Fix random seed of dynamic rnn gradient check

python/paddle/v2/fluid/tests/test_dynrnn_gradient_check.py

@@ -197,7 +197,24 @@ class BaseRNN(object):
         return numpy.array([o.mean() for o in outs.itervalues()]).mean()
 
 
-class TestSimpleMul(unittest.TestCase):
+class SeedFixedTestCase(unittest.TestCase):
+    @classmethod
+    def setUpClass(cls):
+        """Fix random seeds to remove randomness from tests"""
+        cls._np_rand_state = numpy.random.get_state()
+        cls._py_rand_state = random.getstate()
+
+        numpy.random.seed(123)
+        random.seed(124)
+
+    @classmethod
+    def tearDownClass(cls):
+        """Restore random seeds"""
+        numpy.random.set_state(cls._np_rand_state)
+        random.setstate(cls._py_rand_state)
+
+
+class TestSimpleMul(SeedFixedTestCase):
     DATA_NAME = 'X'
     DATA_WIDTH = 32
     PARAM_NAME = 'W'
@@ -263,7 +280,7 @@ class TestSimpleMul(unittest.TestCase):
         self.assertTrue(numpy.allclose(i_g_num, i_g, rtol=0.05))
 
 
-class TestSimpleMulWithMemory(unittest.TestCase):
+class TestSimpleMulWithMemory(SeedFixedTestCase):
     DATA_WIDTH = 32
     HIDDEN_WIDTH = 20
     DATA_NAME = 'X'