STT-tensorflow/tensorflow/python/keras/mixed_precision/autocast_variable.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.
# ==============================================================================
"""Contains AutoCastVariable, a variable which automatically casts itself."""
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function

import threading

from tensorflow.python.distribute import ps_values as ps_distribute_values
from tensorflow.python.distribute import values as distribute_values
from tensorflow.python.eager import context
from tensorflow.python.framework import ops
from tensorflow.python.ops import math_ops
from tensorflow.python.ops import resource_variable_ops
from tensorflow.python.ops import variables
from tensorflow.python.types import core


# _autocast_dtype.dtype is the dtype AutoCastVariables should be cast to, or
# None if AutoCastVariables should not be cast.
_autocast_dtype = threading.local()


def numpy_text(tensor, is_repr=False):
  """Human readable representation of a tensor's numpy value."""
  if tensor.dtype.is_numpy_compatible:
    # pylint: disable=protected-access
    text = repr(tensor._numpy()) if is_repr else str(tensor._numpy())
    # pylint: enable=protected-access
  else:
    text = '<unprintable>'
  if '\n' in text:
    text = '\n' + text
  return text


class AutoCastVariable(variables.Variable, core.Tensor):
  """Variable that will cast itself to a different dtype in applicable contexts.

  This class wraps a floating-point `tf.Variable`. It emulates the variable
  interface and delegates to the wrapped variable, but it additionally will cast
  the wrapped variable under an `enable_auto_cast_variables(dtype)` context
  manager.

  For example:

  >>> v = tf.Variable(1.0, dtype=tf.float32)
  >>> v = AutoCastVariable(v)
  >>> tf.identity(v).dtype
  tf.float32
  >>> with enable_auto_cast_variables(tf.float16):
  ...   tf.identity(v).dtype
  tf.float16

  The purpose of this class is to allow Keras layers to create variables in
  float32, and automatically cast them to float16 or bfloat16 when the layer is
  called.
  """

  def __init__(self, variable):
    """Creates an AutoCastVariable instance.

    Args:
      variable: A floating-point resource variable to wrap.

    Raises:
      ValueError: If `variable` is not a floating-point resource variable
    """
    if not isinstance(variable, variables.Variable):
      raise ValueError('variable must be of type tf.ResourceVariable, but got: '
                       '%s' % variable)
    if not variable.dtype.is_floating:
      raise ValueError('variable must be a floating point variable but has '
                       'type: %s' % variable.dtype.name)
    self._variable = variable
    # 'delegate' means AutoCastVariable.op return self._variable.op, which will
    # raise an AttributeError in Eager (as intended). If set to any other value,
    # AutoCastVariable.op returns that value instead, which is used to set the
    # op attribute in AutoCastVariable.assign().
    self._op = 'delegate'

  def _should_cast(self):
    """Returns True if this variable should be casted when accessed."""
    autocast_dtype = getattr(_autocast_dtype, 'dtype', None)
    return autocast_dtype is not None and self.dtype != autocast_dtype

  @property
  def dtype(self):
    """The dtype of the underlying variable, before any casts are done."""
    return self._variable.dtype

  @property
  def true_dtype(self):
    """Deprecated alias of `dtype`."""
    return self._variable.dtype

  @property
  def _cast_dtype(self):
    dtype = getattr(_autocast_dtype, 'dtype', None)
    return dtype or self._variable.dtype

  def value(self):
    val = self._variable.value()
    if not self._should_cast():
      return val
    return math_ops.cast(val, self._cast_dtype)

  def read_value(self):
    val = self._variable.read_value()
    return math_ops.cast(val, self._cast_dtype)

  def sparse_read(self, indices, name=None):
    """Reads the value of this variable sparsely, using `gather`."""
    val = self._variable.sparse_read(indices, name=name)
    return math_ops.cast(val, self._cast_dtype)

  def gather_nd(self, indices, name=None):
    """Gather slices of the variable into a Tensor."""
    val = self._variable.gather_nd(indices, name=name)
    return math_ops.cast(val, self._cast_dtype)

  def __getattr__(self, name):
    return getattr(self._variable, name)

  def _dense_var_to_tensor(self, dtype=None, name=None, as_ref=False):
    """Converts this variable to a tensor."""
    if not self._should_cast():
      return ops.convert_to_tensor(self._variable, dtype, name, as_ref)
    # TODO(reedwm): Support as_ref?
    assert not as_ref
    if dtype is not None and not dtype.is_compatible_with(self._cast_dtype):
      raise ValueError(
          'Incompatible type conversion requested to type {!r} for '
          'AutoCastVariable which is casted to type {!r}'.format(
              dtype.name, self._cast_dtype.name))
    val = ops.convert_to_tensor_v2_with_dispatch(
        self._variable, dtype=self._variable.dtype, name=name)
    return math_ops.cast(val, self._cast_dtype)

  def _should_act_as_resource_variable(self):
    """Pass resource_variable_ops.is_resource_variable check."""
    pass

  def __repr__(self):
    if context.executing_eagerly() and not self._in_graph_mode:
      repr_str = ("<AutoCastVariable '{v.name}' shape={v.shape} "
                  'dtype={v.dtype.name} dtype_to_cast_to={v._cast_dtype.name}, '
                  'numpy={np_repr}>')
      return repr_str.format(
          v=self, np_repr=numpy_text(self.read_value(), is_repr=True))
    else:
      repr_str = ("<AutoCastVariable '{v.name}' shape={v.shape} "
                  'dtype={v.dtype.name} dtype_to_cast_to={v._cast_dtype.name}>')
      return repr_str.format(v=self)

  # Method delegations: We delegate the following methods to self._variable.
  # Each of these methods simply calls the same method on self._variable. The
  # base Variable raises NotImplementedError for most of these, so we must
  # override them.
  #
  # We do not define the following methods from Variable for the following
  # reasons:
  #   * 'count_up_to': This method only applies to int variables, which cannot
  #     be wrapped with an AutoCastVariable.
  #   * 'ref': Instead we inherit the definition from Variable.
  #     If we defined and delegated to Variable, the ref of an AutoCastVariable
  #     would be the same as the ref of the underlying variable, which would be
  #     strange as they are different Python objects.

  def set_shape(self, shape):
    return self._variable.set_shape(self, shape)

  @property
  def trainable(self):
    return self._variable.trainable

  @property
  def synchronization(self):
    return self._variable.synchronization

  @property
  def aggregation(self):
    return self._variable.aggregation

  def eval(self, session=None):
    return self._variable.eval(session)

  def initialized_value(self):
    return self._variable.initialized_value()

  @property
  def initial_value(self):
    return self._variable.initial_value

  @property
  def constraint(self):
    return self._variable.constraint

  def _apply_assign_update(self,
                           update_fn,
                           value,
                           use_locking=None,
                           name=None,
                           read_value=True):
    # TODO(b/146181571): This logic can be simplified once
    # DistributedVariable.assign returns a DistributedVariable. Currently for
    # MirroredStrategy, it returns a Mirrored value.
    if ops.executing_eagerly_outside_functions():
      assign_op = update_fn(value, use_locking, name, False)
      if read_value:
        # We create a new AutoCastVariable with the same underlying tf.Variable.
        # The new AutoCastVariable is identical except the 'op' attribute is
        # defined. This matches the behavior of tf.Variable.assign.
        var = create_autocast_variable(self._variable)
        var._op = assign_op  # pylint:disable=protected-access
        return var
      return assign_op

    # Fallback to wrapping the returned variable in graph mode if possible
    assign_var = update_fn(value, use_locking, name, read_value)
    if read_value and resource_variable_ops.is_resource_variable(assign_var):
      return create_autocast_variable(assign_var)
    return assign_var

  def _apply_update(self, update_fn, *args, **kwargs):
    update_var = update_fn(*args, **kwargs)
    if ops.executing_eagerly_outside_functions():
      return self

    # Fallback to wrapping the returned variable in graph mode if possible
    if resource_variable_ops.is_resource_variable(update_var):
      return create_autocast_variable(update_var)
    return update_var

  def assign(self, value, use_locking=None, name=None, read_value=True):
    return self._apply_assign_update(self._variable.assign, value, use_locking,
                                     name, read_value)

  def assign_add(self, delta, use_locking=None, name=None, read_value=True):
    return self._apply_assign_update(self._variable.assign_add, delta,
                                     use_locking, name, read_value)

  def assign_sub(self, delta, use_locking=None, name=None, read_value=True):
    return self._apply_assign_update(self._variable.assign_sub, delta,
                                     use_locking, name, read_value)

  def scatter_sub(self, sparse_delta, use_locking=False, name=None):
    return self._apply_update(self._variable.scatter_sub, sparse_delta,
                              use_locking, name)

  def scatter_add(self, sparse_delta, use_locking=False, name=None):
    return self._apply_update(self._variable.scatter_add, sparse_delta,
                              use_locking, name)

  def scatter_max(self, sparse_delta, use_locking=False, name=None):
    return self._apply_update(self._variable.scatter_max, sparse_delta,
                              use_locking, name)

  def scatter_min(self, sparse_delta, use_locking=False, name=None):
    return self._apply_update(self._variable.scatter_min, sparse_delta,
                              use_locking, name)

  def scatter_mul(self, sparse_delta, use_locking=False, name=None):
    return self._apply_update(self._variable.scatter_mul, sparse_delta,
                              use_locking, name)

  def scatter_div(self, sparse_delta, use_locking=False, name=None):
    return self._apply_update(self._variable.scatter_div, sparse_delta,
                              use_locking, name)

  def scatter_update(self, sparse_delta, use_locking=False, name=None):
    return self._apply_update(self._variable.scatter_update, sparse_delta,
                              use_locking, name)

  def batch_scatter_update(self, sparse_delta, use_locking=False, name=None):
    return self._apply_update(self._variable.batch_scatter_update, sparse_delta,
                              use_locking, name)

  def scatter_nd_sub(self, indices, updates, name=None):
    return self._apply_update(self._variable.scatter_nd_sub, indices, updates,
                              name)

  def scatter_nd_add(self, indices, updates, name=None):
    return self._apply_update(self._variable.scatter_nd_add, indices, updates,
                              name)

  def scatter_nd_update(self, indices, updates, name=None):
    return self._apply_update(self._variable.scatter_nd_update, indices,
                              updates, name)

  def load(self, value, session=None):
    return self._variable.load(value, session)

  @property
  def name(self):
    return self._variable.name

  @property
  def _shared_name(self):
    return self._variable._shared_name  # pylint:disable=protected-access

  @property
  def initializer(self):
    return self._variable.initializer

  @property
  def device(self):
    return self._variable.device

  @property
  def op(self):
    if self._op == 'delegate':
      return self._variable.op
    return self._op

  def _as_graph_element(self):
    graph_element = self._variable._as_graph_element()  # pylint:disable=protected-access
    if graph_element is None:
      return self._op
    return graph_element

  @property
  def graph(self):
    return self._variable.graph

  @property
  def shape(self):
    return self._variable.shape

  def get_shape(self):
    return self._variable.get_shape()

  def _gather_saveables_for_checkpoint(self):
    # By delegating this method to the wrapped variable, checkpoints with
    # AutoCastVariables are identical to checkpoints with normal variables.
    # Therefore models checkpointed with AutoCastVariables can be restored on
    # models with normal variables, and vice versa.
    return self._variable._gather_saveables_for_checkpoint()  # pylint:disable=protected-access

  def _map_resources(self, save_options):
    # By delegating this method to the wrapped variable, SavedModel with
    # AutoCastVariables are identical to SavedModel with normal variables.
    obj_map, resource_map = self._variable._map_resources(save_options)  # pylint:disable=protected-access
    obj_map[self] = obj_map[self._variable]
    return obj_map, resource_map

  # TODO(reedwm): Maybe encode the fact the variable is an AutoCastVariable in
  # to_proto().
  def to_proto(self, export_scope=None):
    return self._variable.to_proto(export_scope)

  def from_proto(self, variable_def, import_scope=None):
    return self._variable.from_proto(variable_def, import_scope)

  # Delegate the private attributes _handle_name and _initializer_op to
  # self._variable. SavedModel sets these attributes when loading a model. For
  # example, it sets _handle_name here:
  # https://github.com/tensorflow/tensorflow/blob/db26bd574fa95b5bdd53c08463dd19407cc0297e/tensorflow/python/keras/saving/saved_model/load.py#L211
  # We need to expose these attributes on AutoCastVariable as well for
  # SavedModel to work properly.
  # TODO(reedwm/kathywu): Find a better way to support SavedModel. Exposing
  # private attributes is hacky and difficult to maintain.
  @property
  def _handle_name(self):
    return self._variable._handle_name  # pylint: disable=protected-access

  @_handle_name.setter
  def _handle_name(self, handle_name):
    self._variable._handle_name = handle_name  # pylint: disable=protected-access

  @property
  def _initializer_op(self):
    return self._variable._initializer_op  # pylint: disable=protected-access

  @_initializer_op.setter
  def _initializer_op(self, initializer_op):
    self._variable._initializer_op = initializer_op  # pylint: disable=protected-access

  # Operator overloads:
  # Note we only overload operators that support floating-point types, as
  # non-float variables cannot be wrapped with an AutoCastVariable.
  # Also note: We call read_value() instead of value(), because value() causes
  # gradients not to work properly when TPUStrategy is used: b/143380936

  def __add__(self, o):
    return self.read_value() + o

  def __radd__(self, o):
    return o + self.read_value()

  def __sub__(self, o):
    return self.read_value() - o

  def __rsub__(self, o):
    return o - self.read_value()

  def __mul__(self, o):
    return self.read_value() * o

  def __rmul__(self, o):
    return o * self.read_value()

  def __truediv__(self, o):
    return self.read_value() / o

  def __rtruediv__(self, o):
    return o / self.read_value()

  def __floordiv__(self, o):
    return self.read_value() // o

  def __rfloordiv__(self, o):
    return o // self.read_value()

  def __mod__(self, o):
    return self.read_value() % o

  def __rmod__(self, o):
    return o % self.read_value()

  def __lt__(self, o):
    return self.read_value() < o

  def __le__(self, o):
    return self.read_value() <= o

  def __gt__(self, o):
    return self.read_value() > o

  def __ge__(self, o):
    return self.read_value() >= o

  def __getitem__(self, o):
    return self.read_value()[o]

  def __pow__(self, o, modulo=None):
    return pow(self.read_value(), o, modulo)

  def __rpow__(self, o):
    return pow(o, self.read_value())

  def __neg__(self):
    return -self.read_value()

  def __abs__(self):
    return abs(self.read_value())

  def __div__(self, o):
    try:
      return self.read_value().__div__(o)
    except AttributeError:
      # See https://docs.python.org/3/library/constants.html#NotImplemented
      return NotImplemented

  def __rdiv__(self, o):
    try:
      return self.read_value().__rdiv__(o)
    except AttributeError:
      # See https://docs.python.org/3/library/constants.html#NotImplemented
      return NotImplemented

  def __matmul__(self, o):
    try:
      return self.read_value().__matmul__(o)
    except AttributeError:
      # See https://docs.python.org/3/library/constants.html#NotImplemented
      return NotImplemented

  def __rmatmul__(self, o):
    try:
      return self.read_value().__rmatmul__(o)
    except AttributeError:
      # See https://docs.python.org/3/library/constants.html#NotImplemented
      return NotImplemented

  # pylint: enable=multiple-statements


ops.register_tensor_conversion_function(AutoCastVariable,
                                        AutoCastVariable._dense_var_to_tensor)  # pylint:disable=protected-access


def create_autocast_variable(variable):
  """Creates an AutoCastVariable that wraps another variable.

  This typically just returns `AutoCastVariable(variable)`. But, if the variable
  is a DistributedVariable or one of its subclasses, we instead dynamically
  create a class that subclasses from both AutoCastVariable and
  variable.__class__. This is so the returned variable will still pass
  `isinstance(variable, variable.__class__)`, which is required for
  DistributedVariables and its subclasses to work properly.

  Args:
    variable: A floating-point resource variable to wrap.

  Returns:
    An AutoCastVariable that wraps the variable.
  """
  if not isinstance(variable, (distribute_values.DistributedVariable,
                               ps_distribute_values.AggregatingVariable)):
    return AutoCastVariable(variable)

  class AutoCastDistributedVariable(AutoCastVariable, variable.__class__):
    """An AutoCastVariable that also subclasses from variable.__class__.

    variable.__class__ is either a DistributedVariable or an
    AggregatingVariable.
    """

    def __repr__(self):
      if issubclass(ps_distribute_values.AggregatingVariable,
                    variable.__class__):
        # AggregatingVariable's __repr__ simply calls super.__repr__. So we do
        # the same here for consistency, which calls AutoCastVariable.__repr__.
        return super(AutoCastDistributedVariable, self).__repr__()

      # pylint: disable=missing-format-attribute
      return ('<AutoCastDistributedVariable dtype={v.dtype.name} '
              'dtype_to_cast_to={v._cast_dtype.name} '
              'inner_variable={v._variable}>'
             ).format(v=self)
      # pylint: enable=missing-format-attribute

  return AutoCastDistributedVariable(variable)


class enable_auto_cast_variables(object):  # pylint:disable=invalid-name
  """Context manager which enables the autocasting of `AutoCastVariable`s.

  Under this context manager, `AutoCastVariable`s will be cast to `dtype` if
  `dtype` is floating-point. Otherwise, `AutoCastVariable`s will not be cast.
  """

  __slots__ = ['_dtype', '_prev_dtype']

  def __init__(self, dtype):
    if dtype and not dtype.is_floating:
      dtype = None
    self._dtype = dtype

  def __enter__(self):
    self._prev_dtype = getattr(_autocast_dtype, 'dtype', None)
    _autocast_dtype.dtype = self._dtype

  def __exit__(self, type_arg, value_arg, traceback_arg):
    _autocast_dtype.dtype = self._prev_dtype