Fix tf.test.compute_gradients to exercise imaginary parts

Previously, if x_init_value was not specified, it was set to a random real
value.  This is bad if the dtype is complex, since the imaginary part is left
zero.
Change: 148368889

tensorflow/python/ops/gradient_checker.py

@@ -218,13 +218,9 @@ def _compute_gradient(x,
         x_shape, i_shape)
     x_data = x_init_value
   else:
-    if t == dtypes.float16:
-      dtype = np.float16
-    elif t == dtypes.float32:
-      dtype = np.float32
-    else:
-      dtype = np.float64
-    x_data = np.asfarray(np.random.random_sample(x_shape), dtype=dtype)
+    x_data = np.random.random_sample(x_shape).astype(t.as_numpy_dtype)
+    if t.is_complex:
+      x_data.imag = np.random.random_sample(x_shape)
 
   jacob_t = _compute_theoretical_jacobian(
       x, x_shape, x_data, dy, y_shape, dx, extra_feed_dict=extra_feed_dict)

tensorflow/python/ops/gradient_checker_test.py

@@ -153,7 +153,7 @@ def testComplexConj(self):
                                                                 size)
       correct = np.array([[1, 0], [0, -1]])
       self.assertAllEqual(correct, analytical)
-      self.assertAllClose(correct, numerical, rtol=3e-6)
+      self.assertAllClose(correct, numerical, rtol=2e-5)
       self.assertLess(
           gradient_checker.compute_gradient_error(x, size, y, size), 2e-5)