diff --git a/paddle/fluid/operators/cross_entropy_op.cc b/paddle/fluid/operators/cross_entropy_op.cc
index 66986e14d8f4a25c57c2c90c422171564874239f..8a80619f6636f9f0cab1d0b6332ca05742b9e7f8 100644
--- a/paddle/fluid/operators/cross_entropy_op.cc
+++ b/paddle/fluid/operators/cross_entropy_op.cc
@@ -136,8 +136,8 @@ class CrossEntropyGradientOpBase : public framework::OperatorWithKernel {
                       "Input(Y@Grad) and Input(Y) should have the same rank.");
 
     bool check = true;
-    if ((!ctx->IsRuntime()) && (framework::product(x_dims) <= 0 ||
-                                framework::product(label_dims) <= 0)) {
+    if ((!ctx->IsRuntime()) &&
+        (framework::product(x_dims) <= 0 || framework::product(dy_dims) <= 0)) {
       check = false;
     }
 
diff --git a/python/paddle/fluid/backward.py b/python/paddle/fluid/backward.py
index 1306fbc574e28bfb5619b1504c197e1d90778ddb..860e7092f6958c477668695bf28aec03d876aa9f 100644
--- a/python/paddle/fluid/backward.py
+++ b/python/paddle/fluid/backward.py
@@ -1251,7 +1251,7 @@ def calc_gradient(targets, inputs, target_gradients=None, no_grad_set=None):
             op_desc = _create_op_desc_("fill_constant",
                                        {"ShapeTensor": [target_shape.name]},
                                        {"Out": [grad_name]}, {
-                                           "shape": [],
+                                           "shape": target.shape,
                                            "value": 1.0,
                                            "dtype": target.dtype,
                                        })