Defining new operations

class tf.RegisterGradient

A decorator for registering the gradient function for an op type.

This decorator is only used when defining a new op type. For an op with m inputs and n outputs, the gradient function is a function that takes the original Operation and n Tensor objects (representing the gradients with respect to each output of the op), and returns m Tensor objects (representing the partial gradients with respect to each input of the op).

For example, assuming that operations of type "Sub" take two inputs x and y, and return a single output x - y, the following gradient function would be registered:

@tf.RegisterGradient("Sub")
def _sub_grad(unused_op, grad):
  return grad, tf.neg(grad)

The decorator argument op_type is the string type of an operation. This corresponds to the OpDef.name field for the proto that defines the operation.


tf.RegisterGradient.__init__(op_type) {:#RegisterGradient.init}

Creates a new decorator with op_type as the Operation type.

Args:
  • op_type: The string type of an operation. This corresponds to the OpDef.name field for the proto that defines the operation.

tf.NoGradient(op_type)

Specifies that ops of type op_type do not have a defined gradient.

This function is only used when defining a new op type. It may be used for ops such as tf.size() that are not differentiable. For example:

tf.NoGradient("Size")
Args:
  • op_type: The string type of an operation. This corresponds to the OpDef.name field for the proto that defines the operation.
Raises:
  • TypeError: If op_type is not a string.

class tf.RegisterShape

A decorator for registering the shape function for an op type.

This decorator is only used when defining a new op type. A shape function is a function from an Operation object to a list of TensorShape objects, with one TensorShape for each output of the operation.

For example, assuming that operations of type "Sub" take two inputs x and y, and return a single output x - y, all with the same shape, the following shape function would be registered:

@tf.RegisterShape("Sub")
def _sub_shape(op):
  return [op.inputs[0].get_shape().merge_with(op.inputs[1].get_shape())]

The decorator argument op_type is the string type of an operation. This corresponds to the OpDef.name field for the proto that defines the operation.


tf.RegisterShape.__init__(op_type) {:#RegisterShape.init}

Saves the op_type as the Operation type.


class tf.TensorShape

Represents the shape of a Tensor.

A TensorShape represents a possibly-partial shape specification for a Tensor. It may be one of the following:

  • Fully-known shape: has a known number of dimensions and a known size for each dimension.
  • Partially-known shape: has a known number of dimensions, and an unknown size for one or more dimension.
  • Unknown shape: has an unknown number of dimensions, and an unknown size in all dimensions.

If a tensor is produced by an operation of type "Foo", its shape may be inferred if there is a registered shape function for "Foo". See tf.RegisterShape() for details of shape functions and how to register them. Alternatively, the shape may be set explicitly using Tensor.set_shape().


tf.TensorShape.merge_with(other)

Returns a TensorShape combining the information in self and other.

The dimensions in self and other are merged elementwise, according to the rules defined for Dimension.merge_with().

Args:
  • other: Another TensorShape.
Returns:

A TensorShape containing the combined information of self and other.

Raises:
  • ValueError: If self and other are not compatible.

tf.TensorShape.concatenate(other)

Returns the concatenation of the dimension in self and other.

N.B. If either self or other is completely unknown, concatenation will discard information about the other shape. In future, we might support concatenation that preserves this information for use with slicing.

Args:
  • other: Another TensorShape.
Returns:

A TensorShape whose dimensions are the concatenation of the dimensions in self and other.


tf.TensorShape.ndims

Returns the rank of this shape, or None if it is unspecified.


tf.TensorShape.dims

Returns a list of Dimensions, or None if the shape is unspecified.


tf.TensorShape.as_list()

Returns a list of integers or None for each dimension.

Returns:

A list of integers or None for each dimension.


tf.TensorShape.as_proto()

Returns this shape as a TensorShapeProto.


tf.TensorShape.is_compatible_with(other)

Returns True iff self is compatible with other.

Two possibly-partially-defined shapes are compatible if there exists a fully-defined shape that both shapes can represent. Thus, compatibility allows the shape inference code to reason about partially-defined shapes. For example:

  • TensorShape(None) is compatible with all shapes.

  • TensorShape([None, None]) is compatible with all two-dimensional shapes, such as TensorShape([32, 784]), and also TensorShape(None). It is not compatible with, for example, TensorShape([None]) or TensorShape([None, None, None]).

  • TensorShape([32, None]) is compatible with all two-dimensional shapes with size 32 in the 0th dimension, and also TensorShape([None, None]) and TensorShape(None). It is not compatible with, for example, TensorShape([32]), TensorShape([32, None, 1]) or TensorShape([64, None]).

  • TensorShape([32, 784]) is compatible with itself, and also TensorShape([32, None]), TensorShape([None, 784]), TensorShape([None, None]) and TensorShape(None). It is not compatible with, for example, TensorShape([32, 1, 784]) or TensorShape([None]).

The compatibility relation is reflexive and symmetric, but not transitive. For example, TensorShape([32, 784]) is compatible with TensorShape(None), and TensorShape(None) is compatible with TensorShape([4, 4]), but TensorShape([32, 784]) is not compatible with TensorShape([4, 4]).

Args:
  • other: Another TensorShape.
Returns:

True iff self is compatible with other.


tf.TensorShape.is_fully_defined()

Returns True iff self is fully defined in every dimension.


tf.TensorShape.with_rank(rank)

Returns a shape based on self with the given rank.

This method promotes a completely unknown shape to one with a known rank.

Args:
  • rank: An integer.
Returns:

A shape that is at least as specific as self with the given rank.

Raises:
  • ValueError: If self does not represent a shape with the given rank.

tf.TensorShape.with_rank_at_least(rank)

Returns a shape based on self with at least the given rank.

Args:
  • rank: An integer.
Returns:

A shape that is at least as specific as self with at least the given rank.

Raises:
  • ValueError: If self does not represent a shape with at least the given rank.

tf.TensorShape.with_rank_at_most(rank)

Returns a shape based on self with at most the given rank.

Args:
  • rank: An integer.
Returns:

A shape that is at least as specific as self with at most the given rank.

Raises:
  • ValueError: If self does not represent a shape with at most the given rank.

tf.TensorShape.assert_has_rank(rank)

Raises an exception if self is not compatible with the given rank.

Args:
  • rank: An integer.
Raises:
  • ValueError: If self does not represent a shape with the given rank.

tf.TensorShape.assert_same_rank(other)

Raises an exception if self and other do not have compatible ranks.

Args:
  • other: Another TensorShape.
Raises:
  • ValueError: If self and other do not represent shapes with the same rank.

tf.TensorShape.assert_is_compatible_with(other)

Raises exception if self and other do not represent the same shape.

This method can be used to assert that there exists a shape that both self and other represent.

Args:
  • other: Another TensorShape.
Raises:
  • ValueError: If self and other do not represent the same shape.

tf.TensorShape.assert_is_fully_defined()

Raises an exception if self is not fully defined in every dimension.

Raises:
  • ValueError: If self does not have a known value for every dimension.

Other Methods


tf.TensorShape.__init__(dims) {:#TensorShape.init}

Creates a new TensorShape with the given dimensions.

Args:
  • dims: A list of Dimensions, or None if the shape is unspecified.
  • DEPRECATED: A single integer is treated as a singleton list.
Raises:
  • TypeError: If dims cannot be converted to a list of dimensions.

tf.TensorShape.num_elements()

Returns the total number of elements, or none for incomplete shapes.


class tf.Dimension

Represents the value of one dimension in a TensorShape.


tf.Dimension.__init__(value) {:#Dimension.init}

Creates a new Dimension with the given value.


tf.Dimension.assert_is_compatible_with(other)

Raises an exception if other is not compatible with this Dimension.

Args:
  • other: Another Dimension.
Raises:
  • ValueError: If self and other are not compatible (see is_compatible_with).

tf.Dimension.is_compatible_with(other)

Returns true if other is compatible with this Dimension.

Two known Dimensions are compatible if they have the same value. An unknown Dimension is compatible with all other Dimensions.

Args:
  • other: Another Dimension.
Returns:

True if this Dimension and other are compatible.


tf.Dimension.merge_with(other)

Returns a Dimension that combines the information in self and other.

Dimensions are combined as follows:

Dimension(n)   .merge_with(Dimension(n))    == Dimension(n)
Dimension(n)   .merge_with(Dimension(None)) == Dimension(n)
Dimension(None).merge_with(Dimension(n))    == Dimension(n)
Dimension(None).merge_with(Dimension(None)) == Dimension(None)
Dimension(n)   .merge_with(Dimension(m)) raises ValueError for n != m
Args:
  • other: Another Dimension.
Returns:

A Dimension containing the combined information of self and other.

Raises:
  • ValueError: If self and other are not compatible (see is_compatible_with).

tf.Dimension.value

The value of this dimension, or None if it is unknown.


tf.op_scope(values, name, default_name=None)

Returns a context manager for use when defining a Python op.

This context manager validates that the given values are from the same graph, ensures that graph is the default graph, and pushes a name scope.

For example, to define a new Python op called my_op:

def my_op(a, b, c, name=None):
  with tf.op_scope([a, b, c], name, "MyOp") as scope:
    a = tf.convert_to_tensor(a, name="a")
    b = tf.convert_to_tensor(b, name="b")
    c = tf.convert_to_tensor(c, name="c")
    # Define some computation that uses `a`, `b`, and `c`.
    return foo_op(..., name=scope)
Args:
  • values: The list of Tensor arguments that are passed to the op function.
  • name: The name argument that is passed to the op function.
  • default_name: The default name to use if the name argument is None.
Returns:

A context manager for use in defining Python ops. Yields the name scope.

Raises:
  • ValueError: if neither name nor default_name is provided.

tf.get_seed(op_seed)

Returns the local seeds an operation should use given an op-specific seed.

Given operation-specific seed, op_seed, this helper function returns two seeds derived from graph-level and op-level seeds. Many random operations internally use the two seeds to allow user to change the seed globally for a graph, or for only specific operations.

For details on how the graph-level seed interacts with op seeds, see set_random_seed.

Args:
  • op_seed: integer.
Returns:

A tuple of two integers that should be used for the local seed of this operation.