b291d7591df08418ebb0144ef2e5f33c..bc03285a4c5fe6db2abf2b271d6ddc86e75a9412 100644
--- a/paddle/fluid/imperative/layer.cc
+++ b/paddle/fluid/imperative/layer.cc
@@ -117,14 +117,14 @@ class Autograd {
     std::map<OpBase*, int> dep_counts = ComputeDepCounts(var->PreOp());
 
     while (!ready.empty()) {
-      OpBase* ready_op = ready.back();
-      ready.pop_back();
+      OpBase* ready_op = ready.front();
+      ready.pop_front();
       std::map<std::string, std::vector<VarBase*>> input_grads =
           ready_op->ApplyGrad();
 
       for (auto it = input_grads.rbegin(); it != input_grads.rend(); ++it) {
         const std::vector<VarBase*>& ingrads = it->second;
-        for (int64_t i = ingrads.size() - 1; i >= 0; --i) {
+        for (size_t i = 0; i < ingrads.size(); ++i) {
           if (!ingrads[i]) continue;
           if (ready_op->input_vars_[it->first][i]->IsStopGradient()) {
             continue;
diff --git a/python/paddle/fluid/tests/unittests/test_imperative_transformer.py b/python/paddle/fluid/tests/unittests/test_imperative_transformer.py
index 32abb03e91f380b667f05154de0f5093647feff9..732f0681c4e65006628d51e083a400c0b5bd3d92 100644
--- a/python/paddle/fluid/tests/unittests/test_imperative_transformer.py
+++ b/python/paddle/fluid/tests/unittests/test_imperative_transformer.py
@@ -303,7 +303,7 @@ use_py_reader = False
 sync = False
 
 # how many batches we use
-batch_num = 2
+batch_num = 50
 
 np.random.seed = 1
 src_word_np = np.random.randint(