stateful_random_ops.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.
# ==============================================================================
"""Operations for generating random numbers."""

from tensorflow.python.distribute import distribution_strategy_context as ds_context
from tensorflow.python.distribute import sharded_variable
from tensorflow.python.distribute import values_util
from tensorflow.python.eager import context
from tensorflow.python.framework import config
from tensorflow.python.framework import dtypes
from tensorflow.python.framework import ops
from tensorflow.python.ops import array_ops
from tensorflow.python.ops import array_ops_stack
from tensorflow.python.ops import gen_stateful_random_ops
from tensorflow.python.ops import gen_stateless_random_ops_v2
from tensorflow.python.ops import math_ops
from tensorflow.python.ops import resource_variable_ops
from tensorflow.python.ops import stateless_random_ops
from tensorflow.python.ops import variables
from tensorflow.python.ops.stateless_random_ops import Algorithm
from tensorflow.python.trackable import autotrackable
from tensorflow.python.util import nest
from tensorflow.python.util.tf_export import tf_export


# A seed for random ops (stateful and stateless) will always be 1024
# bits, all of which will be sent to the C++ code. The actual C++
# implementation of some algorithms may only use a lower part of the bits.

UINT64_HALF_SPAN = 2**63
MAX_INT64 = UINT64_HALF_SPAN - 1
MIN_INT64 = -UINT64_HALF_SPAN
UINT64_SPAN = UINT64_HALF_SPAN * 2
# 'Variable' doesn't support uint32 or uint64 yet (due to reasons explained in
# b/111604096 and cl/171681867), so I use signed int here. I choose int64
# instead of int32 here because `VarHandleOp` doesn't support int32 on GPU.
SEED_TYPE = "int64"
SEED_MIN = MIN_INT64
SEED_MAX = MAX_INT64
SEED_UINT_SPAN = UINT64_SPAN
SEED_TYPE_BITS = 64
SEED_BIT_MASK = 0xFFFFFFFFFFFFFFFF
SEED_SIZE = 16  # in units of SEED_TYPE


STATE_TYPE = SEED_TYPE
ALGORITHM_TYPE = STATE_TYPE


# The following sizes are all in unit of uint64.
PHILOX_KEY_SIZE = 1
THREEFRY_KEY_SIZE = 1
PHILOX_COUNTER_SIZE = 2
THREEFRY_COUNTER_SIZE = 1
PHILOX_STATE_SIZE = PHILOX_COUNTER_SIZE + PHILOX_KEY_SIZE
THREEFRY_STATE_SIZE = THREEFRY_COUNTER_SIZE + THREEFRY_KEY_SIZE


RNG_ALG_PHILOX = Algorithm.PHILOX.value
RNG_ALG_THREEFRY = Algorithm.THREEFRY.value


DEFAULT_ALGORITHM = RNG_ALG_PHILOX


def non_deterministic_ints(shape, dtype=dtypes.int64):
  """Non-deterministically generates some integers.

  This op may use some OS-provided source of non-determinism (e.g. an RNG), so
  each execution will give different results.

  Args:
    shape: the shape of the result.
    dtype: (optional) the dtype of the result.

  Returns:
    a tensor whose element values are non-deterministically chosen.
  """
  return gen_stateful_random_ops.non_deterministic_ints(
      shape=shape, dtype=dtype)


def _uint_to_int(n):
  if isinstance(n, int) and n > SEED_MAX:
    n = n - SEED_UINT_SPAN
  return n


def _make_1d_state(state_size, seed):
  """Makes a 1-D RNG state.

  Args:
    state_size: an integer.
    seed: an integer or 1-D tensor.

  Returns:
    a 1-D tensor of shape [state_size] and dtype STATE_TYPE.
  """
  if isinstance(seed, int):
    # chop the Python integer (infinite precision) into chunks of SEED_TYPE
    ls = []
    for _ in range(state_size):
      ls.append(seed & SEED_BIT_MASK)
      seed >>= SEED_TYPE_BITS
    seed = ls
  # to avoid overflow error from ops.convert_to_tensor
  seed = nest.map_structure(_uint_to_int, seed)
  seed = math_ops.cast(seed, STATE_TYPE)
  seed = array_ops.reshape(seed, [-1])
  seed = seed[0:state_size]
  # Padding with zeros on the *left* if too short. Padding on the right would
  # cause a small seed to be used as the "counter" while the "key" is always
  # zero (for counter-based RNG algorithms), because in the current memory
  # layout counter is stored before key. In such a situation two RNGs with
  # two different small seeds may generate overlapping outputs.
  seed_size = seed.shape[0]
  if seed_size is None:
    seed_size = array_ops.shape(seed)[0]
  padding_size = math_ops.maximum(state_size - seed_size, 0)
  padding = array_ops.zeros([padding_size], seed.dtype)
  # can't use `pad` because it doesn't support integer dtypes on GPU
  seed = array_ops.concat([padding, seed], axis=0)
  seed.set_shape([state_size])
  return seed


def _get_counter_size(alg):
  if alg == Algorithm.PHILOX.value:
    return PHILOX_COUNTER_SIZE
  elif alg == Algorithm.THREEFRY.value:
    return THREEFRY_COUNTER_SIZE
  elif alg == Algorithm.AUTO_SELECT.value:
    # For AUTO_SELECT, we'll manage the counter as if it's for Philox.
    return PHILOX_COUNTER_SIZE
  else:
    raise ValueError(stateless_random_ops.unsupported_alg_error_msg(alg))


def _get_state_size(alg):
  if alg == Algorithm.PHILOX.value:
    return PHILOX_STATE_SIZE
  elif alg == Algorithm.THREEFRY.value:
    return THREEFRY_STATE_SIZE
  elif alg == Algorithm.AUTO_SELECT.value:
    # For AUTO_SELECT, we'll manage the state as if it's for Philox.
    return PHILOX_STATE_SIZE
  else:
    raise ValueError(stateless_random_ops.unsupported_alg_error_msg(alg))


def _check_state_shape(shape, alg):
  if isinstance(alg, ops.Tensor) and not context.executing_eagerly():
    return
  shape.assert_is_compatible_with([_get_state_size(int(alg))])


def _make_state_from_seed(seed, alg):
  return _make_1d_state(_get_state_size(alg), seed)


@tf_export("random.create_rng_state", "random.experimental.create_rng_state")
def create_rng_state(seed, alg):
  """Creates a RNG state from an integer or a vector.

  Example:

  >>> tf.random.create_rng_state(
  ...     1234, "philox")
  <tf.Tensor: shape=(3,), dtype=int64, numpy=array([1234,    0,    0])>
  >>> tf.random.create_rng_state(
  ...     [12, 34], "threefry")
  <tf.Tensor: shape=(2,), dtype=int64, numpy=array([12, 34])>

  Args:
    seed: an integer or 1-D numpy array.
    alg: the RNG algorithm. Can be a string, an `Algorithm` or an integer.

  Returns:
    a 1-D numpy array whose size depends on the algorithm.
  """
  alg = stateless_random_ops.convert_alg_to_int(alg)
  return _make_state_from_seed(seed, alg)


def _shape_tensor(shape):
  """Convert to an int32 or int64 tensor, defaulting to int64 if empty."""
  if isinstance(shape, (tuple, list)) and not shape:
    dtype = dtypes.int64
  else:
    dtype = None
  return ops.convert_to_tensor(shape, dtype=dtype, name="shape")


def _convert_to_state_tensor(t):
  # to avoid out-of-range error from ops.convert_to_tensor
  t = nest.map_structure(_uint_to_int, t)
  return math_ops.cast(t, STATE_TYPE)


def get_replica_id():
  rctx = ds_context.get_replica_context()
  if rctx is None:
    return None
  return rctx.replica_id_in_sync_group


@tf_export("random.Generator", "random.experimental.Generator")
class Generator(autotrackable.AutoTrackable):
  """Random-number generator.

  Example:

  Creating a generator from a seed:

  >>> g = tf.random.Generator.from_seed(1234)
  >>> g.normal(shape=(2, 3))
  <tf.Tensor: shape=(2, 3), dtype=float32, numpy=
  array([[ 0.9356609 ,  1.0854305 , -0.93788373],
         [-0.5061547 ,  1.3169702 ,  0.7137579 ]], dtype=float32)>

  Creating a generator from a non-deterministic state:

  >>> g = tf.random.Generator.from_non_deterministic_state()
  >>> g.normal(shape=(2, 3))
  <tf.Tensor: shape=(2, 3), dtype=float32, numpy=...>

  All the constructors allow explicitly choosing an Random-Number-Generation
  (RNG) algorithm. Supported algorithms are `"philox"` and `"threefry"`. For
  example:

  >>> g = tf.random.Generator.from_seed(123, alg="philox")
  >>> g.normal(shape=(2, 3))
  <tf.Tensor: shape=(2, 3), dtype=float32, numpy=
  array([[ 0.8673864 , -0.29899067, -0.9310337 ],
         [-1.5828488 ,  1.2481191 , -0.6770643 ]], dtype=float32)>

  CPU, GPU and TPU with the same algorithm and seed will generate the same
  integer random numbers. Float-point results (such as the output of `normal`)
  may have small numerical discrepancies between different devices.

  This class uses a `tf.Variable` to manage its internal state. Every time
  random numbers are generated, the state of the generator will change. For
  example:

  >>> g = tf.random.Generator.from_seed(1234)
  >>> g.state
  <tf.Variable ... numpy=array([1234,    0,    0])>
  >>> g.normal(shape=(2, 3))
  <...>
  >>> g.state
  <tf.Variable ... numpy=array([2770,    0,    0])>

  The shape of the state is algorithm-specific.

  There is also a global generator:

  >>> g = tf.random.get_global_generator()
  >>> g.normal(shape=(2, 3))
  <tf.Tensor: shape=(2, 3), dtype=float32, numpy=...>

  When creating a generator inside a `tf.distribute.Strategy` scope, each
  replica will get a different stream of random numbers.

  For example, in this code:

  ```
  strat = tf.distribute.MirroredStrategy(devices=["cpu:0", "cpu:1"])
  with strat.scope():
    g = tf.random.Generator.from_seed(1)
    def f():
      return g.normal([])
    results = strat.run(f).values
  ```

  `results[0]` and `results[1]` will have different values.

  If the generator is seeded (e.g. created via `Generator.from_seed`), the
  random numbers will be determined by the seed, even though different replicas
  get different numbers.  One can think of a random number generated on a
  replica as a hash of the replica ID and a "master" random number that may be
  common to all replicas. Hence, the whole system is still deterministic.

  (Note that the random numbers on different replicas are not correlated, even
  if they are deterministically determined by the same seed. They are not
  correlated in the sense that no matter what statistics one calculates on them,
  there won't be any discernable correlation.)

  Generators can be freely saved and restored using `tf.train.Checkpoint`. The
  checkpoint can be restored in a distribution strategy with a different number
  of replicas than the original strategy. If a replica ID is present in both the
  original and the new distribution strategy, its state will be properly
  restored (i.e. the random-number stream from the restored point will be the
  same as that from the saving point) unless the replicas have already diverged
  in their RNG call traces before saving (e.g. one replica has made one RNG call
  while another has made two RNG calls). We don't have such guarantee if the
  generator is saved in a strategy scope and restored outside of any strategy
  scope, or vice versa.

  When a generator is created within the scope of
  `tf.distribute.experimental.ParameterServerStrategy`, the workers
  will share the generator's state (placed on one of the parameter
  servers). In this way the workers will still get different
  random-number streams, as stated above. (This is similar to replicas
  in a `tf.distribute.MirroredStrategy` sequentially accessing a
  generator created outside the strategy.) Each RNG call on a worker
  will incur a round-trip to a parameter server, which may have
  performance impacts. When creating a
  `tf.distribute.experimental.ParameterServerStrategy`, please make
  sure that the `variable_partitioner` argument won't shard small
  variables of shape `[2]` or `[3]` (because generator states must not
  be sharded). Ways to avoid sharding small variables include setting
  `variable_partitioner` to `None` or to
  `tf.distribute.experimental.partitioners.MinSizePartitioner` with a
  large enough `min_shard_bytes` (see
  `tf.distribute.experimental.ParameterServerStrategy`'s documentation
  for more details).
  """

  @classmethod
  def from_state(cls, state, alg):
    """Creates a generator from a state.

    See `__init__` for description of `state` and `alg`.

    Args:
      state: the new state.
      alg: the RNG algorithm.

    Returns:
      The new generator.
    """
    return cls(alg=alg, state=state)

  @classmethod
  def from_seed(cls, seed, alg=None):
    """Creates a generator from a seed.

    A seed is a 1024-bit unsigned integer represented either as a Python
    integer or a vector of integers. Seeds shorter than 1024-bit will be
    padded. The padding, the internal structure of a seed and the way a seed
    is converted to a state are all opaque (unspecified). The only semantics
    specification of seeds is that two different seeds are likely to produce
    two independent generators (but no guarantee).

    Args:
      seed: the seed for the RNG.
      alg: (optional) the RNG algorithm. If None, it will be auto-selected. See
        `__init__` for its possible values.

    Returns:
      The new generator.
    """
    if alg is None:
      # TODO(b/170668986): more sophisticated algorithm selection
      alg = DEFAULT_ALGORITHM
    alg = stateless_random_ops.convert_alg_to_int(alg)
    state = create_rng_state(seed, alg)
    return cls(state=state, alg=alg)

  @classmethod
  def from_non_deterministic_state(cls, alg=None):
    """Creates a generator by non-deterministically initializing its state.

    The source of the non-determinism will be platform- and time-dependent.

    Args:
      alg: (optional) the RNG algorithm. If None, it will be auto-selected. See
        `__init__` for its possible values.

    Returns:
      The new generator.
    """
    if config.is_op_determinism_enabled():
      raise RuntimeError('"from_non_deterministic_state" cannot be called when '  # pylint: disable=g-doc-exception
                         "determinism is enabled.")
    if alg is None:
      # TODO(b/170668986): more sophisticated algorithm selection
      alg = DEFAULT_ALGORITHM
    alg = stateless_random_ops.convert_alg_to_int(alg)
    state = non_deterministic_ints(shape=[_get_state_size(alg)],
                                   dtype=SEED_TYPE)
    return cls(state=state, alg=alg)

  @classmethod
  def from_key_counter(cls, key, counter, alg):
    """Creates a generator from a key and a counter.

    This constructor only applies if the algorithm is a counter-based algorithm.
    See method `key` for the meaning of "key" and "counter".

    Args:
      key: the key for the RNG, a scalar of type STATE_TYPE.
      counter: a vector of dtype STATE_TYPE representing the initial counter for
        the RNG, whose length is algorithm-specific.,
      alg: the RNG algorithm. If None, it will be auto-selected. See
        `__init__` for its possible values.

    Returns:
      The new generator.
    """
    counter = _convert_to_state_tensor(counter)
    key = _convert_to_state_tensor(key)
    alg = stateless_random_ops.convert_alg_to_int(alg)
    counter.shape.assert_is_compatible_with([_get_state_size(alg) - 1])
    key.shape.assert_is_compatible_with([])
    key = array_ops.reshape(key, [1])
    state = array_ops.concat([counter, key], 0)
    return cls(state=state, alg=alg)

  def __init__(self, copy_from=None, state=None, alg=None):
    """Creates a generator.

    The new generator will be initialized by one of the following ways, with
    decreasing precedence:
    (1) If `copy_from` is not None, the new generator is initialized by copying
        information from another generator.
    (2) If `state` and `alg` are not None (they must be set together), the new
        generator is initialized by a state.

    Args:
      copy_from: a generator to be copied from.
      state: a vector of dtype STATE_TYPE representing the initial state of the
        RNG, whose length and semantics are algorithm-specific. If it's a
        variable, the generator will reuse it instead of creating a new
        variable.
      alg: the RNG algorithm. Possible values are
        `tf.random.Algorithm.PHILOX` for the Philox algorithm and
        `tf.random.Algorithm.THREEFRY` for the ThreeFry algorithm
        (see paper 'Parallel Random Numbers: As Easy as 1, 2, 3'
        [https://www.thesalmons.org/john/random123/papers/random123sc11.pdf]).
        The string names `"philox"` and `"threefry"` can also be used.
        Note `PHILOX` guarantees the same numbers are produced (given
        the same random state) across all architectures (CPU, GPU, XLA etc).
    """
    # TODO(b/175072242): Remove distribution-strategy dependencies in this file.
    if ds_context.has_strategy():
      self._distribution_strategy = ds_context.get_strategy()
    else:
      self._distribution_strategy = None
    if copy_from is not None:
      # All other arguments should be None
      assert (alg or state) is None
      self._state_var = self._create_variable(copy_from.state, dtype=STATE_TYPE,
                                              trainable=False)
      self._alg = copy_from.algorithm
    else:
      assert alg is not None and state is not None
      alg = stateless_random_ops.convert_alg_to_int(alg)
      if isinstance(state, variables.Variable):
        _check_state_shape(state.shape, alg)
        self._state_var = state
      else:
        state = _convert_to_state_tensor(state)
        _check_state_shape(state.shape, alg)
        self._state_var = self._create_variable(state, dtype=STATE_TYPE,
                                                trainable=False)
      self._alg = alg

  def _create_variable(self, *args, **kwargs):
    """Creates a variable.

    Args:
      *args: positional arguments passed along to `variables.Variable.
      **kwargs: keyword arguments passed along to `variables.Variable.

    Returns:
      The created variable.
    """
    with ops.name_scope("random_generator"):
      # Make sure we don't change this name since Keras was using this name
      # to filter out the state variable.
      kwargs["name"] = "StateVar"
      v = variables.Variable(*args, **kwargs)
    if isinstance(v, sharded_variable.ShardedVariable):
      # RNG state is an atomic entity representing a 128-bit or
      # 192-bit value, so it mustn't be sharded.
      raise ValueError(
          "tf.random.Generator state is sharded, which is not allowed. When "
          "creating a tf.distribute.experimental.ParameterServerStrategy, "
          "please make sure that the `variable_partitioner` "
          "argument won't shard a "
          "small variable of shape [2] or [3]. Ways to avoid sharding small "
          "variables include setting `variable_partitioner` to None or to "
          "tf.distribute.experimental.partitioners.MinSizePartitioner with a "
          "large enough `min_shard_bytes`.")
    return v

  def reset(self, state):
    """Resets the generator by a new state.

    See `__init__` for the meaning of "state".

    Args:
      state: the new state.
    """
    state = _convert_to_state_tensor(state)
    state.shape.assert_is_compatible_with([_get_state_size(self.algorithm)])
    self._state_var.assign(state)

  def reset_from_seed(self, seed):
    """Resets the generator by a new seed.

    See `from_seed` for the meaning of "seed".

    Args:
      seed: the new seed.
    """
    state = create_rng_state(seed, self.algorithm)
    self._state_var.assign(state)

  def reset_from_key_counter(self, key, counter):
    """Resets the generator by a new key-counter pair.

    See `from_key_counter` for the meaning of "key" and "counter".

    Args:
      key: the new key.
      counter: the new counter.
    """
    counter = _convert_to_state_tensor(counter)
    key = _convert_to_state_tensor(key)
    counter.shape.assert_is_compatible_with(
        [_get_state_size(self.algorithm) - 1])
    key.shape.assert_is_compatible_with([])
    key = array_ops.reshape(key, [1])
    state = array_ops.concat([counter, key], 0)
    self._state_var.assign(state)

  @property
  def state(self):
    """The internal state of the RNG."""
    return self._state_var

  @property
  def algorithm(self):
    """The RNG algorithm id (a Python integer or scalar integer Tensor)."""
    return self._alg

  def _standard_normal(self, shape, dtype):
    key, counter = self._prepare_key_counter(shape)
    return gen_stateless_random_ops_v2.stateless_random_normal_v2(
        shape, key=key, counter=counter, dtype=dtype, alg=self.algorithm)

  @property
  def key(self):
    """The 'key' part of the state of a counter-based RNG.

    For a counter-base RNG algorithm such as Philox and ThreeFry (as
    described in paper 'Parallel Random Numbers: As Easy as 1, 2, 3'
    [https://www.thesalmons.org/john/random123/papers/random123sc11.pdf]),
    the RNG state consists of two parts: counter and key. The output is
    generated via the formula: output=hash(key, counter), i.e. a hashing of
    the counter parametrized by the key. Two RNGs with two different keys can
    be thought as generating two independent random-number streams (a stream
    is formed by increasing the counter).

    Returns:
      A scalar which is the 'key' part of the state, if the RNG algorithm is
        counter-based; otherwise it raises a ValueError.
    """
    alg = self.algorithm
    if alg in (a.value for a in Algorithm):
      return self._state_var[-1]
    else:
      raise ValueError(stateless_random_ops.unsupported_alg_error_msg(alg))

  def _skip_single_var(self, var, delta):
    resource_variable_ops.variable_accessed(var)
    # TODO(wangpeng): Cache the cast algorithm instead of casting everytime.
    return gen_stateful_random_ops.rng_read_and_skip(
        var.handle,
        alg=math_ops.cast(self.algorithm, dtypes.int32),
        delta=math_ops.cast(delta, dtypes.uint64))

  def skip(self, delta):
    """Advance the counter of a counter-based RNG.

    Args:
      delta: the amount of advancement. The state of the RNG after
        `skip(n)` will be the same as that after `normal([n])`
        (or any other distribution). The actual increment added to the
        counter is an unspecified implementation detail.

    Returns:
      A `Tensor` of type `int64`.
    """

    def update_fn(v):
      return self._skip_single_var(v, delta)
    # TODO(b/170515001): Always call strategy.extended.update after calling it
    #   from both replica context and cross-replica context is supported.
    if values_util.is_saving_non_distributed():
      # Assumes replica context with replica_id=0, since we only save the first
      # replica.
      return update_fn(self.state)
    if self._distribution_strategy is not None:
      with ds_context.enter_or_assert_strategy(self._distribution_strategy):
        if ds_context.in_cross_replica_context():
          # Code that operates on all replicas of a variable cannot be saved
          # without retracing.
          values_util.mark_as_unsaveable()
        if (ds_context.in_cross_replica_context() or
            "CentralStorage" in type(self._distribution_strategy).__name__):
          # In cross-replica context we need to use strategy.extended.update.
          # In CentralStorageStrategy we also need to use
          # strategy.extended.update (even for replica context),
          # because variable updates here must be within merge_call.
          return ds_context.get_strategy().extended.update(
              self.state, update_fn)
    return update_fn(self.state)

  def _preprocess_key(self, key):
    if self._distribution_strategy is None:
      return key
    with ds_context.enter_or_assert_strategy(self._distribution_strategy):
      replica_id = get_replica_id()
      if replica_id is not None:
        replica_id = array_ops_stack.stack([replica_id, 0], axis=0)
        replica_id = math_ops.cast(replica_id, dtypes.uint64)
        # Conceptually: key = hash(key, replica_id)
        key = gen_stateless_random_ops_v2.stateless_random_uniform_full_int_v2(
            shape=[1], key=key, counter=replica_id, dtype=dtypes.uint64,
            alg=self.algorithm)
      return key

  def _prepare_key_counter(self, shape):
    delta = math_ops.reduce_prod(shape)
    counter_key = self.skip(delta)
    counter_size = _get_counter_size(self.algorithm)
    counter = array_ops.bitcast(counter_key[:counter_size], dtypes.uint64)
    key = array_ops.bitcast(counter_key[counter_size:counter_size + 1],
                            dtypes.uint64)
    key = self._preprocess_key(key)
    return key, counter

  # The following functions return a tensor and as a side effect update
  # self._state_var.
  def normal(self, shape, mean=0.0, stddev=1.0, dtype=dtypes.float32,
             name=None):
    """Outputs random values from a normal distribution.

    Args:
      shape: A 1-D integer Tensor or Python array. The shape of the output
        tensor.
      mean: A 0-D Tensor or Python value of type `dtype`. The mean of the normal
        distribution.
      stddev: A 0-D Tensor or Python value of type `dtype`. The standard
        deviation of the normal distribution.
      dtype: The type of the output.
      name: A name for the operation (optional).

    Returns:
      A tensor of the specified shape filled with random normal values.
    """
    with ops.name_scope(name, "stateful_normal", [shape, mean, stddev]) as name:
      shape = _shape_tensor(shape)
      mean = ops.convert_to_tensor(mean, dtype=dtype, name="mean")
      stddev = ops.convert_to_tensor(stddev, dtype=dtype, name="stddev")
      rnd = self._standard_normal(shape, dtype=dtype)
      return math_ops.add(rnd * stddev, mean, name=name)

  def _truncated_normal(self, shape, dtype):
    key, counter = self._prepare_key_counter(shape)
    return gen_stateless_random_ops_v2.stateless_truncated_normal_v2(
        shape=shape, key=key, counter=counter, dtype=dtype, alg=self.algorithm)

  def truncated_normal(self, shape,
                       mean=0.0,
                       stddev=1.0,
                       dtype=dtypes.float32,
                       name=None):
    """Outputs random values from a truncated normal distribution.

    The generated values follow a normal distribution with specified mean and
    standard deviation, except that values whose magnitude is more than
    2 standard deviations from the mean are dropped and re-picked.

    Args:
      shape: A 1-D integer Tensor or Python array. The shape of the output
        tensor.
      mean: A 0-D Tensor or Python value of type `dtype`. The mean of the
        truncated normal distribution.
      stddev: A 0-D Tensor or Python value of type `dtype`. The standard
        deviation of the normal distribution, before truncation.
      dtype: The type of the output.
      name: A name for the operation (optional).

    Returns:
      A tensor of the specified shape filled with random truncated normal
        values.
    """
    with ops.name_scope(
        name, "truncated_normal", [shape, mean, stddev]) as name:
      shape_tensor = _shape_tensor(shape)
      mean_tensor = ops.convert_to_tensor(mean, dtype=dtype, name="mean")
      stddev_tensor = ops.convert_to_tensor(stddev, dtype=dtype, name="stddev")
      rnd = self._truncated_normal(shape_tensor, dtype=dtype)
      mul = rnd * stddev_tensor
      return math_ops.add(mul, mean_tensor, name=name)

  def _uniform(self, shape, dtype):
    key, counter = self._prepare_key_counter(shape)
    return gen_stateless_random_ops_v2.stateless_random_uniform_v2(
        shape=shape, key=key, counter=counter, dtype=dtype, alg=self.algorithm)

  def _uniform_full_int(self, shape, dtype, name=None):
    key, counter = self._prepare_key_counter(shape)
    return gen_stateless_random_ops_v2.stateless_random_uniform_full_int_v2(
        shape=shape,
        key=key,
        counter=counter,
        dtype=dtype,
        alg=self.algorithm,
        name=name)

  def uniform(self, shape, minval=0, maxval=None,
              dtype=dtypes.float32, name=None):
    """Outputs random values from a uniform distribution.

    The generated values follow a uniform distribution in the range
    `[minval, maxval)`. The lower bound `minval` is included in the range, while
    the upper bound `maxval` is excluded. (For float numbers especially
    low-precision types like bfloat16, because of
    rounding, the result may sometimes include `maxval`.)

    For floats, the default range is `[0, 1)`.  For ints, at least `maxval` must
    be specified explicitly.

    In the integer case, the random integers are slightly biased unless
    `maxval - minval` is an exact power of two.  The bias is small for values of
    `maxval - minval` significantly smaller than the range of the output (either
    `2**32` or `2**64`).

    For full-range random integers, pass `minval=None` and `maxval=None` with an
    integer `dtype` (for integer dtypes, `minval` and `maxval` must be both
    `None` or both not `None`).

    Args:
      shape: A 1-D integer Tensor or Python array. The shape of the output
        tensor.
      minval: A Tensor or Python value of type `dtype`, broadcastable with
        `shape` (for integer types, broadcasting is not supported, so it needs
        to be a scalar). The lower bound (included) on the range of random
        values to generate. Pass `None` for full-range integers. Defaults to 0.
      maxval: A Tensor or Python value of type `dtype`, broadcastable with
        `shape` (for integer types, broadcasting is not supported, so it needs
        to be a scalar). The upper bound (excluded) on the range of random
        values to generate. Pass `None` for full-range integers. Defaults to 1
        if `dtype` is floating point.
      dtype: The type of the output.
      name: A name for the operation (optional).

    Returns:
      A tensor of the specified shape filled with random uniform values.

    Raises:
      ValueError: If `dtype` is integral and `maxval` is not specified.
    """
    dtype = dtypes.as_dtype(dtype)
    if dtype.is_integer:
      if (minval is None) != (maxval is None):
        raise ValueError("For integer dtype {}, minval and maxval must be both "
                         "`None` or both non-`None`; got minval={} and "
                         "maxval={}".format(dtype, minval, maxval))
    elif maxval is None:
      maxval = 1
    with ops.name_scope(name, "stateful_uniform",
                        [shape, minval, maxval]) as name:
      shape = _shape_tensor(shape)
      if dtype.is_integer and minval is None:
        return self._uniform_full_int(shape=shape, dtype=dtype, name=name)
      minval = ops.convert_to_tensor(minval, dtype=dtype, name="min")
      maxval = ops.convert_to_tensor(maxval, dtype=dtype, name="max")
      if dtype.is_integer:
        key, counter = self._prepare_key_counter(shape)
        return gen_stateless_random_ops_v2.stateless_random_uniform_int_v2(
            shape=shape,
            key=key,
            counter=counter,
            minval=minval,
            maxval=maxval,
            alg=self.algorithm,
            name=name)
      else:
        rnd = self._uniform(shape=shape, dtype=dtype)
        return math_ops.add(rnd * (maxval - minval), minval, name=name)

  def uniform_full_int(self, shape, dtype=dtypes.uint64, name=None):
    """Uniform distribution on an integer type's entire range.

    This method is the same as setting `minval` and `maxval` to `None` in the
    `uniform` method.

    Args:
      shape: the shape of the output.
      dtype: (optional) the integer type, default to uint64.
      name: (optional) the name of the node.

    Returns:
      A tensor of random numbers of the required shape.
    """
    dtype = dtypes.as_dtype(dtype)
    with ops.name_scope(name, "stateful_uniform_full_int",
                        [shape]) as name:
      shape = _shape_tensor(shape)
      return self._uniform_full_int(shape=shape, dtype=dtype, name=name)

  def binomial(self, shape, counts, probs, dtype=dtypes.int32, name=None):
    """Outputs random values from a binomial distribution.

    The generated values follow a binomial distribution with specified count and
    probability of success parameters.

    Example:

    ```python
    counts = [10., 20.]
    # Probability of success.
    probs = [0.8]

    rng = tf.random.Generator.from_seed(seed=234)
    binomial_samples = rng.binomial(shape=[2], counts=counts, probs=probs)


    counts = ... # Shape [3, 1, 2]
    probs = ...  # Shape [1, 4, 2]
    shape = [3, 4, 3, 4, 2]
    rng = tf.random.Generator.from_seed(seed=1717)
    # Sample shape will be [3, 4, 3, 4, 2]
    binomial_samples = rng.binomial(shape=shape, counts=counts, probs=probs)
    ```


    Args:
      shape: A 1-D integer Tensor or Python array. The shape of the output
        tensor.
      counts: Tensor. The counts of the binomial distribution. Must be
        broadcastable with `probs`, and broadcastable with the rightmost
        dimensions of `shape`.
      probs: Tensor. The probability of success for the
        binomial distribution. Must be broadcastable with `counts` and
        broadcastable with the rightmost dimensions of `shape`.
      dtype: The type of the output. Default: tf.int32
      name: A name for the operation (optional).

    Returns:
      samples: A Tensor of the specified shape filled with random binomial
        values.  For each i, each samples[i, ...] is an independent draw from
        the binomial distribution on counts[i] trials with probability of
        success probs[i].
    """
    dtype = dtypes.as_dtype(dtype)
    with ops.name_scope(name, "binomial", [shape, counts, probs]) as name:
      counts = ops.convert_to_tensor(counts, name="counts")
      probs = ops.convert_to_tensor(probs, name="probs")
      shape_tensor = _shape_tensor(shape)
      return gen_stateful_random_ops.stateful_random_binomial(
          self.state.handle,
          self.algorithm,
          shape=shape_tensor,
          counts=counts,
          probs=probs,
          dtype=dtype,
          name=name)

  # TODO(wangpeng): implement other distributions

  def _make_int64_keys(self, shape=()):
    # New independent keys are generated via
    # `new_key[i] = hash(old_key, counter+i)`, which is exactly what
    # `uniform_full_int(dtype=int64)` does for PhiloxRandom_64_128_128 and
    # ThreeFry_64_64_64.
    return self.uniform_full_int(shape=shape, dtype=dtypes.int64)

  def make_seeds(self, count=1):
    """Generates seeds for stateless random ops.

    For example:

    ```python
    seeds = get_global_generator().make_seeds(count=10)
    for i in range(10):
      seed = seeds[:, i]
      numbers = stateless_random_normal(shape=[2, 3], seed=seed)
      ...
    ```

    Args:
      count: the number of seed pairs (note that stateless random ops need a
        pair of seeds to invoke).

    Returns:
      A tensor of shape [2, count] and dtype int64.
    """
    alg = self.algorithm
    if alg in (a.value for a in Algorithm):
      keys = self._make_int64_keys(shape=[count])
      # The two seeds for stateless random ops don't have individual semantics
      # and are scrambled together, so setting one to zero is fine.
      zeros = array_ops.zeros_like(keys)
      return array_ops_stack.stack([keys, zeros])
    else:
      raise ValueError(stateless_random_ops.unsupported_alg_error_msg(alg))

  def split(self, count=1):
    """Returns a list of independent `Generator` objects.

    Two generators are independent of each other in the sense that the
    random-number streams they generate don't have statistically detectable
    correlations. The new generators are also independent of the old one.
    The old generator's state will be changed (like other random-number
    generating methods), so two calls of `split` will return different
    new generators.

    For example:

    ```python
    gens = get_global_generator().split(count=10)
    for gen in gens:
      numbers = gen.normal(shape=[2, 3])
      # ...
    gens2 = get_global_generator().split(count=10)
    # gens2 will be different from gens
    ```

    The new generators will be put on the current device (possible different
    from the old generator's), for example:

    ```python
    with tf.device("/device:CPU:0"):
      gen = Generator(seed=1234)  # gen is on CPU
    with tf.device("/device:GPU:0"):
      gens = gen.split(count=10)  # gens are on GPU
    ```

    Args:
      count: the number of generators to return.

    Returns:
      A list (length `count`) of `Generator` objects independent of each other.
      The new generators have the same RNG algorithm as the old one.
    """
    def _key_to_state(alg, key):
      # Padding with zeros on the left. The zeros will be the counter.
      return [0] * (_get_state_size(alg) - 1) + [key]

    alg = self.algorithm
    if alg in (a.value for a in Algorithm):
      keys = self._make_int64_keys(shape=[count])
      return [Generator(state=_key_to_state(alg, key), alg=alg)
              for key in array_ops_stack.unstack(keys, num=count)]
    else:
      raise ValueError(stateless_random_ops.unsupported_alg_error_msg(alg))


# It's not safe to create TF ops before `init_google` is called, so this is
# initialized to None and get a value the first time `get_global_generator` is
# called.
global_generator = None


@tf_export("random.get_global_generator",
           "random.experimental.get_global_generator")
def get_global_generator():
  """Retrieves the global generator.

  This function will create the global generator the first time it is called,
  and the generator will be placed at the default device at that time, so one
  needs to be careful when this function is first called. Using a generator
  placed on a less-ideal device will incur performance regression.

  Returns:
    The global `tf.random.Generator` object.
  """
  global global_generator
  if global_generator is None:
    if config.is_op_determinism_enabled():
      raise RuntimeError('"get_global_generator" cannot be called if '  # pylint: disable=g-doc-exception
                         "determinism is enabled, unless "
                         '"set_global_generator" has already been called. '
                         'Please call "set_global_generator" first.')
    with ops.init_scope():
      global_generator = Generator.from_non_deterministic_state()
  return global_generator


@tf_export("random.set_global_generator",
           "random.experimental.set_global_generator")
def set_global_generator(generator):
  """Replaces the global generator with another `Generator` object.

  This function replaces the global generator with the provided `generator`
  object.
  A random number generator utilizes a `tf.Variable` object to store its state.
  The user shall be aware of caveats how `set_global_generator` interacts with
  `tf.function`:

  - tf.function puts restrictions on Variable creation thus one cannot freely
    create a new random generator instance inside `tf.function`.
    To call `set_global_generator` inside `tf.function`, the generator instance
    must have already been created eagerly.
  - tf.function captures the Variable during trace-compilation, thus a compiled
    f.function will not be affected `set_global_generator` as demonstrated by
    random_test.py/RandomTest.testResetGlobalGeneratorBadWithDefun .

  For most use cases, avoid calling `set_global_generator` after program
  initialization, and prefer to reset the state of the existing global generator
  instead, such as,

  >>> rng = tf.random.get_global_generator()
  >>> rng.reset_from_seed(30)


  Args:
    generator: the new `Generator` object.
  """
  global global_generator
  global_generator = generator