Python array_ops.broadcast_static_shape函数代码示例

  def _reduce_jacobian_det_over_event(
      self, y, ildj, min_event_ndims, event_ndims):
    """Reduce jacobian over event_ndims - min_event_ndims."""
    if not self.is_constant_jacobian:
      return math_ops.reduce_sum(
          ildj,
          self._get_event_reduce_dims(min_event_ndims, event_ndims))

    # In this case, we need to tile the jacobian over the event and reduce.
    y_rank = array_ops.rank(y)
    y_shape = array_ops.shape(y)[
        y_rank - event_ndims : y_rank - min_event_ndims]

    ones = array_ops.ones(y_shape, ildj.dtype)
    reduced_ildj = math_ops.reduce_sum(
        ones * ildj,
        axis=self._get_event_reduce_dims(min_event_ndims, event_ndims))
    # The multiplication by ones can change the inferred static shape so we try
    # to recover as much as possible.
    if (isinstance(event_ndims, int) and
        y.get_shape().ndims and ildj.get_shape().ndims):
      y_shape = y.get_shape()
      y_shape = y_shape[y_shape.ndims - event_ndims :
                        y_shape.ndims - min_event_ndims]
      ildj_shape = ildj.get_shape()
      broadcast_shape = array_ops.broadcast_static_shape(
          ildj_shape, y_shape)
      reduced_ildj.set_shape(
          broadcast_shape[: broadcast_shape.ndims - (
              event_ndims - min_event_ndims)])

    return reduced_ildj

  def _possibly_broadcast_batch_shape(self, x):
    """Return 'x', possibly after broadcasting the leading dimensions."""
    # If we have no batch shape, our batch shape broadcasts with everything!
    if self._batch_shape_arg is None:
      return x

    # Static attempt:
    #   If we determine that no broadcast is necessary, pass x through
    #   If we need a broadcast, add to an array of zeros.
    #
    # special_shape is the shape that, when broadcast with x's shape, will give
    # the correct broadcast_shape.  Note that
    #   We have already verified the second to last dimension of self.shape
    #   matches x's shape in assert_compatible_matrix_dimensions.
    #   Also, the final dimension of 'x' can have any shape.
    #   Therefore, the final two dimensions of special_shape are 1's.
    special_shape = self.batch_shape.concatenate([1, 1])
    bshape = array_ops.broadcast_static_shape(x.get_shape(), special_shape)
    if special_shape.is_fully_defined():
      # bshape.is_fully_defined iff special_shape.is_fully_defined.
      if bshape == x.get_shape():
        return x
      # Use the built in broadcasting of addition.
      zeros = array_ops.zeros(shape=special_shape, dtype=self.dtype)
      return x + zeros

    # Dynamic broadcast:
    #   Always add to an array of zeros, rather than using a "cond", since a
    #   cond would require copying data from GPU --> CPU.
    special_shape = array_ops.concat((self.batch_shape_dynamic(), [1, 1]), 0)
    zeros = array_ops.zeros(shape=special_shape, dtype=self.dtype)
    return x + zeros

  def benchmarkBatchMatMulBroadcast(self):
    for (a_shape, b_shape) in self.shape_pairs:
      with compat.forward_compatibility_horizon(2019, 4, 26):
        with ops.Graph().as_default(), \
            session.Session(config=benchmark.benchmark_config()) as sess, \
            ops.device("/cpu:0"):
          matrix_a = variables.Variable(
              GetRandomNormalInput(a_shape, np.float32))
          matrix_b = variables.Variable(
              GetRandomNormalInput(b_shape, np.float32))
          variables.global_variables_initializer().run()

          # Use batch matmul op's internal broadcasting.
          self.run_op_benchmark(
              sess,
              math_ops.matmul(matrix_a, matrix_b),
              min_iters=50,
              name="batch_matmul_cpu_{}_{}".format(a_shape, b_shape))

          # Manually broadcast the input matrices using the broadcast_to op.
          broadcasted_batch_shape = array_ops.broadcast_static_shape(
              matrix_a.shape[:-2], matrix_b.shape[:-2])
          broadcasted_a_shape = broadcasted_batch_shape.concatenate(
              matrix_a.shape[-2:])
          broadcasted_b_shape = broadcasted_batch_shape.concatenate(
              matrix_b.shape[-2:])
          self.run_op_benchmark(
              sess,
              math_ops.matmul(
                  array_ops.broadcast_to(matrix_a, broadcasted_a_shape),
                  array_ops.broadcast_to(matrix_b, broadcasted_b_shape)),
              min_iters=50,
              name="batch_matmul_manual_broadcast_cpu_{}_{}".format(
                  a_shape, b_shape))

  def _check_shapes(self):
    """Static check that shapes are compatible."""
    # Broadcast shape also checks that u and v are compatible.
    uv_shape = array_ops.broadcast_static_shape(
        self.u.get_shape(), self.v.get_shape())

    batch_shape = array_ops.broadcast_static_shape(
        self.base_operator.batch_shape, uv_shape[:-2])

    self.base_operator.domain_dimension.assert_is_compatible_with(
        uv_shape[-2])

    if self._diag_update is not None:
      uv_shape[-1].assert_is_compatible_with(self._diag_update.get_shape()[-1])
      array_ops.broadcast_static_shape(
          batch_shape, self._diag_update.get_shape()[:-1])

  def _reduce_jacobian_det_over_event(
      self, y, ildj, min_event_ndims, event_ndims):
    """Reduce jacobian over event_ndims - min_event_ndims."""
    # In this case, we need to tile the Jacobian over the event and reduce.
    y_rank = array_ops.rank(y)
    y_shape = array_ops.shape(y)[
        y_rank - event_ndims : y_rank - min_event_ndims]

    ones = array_ops.ones(y_shape, ildj.dtype)
    reduced_ildj = math_ops.reduce_sum(
        ones * ildj,
        axis=self._get_event_reduce_dims(min_event_ndims, event_ndims))
    # The multiplication by ones can change the inferred static shape so we try
    # to recover as much as possible.
    event_ndims_ = self._maybe_get_static_event_ndims(event_ndims)
    if (event_ndims_ is not None and
        y.shape.ndims is not None and
        ildj.shape.ndims is not None):
      y_shape = y.shape[y.shape.ndims - event_ndims_ :
                        y.shape.ndims - min_event_ndims]
      broadcast_shape = array_ops.broadcast_static_shape(ildj.shape, y_shape)
      reduced_ildj.set_shape(
          broadcast_shape[: broadcast_shape.ndims - (
              event_ndims_ - min_event_ndims)])

    return reduced_ildj

def _broadcast_shape(shape1, shape2):
  """Convenience function which statically broadcasts shape when possible."""
  if (tensor_util.constant_value(shape1) is not None and
      tensor_util.constant_value(shape2) is not None):
    return array_ops.broadcast_static_shape(
        tensor_shape.TensorShape(tensor_util.constant_value(shape1)),
        tensor_shape.TensorShape(tensor_util.constant_value(shape2)))
  return array_ops.broadcast_dynamic_shape(shape1, shape2)

def _static_check_for_broadcastable_batch_shape(operators):
  """ValueError if operators determined to have non-broadcastable shapes."""
  if len(operators) < 2:
    return

  # This will fail if they cannot be broadcast together.
  batch_shape = operators[0].batch_shape
  for op in operators[1:]:
    batch_shape = array_ops.broadcast_static_shape(batch_shape, op.batch_shape)

 def _finish_prob_for_one_fiber(self, y, x, ildj, event_ndims):
   """Finish computation of prob on one element of the inverse image."""
   x = self._maybe_rotate_dims(x, rotate_right=True)
   prob = self.distribution.prob(x)
   if self._is_maybe_event_override:
     prob = math_ops.reduce_prod(prob, self._reduce_event_indices)
   prob *= math_ops.exp(math_ops.cast(ildj, prob.dtype))
   if self._is_maybe_event_override and isinstance(event_ndims, int):
     prob.set_shape(array_ops.broadcast_static_shape(
         y.get_shape().with_rank_at_least(1)[:-event_ndims], self.batch_shape))
   return prob

 def _finish_log_prob_for_one_fiber(self, y, x, ildj):
   """Finish computation of log_prob on one element of the inverse image."""
   x = self._maybe_rotate_dims(x, rotate_right=True)
   log_prob = self.distribution.log_prob(x)
   if self._is_maybe_event_override:
     log_prob = math_ops.reduce_sum(log_prob, self._reduce_event_indices)
   log_prob += math_ops.cast(ildj, log_prob.dtype)
   if self._is_maybe_event_override:
     log_prob.set_shape(array_ops.broadcast_static_shape(
         y.get_shape().with_rank_at_least(1)[:-1], self.batch_shape))
   return log_prob

 def _prob(self, y):
   x, ildj = self.bijector.inverse_and_inverse_log_det_jacobian(y)
   x = self._maybe_rotate_dims(x, rotate_right=True)
   prob = self.distribution.prob(x)
   if self._is_maybe_event_override:
     prob = math_ops.reduce_prod(prob, self._reduce_event_indices)
   prob *= math_ops.exp(ildj)
   if self._is_maybe_event_override:
     prob.set_shape(array_ops.broadcast_static_shape(
         y.get_shape().with_rank_at_least(1)[:-1], self.batch_shape))
   return prob

 def _log_prob(self, y):
   x = self.bijector.inverse(y)
   ildj = self.bijector.inverse_log_det_jacobian(y)
   x = self._maybe_rotate_dims(x, rotate_right=True)
   log_prob = self.distribution.log_prob(x)
   if self._is_maybe_event_override:
     log_prob = math_ops.reduce_sum(log_prob, self._reduce_event_indices)
   log_prob = ildj + log_prob
   if self._is_maybe_event_override:
     log_prob.set_shape(array_ops.broadcast_static_shape(
         y.get_shape().with_rank_at_least(1)[:-1], self.batch_shape))
   return log_prob

def determine_batch_event_shapes(grid, endpoint_affine):
  """Helper to infer batch_shape and event_shape."""
  with ops.name_scope(name="determine_batch_event_shapes"):
    # grid  # shape: [B, k, q]
    # endpoint_affine     # len=k, shape: [B, d, d]
    batch_shape = grid.shape[:-2]
    batch_shape_tensor = array_ops.shape(grid)[:-2]
    event_shape = None
    event_shape_tensor = None

    def _set_event_shape(shape, shape_tensor):
      if event_shape is None:
        return shape, shape_tensor
      return (array_ops.broadcast_static_shape(event_shape, shape),
              array_ops.broadcast_dynamic_shape(
                  event_shape_tensor, shape_tensor))

    for aff in endpoint_affine:
      if aff.shift is not None:
        batch_shape = array_ops.broadcast_static_shape(
            batch_shape, aff.shift.shape[:-1])
        batch_shape_tensor = array_ops.broadcast_dynamic_shape(
            batch_shape_tensor, array_ops.shape(aff.shift)[:-1])
        event_shape, event_shape_tensor = _set_event_shape(
            aff.shift.shape[-1:], array_ops.shape(aff.shift)[-1:])

      if aff.scale is not None:
        batch_shape = array_ops.broadcast_static_shape(
            batch_shape, aff.scale.batch_shape)
        batch_shape_tensor = array_ops.broadcast_dynamic_shape(
            batch_shape_tensor, aff.scale.batch_shape_tensor())
        event_shape, event_shape_tensor = _set_event_shape(
            tensor_shape.TensorShape([aff.scale.range_dimension]),
            aff.scale.range_dimension_tensor()[array_ops.newaxis])

    return batch_shape, batch_shape_tensor, event_shape, event_shape_tensor

def prefer_static_broadcast_shape(
    shape1, shape2, name="prefer_static_broadcast_shape"):
  """Convenience function which statically broadcasts shape when possible.

  Args:
    shape1:  `1-D` integer `Tensor`.  Already converted to tensor!
    shape2:  `1-D` integer `Tensor`.  Already converted to tensor!
    name:  A string name to prepend to created ops.

  Returns:
    The broadcast shape, either as `TensorShape` (if broadcast can be done
      statically), or as a `Tensor`.
  """
  with ops.name_scope(name, values=[shape1, shape2]):
    if (tensor_util.constant_value(shape1) is not None and
        tensor_util.constant_value(shape2) is not None):
      return array_ops.broadcast_static_shape(
          tensor_shape.TensorShape(tensor_util.constant_value(shape1)),
          tensor_shape.TensorShape(tensor_util.constant_value(shape2)))
    return array_ops.broadcast_dynamic_shape(shape1, shape2)

def prefer_static_broadcast_shape(
    shape1, shape2, name="prefer_static_broadcast_shape"):
  """Convenience function which statically broadcasts shape when possible.

  Args:
    shape1:  `1-D` integer `Tensor`.  Already converted to tensor!
    shape2:  `1-D` integer `Tensor`.  Already converted to tensor!
    name:  A string name to prepend to created ops.

  Returns:
    The broadcast shape, either as `TensorShape` (if broadcast can be done
      statically), or as a `Tensor`.
  """
  with ops.name_scope(name, values=[shape1, shape2]):
    def make_shape_tensor(x):
      return ops.convert_to_tensor(x, name="shape", dtype=dtypes.int32)

    def get_tensor_shape(s):
      if isinstance(s, tensor_shape.TensorShape):
        return s
      s_ = tensor_util.constant_value(make_shape_tensor(s))
      if s_ is not None:
        return tensor_shape.TensorShape(s_)
      return None

    def get_shape_tensor(s):
      if not isinstance(s, tensor_shape.TensorShape):
        return make_shape_tensor(s)
      if s.is_fully_defined():
        return make_shape_tensor(s.as_list())
      raise ValueError("Cannot broadcast from partially "
                       "defined `TensorShape`.")

    shape1_ = get_tensor_shape(shape1)
    shape2_ = get_tensor_shape(shape2)
    if shape1_ is not None and shape2_ is not None:
      return array_ops.broadcast_static_shape(shape1_, shape2_)

    shape1_ = get_shape_tensor(shape1)
    shape2_ = get_shape_tensor(shape2)
    return array_ops.broadcast_dynamic_shape(shape1_, shape2_)

def get_broadcast_shape(*tensors):
  """Get broadcast shape as a Python list of integers (preferred) or `Tensor`.

  Args:
    *tensors:  One or more `Tensor` objects (already converted!).

  Returns:
    broadcast shape:  Python list (if shapes determined statically), otherwise
      an `int32` `Tensor`.
  """
  # Try static.
  s_shape = tensors[0].shape
  for t in tensors[1:]:
    s_shape = array_ops.broadcast_static_shape(s_shape, t.shape)
  if s_shape.is_fully_defined():
    return s_shape.as_list()

  # Fallback on dynamic.
  d_shape = array_ops.shape(tensors[0])
  for t in tensors[1:]:
    d_shape = array_ops.broadcast_dynamic_shape(d_shape, array_ops.shape(t))
  return d_shape