259 lines
9.6 KiB
Python
259 lines
9.6 KiB
Python
# Copyright 2022 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.
|
|
# ==============================================================================
|
|
"""Lazily initialized variables, useful for creating a symbolic Keras model."""
|
|
|
|
import threading
|
|
|
|
# isort: off
|
|
from tensorflow.core.framework import attr_value_pb2
|
|
from tensorflow.python.eager import context
|
|
from tensorflow.python.framework import ops
|
|
from tensorflow.python.ops import gen_resource_variable_ops
|
|
from tensorflow.python.ops import resource_variable_ops
|
|
from tensorflow.python.ops import variable_scope
|
|
from tensorflow.python.trackable import base as trackable
|
|
from tensorflow.python.util import compat
|
|
from tensorflow.python.util import tf_contextlib
|
|
|
|
_DISABLE_LAZY_VARIABLE_INIT = threading.local()
|
|
|
|
|
|
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(f"loc:@{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
|
|
|
|
|
|
class LazyInitVariable(resource_variable_ops.BaseResourceVariable):
|
|
"""Lazily initialized variables.
|
|
|
|
The major use case for this class is to serve as a memory efficient
|
|
alternative for tf.Variable. The resource handle of this class is point to
|
|
nothing, which mean it will raise error when its value is fetched in a eager
|
|
context. Having said that, it will perform like a normal tf.Variable when
|
|
using with graph tensor, like KerasTensor produced from tf.keras.Input.
|
|
"""
|
|
|
|
def __init__(
|
|
self,
|
|
initial_value=None,
|
|
trainable=None,
|
|
collections=None,
|
|
validate_shape=True,
|
|
caching_device=None,
|
|
name=None,
|
|
dtype=None,
|
|
variable_def=None,
|
|
import_scope=None,
|
|
constraint=None,
|
|
distribute_strategy=None,
|
|
synchronization=None,
|
|
aggregation=None,
|
|
shape=None,
|
|
**kwargs,
|
|
):
|
|
assert context.executing_eagerly() # To simplify the logic
|
|
assert variable_def is None # Not supported yet.
|
|
assert caching_device is None # Not supported yet
|
|
|
|
if initial_value is None:
|
|
raise ValueError(
|
|
"The `initial_value` arg to `tf.Variable` must "
|
|
"be specified except when you are not providing a "
|
|
"`variable_def`. You provided neither."
|
|
)
|
|
|
|
if (
|
|
isinstance(initial_value, ops.Tensor)
|
|
and hasattr(initial_value, "graph")
|
|
and initial_value.graph.building_function
|
|
):
|
|
raise ValueError(
|
|
f"Argument `initial_value` ({initial_value}) could not "
|
|
"be lifted out of a `tf.function`. "
|
|
f"(Tried to create variable with name='{name}'). "
|
|
"To avoid this error, when constructing `tf.Variable`s "
|
|
"inside of `tf.function` you can create the "
|
|
"`initial_value` tensor in a "
|
|
"`tf.init_scope` or pass a callable `initial_value` "
|
|
"(e.g., `tf.Variable(lambda : "
|
|
"tf.truncated_normal([10, 40]))`). "
|
|
"Please file a feature request if this "
|
|
"restriction inconveniences you."
|
|
)
|
|
|
|
if constraint is not None and not callable(constraint):
|
|
raise ValueError(
|
|
"Argument `constraint` must be None or a callable. "
|
|
f"a callable. Got a {type(constraint)}: {constraint}"
|
|
)
|
|
|
|
self._name = name
|
|
(
|
|
initial_value,
|
|
shape,
|
|
dtype,
|
|
handle,
|
|
handle_name,
|
|
unique_id,
|
|
) = _infer_shape_dtype_and_create_handle(
|
|
initial_value, shape, dtype, name
|
|
)
|
|
|
|
super().__init__(
|
|
distribute_strategy=distribute_strategy,
|
|
initial_value=initial_value,
|
|
shape=shape,
|
|
dtype=dtype,
|
|
name=name,
|
|
unique_id=unique_id,
|
|
handle_name=handle_name,
|
|
constraint=constraint,
|
|
handle=handle,
|
|
graph_element=None,
|
|
trainable=trainable,
|
|
synchronization=synchronization,
|
|
aggregation=aggregation,
|
|
in_graph_mode=False,
|
|
)
|
|
|
|
# TODO(scottzhu): This method and create_and_initialize might be removed if
|
|
# we decide to just use the tf.Variable to replace this class.
|
|
def initialize(self):
|
|
with ops.name_scope(self._name, "Variable", skip_on_eager=False):
|
|
with ops.colocate_with(self._handle), ops.name_scope("Initializer"):
|
|
if callable(self._initial_value):
|
|
initial_value = self._initial_value()
|
|
else:
|
|
initial_value = self._initial_value
|
|
|
|
if not initial_value.shape.is_compatible_with(self._shape):
|
|
raise ValueError(
|
|
"In this `tf.Variable` creation, the initial value's "
|
|
f"shape ({initial_value.shape}) is not compatible with "
|
|
"the explicitly supplied `shape` "
|
|
f"argument ({self._shape})."
|
|
)
|
|
assert self._dtype is initial_value.dtype.base_dtype
|
|
gen_resource_variable_ops.assign_variable_op(
|
|
self._handle, initial_value
|
|
)
|
|
|
|
def create_and_initialize(self):
|
|
if callable(self._initial_value):
|
|
initial_value = self._initial_value()
|
|
|
|
with ops.device(initial_value.device):
|
|
(
|
|
initial_value,
|
|
shape,
|
|
dtype,
|
|
handle,
|
|
handle_name,
|
|
unique_id,
|
|
) = _infer_shape_dtype_and_create_handle(
|
|
initial_value, self._shape, self._dtype, self._name
|
|
)
|
|
self.initialize()
|
|
|
|
super().__init__(
|
|
trainable=self._trainable,
|
|
shape=shape,
|
|
dtype=dtype,
|
|
handle=handle,
|
|
synchronization=self._synchronization,
|
|
constraint=self._constraint,
|
|
aggregation=self._aggregation,
|
|
distribute_strategy=self._distribute_strategy,
|
|
name=self._name,
|
|
unique_id=unique_id,
|
|
handle_name=handle_name,
|
|
graph_element=None,
|
|
initial_value=initial_value,
|
|
initializer_op=None,
|
|
is_initialized_op=None,
|
|
cached_value=None,
|
|
caching_device=None,
|
|
)
|
|
|
|
|
|
def _lazy_init_variable_creator(next_creator, **kwargs):
|
|
if getattr(_DISABLE_LAZY_VARIABLE_INIT, "disabled", False):
|
|
return next_creator(**kwargs)
|
|
else:
|
|
return LazyInitVariable(**kwargs)
|
|
|
|
|
|
@tf_contextlib.contextmanager
|
|
def lazy_init_scope():
|
|
with variable_scope.variable_creator_scope(_lazy_init_variable_creator):
|
|
yield
|
|
|
|
|
|
@tf_contextlib.contextmanager
|
|
def disable_init_variable_creator():
|
|
try:
|
|
global _DISABLE_LAZY_VARIABLE_INIT
|
|
existing_value = getattr(_DISABLE_LAZY_VARIABLE_INIT, "disabled", False)
|
|
_DISABLE_LAZY_VARIABLE_INIT.disabled = True
|
|
yield
|
|
finally:
|
|
_DISABLE_LAZY_VARIABLE_INIT.disabled = existing_value
|