module.py

# Copyright 2019 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.
# ==============================================================================
"""Modules encapsulate building stateful components."""

import re

import six

from tensorflow.python import tf2
from tensorflow.python.framework import ops
from tensorflow.python.ops import variables
from tensorflow.python.trackable import autotrackable
from tensorflow.python.util import nest
from tensorflow.python.util import tf_decorator
from tensorflow.python.util.tf_export import tf_export


@tf_export("Module")
class Module(autotrackable.AutoTrackable):
  """Base neural network module class.

  A module is a named container for `tf.Variable`s, other `tf.Module`s and
  functions which apply to user input. For example a dense layer in a neural
  network might be implemented as a `tf.Module`:

  >>> class Dense(tf.Module):
  ...   def __init__(self, input_dim, output_size, name=None):
  ...     super(Dense, self).__init__(name=name)
  ...     self.w = tf.Variable(
  ...       tf.random.normal([input_dim, output_size]), name='w')
  ...     self.b = tf.Variable(tf.zeros([output_size]), name='b')
  ...   def __call__(self, x):
  ...     y = tf.matmul(x, self.w) + self.b
  ...     return tf.nn.relu(y)

  You can use the Dense layer as you would expect:

  >>> d = Dense(input_dim=3, output_size=2)
  >>> d(tf.ones([1, 3]))
  <tf.Tensor: shape=(1, 2), dtype=float32, numpy=..., dtype=float32)>


  By subclassing `tf.Module` instead of `object` any `tf.Variable` or
  `tf.Module` instances assigned to object properties can be collected using
  the `variables`, `trainable_variables` or `submodules` property:

  >>> d.variables
      (<tf.Variable 'b:0' shape=(2,) dtype=float32, numpy=...,
      dtype=float32)>,
      <tf.Variable 'w:0' shape=(3, 2) dtype=float32, numpy=..., dtype=float32)>)


  Subclasses of `tf.Module` can also take advantage of the `_flatten` method
  which can be used to implement tracking of any other types.

  All `tf.Module` classes have an associated `tf.name_scope` which can be used
  to group operations in TensorBoard and create hierarchies for variable names
  which can help with debugging. We suggest using the name scope when creating
  nested submodules/parameters or for forward methods whose graph you might want
  to inspect in TensorBoard. You can enter the name scope explicitly using
  `with self.name_scope:` or you can annotate methods (apart from `__init__`)
  with `@tf.Module.with_name_scope`.

  >>> class MLP(tf.Module):
  ...   def __init__(self, input_size, sizes, name=None):
  ...     super(MLP, self).__init__(name=name)
  ...     self.layers = []
  ...     with self.name_scope:
  ...       for size in sizes:
  ...         self.layers.append(Dense(input_dim=input_size, output_size=size))
  ...         input_size = size
  ...   @tf.Module.with_name_scope
  ...   def __call__(self, x):
  ...     for layer in self.layers:
  ...       x = layer(x)
  ...     return x

  >>> module = MLP(input_size=5, sizes=[5, 5])
  >>> module.variables
  (<tf.Variable 'mlp/b:0' shape=(5,) dtype=float32, numpy=..., dtype=float32)>,
  <tf.Variable 'mlp/w:0' shape=(5, 5) dtype=float32, numpy=...,
     dtype=float32)>,
  <tf.Variable 'mlp/b:0' shape=(5,) dtype=float32, numpy=..., dtype=float32)>,
  <tf.Variable 'mlp/w:0' shape=(5, 5) dtype=float32, numpy=...,
     dtype=float32)>)
  """

  # AutoTrackable adds object attributes that users will not expect us to
  # include when flattening (these reference dependencies reachable via other
  # object attributes).
  _TF_MODULE_IGNORED_PROPERTIES = frozenset((
      "_self_unconditional_checkpoint_dependencies",
      "_self_unconditional_dependency_names"
  ))

  def __init__(self, name=None):
    if name is None:
      name = camel_to_snake(type(self).__name__)
    else:
      if not valid_identifier(name):
        raise ValueError(
            "%r is not a valid module name. Module names must be valid Python "
            "identifiers (e.g. a valid class name)." % name)

    self._name = name
    if tf2.enabled():
      with ops.name_scope_v2(name) as scope_name:
        self._name_scope = ops.name_scope_v2(scope_name)
    else:
      with ops.name_scope(name, skip_on_eager=False) as scope_name:
        self._scope_name = scope_name

  @property
  def name(self):
    """Returns the name of this module as passed or determined in the ctor.

    NOTE: This is not the same as the `self.name_scope.name` which includes
    parent module names.
    """
    return self._name

  @property
  def name_scope(self):
    """Returns a `tf.name_scope` instance for this class."""
    if tf2.enabled():
      return self._name_scope
    else:
      # In TF1 name_scope is not re-entrant in eager so we cannot memoize it.
      return ops.name_scope(self._scope_name, skip_on_eager=False)

  @property
  def variables(self):
    """Sequence of variables owned by this module and its submodules.

    Note: this method uses reflection to find variables on the current instance
    and submodules. For performance reasons you may wish to cache the result
    of calling this method if you don't expect the return value to change.

    Returns:
      A sequence of variables for the current module (sorted by attribute
      name) followed by variables from all submodules recursively (breadth
      first).
    """
    return tuple(self._flatten(predicate=_is_variable, expand_composites=True))

  @property
  def trainable_variables(self):
    """Sequence of trainable variables owned by this module and its submodules.

    Note: this method uses reflection to find variables on the current instance
    and submodules. For performance reasons you may wish to cache the result
    of calling this method if you don't expect the return value to change.

    Returns:
      A sequence of variables for the current module (sorted by attribute
      name) followed by variables from all submodules recursively (breadth
      first).
    """
    return tuple(
        self._flatten(predicate=_is_trainable_variable, expand_composites=True))

  @property
  def non_trainable_variables(self):
    """Sequence of non-trainable variables owned by this module and its submodules.

    Note: this method uses reflection to find variables on the current instance
    and submodules. For performance reasons you may wish to cache the result
    of calling this method if you don't expect the return value to change.

    Returns:
      A sequence of variables for the current module (sorted by attribute
      name) followed by variables from all submodules recursively (breadth
      first).
    """
    return tuple(self._flatten(predicate=_is_non_trainable_variable))

  @property
  def submodules(self):
    """Sequence of all sub-modules.

    Submodules are modules which are properties of this module, or found as
    properties of modules which are properties of this module (and so on).

    >>> a = tf.Module()
    >>> b = tf.Module()
    >>> c = tf.Module()
    >>> a.b = b
    >>> b.c = c
    >>> list(a.submodules) == [b, c]
    True
    >>> list(b.submodules) == [c]
    True
    >>> list(c.submodules) == []
    True

    Returns:
      A sequence of all submodules.
    """
    return tuple(self._flatten(predicate=_is_module))

  def _flatten(self,
               recursive=True,
               predicate=None,
               attribute_traversal_key=None,
               with_path=False,
               expand_composites=False):
    """Flattened attribute values in sorted order by attribute name.

    Modules are flattened by first walking their attributes in name order.
    Each attribute value is then flattened to find leaf values. If flatten is
    applied `recursive`ly and if the leaf is a `Module` it will also be
    flattened to find leaves. Finally every leaf value is optionally tested
    against the given `predicate` and finally yielded.

    ```
    class Foo(tf.Module):
      def __init__(self):
        super(Foo, self).__init__()
        self.x = [tf.constant('a'), tf.constant('b')]
        self.y = {'i': tf.constant('c'), 'j': tf.constant('d')}
        self.z = tf.constant('e')

      @property
      def tensors(self):
        return tuple(self._flatten(predicate=is_tensor, with_path=True))

    foo = Foo()
    foo.tensors
    # ==> ((('x', 0),   <tf.Tensor: ...'a'>),
    #     (('x', 1),   <tf.Tensor: ...'b'>),
    #     (('y', 'i'), <tf.Tensor: ...'c'>),
    #     (('y', 'j'), <tf.Tensor: ...'d'>),
    #     (('z',),     <tf.Tensor: ...'e'>))
    ```

    `attribute_traversal_key` controls the order object properties are visited.
    If not set objects are visited in ascending order by name.

    Args:
      recursive: Whether to recurse into child modules or not.
      predicate: (Optional) If set then only values matching predicate are
        yielded. A value of `None` (the default) means no items will be
        filtered.
      attribute_traversal_key: (Optional) Method to rekey object attributes
        before they are sorted. Contract is the same as `key` argument to
        builtin `sorted` and only applies to object properties.
      with_path: (Optional) Whether to include the path to the object as well
        as the object itself. If `with_path` is `True` then leaves will not be
        de-duplicated (e.g. if the same leaf instance is reachable via multiple
        modules then it will be yielded multiple times with different paths).
      expand_composites: If true, then composite tensors are expanded into their
        component tensors.

    Returns:
      Flat generator for leaves of the current module and optionally all
      submodules.
    """
    if predicate is None:
      predicate = lambda _: True

    return _flatten_module(
        self,
        recursive=recursive,
        predicate=predicate,
        attributes_to_ignore=self._TF_MODULE_IGNORED_PROPERTIES,
        attribute_traversal_key=attribute_traversal_key,
        with_path=with_path,
        expand_composites=expand_composites)

  @classmethod
  def with_name_scope(cls, method):
    """Decorator to automatically enter the module name scope.

    >>> class MyModule(tf.Module):
    ...   @tf.Module.with_name_scope
    ...   def __call__(self, x):
    ...     if not hasattr(self, 'w'):
    ...       self.w = tf.Variable(tf.random.normal([x.shape[1], 3]))
    ...     return tf.matmul(x, self.w)

    Using the above module would produce `tf.Variable`s and `tf.Tensor`s whose
    names included the module name:

    >>> mod = MyModule()
    >>> mod(tf.ones([1, 2]))
    <tf.Tensor: shape=(1, 3), dtype=float32, numpy=..., dtype=float32)>
    >>> mod.w
    <tf.Variable 'my_module/Variable:0' shape=(2, 3) dtype=float32,
    numpy=..., dtype=float32)>

    Args:
      method: The method to wrap.

    Returns:
      The original method wrapped such that it enters the module's name scope.
    """
    def method_with_name_scope(self, *args, **kwargs):
      with self.name_scope:
        return method(self, *args, **kwargs)

    return tf_decorator.make_decorator(method, method_with_name_scope)


def _is_variable(obj):
  return isinstance(obj, variables.Variable)


def _is_trainable_variable(obj):
  return _is_variable(obj) and getattr(obj, "trainable", False)


def _is_non_trainable_variable(obj):
  return _is_variable(obj) and not getattr(obj, "trainable", False)


def _is_module(obj):
  return isinstance(obj, Module)

_CAMEL_TO_SNAKE_R = re.compile(r"((?<=[a-z0-9])[A-Z]|(?!^)[A-Z](?=[a-z]))")
_VALID_IDENTIFIER = re.compile(r"^[a-zA-Z_]([a-zA-Z0-9_])*$")


def valid_identifier(name):
  return bool(_VALID_IDENTIFIER.match(name))


def camel_to_snake(value):
  return _CAMEL_TO_SNAKE_R.sub(r"_\1", value).lower()


def _flatten_module(module,
                    recursive,
                    predicate,
                    attribute_traversal_key,
                    attributes_to_ignore,
                    with_path,
                    expand_composites,
                    module_path=(),
                    seen=None,
                    recursion_stack=None):
  """Implementation of `flatten`.

  Args:
    module: Current module to process.
    recursive: Whether to recurse into child modules or not.
    predicate: (Optional) If set then only values matching predicate are
      yielded. A value of `None` (the default) means no items will be
      filtered.
    attribute_traversal_key: (Optional) Method to rekey object attributes
      before they are sorted. Contract is the same as `key` argument to
      builtin `sorted` and only applies to object properties.
    attributes_to_ignore: object attributes to ignored.
    with_path: (Optional) Whether to include the path to the object as well
      as the object itself. If `with_path` is `True` then leaves will not be
      de-duplicated (e.g. if the same leaf instance is reachable via multiple
      modules then it will be yielded multiple times with different paths).
    expand_composites: If true, then composite tensors are expanded into their
      component tensors.
    module_path: The path to the current module as a tuple.
    seen: A set containing all leaf IDs seen so far.
    recursion_stack: A list containing all module IDs associated with the
      current call stack.

  Yields:
    Matched leaves with the optional corresponding paths of the current module
    and optionally all its submodules.
  """
  module_id = id(module)
  if seen is None:
    seen = set([module_id])

  module_dict = vars(module)
  submodules = []

  if recursion_stack is None:
    recursion_stack = []

  # When calling `_flatten_module` with `with_path=False`, the global lookup
  # table `seen` guarantees the uniqueness of the matched objects.
  # In the case of `with_path=True`, there might be multiple paths associated
  # with the same predicate, so we don't stop traversing according to `seen`
  # to make sure all these paths are returned.
  # When there are cycles connecting submodules, we break cycles by avoiding
  # following back edges (links pointing to a node in `recursion_stack`).
  if module_id in recursion_stack:
    recursive = False

  for key in sorted(module_dict, key=attribute_traversal_key):
    if key in attributes_to_ignore:
      continue

    prop = module_dict[key]
    try:
      leaves = nest.flatten_with_tuple_paths(
          prop, expand_composites=expand_composites)
    except Exception as cause:  # pylint: disable=broad-except
      six.raise_from(
          ValueError(
              "Error processing property {!r} of {!r}".format(key, prop)),
          cause)

    for leaf_path, leaf in leaves:
      leaf_path = (key,) + leaf_path

      if not with_path:
        leaf_id = id(leaf)
        if leaf_id in seen:
          continue
        seen.add(leaf_id)

      if predicate(leaf):
        if with_path:
          yield module_path + leaf_path, leaf
        else:
          yield leaf

      if recursive and _is_module(leaf):
        # Walk direct properties first then recurse.
        submodules.append((module_path + leaf_path, leaf))

  recursion_stack.append(module_id)

  for submodule_path, submodule in submodules:
    subvalues = _flatten_module(
        submodule,
        recursive=recursive,
        predicate=predicate,
        attribute_traversal_key=attribute_traversal_key,
        attributes_to_ignore=submodule._TF_MODULE_IGNORED_PROPERTIES,  # pylint: disable=protected-access
        with_path=with_path,
        expand_composites=expand_composites,
        module_path=submodule_path,
        seen=seen,
        recursion_stack=recursion_stack)

    for subvalue in subvalues:
      # Predicate is already tested for these values.
      yield subvalue

  recursion_stack.pop()