Update make_tensor_proto to support nested nparray values when dtype is specified.

This removes an inconsistency between graph & eager mode.  In particular, prior to this CL, the following succeeds in eager mode but fails in graph mode:

  tf.convert_to_tensor([22, np.array(1)], dtype=tf.int32)

PiperOrigin-RevId: 310212356
Change-Id: Ib836171b7ebc6ef5974c364b8ded172503120eba

tensorflow/python/framework/tensor_util.py

@@ -260,8 +260,12 @@ def _check_quantized(values):
 
 def _generate_isinstance_check(expected_types):
   def inner(values):
-    _ = [_check_failed(v) for v in nest.flatten(values)
-         if not isinstance(v, expected_types)]
+    for v in nest.flatten(values):
+      if not (isinstance(v, expected_types) or
+              (isinstance(v, np.ndarray) and
+               issubclass(v.dtype.type, expected_types))):
+        _check_failed(v)
+
   return inner
 
 _check_int = _generate_isinstance_check(

tensorflow/python/framework/tensor_util_test.py

@@ -713,6 +713,19 @@ class TensorUtilTest(test.TestCase):
     self.assertAllEqual(
         np.array([[(1 + 2j), (3 + 4j)], [(5 + 6j), (7 + 8j)]]), a)
 
+  def testNestedNumpyArrayWithoutDType(self):
+    t = tensor_util.make_tensor_proto([10.0, 20.0, np.array(30.0)])
+    a = tensor_util.MakeNdarray(t)
+    self.assertEqual(np.float32, a.dtype)
+    self.assertAllClose(np.array([10.0, 20.0, 30.0], dtype=np.float32), a)
+
+  def testNestedNumpyArrayWithDType(self):
+    t = tensor_util.make_tensor_proto([10.0, 20.0, np.array(30.0)],
+                                      dtype=dtypes.float32)
+    a = tensor_util.MakeNdarray(t)
+    self.assertEqual(np.float32, a.dtype)
+    self.assertAllClose(np.array([10.0, 20.0, 30.0], dtype=np.float32), a)
+
   def testUnsupportedDTypes(self):
     with self.assertRaises(TypeError):
       tensor_util.make_tensor_proto(np.array([1]), 0)