def create_file_writer_v2(logdir,
max_queue=None,
flush_millis=None,
filename_suffix=None,
name=None):
"""Creates a summary file writer for the given log directory.
Args:
logdir: a string specifying the directory in which to write an event file.
max_queue: the largest number of summaries to keep in a queue; will
flush once the queue gets bigger than this. Defaults to 10.
flush_millis: the largest interval between flushes. Defaults to 120,000.
filename_suffix: optional suffix for the event file name. Defaults to `.v2`.
name: a name for the op that creates the writer.
Returns:
A SummaryWriter object.
"""
if logdir is None:
raise ValueError("logdir cannot be None")
inside_function = ops.inside_function()
with ops.name_scope(name,
"create_file_writer") as scope, ops.device("cpu:0"):
# Run init inside an init_scope() to hoist it out of tf.functions.
with ops.init_scope():
if context.executing_eagerly():
_check_create_file_writer_args(inside_function,
logdir=logdir,
max_queue=max_queue,
flush_millis=flush_millis,
filename_suffix=filename_suffix)
logdir = ops.convert_to_tensor(logdir, dtype=dtypes.string)
if max_queue is None:
max_queue = constant_op.constant(10)
if flush_millis is None:
flush_millis = constant_op.constant(2 * 60 * 1000)
if filename_suffix is None:
filename_suffix = constant_op.constant(".v2")
# Prepend the PID and a process-local UID to the filename suffix to avoid
# filename collisions within the machine (the filename already contains
# the hostname to avoid cross-machine collisions).
unique_prefix = constant_op.constant(".%s.%s" %
(os.getpid(), ops.uid()))
filename_suffix = unique_prefix + filename_suffix
# Use a unique shared_name to prevent resource sharing.
if context.executing_eagerly():
shared_name = context.shared_name()
else:
shared_name = ops.name_from_scope_name(scope) # pylint: disable=protected-access
return ResourceSummaryWriter(
shared_name=shared_name,
init_op_fn=functools.partial(
gen_summary_ops.create_summary_file_writer,
logdir=logdir,
max_queue=max_queue,
flush_millis=flush_millis,
filename_suffix=filename_suffix),
name=name,
v2=True,
metadata={"logdir": logdir})
def variable_creator_scope(self, next_creator, **kwargs):
"""Creates variables & adds them to collections to match legacy code."""
collections = kwargs.pop("collections", None)
v = None
# Get expected variable name.
with ops.name_scope(kwargs.get("name", None), "Variable") as name:
variable_name = ops.name_from_scope_name(name)
kwargs["name"] = name
if self._share_variables:
v = self._variables_by_name.get(variable_name, None)
if v is None:
v = next_creator(**kwargs)
self._variables_by_name[variable_name] = v
if collections is None:
collections = [ops.GraphKeys.GLOBAL_VARIABLES]
if v.trainable and ops.GraphKeys.TRAINABLE_VARIABLES not in collections:
collections = list(collections) + [
ops.GraphKeys.TRAINABLE_VARIABLES
]
ops.add_to_collections(collections, v)
return v
def variable_creator_scope(self, next_creator, **kwargs):
"""Creates variables & adds them to collections to match legacy code."""
collections = kwargs.pop("collections", None)
v = None
# Get expected variable name.
with ops.name_scope(kwargs.get("name", None), "Variable") as name:
variable_name = ops.name_from_scope_name(name)
kwargs["name"] = name
if self._share_variables:
v = self._variables_by_name.get(variable_name, None)
if v is None:
v = next_creator(**kwargs)
self._variables_by_name[variable_name] = v
if collections is None:
collections = [ops.GraphKeys.GLOBAL_VARIABLES]
if v.trainable and ops.GraphKeys.TRAINABLE_VARIABLES not in collections:
collections = list(collections) + [ops.GraphKeys.TRAINABLE_VARIABLES]
ops.add_to_collections(collections, v)
return v
def create_file_writer_v2(logdir,
max_queue=None,
flush_millis=None,
filename_suffix=None,
name=None):
"""Creates a summary file writer for the given log directory.
Args:
logdir: a string specifying the directory in which to write an event file.
max_queue: the largest number of summaries to keep in a queue; will
flush once the queue gets bigger than this. Defaults to 10.
flush_millis: the largest interval between flushes. Defaults to 120,000.
filename_suffix: optional suffix for the event file name. Defaults to `.v2`.
name: a name for the op that creates the writer.
Returns:
A SummaryWriter object.
"""
if logdir is None:
raise ValueError("logdir cannot be None")
inside_function = ops.inside_function()
with ops.name_scope(name, "create_file_writer") as scope, ops.device("cpu:0"):
# Run init inside an init_scope() to hoist it out of tf.functions.
with ops.init_scope():
if context.executing_eagerly():
_check_create_file_writer_args(
inside_function,
logdir=logdir,
max_queue=max_queue,
flush_millis=flush_millis,
filename_suffix=filename_suffix)
logdir = ops.convert_to_tensor(logdir, dtype=dtypes.string)
if max_queue is None:
max_queue = constant_op.constant(10)
if flush_millis is None:
flush_millis = constant_op.constant(2 * 60 * 1000)
if filename_suffix is None:
filename_suffix = constant_op.constant(".v2")
# Prepend the PID and a process-local UID to the filename suffix to avoid
# filename collisions within the machine (the filename already contains
# the hostname to avoid cross-machine collisions).
unique_prefix = constant_op.constant(".%s.%s" % (os.getpid(), ops.uid()))
filename_suffix = unique_prefix + filename_suffix
# Use a unique shared_name to prevent resource sharing.
if context.executing_eagerly():
shared_name = context.shared_name()
else:
shared_name = ops.name_from_scope_name(scope) # pylint: disable=protected-access
return ResourceSummaryWriter(
shared_name=shared_name,
init_op_fn=functools.partial(
gen_summary_ops.create_summary_file_writer,
logdir=logdir,
max_queue=max_queue,
flush_millis=flush_millis,
filename_suffix=filename_suffix),
name=name,
v2=True)
def create_fn():
# Use unique shared_name to prevent resource sharing in eager mode, but
# otherwise use a fixed shared_name to allow SavedModel TF 1.x loading.
if context.executing_eagerly():
shared_name = context.anonymous_name()
else:
shared_name = ops.name_from_scope_name(scope) # pylint: disable=protected-access
return gen_summary_ops.summary_writer(
shared_name=shared_name, name=name)
def __init__(self, variables, *, enable_packed_handle=False):
if enable_packed_handle and not ops.executing_eagerly_outside_functions(
):
raise ValueError(
"Argument `enable_packed_handle` is true, but packed handle is only "
"supported in eager mode. Please make sure eager execution is "
"enabled.")
self._variables = variables
if enable_packed_handle:
self._packed_handle = ops.pack_eager_tensors(
[v.handle for v in variables])
else:
self._packed_handle = None
for v in variables:
v._distributed_container = weakref.ref(self) # pylint: disable=protected-access
self._device_to_handle = {v.device: v.handle for v in variables}
self._primary_handle = variables[0].handle
with ops.init_scope(), \
ops.name_scope("DistributedVariable", skip_on_eager=False) as name:
handle_name = ops.name_from_scope_name(name)
self._unique_id = "%s_%d" % (handle_name, ops.uid())
if context.executing_eagerly():
initial_value = None
initializer = None
else:
initial_value = variables[0].initial_value
initializer = control_flow_ops.group(
[v.initializer for v in variables])
super().__init__(
trainable=variables[0].trainable,
shape=variables[0].shape,
dtype=variables[0].dtype,
handle=None,
synchronization=variables[0].synchronization,
constraint=variables[0].constraint,
aggregation=variables[0].aggregation,
distribute_strategy=variables[0]._distribute_strategy,
name=variables[0].name,
unique_id=self._unique_id,
handle_name=handle_name,
graph_element=variables[0]._graph_element,
initial_value=initial_value,
initializer_op=initializer,
is_initialized_op=None,
cached_value=None,
handle_deleter=_DummyResourceDeleter(),
caching_device=None,
is_variables=True)
def __init__(self, distributed_variables=None, name=None, **unused_kwargs):
"""Packs a list of variables which are distributed across devices.
Args:
distributed_variables: A list of distributed Variables to pack.
name: Optional name for the variable. Defaults to `'Variable'` and gets
uniquified automatically.
"""
if not ops.executing_eagerly_outside_functions():
raise ValueError(
"PackedDistributedVariable should be created in eager mode.")
if not distributed_variables:
raise ValueError("Expect a non-empty list of variables to pack.")
for i, var in enumerate(distributed_variables):
if not resource_variable_ops.is_resource_variable(var):
raise ValueError("Expect a list of ResourceVariables to pack, "
"but the %d-th variable is %s" %
(i, type(var)))
self._distributed_variables = distributed_variables
self._devices = [v.device for v in distributed_variables]
with ops.init_scope():
with ops.name_scope(name, "Variable", skip_on_eager=False) as name:
handle = ops.pack_eager_tensors(
[var.handle for var in distributed_variables])
handle_name = ops.name_from_scope_name(name)
unique_id = "%s_%d" % (handle_name, ops.uid())
super(PackedDistributedVariable, self).__init__(
trainable=distributed_variables[0].trainable,
shape=distributed_variables[0].shape,
dtype=distributed_variables[0].dtype,
handle=handle,
synchronization=distributed_variables[0].synchronization,
constraint=distributed_variables[0].constraint,
aggregation=distributed_variables[0].aggregation,
distribute_strategy=distributed_variables[0].
_distribute_strategy, # pylint: disable=protected-access
name=name,
unique_id=unique_id,
handle_name=handle_name,
graph_element=None,
initial_value=None,
initializer_op=None,
is_initialized_op=None,
cached_value=None,
caching_device=None,
is_distributed_variables=True)
def _infer_shape_dtype_and_create_handle(initial_value, shape, dtype, name):
"""Infer shape and dtype from initial_value and create a variable handle."""
with ops.name_scope(name, "Variable", skip_on_eager=False) as name:
handle_name = ops.name_from_scope_name(name)
unique_id = "%s_%d" % (handle_name, ops.uid())
# Use attr_scope and device(None) to simulate the behavior of
# colocate_with when the variable we want to colocate with doesn't
# yet exist.
device_context_manager = ops.NullContextmanager
attr = attr_value_pb2.AttrValue(
list=attr_value_pb2.AttrValue.ListValue(
s=[compat.as_bytes("loc:@%s" % handle_name)]))
with ops.get_default_graph()._attr_scope({"_class": attr}):
with ops.name_scope("Initializer"), device_context_manager(None):
if not callable(initial_value):
if isinstance(initial_value,
trackable.CheckpointInitialValue):
raise NotImplementedError(
"CheckpointInitialValue is not supported to be the "
"initial value of a lazy variable.")
initial_value = ops.convert_to_tensor(initial_value,
name="initial_value",
dtype=dtype)
assert not callable(initial_value)
assert initial_value.shape.is_compatible_with(shape)
dtype = dtype or initial_value.dtype.base_dtype
shape = shape or initial_value.shape
assert dtype
assert shape
handle = (
resource_variable_ops._variable_handle_from_shape_and_dtype(
shape=shape,
dtype=dtype,
shared_name=None, # Never shared
name=name,
graph_mode=False,
initial_value=None,
))
# initial_value=initial_value if not callable(initial_value) else
# None)
return initial_value, shape, dtype, handle, handle_name, unique_id
def __init__(
self, # pylint: disable=super-init-not-called
initial_value=None,
trainable=None,
caching_device=None,
name=None,
dtype=None,
constraint=None,
add_initializers_to=None,
lifted_initializer_graph=None,
synchronization=None,
aggregation=None,
**unused_kwargs):
"""Creates a variable.
Args:
initial_value: A `Tensor`, or Python object convertible to a `Tensor`,
which is the initial value for the Variable. The initial value must have
a shape specified unless `validate_shape` is set to False. Can also be a
callable with no argument that returns the initial value when called.
(Note that initializer functions from init_ops.py must first be bound
to a shape before being used here.)
trainable: If `True`, GradientTapes automatically watch uses of this
Variable.
caching_device: Optional device string or function describing where the
Variable should be cached for reading. Defaults to the Variable's
device. If not `None`, caches on another device. Typical use is to
cache on the device where the Ops using the Variable reside, to
deduplicate copying through `Switch` and other conditional statements.
name: Optional name for the variable. Defaults to `'Variable'` and gets
uniquified automatically.
dtype: If set, initial_value will be converted to the given type.
If None, either the datatype will be kept (if initial_value is
a Tensor) or float32 will be used (if it is a Python object convertible
to a Tensor).
constraint: An optional projection function to be applied to the variable
after being updated by an `Optimizer` (e.g. used to implement norm
constraints or value constraints for layer weights). The function must
take as input the unprojected Tensor representing the value of the
variable and return the Tensor for the projected value
(which must have the same shape). Constraints are not safe to
use when doing asynchronous distributed training.
add_initializers_to: if not None and not in legacy graph mode, the
initializer tensor will be added to this map in addition to adding the
assignment to the function.
lifted_initializer_graph: FuncGraph to try to lift initializers to.
synchronization: Indicates when a distributed a variable will be
aggregated. Accepted values are constants defined in the class
`tf.VariableSynchronization`. By default the synchronization is set to
`AUTO` and the current `DistributionStrategy` chooses
when to synchronize. If `synchronization` is set to `ON_READ`,
`trainable` must not be set to `True`.
aggregation: Indicates how a distributed variable will be aggregated.
Accepted values are constants defined in the class
`tf.VariableAggregation`.
Raises:
ValueError: If the initial value is not specified, or does not have a
shape and `validate_shape` is `True`.
RuntimeError: If called outside of a function definition.
"""
if not ops.inside_function():
# If we've been init_scope()d out of the function definition nothing to do
# here; we can't really do the capturing or conditional logic.
resource_variable_ops.ResourceVariable.__init__(
self,
initial_value=initial_value,
trainable=trainable,
caching_device=caching_device,
name=name,
dtype=dtype,
constraint=constraint)
return
with ops.init_scope():
self._in_graph_mode = not context.executing_eagerly()
if initial_value is None:
raise ValueError("initial_value must be specified.")
init_from_fn = callable(initial_value)
if constraint is not None and not callable(constraint):
raise ValueError("The `constraint` argument must be a callable.")
if isinstance(initial_value, trackable.CheckpointInitialValue):
self._maybe_initialize_trackable()
self._update_uid = initial_value.checkpoint_position.restore_uid
initial_value = initial_value.wrapped_value
synchronization, aggregation, trainable = (
variables.validate_synchronization_aggregation_trainable(
synchronization, aggregation, trainable, name))
self._trainable = trainable
self._synchronization = synchronization
self._aggregation = aggregation
self._save_slice_info = None
self._initial_value = None
self._initializer_op = None
self._is_initialized_op = None
self._graph_element = None
self._cached_value = None
# Store the graph key so optimizers know how to only retrieve variables from
# this graph. Guaranteed to be the same as the eager graph_key.
self._graph_key = ops.get_default_graph()._graph_key # pylint: disable=protected-access
with ops.name_scope(name, "Variable",
[] if init_from_fn else [initial_value]) as name:
# pylint: disable=protected-access
with ops.init_scope():
handle_name = ops.name_from_scope_name(name)
unique_id = "%s_%d" % (handle_name, ops.uid())
shared_name = context.shared_name(unique_id)
with ops.name_scope("Initializer"), ops.device(None):
initial_value = ops.convert_to_tensor(
initial_value() if init_from_fn else initial_value,
name="initial_value",
dtype=dtype)
with ops.init_scope():
self._handle = resource_variable_ops.eager_safe_variable_handle(
initial_value=initial_value,
shared_name=shared_name,
name=name,
graph_mode=self._in_graph_mode)
self._shape = initial_value.shape
self._unique_id = unique_id
self._handle_name = handle_name + ":0"
self._dtype = initial_value.dtype.base_dtype
self._constraint = constraint
assert initial_value is not None
if self._in_graph_mode:
with ops.init_scope():
outer_graph = ops.get_default_graph()
func_graph = ops.get_default_graph()
function_placeholders = (func_graph.inputs +
func_graph.internal_captures)
placeholder_ops = set(
[tensor.op for tensor in function_placeholders])
lifted_initializer = lift_to_graph.lift_to_graph(
[initial_value],
outer_graph,
disallowed_placeholders=placeholder_ops)[initial_value]
with ops.init_scope():
self._initial_value = lifted_initializer
with ops.name_scope("IsInitialized"):
self._is_initialized_op = (
resource_variable_ops.var_is_initialized_op(
self._handle))
if initial_value is not None:
with ops.name_scope("Assign") as n, ops.colocate_with(
self._handle):
self._initializer_op = resource_variable_ops.assign_variable_op(
self._handle, lifted_initializer, name=n)
with ops.name_scope("Read"), ops.colocate_with(
self._handle):
# Manually assign reads to the handle's device to avoid log
# messages.
with ops.device(self._handle.device):
value = self._read_variable_op()
self._graph_element = value
ops.add_to_collection(ops.GraphKeys.GLOBAL_VARIABLES, self)
else:
if add_initializers_to is not None:
add_initializers_to[self] = initial_value
def assign_fn():
with ops.name_scope("Assign") as n, ops.colocate_with(
self._handle):
resource_variable_ops.assign_variable_op(self._handle,
initial_value,
name=n)
# Returning values to keep tf.cond happy.
return ops.convert_to_tensor(1)
def not_assign_fn():
return ops.convert_to_tensor(0)
# Note: this cond is always guaranteed to run because we're inside a
# defun which will insert automatic control dependencies.
control_flow_ops.cond(
resource_variable_ops.var_is_initialized_op(self._handle),
not_assign_fn, assign_fn)
# After the handle has been created, set up a way to clean it up when
# executing eagerly. We'll hold the only reference to the deleter, so that
# when this object is garbage collected the deleter will be too. This
# means ResourceVariables can be part of reference cycles without those
# cycles being uncollectable.
if not self._in_graph_mode:
self._handle_deleter = resource_variable_ops.EagerResourceDeleter(
handle=self._handle, handle_device=self._handle.device)
self._cached_shape_as_list = None
def __init__(self, # pylint: disable=super-init-not-called
initial_value=None,
trainable=True,
name=None,
dtype=None,
constraint=None,
initialize=True,
**unused_kwargs):
"""Creates a variable.
Args:
initial_value: A `Tensor`, or Python object convertible to a `Tensor`,
which is the initial value for the Variable. The initial value must have
a shape specified unless `validate_shape` is set to False. Can also be a
callable with no argument that returns the initial value when called.
(Note that initializer functions from init_ops.py must first be bound
to a shape before being used here.)
trainable: If `True`, automatically watches this variable on GradientTape
whenever it's used.
name: Optional name for the variable. Defaults to `'Variable'` and gets
uniquified automatically.
dtype: If set, initial_value will be converted to the given type.
If None, either the datatype will be kept (if initial_value is
a Tensor) or float32 will be used (if it is a Python object convertible
to a Tensor).
constraint: An optional projection function to be applied to the variable
after being updated by an `Optimizer` (e.g. used to implement norm
constraints or value constraints for layer weights). The function must
take as input the unprojected Tensor representing the value of the
variable and return the Tensor for the projected value
(which must have the same shape). Constraints are not safe to
use when doing asynchronous distributed training.
initialize: if True, runs initialization in eager execution; leaves the
variable uninitialized otherwise.
Raises:
ValueError: If the initial value is not specified, or does not have a
shape and `validate_shape` is `True`.
"""
if initial_value is None:
raise ValueError("initial_value must be specified.")
init_from_fn = callable(initial_value)
if isinstance(initial_value, ops.Tensor) and hasattr(
initial_value, "graph") and initial_value.graph.building_function:
raise ValueError("Tensor-typed variable initializers must either be "
"wrapped in an init_scope or callable "
"(e.g., `tf.Variable(lambda : "
"tf.truncated_normal([10, 40]))`) when building "
"functions. Please file a feature request if this "
"restriction inconveniences you.")
if constraint is not None and not callable(constraint):
raise ValueError("The `constraint` argument must be a callable.")
if isinstance(initial_value, trackable.CheckpointInitialValue):
self._maybe_initialize_trackable()
self._update_uid = initial_value.checkpoint_position.restore_uid
initial_value = initial_value.wrapped_value
self._trainable = trainable
self._save_slice_info = None
# Store the graph key so optimizers know how to only retrieve variables from
# this graph.
self._graph_key = ops.get_default_graph()._graph_key # pylint: disable=protected-access
with ops.init_scope():
self._in_graph_mode = not context.executing_eagerly()
with ops.name_scope(name, "Variable", []
if init_from_fn else [initial_value]) as name:
# pylint: disable=protected-access
handle_name = ops.name_from_scope_name(name)
shared_name = handle_name
if init_from_fn:
# Use attr_scope and device(None) to simulate the behavior of
# colocate_with when the variable we want to colocate with doesn't
# yet exist.
if self._in_graph_mode:
with ops.name_scope("Initializer"), ops.device(None):
initial_value = ops.convert_to_tensor(
initial_value(), name="initial_value", dtype=dtype)
self._handle = _eager_safe_variable_handle(
shape=initial_value.get_shape(),
dtype=initial_value.dtype.base_dtype,
shared_name=shared_name,
name=name,
graph_mode=self._in_graph_mode)
self._shape = initial_value.get_shape()
else:
initial_value = initial_value()
with ops.name_scope("Initializer"):
initial_value = ops.convert_to_tensor(
initial_value, name="initial_value", dtype=dtype)
self._handle = _eager_safe_variable_handle(
shape=initial_value.get_shape(),
dtype=initial_value.dtype.base_dtype,
shared_name=shared_name,
name=name,
graph_mode=False)
self._shape = initial_value.get_shape()
# pylint: enable=protected-access
# Or get the initial value from a Tensor or Python object.
else:
with ops.name_scope("Initializer"):
initial_value = ops.convert_to_tensor(
initial_value, name="initial_value", dtype=dtype)
# pylint: disable=protected-access
if (self._in_graph_mode and initial_value is not None and
initial_value.op._get_control_flow_context() is not None):
raise ValueError(
"Initializer for variable %s is from inside a control-flow "
"construct, such as a loop or conditional. When creating a "
"variable inside a loop or conditional, use a lambda as the "
"initializer." % name)
# pylint: enable=protected-access
self._handle = _eager_safe_variable_handle(
shape=initial_value.get_shape(),
dtype=initial_value.dtype.base_dtype,
shared_name=shared_name,
name=name,
graph_mode=self._in_graph_mode)
self._shape = initial_value.get_shape()
self._unique_id = shared_name
self._initial_value = initial_value if self._in_graph_mode else None
self._handle_name = handle_name + ":0"
self._dtype = initial_value.dtype.base_dtype
self._constraint = constraint
if self._in_graph_mode:
with ops.name_scope("IsInitialized"):
self._is_initialized_op = (
resource_variable_ops.var_is_initialized_op(self._handle))
if initial_value is not None:
with ops.name_scope("Assign") as n, ops.colocate_with(self._handle):
self._initializer_op = (
resource_variable_ops.assign_variable_op(
self._handle,
self._try_guard_against_uninitialized_dependencies(
initial_value),
name=n))
with ops.name_scope("Read"), ops.colocate_with(self._handle):
# Manually assign reads to the handle's device to avoid log
# messages.
with ops.device(self._handle.device):
value = self._read_variable_op()
self._graph_element = value
self._cached_value = None
else:
if initialize:
resource_variable_ops.assign_variable_op(self._handle,
initial_value)
self._is_initialized_op = None
self._initializer_op = None
self._graph_element = None
self._cached_value = None
self._handle_deleter = None
self._cached_shape_as_list = None
def _init_from_args(self,
embedding_dim,
initializer=None,
trainable=True,
collections=None,
caching_device=None,
name=None,
ktype=None,
vtype=None,
constraint=None,
synchronization=None,
aggregation=None,
distribute_strategy=None,
invalid_key=-1):
"""Creates a variable.
Args:
initial_value: A `Tensor`, or Python object convertible to a `Tensor`,
which is the initial value for the EmbeddingVariable. Can also be a
callable with no argument that returns the initial value when called.
(Note that initializer functions from init_ops.py must first be bound
to a shape before being used here.)
trainable: If `True`, the default, also adds the variable to the graph
collection `GraphKeys.TRAINABLE_VARIABLES`. This collection is used as
the default list of variables to use by the `Optimizer` classes.
collections: List of graph collections keys. The new variable is added to
these collections. Defaults to `[GraphKeys.GLOBAL_VARIABLES]`.
name: Optional name for the variable. Defaults to `'EmbeddingVariable'` and gets
uniquified automatically.
ktype: If set, EV's key will be converted to the given type.
If None, int32 will be used.
vtype: If set, initial_value will be converted to the given type.
If None, either the datatype will be kept (if initial_value is
a Tensor) or float32 will be used (if it is a Python object convertible
to a Tensor).
constraint: An optional projection function to be applied to the variable
after being updated by an `Optimizer` (e.g. used to implement norm
constraints or value constraints for layer weights). The function must
take as input the unprojected Tensor representing the value of the
variable and return the Tensor for the projected value
(which must have the same shape). Constraints are not safe to
use when doing asynchronous distributed training.
@compatibility(eager)
When Eager Execution is enabled, variables are never added to collections.
It is not implicitly added to the GLOBAL_VARIABLES or TRAINABLE_VARIABLES
collections, and the `collections` argument is ignored.
@end_compatibility
"""
if isinstance(embedding_dim, tensor_shape.TensorShape):
embedding_shape = embedding_dim
elif isinstance(embedding_dim, six.integer_types):
embedding_shape = [embedding_dim]
initial_value = initializer(shape=embedding_shape)
init_from_fn = callable(initial_value)
if ktype is None:
ktype = dtypes.int32
if collections is None:
collections = [ops.GraphKeys.GLOBAL_VARIABLES]
if not isinstance(collections, (list, tuple, set)):
raise ValueError(
"collections argument to EmbeddingVariable constructor must be a list, tuple, "
"or set. Got %s of type %s" % (collections, type(collections)))
if constraint is not None and not callable(constraint):
raise ValueError("The `constraint` argument must be a callable.")
self._initializer = initializer
if trainable and ops.GraphKeys.TRAINABLE_VARIABLES not in collections:
collections = list(collections) + [
ops.GraphKeys.TRAINABLE_VARIABLES
]
with ops.init_scope():
self._in_graph_mode = not context.executing_eagerly()
with ops.name_scope(name,
"EmbeddingVariable",
[] if init_from_fn else [initial_value],
skip_on_eager=False) as name:
# pylint: disable=protected-access
self._invalid_key = invalid_key
self._ktype = ktype
handle_name = ops.name_from_scope_name(name)
if self._in_graph_mode:
shared_name = handle_name
unique_id = shared_name
else:
# When in eager mode use a uid for the shared_name, to prevent
# accidental sharing.
unique_id = "%s_%d" % (handle_name, ops.uid())
shared_name = unique_id
# Use attr_scope and device(None) to simulate the behavior of
# colocate_with when the variable we want to colocate with doesn't
# yet exist.
device_context_manager = (ops.device if self._in_graph_mode
else ops.NullContextmanager)
attr = attr_value_pb2.AttrValue(
list=attr_value_pb2.AttrValue.ListValue(
s=[compat.as_bytes("loc:@%s" % handle_name)]))
with ops.get_default_graph()._attr_scope({"_class": attr}):
with ops.name_scope("Initializer"), device_context_manager(
None):
if init_from_fn:
initial_value = initial_value()
if isinstance(initial_value,
trackable.CheckpointInitialValue):
self._maybe_initialize_trackable()
self._update_uid = initial_value.checkpoint_position.restore_uid
initial_value = initial_value.wrapped_value
initial_value = ops.convert_to_tensor(
initial_value, name="initial_value", dtype=vtype)
shape = initial_value.shape
handle = self._embedding_variable_handle(
shape=initial_value.get_shape(),
dtype=initial_value.dtype.base_dtype,
shared_name=shared_name,
name=name,
graph_mode=self._in_graph_mode)
# pylint: disable=protected-access
if (self._in_graph_mode and initial_value is not None
and initial_value.op._get_control_flow_context()
is not None):
raise ValueError(
"Initializer for variable %s is from inside a control-flow "
"construct, such as a loop or conditional. When creating a "
"variable inside a loop or conditional, use a lambda as the "
"initializer." % name)
# pylint: enable=protected-access
vtype = initial_value.dtype.base_dtype
if self._in_graph_mode:
with ops.name_scope("IsInitialized"):
is_initialized_op = (gen_ev_ops.ev_is_initialized_op(
handle, Tkey=self._ktype, Tvalue=vtype))
if initial_value is not None:
# pylint: disable=g-backslash-continuation
with ops.name_scope("Initialize") as n, \
ops.colocate_with(None, ignore_existing=True), \
ops.device(handle.device):
# pylint: disable=protected-access
initializer_op = (gen_ev_ops.initialize_ev_op(
handle,
variables.
_try_guard_against_uninitialized_dependencies(
name, initial_value),
ops.convert_to_tensor(invalid_key,
dtype=self._ktype),
shape=initial_value.get_shape(),
name=n))
cached_value = None
graph_element = None
else:
gen_ev_ops.initialize_ev_op(
handle,
initial_value,
ops.convert_to_tensor(invalid_key, dtype=self._ktype),
shape=initial_value.get_shape())
is_initialized_op = None
initializer_op = None
graph_element = None
cached_value = None
if not context.executing_eagerly():
# Eager variables are only added to collections if they are part of an
# eager variable store (otherwise in an interactive session they would
# hog memory and cause OOM). This is done in ops/variable_scope.py.
ops.add_to_collections(collections, self)
elif ops.GraphKeys.GLOBAL_STEP in collections:
ops.add_to_collections(ops.GraphKeys.GLOBAL_STEP, self)
initial_value = initial_value if self._in_graph_mode else None
super(resource_variable_ops.ResourceVariable,
self).__init__(trainable=trainable,
shape=shape,
dtype=vtype,
handle=handle,
synchronization=synchronization,
constraint=constraint,
aggregation=aggregation,
distribute_strategy=distribute_strategy,
name=name,
unique_id=unique_id,
handle_name=handle_name,
graph_element=graph_element,
initial_value=initial_value,
initializer_op=initializer_op,
is_initialized_op=is_initialized_op,
cached_value=cached_value,
caching_device=caching_device)
def _init_from_args(self,
initial_value=None,
trainable=None,
collections=None,
caching_device=None,
name=None,
dtype=None,
constraint=None,
synchronization=None,
aggregation=None,
distribute_strategy=None,
shape=None):
"""Creates a variable.
Args:
initial_value: A `Tensor`, or Python object convertible to a `Tensor`,
which is the initial value for the Variable. The initial value must have
a shape specified unless `validate_shape` is set to False. Can also be a
callable with no argument that returns the initial value when called.
(Note that initializer functions from init_ops.py must first be bound
to a shape before being used here.)
trainable: If `True`, the default, also adds the variable to the graph
collection `GraphKeys.TRAINABLE_VARIABLES`. This collection is used as
the default list of variables to use by the `Optimizer` classes.
Defaults to `True`, unless `synchronization` is set to `ON_READ`, in
which case it defaults to `False`.
collections: List of graph collections keys. The new variable is added to
these collections. Defaults to `[GraphKeys.GLOBAL_VARIABLES]`.
caching_device: Optional device string or function describing where the
Variable should be cached for reading. Defaults to the Variable's
device. If not `None`, caches on another device. Typical use is to
cache on the device where the Ops using the Variable reside, to
deduplicate copying through `Switch` and other conditional statements.
name: Optional name for the variable. Defaults to `'Variable'` and gets
uniquified automatically.
dtype: If set, initial_value will be converted to the given type.
If None, either the datatype will be kept (if initial_value is
a Tensor) or float32 will be used (if it is a Python object convertible
to a Tensor).
constraint: An optional projection function to be applied to the variable
after being updated by an `Optimizer` (e.g. used to implement norm
constraints or value constraints for layer weights). The function must
take as input the unprojected Tensor representing the value of the
variable and return the Tensor for the projected value
(which must have the same shape). Constraints are not safe to
use when doing asynchronous distributed training.
synchronization: Indicates when a distributed a variable will be
aggregated. Accepted values are constants defined in the class
`tf.VariableSynchronization`. By default the synchronization is set to
`AUTO` and the current `DistributionStrategy` chooses
when to synchronize.
aggregation: Indicates how a distributed variable will be aggregated.
Accepted values are constants defined in the class
`tf.VariableAggregation`.
distribute_strategy: DistributionStrategy under which this variable
was created.
shape: (optional) The shape of this variable. If None, the shape of
`initial_value` will be used. When setting this argument to
`tf.TensorShape(None)` (representing an unspecified shape), the variable
can be assigned with values of different shapes.
Raises:
ValueError: If the initial value is not specified, or does not have a
shape and `validate_shape` is `True`.
@compatibility(eager)
When Eager Execution is enabled, variables are never added to collections.
It is not implicitly added to the `GLOBAL_VARIABLES` or
`TRAINABLE_VARIABLES` collections, and the `collections` argument is
ignored.
@end_compatibility
"""
synchronization, aggregation, trainable = (
variables.validate_synchronization_aggregation_trainable(
synchronization, aggregation, trainable, name))
if initial_value is None:
raise ValueError("initial_value must be specified.")
init_from_fn = callable(initial_value)
if isinstance(initial_value, ops.Tensor) and hasattr(
initial_value,
"graph") and initial_value.graph.building_function:
raise ValueError(
"Tensor-typed variable initializers must either be "
"wrapped in an init_scope or callable "
"(e.g., `tf.Variable(lambda : "
"tf.truncated_normal([10, 40]))`) when building "
"functions. Please file a feature request if this "
"restriction inconveniences you.")
if collections is None:
collections = [ops.GraphKeys.GLOBAL_VARIABLES]
if not isinstance(collections, (list, tuple, set)):
raise ValueError(
"collections argument to Variable constructor must be a list, tuple, "
"or set. Got %s of type %s" % (collections, type(collections)))
if constraint is not None and not callable(constraint):
raise ValueError("The `constraint` argument must be a callable.")
if isinstance(initial_value, trackable.CheckpointInitialValue):
self._maybe_initialize_trackable()
self._update_uid = initial_value.checkpoint_position.restore_uid
initial_value = initial_value.wrapped_value
if trainable and ops.GraphKeys.TRAINABLE_VARIABLES not in collections:
collections = list(collections) + [
ops.GraphKeys.TRAINABLE_VARIABLES
]
with ops.init_scope():
self._in_graph_mode = not context.executing_eagerly()
with ops.name_scope(
name, "TrainableWrapper",
[] if init_from_fn else [initial_value]) as name:
# pylint: disable=protected-access
handle_name = ops.name_from_scope_name(name)
handle_name = (handle_name or "TrainableWrapperHandle")
if self._in_graph_mode:
shared_name = handle_name
unique_id = shared_name
else:
# When in eager mode use a uid for the shared_name, to prevent
# accidental sharing.
unique_id = "%s_%d" % (handle_name, ops.uid())
shared_name = None #context.shared_name()
# Use attr_scope and device(None) to simulate the behavior of
# colocate_with when the variable we want to colocate with doesn't
# yet exist.
device_context_manager = (ops.device if self._in_graph_mode
else ops.NullContextmanager)
attr = attr_value_pb2.AttrValue(
list=attr_value_pb2.AttrValue.ListValue(
s=[compat.as_bytes("loc:@%s" % handle_name)]))
with ops.get_default_graph()._attr_scope({"_class": attr}):
with ops.name_scope("Initializer"), device_context_manager(
None):
initial_value = ops.convert_to_tensor(
initial_value() if init_from_fn else initial_value,
name="initial_value",
dtype=dtype)
if shape is None:
shape = initial_value.shape
handle = resource_variable_ops.eager_safe_variable_handle(
initial_value=initial_value,
shape=None, # shape,
shared_name=shared_name,
name=name,
graph_mode=self._in_graph_mode)
# pylint: disable=protected-access
if (self._in_graph_mode and initial_value is not None
and initial_value.op._get_control_flow_context()
is not None):
raise ValueError(
"Initializer for variable %s is from inside a control-flow "
"construct, such as a loop or conditional. When creating a "
"variable inside a loop or conditional, use a lambda as the "
"initializer." % name)
# pylint: enable=protected-access
dtype = initial_value.dtype.base_dtype
if self._in_graph_mode:
with ops.name_scope("IsInitialized"):
is_initialized_op = (gen_resource_variable_ops.
var_is_initialized_op(handle))
if initial_value is not None:
# pylint: disable=g-backslash-continuation
with ops.name_scope("Assign") as n, \
ops.colocate_with(None, ignore_existing=True), \
ops.device(handle.device):
# pylint: disable=protected-access
initializer_op = (
gen_resource_variable_ops.assign_variable_op(
handle,
variables.
_try_guard_against_uninitialized_dependencies(
name, initial_value),
name=n))
# pylint: enable=protected-access
# pylint: enable=g-backslash-continuation
with ops.name_scope("Read"):
# Manually assign reads to the handle's device to avoid log
# messages.
with ops.device(handle.device):
with ops.control_dependencies([
gen_resource_variable_ops.
assign_variable_op(
handle,
self.prefetch_values(),
name="AssignBeforeInitRead")
]):
value = gen_resource_variable_ops.read_variable_op(
handle, dtype)
graph_element = value
if caching_device is not None:
# Variables may be created in a tf.device() or ops.colocate_with()
# context. At the same time, users would expect caching device to
# be independent of this context, and/or would not expect the
# current device context to be merged with the caching device
# spec. Therefore we reset the colocation stack before creating
# the cached value. Note that resetting the colocation stack will
# also reset the device stack.
with ops.colocate_with(None, ignore_existing=True):
with ops.device(caching_device):
cached_value = array_ops.identity(value)
else:
cached_value = None
else:
gen_resource_variable_ops.assign_variable_op(
handle, initial_value)
is_initialized_op = None
initializer_op = None
graph_element = None
if caching_device:
with ops.device(caching_device):
with ops.control_dependencies([
gen_resource_variable_ops.
assign_variable_op(
handle,
self.prefetch_values(),
name="AssignBeforeInitRead")
]):
cached_value = gen_resource_variable_ops.read_variable_op(
handle, dtype)
else:
cached_value = None
if not context.executing_eagerly():
# Eager variables are only added to collections if they are part of an
# eager variable store (otherwise in an interactive session they would
# hog memory and cause OOM). This is done in ops/variable_scope.py.
ops.add_to_collections(collections, self)
elif ops.GraphKeys.GLOBAL_STEP in collections:
ops.add_to_collections(ops.GraphKeys.GLOBAL_STEP, self)
initial_value = initial_value if self._in_graph_mode else None
super(resource_variable_ops.ResourceVariable,
self).__init__(trainable=trainable,
shape=shape,
dtype=dtype,
handle=handle,
synchronization=synchronization,
constraint=constraint,
aggregation=aggregation,
distribute_strategy=distribute_strategy,
name=name,
unique_id=unique_id,
handle_name=handle_name,
graph_element=graph_element,
initial_value=initial_value,
initializer_op=initializer_op,
is_initialized_op=is_initialized_op,
cached_value=cached_value)
def __init__( # pylint: disable=super-init-not-called
self,
trainable=None,
caching_device=None,
name=None,
shape=None,
dtype=None,
constraint=None,
synchronization=None,
aggregation=None,
extra_handle_data=None,
distribute_strategy=None,
**unused_kwargs):
"""Creates the variable handle.
Args:
trainable: If `True`, GradientTapes automatically watch uses of this
Variable.
caching_device: Optional device string or function describing where the
Variable should be cached for reading. Defaults to the Variable's
device. If not `None`, caches on another device. Typical use is to
cache on the device where the Ops using the Variable reside, to
deduplicate copying through `Switch` and other conditional statements.
name: Optional name for the variable. Defaults to `'Variable'` and gets
uniquified automatically.
shape: The variable's shape.
dtype: The variable's dtype.
constraint: An optional projection function to be applied to the variable
after being updated by an `Optimizer` (e.g. used to implement norm
constraints or value constraints for layer weights). The function must
take as input the unprojected Tensor representing the value of the
variable and return the Tensor for the projected value
(which must have the same shape). Constraints are not safe to
use when doing asynchronous distributed training.
synchronization: Indicates when a distributed a variable will be
aggregated. Accepted values are constants defined in the class
`tf.VariableSynchronization`. By default the synchronization is set to
`AUTO` and the current `DistributionStrategy` chooses
when to synchronize.
aggregation: Indicates how a distributed variable will be aggregated.
Accepted values are constants defined in the class
`tf.VariableAggregation`.
extra_handle_data: Optional, another resource handle or Tensor with handle
data to merge with `shape` and `dtype`.
distribute_strategy: The tf.distribute.Strategy this variable is being
created inside of.
"""
with ops.init_scope():
self._in_graph_mode = not context.executing_eagerly()
if not self._in_graph_mode:
raise ValueError(
"Can't create UninitializedValue in non-graph mode")
with ops.name_scope(name, "Variable") as name:
handle_name = ops.name_from_scope_name(name)
if self._in_graph_mode:
shared_name = handle_name
unique_id = shared_name
else:
unique_id = "%s_%d" % (handle_name, ops.uid())
shared_name = context.shared_name(unique_id)
handle = variable_handle_from_shape_and_dtype(
shape=shape,
dtype=dtype,
shared_name=shared_name,
name=name,
graph_mode=self._in_graph_mode,
extra_handle_data=extra_handle_data)
if not context.executing_eagerly():
with ops.name_scope("Read"):
# Manually assign reads to the handle's device to avoid log
# messages.
with ops.device(handle.device):
value = gen_resource_variable_ops.read_variable_op(
handle, dtype)
_maybe_set_handle_data(dtype, handle, value)
graph_element = value
ops.add_to_collection(ops.GraphKeys.GLOBAL_VARIABLES, self)
# Do *not* add to TRAINABLE_VARIABLES here, even if self._trainable,
# because retraining or frozen use of imported SavedModels is
# controlled at higher levels of model building.
else:
graph_element = None
super(UninitializedVariable,
self).__init__(distribute_strategy=distribute_strategy,
shape=shape,
dtype=dtype,
unique_id=unique_id,
handle_name=handle_name,
constraint=constraint,
handle=handle,
graph_element=graph_element,
trainable=trainable,
synchronization=synchronization,
aggregation=aggregation)
def __init__(self, # pylint: disable=super-init-not-called
initial_value=None,
trainable=None,
caching_device=None,
name=None,
dtype=None,
constraint=None,
add_initializers_to=None,
lifted_initializer_graph=None,
synchronization=None,
aggregation=None,
**unused_kwargs):
"""Creates a variable.
Args:
initial_value: A `Tensor`, or Python object convertible to a `Tensor`,
which is the initial value for the Variable. The initial value must have
a shape specified unless `validate_shape` is set to False. Can also be a
callable with no argument that returns the initial value when called.
(Note that initializer functions from init_ops.py must first be bound
to a shape before being used here.)
trainable: If `True`, GradientTapes automatically watch uses of this
Variable.
caching_device: Optional device string or function describing where the
Variable should be cached for reading. Defaults to the Variable's
device. If not `None`, caches on another device. Typical use is to
cache on the device where the Ops using the Variable reside, to
deduplicate copying through `Switch` and other conditional statements.
name: Optional name for the variable. Defaults to `'Variable'` and gets
uniquified automatically.
dtype: If set, initial_value will be converted to the given type.
If None, either the datatype will be kept (if initial_value is
a Tensor) or float32 will be used (if it is a Python object convertible
to a Tensor).
constraint: An optional projection function to be applied to the variable
after being updated by an `Optimizer` (e.g. used to implement norm
constraints or value constraints for layer weights). The function must
take as input the unprojected Tensor representing the value of the
variable and return the Tensor for the projected value
(which must have the same shape). Constraints are not safe to
use when doing asynchronous distributed training.
add_initializers_to: if not None and not in legacy graph mode, the
initializer tensor will be added to this map in addition to adding the
assignment to the function.
lifted_initializer_graph: FuncGraph to try to lift initializers to.
synchronization: Indicates when a distributed a variable will be
aggregated. Accepted values are constants defined in the class
`tf.VariableSynchronization`. By default the synchronization is set to
`AUTO` and the current `DistributionStrategy` chooses
when to synchronize. If `synchronization` is set to `ON_READ`,
`trainable` must not be set to `True`.
aggregation: Indicates how a distributed variable will be aggregated.
Accepted values are constants defined in the class
`tf.VariableAggregation`.
Raises:
ValueError: If the initial value is not specified, or does not have a
shape and `validate_shape` is `True`.
RuntimeError: If called outside of a function definition.
"""
if not ops.inside_function():
# If we've been init_scope()d out of the function definition nothing to do
# here; we can't really do the capturing or conditional logic.
resource_variable_ops.ResourceVariable.__init__(
self, initial_value=initial_value, trainable=trainable,
caching_device=caching_device, name=name, dtype=dtype,
constraint=constraint)
return
with ops.init_scope():
self._in_graph_mode = not context.executing_eagerly()
if initial_value is None:
raise ValueError("initial_value must be specified.")
init_from_fn = callable(initial_value)
if constraint is not None and not callable(constraint):
raise ValueError("The `constraint` argument must be a callable.")
if isinstance(initial_value, trackable.CheckpointInitialValue):
self._maybe_initialize_trackable()
self._update_uid = initial_value.checkpoint_position.restore_uid
initial_value = initial_value.wrapped_value
synchronization, aggregation, trainable = (
variables.validate_synchronization_aggregation_trainable(
synchronization, aggregation, trainable, name))
self._trainable = trainable
self._synchronization = synchronization
self._aggregation = aggregation
self._save_slice_info = None
self._initial_value = None
self._initializer_op = None
self._is_initialized_op = None
self._graph_element = None
self._cached_value = None
# Store the graph key so optimizers know how to only retrieve variables from
# this graph. Guaranteed to be the same as the eager graph_key.
self._graph_key = ops.get_default_graph()._graph_key # pylint: disable=protected-access
with ops.name_scope(name, "Variable", []
if init_from_fn else [initial_value]) as name:
# pylint: disable=protected-access
with ops.init_scope():
handle_name = ops.name_from_scope_name(name)
unique_id = "%s_%d" % (handle_name, ops.uid())
shared_name = context.shared_name(unique_id)
with ops.name_scope("Initializer"), ops.device(None):
initial_value = ops.convert_to_tensor(
initial_value() if init_from_fn else initial_value,
name="initial_value", dtype=dtype)
with ops.init_scope():
self._handle = resource_variable_ops.eager_safe_variable_handle(
initial_value=initial_value,
shared_name=shared_name,
name=name,
graph_mode=self._in_graph_mode)
self._shape = initial_value.shape
self._unique_id = unique_id
self._handle_name = handle_name + ":0"
self._dtype = initial_value.dtype.base_dtype
self._constraint = constraint
assert initial_value is not None
if self._in_graph_mode:
with ops.init_scope():
outer_graph = ops.get_default_graph()
func_graph = ops.get_default_graph()
function_placeholders = (
func_graph.inputs + func_graph.internal_captures)
placeholder_ops = set(
[tensor.op for tensor in function_placeholders])
lifted_initializer = lift_to_graph.lift_to_graph(
[initial_value], outer_graph,
disallowed_placeholders=placeholder_ops)[initial_value]
with ops.init_scope():
self._initial_value = lifted_initializer
with ops.name_scope("IsInitialized"):
self._is_initialized_op = (
resource_variable_ops.var_is_initialized_op(self._handle))
if initial_value is not None:
with ops.name_scope("Assign") as n, ops.colocate_with(self._handle):
self._initializer_op = resource_variable_ops.assign_variable_op(
self._handle, lifted_initializer, name=n)
with ops.name_scope("Read"), ops.colocate_with(self._handle):
# Manually assign reads to the handle's device to avoid log
# messages.
with ops.device(self._handle.device):
value = self._read_variable_op()
self._graph_element = value
ops.add_to_collection(ops.GraphKeys.GLOBAL_VARIABLES, self)
else:
if add_initializers_to is not None:
add_initializers_to[self] = initial_value
def assign_fn():
with ops.name_scope("Assign") as n, ops.colocate_with(self._handle):
resource_variable_ops.assign_variable_op(
self._handle,
initial_value,
name=n)
# Returning values to keep tf.cond happy.
return ops.convert_to_tensor(1)
def not_assign_fn():
return ops.convert_to_tensor(0)
# Note: this cond is always guaranteed to run because we're inside a
# defun which will insert automatic control dependencies.
control_flow_ops.cond(
resource_variable_ops.var_is_initialized_op(self._handle),
not_assign_fn, assign_fn)
# After the handle has been created, set up a way to clean it up when
# executing eagerly. We'll hold the only reference to the deleter, so that
# when this object is garbage collected the deleter will be too. This
# means ResourceVariables can be part of reference cycles without those
# cycles being uncollectable.
if not self._in_graph_mode:
self._handle_deleter = resource_variable_ops.EagerResourceDeleter(
handle=self._handle, handle_device=self._handle.device)
self._cached_shape_as_list = None
def __init__(self, # pylint: disable=super-init-not-called
initial_value=None,
trainable=True,
name=None,
dtype=None,
constraint=None,
initialize=True,
**unused_kwargs):
"""Creates a variable.
Args:
initial_value: A `Tensor`, or Python object convertible to a `Tensor`,
which is the initial value for the Variable. The initial value must have
a shape specified unless `validate_shape` is set to False. Can also be a
callable with no argument that returns the initial value when called.
(Note that initializer functions from init_ops.py must first be bound
to a shape before being used here.)
trainable: If `True`, automatically watches this variable on GradientTape
whenever it's used.
name: Optional name for the variable. Defaults to `'Variable'` and gets
uniquified automatically.
dtype: If set, initial_value will be converted to the given type.
If None, either the datatype will be kept (if initial_value is
a Tensor) or float32 will be used (if it is a Python object convertible
to a Tensor).
constraint: An optional projection function to be applied to the variable
after being updated by an `Optimizer` (e.g. used to implement norm
constraints or value constraints for layer weights). The function must
take as input the unprojected Tensor representing the value of the
variable and return the Tensor for the projected value
(which must have the same shape). Constraints are not safe to
use when doing asynchronous distributed training.
initialize: if True, runs initialization in eager execution; leaves the
variable uninitialized otherwise.
Raises:
ValueError: If the initial value is not specified, or does not have a
shape and `validate_shape` is `True`.
"""
if initial_value is None:
raise ValueError("initial_value must be specified.")
init_from_fn = callable(initial_value)
if isinstance(initial_value, ops.Tensor) and hasattr(
initial_value, "graph") and initial_value.graph.building_function:
raise ValueError("Tensor-typed variable initializers must either be "
"wrapped in an init_scope or callable "
"(e.g., `tf.Variable(lambda : "
"tf.truncated_normal([10, 40]))`) when building "
"functions. Please file a feature request if this "
"restriction inconveniences you.")
if constraint is not None and not callable(constraint):
raise ValueError("The `constraint` argument must be a callable.")
if isinstance(initial_value, trackable.CheckpointInitialValue):
self._maybe_initialize_trackable()
self._update_uid = initial_value.checkpoint_position.restore_uid
initial_value = initial_value.wrapped_value
self._trainable = trainable
self._save_slice_info = None
# Store the graph key so optimizers know how to only retrieve variables from
# this graph.
self._graph_key = ops.get_default_graph()._graph_key # pylint: disable=protected-access
with ops.init_scope():
self._in_graph_mode = not context.executing_eagerly()
with ops.name_scope(name, "Variable", []
if init_from_fn else [initial_value]) as name:
# pylint: disable=protected-access
handle_name = ops.name_from_scope_name(name)
shared_name = handle_name
if init_from_fn:
# Use attr_scope and device(None) to simulate the behavior of
# colocate_with when the variable we want to colocate with doesn't
# yet exist.
if self._in_graph_mode:
with ops.name_scope("Initializer"), ops.device(None):
initial_value = ops.convert_to_tensor(
initial_value(), name="initial_value", dtype=dtype)
self._handle = _eager_safe_variable_handle(
shape=initial_value.get_shape(),
dtype=initial_value.dtype.base_dtype,
shared_name=shared_name,
name=name,
graph_mode=self._in_graph_mode)
self._shape = initial_value.get_shape()
else:
initial_value = initial_value()
with ops.name_scope("Initializer"):
initial_value = ops.convert_to_tensor(
initial_value, name="initial_value", dtype=dtype)
self._handle = _eager_safe_variable_handle(
shape=initial_value.get_shape(),
dtype=initial_value.dtype.base_dtype,
shared_name=shared_name,
name=name,
graph_mode=False)
self._shape = initial_value.get_shape()
# pylint: enable=protected-access
# Or get the initial value from a Tensor or Python object.
else:
with ops.name_scope("Initializer"):
initial_value = ops.convert_to_tensor(
initial_value, name="initial_value", dtype=dtype)
# pylint: disable=protected-access
if (self._in_graph_mode and initial_value is not None and
initial_value.op._get_control_flow_context() is not None):
raise ValueError(
"Initializer for variable %s is from inside a control-flow "
"construct, such as a loop or conditional. When creating a "
"variable inside a loop or conditional, use a lambda as the "
"initializer." % name)
# pylint: enable=protected-access
self._handle = _eager_safe_variable_handle(
shape=initial_value.get_shape(),
dtype=initial_value.dtype.base_dtype,
shared_name=shared_name,
name=name,
graph_mode=self._in_graph_mode)
self._shape = initial_value.get_shape()
self._unique_id = shared_name
self._initial_value = initial_value if self._in_graph_mode else None
self._handle_name = handle_name + ":0"
self._dtype = initial_value.dtype.base_dtype
self._constraint = constraint
if self._in_graph_mode:
with ops.name_scope("IsInitialized"):
self._is_initialized_op = (
resource_variable_ops.var_is_initialized_op(self._handle))
if initial_value is not None:
with ops.name_scope("Assign") as n, ops.colocate_with(self._handle):
self._initializer_op = (
resource_variable_ops.assign_variable_op(
self._handle,
self._try_guard_against_uninitialized_dependencies(
initial_value),
name=n))
with ops.name_scope("Read"), ops.colocate_with(self._handle):
# Manually assign reads to the handle's device to avoid log
# messages.
with ops.device(self._handle.device):
value = self._read_variable_op()
self._graph_element = value
self._cached_value = None
else:
if initialize:
resource_variable_ops.assign_variable_op(self._handle,
initial_value)
self._is_initialized_op = None
self._initializer_op = None
self._graph_element = None
self._cached_value = None
self._handle_deleter = None
self._cached_shape_as_list = None
