def testGetTensorFromInfoRaisesErrors(self):
expected = array_ops.placeholder(dtypes.float32, 1, name="x")
tensor_info = utils.build_tensor_info(expected)
tensor_info.name = "blah:0" # Nonexistent name.
with self.assertRaises(KeyError):
utils.get_tensor_from_tensor_info(tensor_info)
tensor_info.ClearField("name") # Malformed (missing encoding).
with self.assertRaises(ValueError):
utils.get_tensor_from_tensor_info(tensor_info)
def testGetTensorFromInfoRaisesErrors(self):
expected = array_ops.placeholder(dtypes.float32, 1, name="x")
tensor_info = utils.build_tensor_info(expected)
tensor_info.name = "blah:0" # Nonexistant name.
with self.assertRaises(KeyError):
utils.get_tensor_from_tensor_info(tensor_info)
tensor_info.ClearField("name") # Malformed (missing encoding).
with self.assertRaises(ValueError):
utils.get_tensor_from_tensor_info(tensor_info)
def testGetTensorFromInfoRagged(self): expected = ragged_factory_ops.constant([[1, 2], [3]], name="x") tensor_info = utils.build_tensor_info(expected) actual = utils.get_tensor_from_tensor_info(tensor_info) self.assertIsInstance(actual, ragged_tensor.RaggedTensor) self.assertEqual(expected.values.name, actual.values.name) self.assertEqual(expected.row_splits.name, actual.row_splits.name)
def testGetTensorFromInfoSparse(self):
expected = array_ops.sparse_placeholder(dtypes.float32, name="x")
tensor_info = utils.build_tensor_info(expected)
actual = utils.get_tensor_from_tensor_info(tensor_info)
self.assertIsInstance(actual, sparse_tensor.SparseTensor)
self.assertEqual(expected.values.name, actual.values.name)
self.assertEqual(expected.indices.name, actual.indices.name)
self.assertEqual(expected.dense_shape.name, actual.dense_shape.name)
def testGetTensorFromInfoSparse(self): expected = array_ops.sparse_placeholder(dtypes.float32, name="x") tensor_info = utils.build_tensor_info(expected) actual = utils.get_tensor_from_tensor_info(tensor_info) self.assertIsInstance(actual, sparse_tensor.SparseTensor) self.assertEqual(expected.values.name, actual.values.name) self.assertEqual(expected.indices.name, actual.indices.name) self.assertEqual(expected.dense_shape.name, actual.dense_shape.name)
def testGetTensorFromInfoInOtherGraph(self):
with ops.Graph().as_default() as expected_graph:
expected = array_ops.placeholder(dtypes.float32, 1, name="right")
tensor_info = utils.build_tensor_info(expected)
with ops.Graph().as_default(): # Some other graph.
array_ops.placeholder(dtypes.float32, 1, name="other")
actual = utils.get_tensor_from_tensor_info(tensor_info,
graph=expected_graph)
self.assertIsInstance(actual, ops.Tensor)
self.assertIs(actual.graph, expected_graph)
self.assertEqual(expected.name, actual.name)
def testGetTensorFromInfoInOtherGraph(self):
with ops.Graph().as_default() as expected_graph:
expected = array_ops.placeholder(dtypes.float32, 1, name="right")
tensor_info = utils.build_tensor_info(expected)
with ops.Graph().as_default(): # Some other graph.
array_ops.placeholder(dtypes.float32, 1, name="other")
actual = utils.get_tensor_from_tensor_info(tensor_info,
graph=expected_graph)
self.assertIsInstance(actual, ops.Tensor)
self.assertIs(actual.graph, expected_graph)
self.assertEqual(expected.name, actual.name)
def testGetTensorFromInfoInScope(self):
# Build a TensorInfo with name "bar/x:0".
with ops.Graph().as_default():
with ops.name_scope("bar"):
unscoped = array_ops.placeholder(dtypes.float32, 1, name="x")
tensor_info = utils.build_tensor_info(unscoped)
self.assertEqual("bar/x:0", tensor_info.name)
# Build a graph with node "foo/bar/x:0", akin to importing into scope foo.
with ops.Graph().as_default():
with ops.name_scope("foo"):
with ops.name_scope("bar"):
expected = array_ops.placeholder(dtypes.float32, 1, name="x")
self.assertEqual("foo/bar/x:0", expected.name)
# Test that tensor is found by prepending the import scope.
actual = utils.get_tensor_from_tensor_info(tensor_info,
import_scope="foo")
self.assertEqual(expected.name, actual.name)
def testGetTensorFromInfoInScope(self):
# Build a TensorInfo with name "bar/x:0".
with ops.Graph().as_default():
with ops.name_scope("bar"):
unscoped = array_ops.placeholder(dtypes.float32, 1, name="x")
tensor_info = utils.build_tensor_info(unscoped)
self.assertEqual("bar/x:0", tensor_info.name)
# Build a graph with node "foo/bar/x:0", akin to importing into scope foo.
with ops.Graph().as_default():
with ops.name_scope("foo"):
with ops.name_scope("bar"):
expected = array_ops.placeholder(dtypes.float32, 1, name="x")
self.assertEqual("foo/bar/x:0", expected.name)
# Test that tensor is found by prepending the import scope.
actual = utils.get_tensor_from_tensor_info(tensor_info,
import_scope="foo")
self.assertEqual(expected.name, actual.name)
def _ensure_servable(input_tensors, names_to_output_tensor_infos):
"""Check that the signature outputs don't depend on unreachable placeholders.
Args:
input_tensors: An iterable of `Tensor`s specified as the signature's inputs.
names_to_output_tensor_infos: An mapping from output names to respective
`TensorInfo`s corresponding to the signature's output tensors.
Raises:
ValueError: If any of the signature's outputs depend on placeholders not
provided as signature's inputs.
"""
plain_input_tensors = nest.flatten(input_tensors, expand_composites=True)
graph = op_selector.get_unique_graph(plain_input_tensors)
output_tensors = [
utils.get_tensor_from_tensor_info(tensor, graph=graph)
for tensor in names_to_output_tensor_infos.values()
]
plain_output_tensors = nest.flatten(output_tensors, expand_composites=True)
dependency_ops = op_selector.get_backward_walk_ops(
plain_output_tensors, stop_at_ts=plain_input_tensors)
fed_tensors = object_identity.ObjectIdentitySet(plain_input_tensors)
for dependency_op in dependency_ops:
if _must_be_fed(dependency_op) and (not all(
output in fed_tensors for output in dependency_op.outputs)):
input_tensor_names = [tensor.name for tensor in plain_input_tensors]
output_tensor_keys = list(names_to_output_tensor_infos.keys())
output_tensor_names = [tensor.name for tensor in plain_output_tensors]
dependency_path = op_selector.show_path(dependency_op,
plain_output_tensors,
plain_input_tensors)
raise ValueError(
f'The signature\'s input tensors {input_tensor_names} are '
f'insufficient to compute its output keys {output_tensor_keys} '
f'(respectively, tensors {output_tensor_names}) because of the '
f'dependency on `{dependency_op.name}` which is not given as '
'a signature input, as illustrated by the following dependency path: '
f'{dependency_path}')
def testGetTensorFromInfoDense(self):
expected = array_ops.placeholder(dtypes.float32, 1, name="x")
tensor_info = utils.build_tensor_info(expected)
actual = utils.get_tensor_from_tensor_info(tensor_info)
self.assertIsInstance(actual, ops.Tensor)
self.assertEqual(expected.name, actual.name)
def testGetTensorFromInfoDense(self): expected = array_ops.placeholder(dtypes.float32, 1, name="x") tensor_info = utils.build_tensor_info(expected) actual = utils.get_tensor_from_tensor_info(tensor_info) self.assertIsInstance(actual, ops.Tensor) self.assertEqual(expected.name, actual.name)
