diff --git a/python/paddle/fluid/tests/unittests/test_hsigmoid_op.py b/python/paddle/fluid/tests/unittests/test_hsigmoid_op.py
index 33c65d00e0992a7ef56f94960179ea4b7256518d..a5673e1e186d3d41e05427ad9aefad5e5fca8571 100644
--- a/python/paddle/fluid/tests/unittests/test_hsigmoid_op.py
+++ b/python/paddle/fluid/tests/unittests/test_hsigmoid_op.py
@@ -71,7 +71,7 @@ def hsigmoid(x, w, label, bias, num_classes):
     code_table = [0 for _ in range(code_length)]
     pre_output = np.zeros((batch_size, code_length))
     pre_sum = np.zeros((batch_size, 1))
-    out = np.zeros((batch_size, 1)).astype("float32")
+    out = np.zeros((batch_size, 1))
     for i in range(batch_size):
         code_table = CodeTable(num_classes, label[i])
         length = code_table.get_length()
@@ -102,6 +102,30 @@ def hsigmoid(x, w, label, bias, num_classes):
     return pre_output, out
 
 
+def hsigmoid_grad(x, w, label, bias, num_classes):
+    batch_size = x.shape[0]
+    dx = np.zeros(x.shape)
+    dw = np.zeros(w.shape)
+    db = np.zeros(bias.shape)
+    for i in range(batch_size):
+        code_table = CodeTable(num_classes, label[i])
+        length = code_table.get_length()
+        for j in range(length):
+            idx = code_table.cal_index(j)
+            t = 1 / (1 + np.exp(-(np.dot(w[idx], x[i]) + bias[idx])))
+            dx[i] = dx[i] + t * w[idx]
+            dw[idx] += t * x[i]
+            db[idx] += t
+            if code_table.cal_bit(j):
+                dx[i] = dx[i] - w[idx]
+                dw[idx] -= x[i]
+                db[idx] -= 1
+    dx /= batch_size
+    dw /= batch_size
+    db /= batch_size
+    return [dx, dw, db]
+
+
 def hsigmoidWithCustomTree(x, w, path_table, path_code, label, bias,
                            num_classes):
     batch_size = x.shape[0]
@@ -110,7 +134,7 @@ def hsigmoidWithCustomTree(x, w, path_table, path_code, label, bias,
     # init pre_out with shape [N, code_length]
     pre_output = np.zeros((batch_size, code_length))
     pre_sum = np.zeros((batch_size, 1))
-    out = np.zeros((batch_size, 1)).astype("float32")
+    out = np.zeros((batch_size, 1))
     if isinstance(bias, np.ndarray):
         for i in range(batch_size):
             code_table = CodeTableWithCustomTree(path_table, path_code, i)
@@ -145,28 +169,30 @@ def hsigmoidWithCustomTree(x, w, path_table, path_code, label, bias,
 class TestHSigmoidOp(OpTest):
     def setUp(self):
         self.op_type = "hierarchical_sigmoid"
-        num_classes = 6
-        feature_size = 8
-        batch_size = 15
-        x = np.random.random((batch_size, feature_size)).astype("float32") * 2
-        w = np.random.random(
-            (num_classes - 1, feature_size)).astype("float32") * 2
+        num_classes = 101
+        feature_size = 5
+        batch_size = 20
+        x = np.random.uniform(-1, 1, (batch_size, feature_size))
+        w = np.random.uniform(-1, 1, (num_classes - 1, feature_size))
         label = np.random.randint(0, num_classes, (batch_size, 1))
-        bias = np.random.random((num_classes - 1, 1)).astype("float32")
+        bias = np.random.uniform(-1, 1, (num_classes - 1, 1))
         self.attrs = {'num_classes': num_classes, 'is_sparse': False}
         self.inputs = {'X': x, 'W': w, 'Label': label, 'Bias': bias}
         pre_output, out = hsigmoid(x, w, label, bias, num_classes)
         self.outputs = {'PreOut': pre_output, 'Out': out}
+        self.user_grads = hsigmoid_grad(x, w, label, bias, num_classes)
 
     def test_check_output(self):
         self.check_output()
 
     def test_check_grad(self):
-        self.check_grad(['Bias', 'X', 'W'], ['Out'], no_grad_set=set('Label'))
+        self.check_grad(
+            ['X', 'W', 'Bias'], ['Out'], user_defined_grads=self.user_grads)
+        #self.check_grad(['X', 'W', 'Bias'], ['Out'])
 
 
 @skip_check_grad_ci(
-    reason="For 'TestHSigmoidOpSparse', check_grad is 'TestHSigmoidOpWithSparseGrad'."
+    reason="For 'TestHSigmoidOpSparse', check_grad is is separately calculated by 'TestHSigmoidOpWithSparseGrad'."
 )
 class TestHSigmoidOpSparse(OpTest):
     def setUp(self):
@@ -174,8 +200,8 @@ class TestHSigmoidOpSparse(OpTest):
         num_classes = 6  #using 1,2,3,4,5,6 to build a huffman tree and select 1,2,5,6 as sample
         feature_size = 8
         batch_size = 4
-        x = np.random.random((batch_size, feature_size)).astype("float32")
-        w = np.random.random((num_classes - 1, feature_size)).astype("float32")
+        x = np.random.random((batch_size, feature_size))
+        w = np.random.random((num_classes - 1, feature_size))
         label = np.array([0, 1, 4, 5])
         path_table = np.array(
             [(0, 2, -1, -1, -1), (0, 1, 3, -1, -1), (0, 1, 4, -1, -1),
@@ -183,7 +209,7 @@ class TestHSigmoidOpSparse(OpTest):
               -1)])  #np.array to store 1,2,5,6s' non-leaf path(root -> leaf)
         path_code = np.array([(0, 0, -1, -1, -1), (1, 1, 1, -1, -1), (
             1, 0, 0, -1, -1), (0, 1, -1, -1, -1)])  #np.array to store 
-        bias = np.random.random((num_classes - 1, 1)).astype("float32")
+        bias = np.random.random((num_classes - 1, 1))
         self.attrs = {'num_classes': num_classes, 'is_sparse': True}
         self.inputs = {
             'X': x,
@@ -269,15 +295,17 @@ class TestHSigmoidOpWithSparseGrad(unittest.TestCase):
         assert (dense_result == sparse_result)
 
 
+@skip_check_grad_ci(
+    reason="[skip shape check] The huffman tree is structed separately. It will be complicated if use large shape."
+)
 class TestHSigmoidOpWithCostumTree(OpTest):
     def setUp(self):
         self.op_type = "hierarchical_sigmoid"
         num_classes = 6  #using 1,2,3,4,5,6 to build a huffman tree and select 1,2,5,6 as sample
         feature_size = 8
         batch_size = 4
-        x = np.random.random((batch_size, feature_size)).astype("float32") * 2
-        w = np.random.random(
-            (num_classes - 1, feature_size)).astype("float32") * 2
+        x = np.random.uniform(-1, 1, (batch_size, feature_size))
+        w = np.random.uniform(-1, 1, (num_classes - 1, feature_size))
         label = np.array([0, 1, 4, 5])
         path_table = np.array(
             [(0, 2, -1, -1, -1), (0, 1, 3, -1, -1), (0, 1, 4, -1, -1),
@@ -285,7 +313,7 @@ class TestHSigmoidOpWithCostumTree(OpTest):
               -1)])  #np.array to store 1,2,5,6s' non-leaf path(root -> leaf)
         path_code = np.array([(0, 0, -1, -1, -1), (1, 1, 1, -1, -1), (
             1, 0, 0, -1, -1), (0, 1, -1, -1, -1)])  #np.array to store 
-        bias = np.random.random((num_classes - 1, 1)).astype("float32")
+        bias = np.random.random((num_classes - 1, 1))
         self.attrs = {'num_classes': num_classes, 'is_sparse': False}
         self.inputs = {
             'X': x,
@@ -306,15 +334,17 @@ class TestHSigmoidOpWithCostumTree(OpTest):
         self.check_grad(['Bias', 'X', 'W'], ['Out'], no_grad_set=set('Label'))
 
 
+@skip_check_grad_ci(
+    reason="[skip shape check] The huffman tree is structed separately. It will be complicated if use large shape."
+)
 class TestHSigmoidOpWithCostumTreeWithoutBias(OpTest):
     def setUp(self):
         self.op_type = "hierarchical_sigmoid"
         num_classes = 6  #using 1,2,3,4,5,6 to build a huffman tree and select 1,2,5,6 as sample
         feature_size = 8
         batch_size = 4
-        x = np.random.random((batch_size, feature_size)).astype("float32") * 2
-        w = np.random.random(
-            (num_classes - 1, feature_size)).astype("float32") * 2
+        x = np.random.uniform(-1, 1, (batch_size, feature_size))
+        w = np.random.uniform(-1, 1, (num_classes - 1, feature_size))
         label = np.array([0, 1, 4, 5])
         path_table = np.array(
             [(0, 2, -1, -1, -1), (0, 1, 3, -1, -1), (0, 1, 4, -1, -1),
diff --git a/python/paddle/fluid/tests/unittests/white_list/check_shape_white_list.py b/python/paddle/fluid/tests/unittests/white_list/check_shape_white_list.py
index 2a2f7c84fb73c05e111990f4af0db69f061526b3..a43cd8f80dde889b3c338d58809c7b16334428d0 100644
--- a/python/paddle/fluid/tests/unittests/white_list/check_shape_white_list.py
+++ b/python/paddle/fluid/tests/unittests/white_list/check_shape_white_list.py
@@ -20,7 +20,6 @@ NEED_TO_FIX_OP_LIST = [
     'conv2d_transpose',
     'depthwise_conv2d_transpose',
     'grid_sampler',
-    'hierarchical_sigmoid',
     'lstmp',
     'margin_rank_loss',
     'matmul',
diff --git a/python/paddle/fluid/tests/unittests/white_list/op_accuracy_white_list.py b/python/paddle/fluid/tests/unittests/white_list/op_accuracy_white_list.py
index e5d65db236c05b43df45caf809b75c1f4dff2d57..44e0c7b5e9008b5125f374bf076b5c8f57c8de73 100644
--- a/python/paddle/fluid/tests/unittests/white_list/op_accuracy_white_list.py
+++ b/python/paddle/fluid/tests/unittests/white_list/op_accuracy_white_list.py
@@ -32,7 +32,6 @@ NO_FP64_CHECK_GRAD_OP_LIST = [
     'depthwise_conv2d_transpose', \
     'dropout', \
     'fused_elemwise_activation', \
-    'hierarchical_sigmoid', \
     'hinge_loss', \
     'huber_loss', \
     'im2sequence', \