Revert `no_grad` change and add new implementation (#2513)

doc/fluid/api_cn/dygraph_cn/no_grad_cn.rst

@@ -7,8 +7,6 @@ no_grad
 .. py:class:: paddle.fluid.dygraph.no_grad
 
 :api_attr: 命令式编程模式（动态图)
-:alias_main: paddle.no_grad
-:alias: paddle.no_grad
 :old_api: paddle.fluid.dygraph.no_grad
 
 
@@ -21,21 +19,19 @@ no_grad
 ..  code-block:: python
 
     import numpy as np
-    import paddle
-
-    paddle.disable_static()
+    import paddle.fluid as fluid
 
     paddle.enable_imperative()
 
     # 用作生成器
 
     data = np.array([[2, 3], [4, 5]]).astype('float32')
-    l0 = paddle.nn.Linear(2, 2)  # l0.weight.gradient() is None
-    l1 = paddle.nn.Linear(2, 2)
-    with paddle.no_grad():
+    l0 = fluid.Linear(2, 2)  # l0.weight.gradient() is None
+    l1 = fluid.Linear(2, 2)
+    with fluid.no_grad():
         # l1.weight.stop_gradient is False
         tmp = l1.weight * 2  # tmp.stop_gradient is True
-    x = paddle.to_tensor(data)
+    x = fluid.dygraph.to_variable(data)
     y = l0(x) + tmp
     o = l1(y)
     o.backward()
@@ -43,12 +39,13 @@ no_grad
     print(l0.weight.gradient() is None)  # False
 
     # 用作装饰器
-    @paddle.no_grad()
+
+    @fluid.no_grad()
     def test_layer():
         inp = np.ones([3, 1024], dtype='float32')
-        t = paddle.to_tensor(inp)
-        linear1 = paddle.nn.Linear(1024, 4, bias_attr=False)
-        linear2 = paddle.nn.Linear(4, 4)
+        t = fluid.dygraph.base.to_variable(inp)
+        linear1 = fluid.Linear(1024, 4, bias_attr=False)
+        linear2 = fluid.Linear(4, 4)
         ret = linear1(t)
         dy_ret = linear2(ret)
 

doc/fluid/api_cn/paddle_cn/no_grad_cn.rst

-.. _cn_api_paddle_cn_name_scope:
+.. _cn_api_paddle_no_grad:
 
-name_scope
+no_grad
 -------------------------------
-:doc_source: paddle.fluid.dygraph.no_grad
+
+.. py:class:: paddle.fluid.dygraph.no_grad_
+
+:api_attr: 命令式编程模式（动态图)
+
+
+创建一个上下文来禁用动态图梯度计算。在此模式下，每次计算的结果都将具有stop_gradient=True。
+
+也可以用作一个装饰器（需要创建实例对象作为装饰器）。
+
+**代码示例**
+
+..  code-block:: python
+
+    import numpy as np
+    import paddle
+
+    paddle.disable_static()
+
+    # 用作生成器
+
+    data = np.array([[2, 3], [4, 5]]).astype('float32')
+    l0 = paddle.nn.Linear(2, 2)  # l0.weight.gradient() is None
+    l1 = paddle.nn.Linear(2, 2)
+    with paddle.no_grad():
+        # l1.weight.stop_gradient is False
+        tmp = l1.weight * 2  # tmp.stop_gradient is True
+    x = paddle.to_tensor(data)
+    y = l0(x) + tmp
+    o = l1(y)
+    o.backward()
+    print(tmp.gradient() is None)  # True
+    print(l0.weight.gradient() is None)  # False
+
+    # 用作装饰器
+    @paddle.no_grad()
+    def test_layer():
+        inp = np.ones([3, 1024], dtype='float32')
+        t = paddle.to_tensor(inp)
+        linear1 = paddle.nn.Linear(1024, 4, bias_attr=False)
+        linear2 = paddle.nn.Linear(4, 4)
+        ret = linear1(t)
+        dy_ret = linear2(ret)
+
+    test_layer()