diff --git a/paddle/fluid/imperative/layer.h b/paddle/fluid/imperative/layer.h
index 46107341a4edc3ec6bb5569c3fa50bbd9243bd26..78205486c5534ac0c61cc6d545bdafa4dfc95695 100644
--- a/paddle/fluid/imperative/layer.h
+++ b/paddle/fluid/imperative/layer.h
@@ -152,11 +152,13 @@ class VarBase {
 
   void ClearGradient() {
     VLOG(1) << "clear gradient of " << var_desc_->Name();
-    auto grads_t = grads_->var_->GetMutable<framework::LoDTensor>();
-    operators::math::set_constant(
-        *(platform::DeviceContextPool::Instance().Get(
-            grads_->var_->Get<framework::LoDTensor>().place())),
-        grads_t, 0.0);
+    if (grads_ && grads_->var_ && grads_->var_->IsInitialized()) {
+      auto grads_t = grads_->var_->GetMutable<framework::LoDTensor>();
+      operators::math::set_constant(
+          *(platform::DeviceContextPool::Instance().Get(
+              grads_->var_->Get<framework::LoDTensor>().place())),
+          grads_t, 0.0);
+    }
   }
 
   framework::LoDTensor& GradValue();
diff --git a/python/paddle/fluid/imperative/layers.py b/python/paddle/fluid/imperative/layers.py
index c338c65a76b86cd28ee7cc8e38a422c4e382a9de..71ff95bdea36967c1fa6b5c94cc7ca305e7a544a 100644
--- a/python/paddle/fluid/imperative/layers.py
+++ b/python/paddle/fluid/imperative/layers.py
@@ -52,8 +52,7 @@ class Layer(core.Layer):
 
     def clear_gradients(self):
         for p in self.parameters():
-            if not p._stop_gradient:
-                p._clear_gradient()
+            p._clear_gradient()
 
     def _build_once(self, inputs):
         pass