186 lines
6.4 KiB
Python
186 lines
6.4 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.
|
||
|
# ==============================================================================
|
||
|
"""Gradient support for Composite Tensors."""
|
||
|
|
||
|
import abc
|
||
|
import sys
|
||
|
|
||
|
from tensorflow.python.framework import composite_tensor
|
||
|
from tensorflow.python.util import nest
|
||
|
|
||
|
|
||
|
# pylint:disable=g-import-not-at-top
|
||
|
if sys.version_info >= (3, 8):
|
||
|
from typing import Protocol
|
||
|
from typing import runtime_checkable
|
||
|
else:
|
||
|
from typing_extensions import Protocol
|
||
|
from typing_extensions import runtime_checkable
|
||
|
# pylint:enable=g-import-not-at-top
|
||
|
|
||
|
|
||
|
# TODO(xjun): Add CompositeTensorGradient support for SparseTensor,
|
||
|
# StructuredTensor, and MaskedTensor.
|
||
|
class CompositeTensorGradient(object, metaclass=abc.ABCMeta):
|
||
|
"""Class used to help compute gradients for CompositeTensors.
|
||
|
|
||
|
This abstract base class defines two methods: `get_gradient_components`, which
|
||
|
returns the components of a value that should be included in gradients; and
|
||
|
`replace_gradient_components`, which replaces the gradient components in a
|
||
|
value. These methods can be used to compute the gradient of a `y` with
|
||
|
respect to `x` (`grad(y, x)`) as follows:
|
||
|
|
||
|
* If `y` is a `CompositeTensor` with `CompositeTensorGradient` `cg` =
|
||
|
`y.__composite_gradient__`, then `grad(y, x)` =
|
||
|
`grad(cg.get_gradient_components(y), x)`.
|
||
|
|
||
|
* If `x` is a `CompositeTensor` with `CompositeTensorGradient` `cg` =
|
||
|
'x.__composite_gradient__', then `grad(y, x)` =
|
||
|
`cg.replace_gradient_components(x, grad(y, cg.get_gradient_components(x))`.
|
||
|
"""
|
||
|
|
||
|
@abc.abstractmethod
|
||
|
def get_gradient_components(self, value):
|
||
|
"""Returns the components of `value` that should be included in gradients.
|
||
|
|
||
|
This method may not call TensorFlow ops, since any new ops added to the
|
||
|
graph would not be propertly tracked by the gradient mechanisms.
|
||
|
|
||
|
Args:
|
||
|
value: A `CompositeTensor` value.
|
||
|
|
||
|
Returns:
|
||
|
A nested structure of `Tensor` or `IndexedSlices`.
|
||
|
"""
|
||
|
raise NotImplementedError(
|
||
|
f"{type(self).__name__}.get_gradient_components()")
|
||
|
|
||
|
@abc.abstractmethod
|
||
|
def replace_gradient_components(self, value, component_grads):
|
||
|
"""Replaces the gradient components in `value` with `component_grads`.
|
||
|
|
||
|
Args:
|
||
|
value: A value with its gradient components compatible with
|
||
|
`component_grads`.
|
||
|
component_grads: A nested structure of `Tensor` or `IndexedSlices` or
|
||
|
`None` (for unconnected gradients).
|
||
|
|
||
|
Returns:
|
||
|
A copy of `value`, where the components that should be included in
|
||
|
gradients have been replaced by `component_grads`; or `None` (if
|
||
|
`component_grads` includes `None`).
|
||
|
"""
|
||
|
raise NotImplementedError(
|
||
|
f"{type(self).__name__}.replace_gradient_components()")
|
||
|
|
||
|
|
||
|
@runtime_checkable
|
||
|
class CompositeTensorGradientProtocol(Protocol):
|
||
|
"""Protocol for adding gradient support to CompositeTensors."""
|
||
|
__composite_gradient__: CompositeTensorGradient
|
||
|
|
||
|
|
||
|
class WithValuesCompositeTensorGradient(CompositeTensorGradient):
|
||
|
"""CompositeTensorGradient based on `T.values` and `T.with_values`."""
|
||
|
|
||
|
def get_gradient_components(self, value):
|
||
|
return value.values
|
||
|
|
||
|
def replace_gradient_components(self, value, component_grads):
|
||
|
return value.with_values(component_grads)
|
||
|
|
||
|
|
||
|
def _get_tensors_for_gradient(x):
|
||
|
"""Returns the Tensors in `x` that should be differentiated.
|
||
|
|
||
|
Args:
|
||
|
x: A `Tensor` or `CompositeTensor`.
|
||
|
|
||
|
Returns:
|
||
|
A `Tensor` or a nested structure of `Tensor`.
|
||
|
"""
|
||
|
if not isinstance(x, composite_tensor.CompositeTensor):
|
||
|
return x
|
||
|
|
||
|
if not isinstance(x, CompositeTensorGradientProtocol):
|
||
|
raise ValueError(
|
||
|
f"Type {type(x).__name__} is not supported as a gradient source or "
|
||
|
"gradient target.")
|
||
|
composite_gradient = x.__composite_gradient__
|
||
|
gradient_components = composite_gradient.get_gradient_components(x)
|
||
|
if gradient_components is x:
|
||
|
return x
|
||
|
return nest.map_structure(_get_tensors_for_gradient, gradient_components)
|
||
|
|
||
|
|
||
|
def _replace_tensors_for_gradient(x, grad):
|
||
|
"""Replaces the tensors in `x` that should be differentiated with `grad`.
|
||
|
|
||
|
Args:
|
||
|
x: A `Tensor` or `CompositeTensor`.
|
||
|
grad: A nested structure of `Tensor`, with the same structure as the value
|
||
|
returned by `_get_tensors_for_gradient(x)`.
|
||
|
|
||
|
Returns:
|
||
|
A `Tensor` or `CompositeTensor`.
|
||
|
"""
|
||
|
if not isinstance(x, composite_tensor.CompositeTensor):
|
||
|
return grad
|
||
|
|
||
|
if not isinstance(x, CompositeTensorGradientProtocol):
|
||
|
raise ValueError(
|
||
|
f"Type {type(x).__name__} is not supported as a gradient source.")
|
||
|
|
||
|
composite_gradient = x.__composite_gradient__
|
||
|
x_components = composite_gradient.get_gradient_components(x)
|
||
|
if x_components is x:
|
||
|
grad_components = grad
|
||
|
else:
|
||
|
grad_components = nest.map_structure_up_to(x_components,
|
||
|
_replace_tensors_for_gradient,
|
||
|
x_components, grad)
|
||
|
if grad_components is None:
|
||
|
return None
|
||
|
return composite_gradient.replace_gradient_components(x, grad_components)
|
||
|
|
||
|
|
||
|
def get_flat_tensors_for_gradients(xs):
|
||
|
"""Returns a flat list of Tensors that should be differentiated for `xs`.
|
||
|
|
||
|
Args:
|
||
|
xs: A list of `Tensor`s or `CompositeTensor`s.
|
||
|
|
||
|
Returns:
|
||
|
A flat list of `Tensor`s constructed from `xs`, where `Tensor` values are
|
||
|
left as-is, and `CompositeTensor`s are replaced with
|
||
|
`_get_tensors_for_gradient(x)`.
|
||
|
"""
|
||
|
return nest.flatten([_get_tensors_for_gradient(x) for x in xs])
|
||
|
|
||
|
|
||
|
def replace_flat_tensors_for_gradients(xs, flat_grads):
|
||
|
"""Replaces Tensors that should be differentiated in `xs` with `flat_grads`.
|
||
|
|
||
|
Args:
|
||
|
xs: A list of `Tensor`s or `CompositeTensor`s.
|
||
|
flat_grads: A list of `Tensor`.
|
||
|
|
||
|
Returns:
|
||
|
A list of `Tensor` or `CompositeTensor`.
|
||
|
"""
|
||
|
xs_structure = [_get_tensors_for_gradient(x) for x in xs]
|
||
|
grads = nest.pack_sequence_as(xs_structure, flat_grads)
|
||
|
return [_replace_tensors_for_gradient(x, grad) for x, grad in zip(xs, grads)]
|