gradient_checker: Convert arguments to tensors.

Without this change, the test added to gradient_checker_v2_test.py
would fail with an error like:

AttributeError: 'numpy.ndarray' object has no attribute '_id'

in the call to "tape.watch(x)".

Since gradient_checker_v2.compute_gradients() requires that the "x" argument be
a list of Tensors, it seems reasonable to ensure that the "f" argument is
provided with Tensors.

While at it, also check that the "x" argument is a list to provide a clearer
error message than would be obtained without this check.

PiperOrigin-RevId: 223820537

tensorflow/python/ops/gradient_checker_v2.py

@@ -88,14 +88,18 @@ def _prepare(f, xs_dtypes):
     a function that will be evaluated in both graph and eager mode
   """
   if context.executing_eagerly():
-    return f
+
+    def decorated_eager(*xs_data):
+      return f(*map(ops.convert_to_tensor, xs_data))
+
+    return decorated_eager
   xs = [array_ops.placeholder(x_dtype) for x_dtype in xs_dtypes]
   y = f(*xs)
   sess = ops.get_default_session()
-  def decorated(*xs_data):
+  def decorated_graph(*xs_data):
     xs_data = [_to_numpy(a) for a in xs_data]
     return sess.run(y, feed_dict=dict(zip(xs, xs_data)))
-  return decorated
+  return decorated_graph
 
 
 def _compute_theoretical_jacobian(f, y_shape, y_dtype, xs, param):
@@ -288,6 +292,9 @@ def compute_gradient(f, x, delta=1e-3):
   Raises:
     ValueError: If result is empty but the gradient is nonzero.
   """
+  if not isinstance(x, list):
+    raise ValueError(
+        "`x` must be a list of Tensors (arguments to `f`), not a %s" % type(x))
   return _compute_gradient_list(f, x, delta)
 
 

tensorflow/python/ops/gradient_checker_v2_test.py

@@ -21,6 +21,7 @@ from __future__ import print_function
 
 import numpy as np
 
+from tensorflow.python.eager import backprop
 from tensorflow.python.eager import context
 from tensorflow.python.framework import constant_op
 from tensorflow.python.framework import dtypes
@@ -177,6 +178,19 @@ class GradientCheckerTest(test.TestCase):
     with self.assertRaisesRegexp(AssertionError, "False is not true"):
       self.assertTrue(error < 1.0)
 
+  def testGradGrad(self):
+
+    def f(x):
+      with backprop.GradientTape() as tape:
+        tape.watch(x)
+        y = math_ops.square(x)
+        z = math_ops.square(y)
+      return tape.gradient(z, x)
+
+    analytical, numerical = gradient_checker.compute_gradient(f, [2.0])
+    self.assertAllEqual([[[48.]]], analytical)
+    self.assertAllClose([[[48.]]], numerical, rtol=1e-4)
+
 
 @test_util.run_all_in_graph_and_eager_modes
 class MiniMNISTTest(test.TestCase):