Python array_ops.stop_gradient方法代码示例

# 需要导入模块: from tensorflow.python.ops import array_ops [as 别名]
# 或者: from tensorflow.python.ops.array_ops import stop_gradient [as 别名]
def get_score_function_with_advantage(advantage_fn=None,
                                      name="ScoreFunctionWithAdvantage"):
  """Score function estimator with advantage function.

  Args:
    advantage_fn: callable that takes the `StochasticTensor` and the
      downstream `loss` and returns a `Tensor` advantage
      (e.g. `loss - baseline`).
    name: name to prepend ops with.

  Returns:
    Callable score function estimator that takes the `StochasticTensor`, the
    sampled `value`, and the downstream `loss`, and uses the provided advantage.
  """

  def score_function_with_advantage(stochastic_tensor, value, loss):
    with ops.name_scope(name, values=[value, loss]):
      advantage = advantage_fn(stochastic_tensor, loss)
      advantage = array_ops.stop_gradient(advantage)
      return stochastic_tensor.distribution.log_prob(value) * advantage

  return score_function_with_advantage 

# 需要导入模块: from tensorflow.python.ops import array_ops [as 别名]
# 或者: from tensorflow.python.ops.array_ops import stop_gradient [as 别名]
def loss(self, final_loss, name="Loss"):
    # Return a loss based on final_loss and the distribution. Returns
    # None if pathwise derivatives are supported, if the loss_fn
    # was explicitly set to None, or if the value type is MeanValue.
    if self._loss_fn is None:
      return None

    if (self._dist.reparameterization_type == distribution.FULLY_REPARAMETERIZED
        and not self._value_type.stop_gradient):
      # Can perform pathwise-derivative on this one; no additional loss needed.
      return None

    with ops.name_scope(self.name, values=[final_loss]):
      with ops.name_scope(name):
        if (self._value_type.stop_gradient or
            isinstance(self._value_type, SampleValue)):
          return self._loss_fn(self, self._value, final_loss)
        elif isinstance(self._value_type, MeanValue):
          return None  # MeanValue generally provides its own gradient
        else:
          raise TypeError("Unrecognized Distribution Value Type: %s",
                          self._value_type) 

# 需要导入模块: from tensorflow.python.ops import array_ops [as 别名]
# 或者: from tensorflow.python.ops.array_ops import stop_gradient [as 别名]
def loss(self, sample_loss):
    """Returns the term to add to the surrogate loss.

    This method is called by `surrogate_loss`.  The input `sample_loss` should
    have already had `stop_gradient` applied to it.  This is because the
    surrogate_loss usually provides a Monte Carlo sample term of the form
    `differentiable_surrogate * sample_loss` where `sample_loss` is considered
    constant with respect to the input for purposes of the gradient.

    Args:
      sample_loss: `Tensor`, sample loss downstream of this `StochasticTensor`.

    Returns:
      Either `None` or a `Tensor`.
    """
    raise NotImplementedError("surrogate_loss not implemented") 

# 需要导入模块: from tensorflow.python.ops import array_ops [as 别名]
# 或者: from tensorflow.python.ops.array_ops import stop_gradient [as 别名]
def _create_value(self):
    """Create the value Tensor based on the value type, store as self._value."""

    if isinstance(self._value_type, MeanValue):
      value_tensor = self._dist.mean()
    elif isinstance(self._value_type, SampleValue):
      value_tensor = self._dist.sample(self._value_type.shape)
    else:
      raise TypeError("Unrecognized Distribution Value Type: %s",
                      self._value_type)

    if self._value_type.stop_gradient:
      # stop_gradient is being enforced by the value type
      return array_ops.stop_gradient(value_tensor)

    if isinstance(self._value_type, MeanValue):
      return value_tensor  # Using pathwise-derivative for this one.
    if self._dist.is_continuous and self._dist.is_reparameterized:
      return value_tensor  # Using pathwise-derivative for this one.
    else:
      # Will have to perform some variant of score function
      # estimation.  Call stop_gradient on the sampler just in case we
      # may accidentally leak some gradient from it.
      return array_ops.stop_gradient(value_tensor) 

# 需要导入模块: from tensorflow.python.ops import array_ops [as 别名]
# 或者: from tensorflow.python.ops.array_ops import stop_gradient [as 别名]
def _create_value(self):
    """Create the value Tensor based on the value type, store as self._value."""

    if isinstance(self._value_type, MeanValue):
      value_tensor = self._dist.mean()
    elif isinstance(self._value_type, SampleValue):
      value_tensor = self._dist.sample(self._value_type.shape)
    else:
      raise TypeError(
          "Unrecognized Distribution Value Type: %s", self._value_type)

    if self._value_type.stop_gradient:
      # stop_gradient is being enforced by the value type
      return array_ops.stop_gradient(value_tensor)

    if isinstance(self._value_type, MeanValue):
      return value_tensor  # Using pathwise-derivative for this one.
    if self._dist.is_continuous and self._dist.is_reparameterized:
      return value_tensor  # Using pathwise-derivative for this one.
    else:
      # Will have to perform some variant of score function
      # estimation.  Call stop_gradient on the sampler just in case we
      # may accidentally leak some gradient from it.
      return array_ops.stop_gradient(value_tensor) 

# 需要导入模块: from tensorflow.python.ops import array_ops [as 别名]
# 或者: from tensorflow.python.ops.array_ops import stop_gradient [as 别名]
def _argmax_or_mcsearch(embedding, output_projection=None, update_embedding=True, mc_search=False):
    def loop_function(prev, _):
        if output_projection is not None:
            prev = nn_ops.xw_plus_b(prev, output_projection[0], output_projection[1])


        if isinstance(mc_search, bool):
            prev_symbol = tf.reshape(tf.multinomial(prev, 1), [-1]) if mc_search else math_ops.argmax(prev, 1)
        else:
            prev_symbol = tf.cond(mc_search, lambda: tf.reshape(tf.multinomial(prev, 1), [-1]), lambda: tf.argmax(prev, 1))


        emb_prev = embedding_ops.embedding_lookup(embedding, prev_symbol)
        if not update_embedding:
            emb_prev = array_ops.stop_gradient(emb_prev)
        return emb_prev
    return loop_function 

# 需要导入模块: from tensorflow.python.ops import array_ops [as 别名]
# 或者: from tensorflow.python.ops.array_ops import stop_gradient [as 别名]
def _extract_argmax_and_embed(embedding, output_projection=None, update_embedding=True):
  """Get a loop_function that extracts the previous symbol and embeds it.

  Args:
    embedding: embedding tensor for symbols.
    output_projection: None or a pair (W, B). If provided, each fed previous
      output will first be multiplied by W and added B.
    update_embedding: Boolean; if False, the gradients will not propagate
      through the embeddings.

  Returns:
    A loop function.
  """
  def loop_function(prev, _):
    if output_projection is not None:
      prev = nn_ops.xw_plus_b(
          prev, output_projection[0], output_projection[1])
    prev_symbol = math_ops.argmax(prev, 1)
    # Note that gradients will not propagate through the second parameter of
    # embedding_lookup.
    emb_prev = embedding_ops.embedding_lookup(embedding, prev_symbol)
    if not update_embedding:
      emb_prev = array_ops.stop_gradient(emb_prev)
    return emb_prev
  return loop_function 

# 需要导入模块: from tensorflow.python.ops import array_ops [as 别名]
# 或者: from tensorflow.python.ops.array_ops import stop_gradient [as 别名]
def evaluate_on_sample(self, seed=None):
        """Evaluates the log probability on a random sample.
        Args:
          seed: int or None. Random seed for this draw from the distribution.
        Returns:
          Log probability of sampled targets, summed across examples.
        """
        if seed is None:
            seed = self._default_seed
        # We treat the targets as "constant".  It's only the inputs that get
        # "back-propped" through.
        return self._evaluate(array_ops.stop_gradient(self.sample(seed)))


# TODO(jamesmartens): should this just inherit from object to avoid "diamond"
# inheritance, or is there a better way? 

# 需要导入模块: from tensorflow.python.ops import array_ops [as 别名]
# 或者: from tensorflow.python.ops.array_ops import stop_gradient [as 别名]
def _extract_argmax_and_embed(embedding, output_projection=None,
                              update_embedding=True):
  """Get a loop_function that extracts the previous symbol and embeds it.
  Args:
    embedding: embedding tensor for symbols.
    output_projection: None or a pair (W, B). If provided, each fed previous
      output will first be multiplied by W and added B.
    update_embedding: Boolean; if False, the gradients will not propagate
      through the embeddings.
  Returns:
    A loop function.
  """
  def loop_function(prev, _):
    if output_projection is not None:
      prev = nn_ops.xw_plus_b(
          prev, output_projection[0], output_projection[1])
    prev_symbol = math_ops.argmax(prev, 1)
    # Note that gradients will not propagate through the second parameter of
    # embedding_lookup.
    emb_prev = embedding_ops.embedding_lookup(embedding, prev_symbol)
    if not update_embedding:
      emb_prev = array_ops.stop_gradient(emb_prev)
    return emb_prev
  return loop_function 

# 需要导入模块: from tensorflow.python.ops import array_ops [as 别名]
# 或者: from tensorflow.python.ops.array_ops import stop_gradient [as 别名]
def stop_gradient(variables):
  """Returns `variables` but with zero gradient w.r.t. every other variable.

  Arguments:
      variables: Tensor or list of tensors to consider constant with respect
        to any other variable.


  Returns:
      A single tensor or a list of tensors (depending on the passed argument)
      that has no gradient with respect to any other variable.
  """
  if isinstance(variables, (list, tuple)):
    return map(array_ops.stop_gradient, variables)
  return array_ops.stop_gradient(variables)


# CONTROL FLOW 

# 需要导入模块: from tensorflow.python.ops import array_ops [as 别名]
# 或者: from tensorflow.python.ops.array_ops import stop_gradient [as 别名]
def _force_data_dependency(first_compute, then_compute):
  """Force all of `then_compute` to depend on all of `first_compute`.

  Uses a dummy data dependency, which is useful when running on TPUs because
  XLA ignores control dependencies. Only supports float arguments.

  Args:
    first_compute: `list<Tensor>`. These will be made to run before the
      `Tensor`s `then_compute`.
    then_compute: `list<Tensor>`. These will run after all the `Tensor`s in
      `first_compute`.

  Returns:
    `list<Tensor>`, same length as `then_compute`.

  Raises:
    ValueError: if ranks are unknown or types are not floating.
  """

  def _first_element(x):
    if x.get_shape().ndims is None:
      raise ValueError("Rank of Tensor %s must be known" % x)
    ndims = x.get_shape().ndims
    begin = framework_ops.convert_to_tensor([0] * ndims, dtype=dtypes.int32)
    size = framework_ops.convert_to_tensor([1] * ndims, dtype=dtypes.int32)
    return array_ops.reshape(array_ops.slice(x, begin, size), [])

  first_compute_sum = math_ops.add_n(
      [_first_element(x) for x in first_compute if x is not None])
  dtype = first_compute_sum.dtype
  if not dtype.is_floating:
    raise ValueError("_force_data_dependency only supports floating dtypes.")
  epsilon = np.finfo(dtype.as_numpy_dtype).tiny
  zero = array_ops.stop_gradient(epsilon * first_compute_sum)

  return [
      array_ops.identity(x) + zero if x is not None else None
      for x in then_compute
  ] 

# 需要导入模块: from tensorflow.python.ops import array_ops [as 别名]
# 或者: from tensorflow.python.ops.array_ops import stop_gradient [as 别名]
def _AvgPool3DGradGrad(op, grad):
  return (array_ops.stop_gradient(op.inputs[0]), gen_nn_ops.avg_pool3d(
      grad,
      op.get_attr("ksize"),
      op.get_attr("strides"),
      op.get_attr("padding"),
      data_format=op.get_attr("data_format"))) 

# 需要导入模块: from tensorflow.python.ops import array_ops [as 别名]
# 或者: from tensorflow.python.ops.array_ops import stop_gradient [as 别名]
def _AvgPoolGradGrad(op, grad):
  return (array_ops.stop_gradient(op.inputs[0]), gen_nn_ops._avg_pool(
      grad,
      op.get_attr("ksize"),
      op.get_attr("strides"),
      op.get_attr("padding"),
      data_format=op.get_attr("data_format"))) 

# 需要导入模块: from tensorflow.python.ops import array_ops [as 别名]
# 或者: from tensorflow.python.ops.array_ops import stop_gradient [as 别名]
def stop_gradient(variables):
  """Returns `variables` but with zero gradient w.r.t. every other variable.

  Arguments:
      variables: List of variables.

  Returns:
      The same list of variables.
  """
  return array_ops.stop_gradient(variables)


# CONTROL FLOW 

# 需要导入模块: from tensorflow.python.ops import array_ops [as 别名]
# 或者: from tensorflow.python.ops.array_ops import stop_gradient [as 别名]
def _extract_argmax_and_embed(embedding,
                              output_projection=None,
                              update_embedding=True):
  """Get a loop_function that extracts the previous symbol and embeds it.

  Args:
    embedding: embedding tensor for symbols.
    output_projection: None or a pair (W, B). If provided, each fed previous
      output will first be multiplied by W and added B.
    update_embedding: Boolean; if False, the gradients will not propagate
      through the embeddings.

  Returns:
    A loop function.
  """

  def loop_function(prev, _):
    if output_projection is not None:
      prev = nn_ops.xw_plus_b(prev, output_projection[0], output_projection[1])
    prev_symbol = math_ops.argmax(prev, 1)
    # Note that gradients will not propagate through the second parameter of
    # embedding_lookup.
    emb_prev = embedding_ops.embedding_lookup(embedding, prev_symbol)
    if not update_embedding:
      emb_prev = array_ops.stop_gradient(emb_prev)
    return emb_prev

  return loop_function