GradientTape.jacobian/batch_jacobian: allow calls inside the GradientTape's scope

Since they call GradientTape.gradient eventually anyway, they'll inherit the warning about persistent tapes recording themselves taking extra memory.

Fixes #34260.
Fixes #41510.

PiperOrigin-RevId: 357251843
Change-Id: I913d5dee058b01a1e3fbb830c246d4e7a2133cd5

tensorflow/python/eager/backprop.py

@@ -873,6 +873,18 @@ class GradientTape(object):
     tape.pop_tape(self._tape)
     self._recording = False
 
+  @tf_contextlib.contextmanager
+  def _ensure_recording(self):
+    """Ensures that this tape is recording."""
+    if not self._recording:
+      try:
+        self._push_tape()
+        yield
+      finally:
+        self._pop_tape()
+    else:
+      yield
+
   def watch(self, tensor):
     """Ensures that `tensor` is being traced by this tape.
 
@@ -1144,14 +1156,12 @@ class GradientTape(object):
     target_shape = array_ops.shape(target)
     # Note that we push and pop the tape here and below. This is needed since we
     # need gradients through the enclosed operations.
-    self._push_tape()
+    with self._ensure_recording():
-    target = array_ops.reshape(target, [-1])
+      target = array_ops.reshape(target, [-1])
-    self._pop_tape()
 
     def loop_fn(i):
-      self._push_tape()
+      with self._ensure_recording():
-      y = array_ops.gather(target, i)
+        y = array_ops.gather(target, i)
-      self._pop_tape()
       return self.gradient(y, flat_sources,
                            unconnected_gradients=unconnected_gradients)
 
@@ -1285,16 +1295,14 @@ class GradientTape(object):
     # Flatten target to 2-D.
     # Note that we push and pop the tape here and below. This is needed since we
     # need gradients through the enclosed operations.
-    self._push_tape()
+    with self._ensure_recording():
-    with ops.control_dependencies(
+      with ops.control_dependencies(
-        [check_ops.assert_equal(batch_size, source_shape[0])]):
+          [check_ops.assert_equal(batch_size, source_shape[0])]):
-      target = array_ops.reshape(target, [batch_size, target_row_size])
+        target = array_ops.reshape(target, [batch_size, target_row_size])
-    self._pop_tape()
 
     def loop_fn(i):
-      self._push_tape()
+      with self._ensure_recording():
-      y = array_ops.gather(target, i, axis=1)
+        y = array_ops.gather(target, i, axis=1)
-      self._pop_tape()
       return self.gradient(y, source,
                            unconnected_gradients=unconnected_gradients)
 

tensorflow/python/eager/backprop_test.py

@@ -854,6 +854,24 @@ class BackpropTest(test.TestCase, parameterized.TestCase):
         RuntimeError, 'A non-persistent GradientTape can only'):
       g.jacobian(y, [x])
 
+  @test_util.assert_no_new_tensors
+  def testJacobianInsideGradientTapeScope(self):
+    with backprop.GradientTape() as g:
+      x = constant_op.constant(3.0)
+      g.watch(x)
+      y = x * x
+      z = y * y
+      self.assertAllClose(4. * 3. ** 3., g.jacobian(z, x))
+
+  @test_util.assert_no_new_tensors
+  def testBatchJacobianInsideGradientTapeScope(self):
+    with backprop.GradientTape(persistent=True) as g:
+      x = constant_op.constant([[3.0]])
+      g.watch(x)
+      y = x * x
+      z = y * y
+      self.assertAllClose([[[4. * 3. ** 3.]]], g.batch_jacobian(z, x))
+
   @test_util.assert_no_new_tensors
   def testGradientTapeBatchJacobianCalledMultipleTimes(self):
     with backprop.GradientTape() as g: