Skeleton for PolymorphicFunction serialization

Missing things like variables, function/argument names, support for arguments that aren't a flat list of Tensors, and many other things. But it does manage to save, restore, and call a function.

Starts saving a bit of extra metadata when a new function trace is created. Since this does not have to be computed each time the function is called, I expect the performance impact to be minimal.

PiperOrigin-RevId: 225072712

tensorflow/python/eager/def_function.py

@@ -242,6 +242,7 @@ class PolymorphicFunction(object):
       raise NotImplementedError()
     self._created_variables = None
     self._stateful_fn = None
+    self._stateless_fn = None
     self._descriptor_cache = weakref.WeakKeyDictionary()
     self._name = name
 
@@ -382,6 +383,26 @@ class PolymorphicFunction(object):
 
     return initialize_variables.get_concrete_function()
 
+  @property
+  def _cached_input_signatures(self):
+    """All input signatures used to call this PolymorphicFunction."""
+    seen = set()
+    # Preserves signature ordering rather than returning a set() so that we
+    # don't need to re-sort signatures later to work around Python 2's set
+    # nondeterminism.
+    # pylint: disable=protected-access
+    concrete_functions = []
+    if self._stateful_fn:
+      concrete_functions.extend(self._stateful_fn._function_cache.values())
+    if self._stateless_fn:
+      concrete_functions.extend(self._stateless_fn._function_cache.values())
+    for concrete_function in concrete_functions:
+      signature = concrete_function._python_call_signature
+      if signature not in seen:
+        yield signature
+        seen.add(signature)
+    # pylint: enable=protected-access
+
   def get_concrete_function(self, *args, **kwargs):
     """Returns a `Function` object specialized to inputs and execution context.
 

tensorflow/python/eager/def_function_test.py

@@ -238,6 +238,25 @@ class DefFunctionTest(test.TestCase):
     concrete = compute.get_concrete_function(
         tensor_spec.TensorSpec(None, dtypes.float32))
     self.assertAllClose(4., concrete(constant_op.constant(2.)))
+    input_signature, = compute._cached_input_signatures
+    self.assertEqual(
+        tuple(input_signature),
+        (tensor_spec.TensorSpec(None, dtypes.float32),))
+
+  def test_serialization_signature_cache(self):
+
+    @def_function.function
+    def f(x, y):
+      return x, y
+
+    f(constant_op.constant([[3., 4.]]), constant_op.constant([2.]))
+    f(constant_op.constant([[3, 4, 5]]), constant_op.constant([2]))
+    self.assertEqual(
+        set(f._cached_input_signatures),
+        set(((tensor_spec.TensorSpec([1, 2], dtypes.float32),
+              tensor_spec.TensorSpec([1], dtypes.float32)),
+             (tensor_spec.TensorSpec([1, 3], dtypes.int32),
+              tensor_spec.TensorSpec([1], dtypes.int32)))))
 
 
 if __name__ == '__main__':

tensorflow/python/eager/function.py

@@ -748,6 +748,19 @@ class Function(object):
     return ret
 
 
+class UnknownArgument(object):
+  """Signifies an argument which is not currently handled."""
+  pass
+
+
+def _encode_arg_for_serialization(arg):
+  """A representation for this argument, for serializing signatures."""
+  if isinstance(arg, ops.Tensor):
+    return tensor_spec.TensorSpec(arg.shape, arg.dtype)
+  else:
+    return UnknownArgument()
+
+
 pywrap_tensorflow.RegisterType("Tensor", ops.Tensor)
 pywrap_tensorflow.RegisterType("IndexedSlices", ops.IndexedSlices)
 
@@ -1163,6 +1176,14 @@ class PolymorphicFunction(object):
                 autograph=self._autograph,
                 arg_names=arg_names),
             self._function_attributes)
+        if self._input_signature:
+          python_call_signature = self._input_signature
+        else:
+          python_call_signature = tuple(
+              _encode_arg_for_serialization(arg) for arg in args)
+        # Save information about non-Tensor arguments with the concrete
+        # function. Used to serialize PolymorphicFunctions.
+        graph_function._python_call_signature = python_call_signature  # pylint: disable=protected-access
         self._function_cache[cache_key] = graph_function
       return graph_function, args, kwargs
 

tensorflow/python/framework/function.py

@@ -993,9 +993,10 @@ def _call(sig, *inputs, **kwargs):
   name = kwargs.pop("name", None)
   g = ops.get_default_graph()
   func_name = sig.name
+  if name is None:
+    name = func_name
   attrs = _parse_kwargs_as_attrs(func_name, **kwargs)
   output_types = [dtypes.DType(x.type) for x in sig.output_arg]
-  with ops.name_scope(name, func_name, inputs) as name:
   op = g.create_op(
       func_name,
       list(inputs),
@@ -1046,12 +1047,13 @@ def _from_definition(fdef, grad_func=None):
   c_func = c_api.TF_FunctionImportFunctionDef(serialized)
   result._c_func = c_api_util.ScopedTFFunction(c_func)
   result._extra_inputs = []
+  result._op_def = fdef.signature
   # pylint: enable=protected-access
 
   return result
 
 
-def _from_library(lib):
+def from_library(lib):
   """Creates _DefinedFunctions initialized from a FunctionDefLibrary proto.
 
   This method handles assigning the correct gradient functions to each

tensorflow/python/framework/function_test.py

@@ -1287,7 +1287,7 @@ class FunctionsFromProtos(test.TestCase):
       gradients_impl.gradients([f1, f2, f3, f4], c)
 
     library = g.as_graph_def().library
-    new_funcs = function._from_library(library)
+    new_funcs = function.from_library(library)
 
     def CheckNewFunc(func):
       new_func = [f for f in new_funcs if f.name == func.name]
@@ -1303,7 +1303,7 @@ class FunctionsFromProtos(test.TestCase):
 
   def testFromLibraryEmptyLib(self):
     library = function_pb2.FunctionDefLibrary()
-    self.assertEqual(len(function._from_library(library)), 0)
+    self.assertEqual(len(function.from_library(library)), 0)
 
   def testFromLibraryMissingFuncDef(self):
 
@@ -1327,7 +1327,7 @@ class FunctionsFromProtos(test.TestCase):
     with self.assertRaisesRegexp(
         ValueError,
         "FunctionDefLibrary missing 'G1_[0-9a-zA-Z]{8,11}' FunctionDef"):
-      function._from_library(library)
+      function.from_library(library)
 
     # Create invalid function def that is missing F1 function def
     library = function_pb2.FunctionDefLibrary()
@@ -1337,7 +1337,7 @@ class FunctionsFromProtos(test.TestCase):
     with self.assertRaisesRegexp(
         ValueError,
         "FunctionDefLibrary missing 'F1_[0-9a-zA-Z]{8,11}' FunctionDef"):
-      function._from_library(library)
+      function.from_library(library)
 
   def testFromLibraryCyclicGradFuncs(self):
 
@@ -1366,7 +1366,7 @@ class FunctionsFromProtos(test.TestCase):
 
     with self.assertRaisesRegexp(
         ValueError, "FunctionDefLibrary contains cyclic gradient functions!"):
-      function._from_library(library)
+      function.from_library(library)
 
   def testExperimentalAttrs(self):
 

tensorflow/python/framework/importer.py

@@ -442,11 +442,9 @@ def import_graph_def(graph_def,
     _ProcessNewOps(graph)
 
   if graph_def.library and graph_def.library.function:
-    # pylint: disable=protected-access
+    functions = function.from_library(graph_def.library)
-    functions = function._from_library(graph_def.library)
     for f in functions:
       f.add_to_graph(graph)
-    # pylint: enable=protected-access
 
   # Treat input mappings that don't appear in the graph as an error, because
   # they are likely to be due to a typo.

tensorflow/python/saved_model/BUILD

@@ -287,7 +287,7 @@ py_library(
     deps = [
         ":builder",
         ":constants",
-        ":loader",
+        ":function_serialization",
         ":saved_object_graph_py",
         ":signature_constants",
         ":signature_def_utils",
@@ -295,15 +295,20 @@ py_library(
         ":utils",
         "//tensorflow/core:protos_all_py",
         "//tensorflow/python:array_ops",
+        "//tensorflow/python:constant_op",
+        "//tensorflow/python:control_flow_ops",
+        "//tensorflow/python:dtypes",
         "//tensorflow/python:framework",
         "//tensorflow/python:framework_ops",
         "//tensorflow/python:lib",
         "//tensorflow/python:resource_variable_ops",
+        "//tensorflow/python:tensor_spec",
         "//tensorflow/python:util",
         "//tensorflow/python/eager:context",
         "//tensorflow/python/eager:def_function",
         "//tensorflow/python/eager:function",
         "//tensorflow/python/training/checkpointable:base",
+        "//tensorflow/python/training/checkpointable:tracking",
         "//tensorflow/python/training/checkpointable:util",
     ],
 )
@@ -330,8 +335,12 @@ py_library(
     ],
     srcs_version = "PY2AND3",
     deps = [
+        ":constants",
+        ":function_deserialization",
         ":loader",
         ":saved_object_graph_py",
+        ":utils",
+        "//tensorflow/python:function",
         "//tensorflow/python:lib",
         "//tensorflow/python:util",
         "//tensorflow/python/training/checkpointable:tracking",
@@ -345,10 +354,34 @@ py_test(
     deps = [
         ":load",
         ":save",
+        "//tensorflow/python:constant_op",
         "//tensorflow/python:dtypes",
+        "//tensorflow/python:lib",
         "//tensorflow/python:tensor_spec",
         "//tensorflow/python/eager:def_function",
         "//tensorflow/python/eager:test",
         "//tensorflow/python/training/checkpointable:tracking",
     ],
 )
+
+py_library(
+    name = "function_serialization",
+    srcs = [
+        "function_serialization.py",
+    ],
+    srcs_version = "PY2AND3",
+    deps = [
+        ":saved_object_graph_py",
+        "//tensorflow/python/eager:def_function",
+        "//tensorflow/python/eager:function",
+    ],
+)
+
+py_library(
+    name = "function_deserialization",
+    srcs = [
+        "function_deserialization.py",
+    ],
+    srcs_version = "PY2AND3",
+    deps = ["//tensorflow/python/eager:def_function"],
+)

tensorflow/python/saved_model/function_deserialization.py

@@ -0,0 +1,46 @@
+# Copyright 2018 The TensorFlow Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+# ==============================================================================
+"""Tools for deserializing PolymorphicFunctions."""
+
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+from tensorflow.python.eager import def_function
+
+
+def recreate_polymorphic_function(
+    saved_polymorphic_function, defined_functions):
+  """Creates a PolymorphicFunction which runs restored function definitions."""
+  @def_function.function
+  def restored_function(*args):
+    """Calls a restored function."""
+    # Try calling each function, return a value from the first one whose
+    # signature matches.
+    # TODO(allenl): Consider re-populating the function cache directly.
+    # TODO(allenl): Functions saved with input_signatures should revive with
+    # input_signatures.
+    for monomorphic_function in saved_polymorphic_function.monomorphic_function:
+      try:
+        # TODO(allenl): Passing an explicit name here prevents invalid name
+        # errors. We should replace this with something based on the actual
+        # Python function name.
+        return defined_functions[monomorphic_function.concrete_function](
+            *args, name="imported_function")
+      except ValueError:
+        continue
+    raise AssertionError(
+        "Could not find matching function to call for arguments: %s" % (args,))
+  return restored_function

tensorflow/python/saved_model/function_serialization.py

@@ -0,0 +1,71 @@
+# Copyright 2018 The TensorFlow Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+# ==============================================================================
+"""Tools for serializing PolymorphicFunctions."""
+
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+from tensorflow.python.eager import def_function
+from tensorflow.python.eager import function as defun_lib
+from tensorflow.python.saved_model import saved_object_graph_pb2
+
+
+def _serialize_polymorphic_function(function):
+  """Represents a PolymorphicFunction in a SavedModel.
+
+  Adds `function`'s concrete functions to the current graph.
+
+  Args:
+    function: A `PolymorphicFunction` to serialize.
+
+  Returns:
+    An unserialized `SavedPolymorphicFunction` protocol buffer object.
+  """
+  monomorphic_functions = []
+  for signature in function._cached_input_signatures:  # pylint: disable=protected-access
+    if any(isinstance(arg, defun_lib.UnknownArgument) for arg in signature):
+      continue
+    concrete_function = function.get_concrete_function(*signature)
+    concrete_function.add_to_graph()
+    monomorphic_functions.append(
+        saved_object_graph_pb2.SavedMonomorphicFunction(
+            concrete_function=concrete_function.name))
+  return saved_object_graph_pb2.SavedPolymorphicFunction(
+      monomorphic_function=monomorphic_functions)
+
+
+def add_polymorphic_functions_to_object_graph_proto(
+    checkpointable_objects, saved_object_graph):
+  """Finds PolymorphicFunctions attached to objects and saves them."""
+  existing_objects = list(zip(checkpointable_objects, saved_object_graph.nodes))
+  for obj, obj_proto in existing_objects:
+    for attribute_name in dir(obj):
+      try:
+        attribute_value = getattr(obj, attribute_name, None)
+      except:  # pylint: disable=bare-except
+        # We really don't want to throw an exception just because some object's
+        # attribute accessor is broken.
+        attribute_value = None
+      # TODO(allenl): Consider de-duplicating functions which are referenced
+      # from multiple attributes.
+      if isinstance(attribute_value, def_function.PolymorphicFunction):
+        function_node_id = len(saved_object_graph.nodes)
+        function_node = saved_object_graph.nodes.add()
+        function_node.function.CopyFrom(
+            _serialize_polymorphic_function(attribute_value))
+        reference = obj_proto.children.add()
+        reference.node_id = function_node_id
+        reference.local_name = attribute_name

tensorflow/python/saved_model/load.py

@@ -20,8 +20,10 @@ from __future__ import print_function
 
 import os
 
+from tensorflow.python.framework import function as function_lib
 from tensorflow.python.lib.io import file_io
 from tensorflow.python.saved_model import constants
+from tensorflow.python.saved_model import function_deserialization
 from tensorflow.python.saved_model import loader_impl
 from tensorflow.python.saved_model import saved_object_graph_pb2
 from tensorflow.python.saved_model import utils_impl as saved_model_utils
@@ -33,9 +35,17 @@ class _Loader(object):
   """Helper class to load an object-based SavedModel."""
 
   def __init__(self, object_graph_proto, saved_model_proto, export_dir):
-    self._asset_file_def = saved_model_proto.meta_graphs[0].asset_file_def
+    meta_graph = saved_model_proto.meta_graphs[0]
+    self._asset_file_def = meta_graph.asset_file_def
     self._proto = object_graph_proto
     self._export_dir = export_dir
+    self._defined_functions = {}
+    for defined_function in function_lib.from_library(
+        meta_graph.graph_def.library):
+      # TODO(allenl): Do we need to do name mapping here? Not quite sure what
+      # happens when loaded names collide with existing names.
+      defined_function.add_to_graph(None)
+      self._defined_functions[defined_function.name] = defined_function
     self._load_all()
 
   def _load_all(self):
@@ -52,6 +62,7 @@ class _Loader(object):
     factory = {
         "user_object": lambda: self._recreate_user_object(proto.user_object),
         "asset": lambda: self._recreate_asset(proto.asset),
+        "function": lambda: self._recreate_function(proto.function)
     }
     kind = proto.WhichOneof("kind")
     if kind not in factory:
@@ -68,6 +79,10 @@ class _Loader(object):
         self._asset_file_def[proto.asset_file_def_index].filename)
     return tracking.TrackableAsset(filename)
 
+  def _recreate_function(self, proto):
+    return function_deserialization.recreate_polymorphic_function(
+        proto, self._defined_functions)
+
 
 def _load_saved_object_graph_proto(filename):
   with file_io.FileIO(filename, "rb") as f:
@@ -92,5 +107,4 @@ def load(export_dir):
     raise NotImplementedError(
         "Currently only SavedModels exported with `tf.saved_model.save` may be "
         "imported. Other SavedModels may eventually be supported via load().")
-  # TODO(allenl): load functions from the SavedModel into the eager context
   return root

tensorflow/python/saved_model/load_test.py

@@ -23,6 +23,7 @@ import tempfile
 
 from tensorflow.python.eager import def_function
 from tensorflow.python.eager import test
+from tensorflow.python.framework import constant_op
 from tensorflow.python.framework import dtypes
 from tensorflow.python.framework import tensor_spec
 from tensorflow.python.lib.io import file_io
@@ -47,6 +48,7 @@ class LoadTest(test.TestCase):
     imported = load.load(save_dir)
     self.assertIs(imported.dep_three, imported.dep_two.dep)
     self.assertIsNot(imported.dep_one, imported.dep_two)
+    self.assertEqual(4., imported.f(constant_op.constant(2.)).numpy())
 
   def _make_asset(self, contents):
     filename = tempfile.mktemp(prefix=self.get_temp_dir())

tensorflow/python/saved_model/save.py

@@ -37,6 +37,7 @@ from tensorflow.python.ops import control_flow_ops
 from tensorflow.python.ops import resource_variable_ops
 from tensorflow.python.saved_model import builder_impl
 from tensorflow.python.saved_model import constants
+from tensorflow.python.saved_model import function_serialization
 from tensorflow.python.saved_model import saved_object_graph_pb2
 from tensorflow.python.saved_model import signature_constants
 from tensorflow.python.saved_model import signature_def_utils
@@ -511,6 +512,9 @@ def _write_object_graph(root, export_dir, asset_file_def_index):
   for obj, obj_proto in zip(checkpointable_objects, proto.nodes):
     _write_object_proto(obj, obj_proto, asset_file_def_index)
 
+  function_serialization.add_polymorphic_functions_to_object_graph_proto(
+      checkpointable_objects, proto)
+
   extra_asset_dir = os.path.join(
       compat.as_bytes(export_dir),
       compat.as_bytes(constants.EXTRA_ASSETS_DIRECTORY))

tensorflow/python/saved_model/saved_object_graph.proto

@@ -48,6 +48,7 @@ message SavedObject {
   oneof kind {
     SavedUserObject user_object = 4;
     SavedAsset asset = 5;
+    SavedPolymorphicFunction function = 6;
   }
 }
 
@@ -71,3 +72,13 @@ message SavedAsset {
   // `AssetFileDef.tensor_info`, MUST be ignored.
   uint32 asset_file_def_index = 1;
 }
+
+// A function with multiple signatures, possibly with non-Tensor arguments.
+message SavedPolymorphicFunction {
+  repeated SavedMonomorphicFunction monomorphic_function = 1;
+}
+
+message SavedMonomorphicFunction {
+  // A reference to a TensorFlow function in the MetaGraph's FunctionDefLibrary
+  string concrete_function = 1;
+}