Better error message from trying to differentiate PreventGradient.

Also fix small lint error in gradients_impl.py
Change: 146822514

7 files changed, 21 insertions, 7 deletions

diff --git a/tensorflow/python/kernel_tests/ctc_loss_op_test.py b/tensorflow/python/kernel_tests/ctc_loss_op_test.py
index 434faefafa..5b93f90a79 100644
--- a/tensorflow/python/kernel_tests/ctc_loss_op_test.py
+++ b/tensorflow/python/kernel_tests/ctc_loss_op_test.py
@@ -257,7 +257,7 @@ class CTCLossTest(test.TestCase):
       # Taking ths second gradient should fail, since it is not
       # yet supported.
       with self.assertRaisesRegexp(LookupError,
-                                   ".*No gradient defined.*PreventGradient.*"):
+                                   "explicitly disabled"):
         _ = gradients_impl._hessian_vector_product(loss, [inputs_t], v)
 
 
diff --git a/tensorflow/python/kernel_tests/sparse_xent_op_test.py b/tensorflow/python/kernel_tests/sparse_xent_op_test.py
index bff9ac4ec0..cd5b711a0e 100644
--- a/tensorflow/python/kernel_tests/sparse_xent_op_test.py
+++ b/tensorflow/python/kernel_tests/sparse_xent_op_test.py
@@ -214,7 +214,7 @@ class SparseXentTest(test.TestCase):
     # Taking ths second gradient should fail, since it is not
     # yet supported.
     with self.assertRaisesRegexp(LookupError,
-                                 ".*No gradient defined.*PreventGradient.*"):
+                                 "explicitly disabled"):
       _ = gradients_impl.hessians(loss, [weights])
 
   def _testHighDim(self, features, labels):
diff --git a/tensorflow/python/kernel_tests/xent_op_test.py b/tensorflow/python/kernel_tests/xent_op_test.py
index c8c4e2bd1e..d037ceac61 100644
--- a/tensorflow/python/kernel_tests/xent_op_test.py
+++ b/tensorflow/python/kernel_tests/xent_op_test.py
@@ -191,7 +191,7 @@ class XentTest(test.TestCase):
     # Taking ths second gradient should fail, since it is not
     # yet supported.
     with self.assertRaisesRegexp(LookupError,
-                                 ".*No gradient defined.*PreventGradient.*"):
+                                 "explicitly disabled"):
       _ = gradients_impl.hessians(loss, [f])
 
   def testWrapper(self):
diff --git a/tensorflow/python/ops/array_grad.py b/tensorflow/python/ops/array_grad.py
index 50aab4fad5..cb545f9fca 100644
--- a/tensorflow/python/ops/array_grad.py
+++ b/tensorflow/python/ops/array_grad.py
@@ -343,6 +343,12 @@ def _FillGrad(_, grad):
 ops.NotDifferentiable("ZerosLike")
 
 
+@ops.RegisterGradient("PreventGradient")
+def _PreventGradientGrad(op, _):
+  raise LookupError(
+      "Gradient explicitly disabled. Reason: %s" % op.get_attr("message"))
+
+
 @ops.RegisterGradient("Gather")
 def _GatherGrad(op, grad):
   """Gradient for Gather op."""
diff --git a/tensorflow/python/ops/ctc_ops.py b/tensorflow/python/ops/ctc_ops.py
index 1ce5597e13..b0a1fc3dd1 100644
--- a/tensorflow/python/ops/ctc_ops.py
+++ b/tensorflow/python/ops/ctc_ops.py
@@ -165,7 +165,10 @@ def _CTCLossGrad(op, grad_loss, _):
   # due to the fused implementation's interaction with tf.gradients(),
   # so we make sure we prevent silently incorrect results by raising
   # an error if the second derivative is requested via prevent_gradient.
-  grad_without_gradient = array_ops.prevent_gradient(op.outputs[1])
+  grad_without_gradient = array_ops.prevent_gradient(
+      op.outputs[1], message="Currently there is no way to take the second "
+      " derivative of ctc_loss due to the fused implementation's interaction "
+      " with tf.gradients()")
   # Return gradient for inputs and None for
   # labels_indices, labels_values and sequence_length
   return [_BroadcastMul(grad_loss, grad_without_gradient), None, None, None]
diff --git a/tensorflow/python/ops/gradients_test.py b/tensorflow/python/ops/gradients_test.py
index 453313b4ac..cfd463283d 100644
--- a/tensorflow/python/ops/gradients_test.py
+++ b/tensorflow/python/ops/gradients_test.py
@@ -418,7 +418,7 @@ class PreventGradientTest(test_util.TensorFlowTestCase):
     with ops.Graph().as_default():
       inp = constant(1.0, shape=[100, 32], name="in")
       out = array_ops.prevent_gradient(inp)
-      with self.assertRaisesRegexp(LookupError, "No gradient defined"):
+      with self.assertRaisesRegexp(LookupError, "explicitly disabled"):
         _ = gradients.gradients(out, inp)
 
 
diff --git a/tensorflow/python/ops/nn_grad.py b/tensorflow/python/ops/nn_grad.py
index 6147cdf221..397c522dbe 100644
--- a/tensorflow/python/ops/nn_grad.py
+++ b/tensorflow/python/ops/nn_grad.py
@@ -331,7 +331,10 @@ def _SoftmaxCrossEntropyWithLogitsGrad(op, grad_0, _):
   # due to the fused implementation's interaction with tf.gradients(),
   # so we make sure we prevent silently incorrect results by raising
   # an error if the second derivative is requested via prevent_gradient.
-  softmax_grad_without_gradient = array_ops.prevent_gradient(op.outputs[1])
+  softmax_grad_without_gradient = array_ops.prevent_gradient(
+      op.outputs[1], message="Currently there is no way to take the second "
+      "derivative of softmax_cross_entropy_with_logits due to the fused "
+      " implementation's interaction with tf.gradients()")
   return _BroadcastMul(grad_0, softmax_grad_without_gradient), None
 
 
@@ -347,7 +350,9 @@ def _SparseSoftmaxCrossEntropyWithLogitsGrad(op, grad_0, _):
   # so we make sure we prevent silently incorrect results by raising
   # an error if the second derivative is requested via prevent_gradient.
   sparse_softmax_grad_without_gradient = array_ops.prevent_gradient(
-      op.outputs[1])
+      op.outputs[1], message="Currently there is no way to take the second "
+      "derivative of sparse_softmax_cross_entropy_with_logits due to the fused "
+      "implementation's interaction with tf.gradients()")
   return _BroadcastMul(grad_0, sparse_softmax_grad_without_gradient), None