# Copyright 2015 Google Inc. 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.

# ==============================================================================

"""Classes and functions used to construct graphs."""

# pylint: disable=g-bad-name

from __future__ import absolute_import

from __future__ import division

from __future__ import print_function

import collections

import contextlib

import copy

import linecache

import re

import sys

import threading

import weakref

import tensorflow.python.platform

import six

from tensorflow.core.framework import attr_value_pb2

from tensorflow.core.framework import graph_pb2

from tensorflow.python.framework import device as pydev

from tensorflow.python.framework import dtypes

from tensorflow.python.framework import registry

from tensorflow.python.framework import tensor_shape

from tensorflow.python.util import compat

def _convert_stack(stack):

"""Converts a stack extracted using _extract_stack() to a traceback stack.

Args:

stack: A list of n 4-tuples, (filename, lineno, name, frame_globals).

Returns:

A list of n 4-tuples (filename, lineno, name, code), where the code tuple

element is calculated from the corresponding elements of the input tuple.

"""

ret = []

for filename, lineno, name, frame_globals in stack:

linecache.checkcache(filename)

line = linecache.getline(filename, lineno, frame_globals)

if line:

line = line.strip()

else:

line = None

ret.append((filename, lineno, name, line))

return ret

# pylint: disable=line-too-long

def _extract_stack():

"""A lightweight re-implementation of traceback.extract_stack.

NOTE(mrry): traceback.extract_stack eagerly retrieves the line of code for

each stack frame using linecache, which results in an abundance of stat()

calls. This implementation does not retrieve the code, and any consumer

should apply _convert_stack to the result to obtain a traceback that can

be formatted etc. using traceback methods.

Returns:

A list of 4-tuples (filename, lineno, name, frame_globals) corresponding to

the call stack of the current thread.

"""

# pylint: enable=line-too-long

try:

raise ZeroDivisionError

except ZeroDivisionError:

f = sys.exc_info()[2].tb_frame.f_back

ret = []

while f is not None:

lineno = f.f_lineno

co = f.f_code

filename = co.co_filename

name = co.co_name

frame_globals = f.f_globals

ret.append((filename, lineno, name, frame_globals))

f = f.f_back

ret.reverse()

return ret

class Tensor(object):

"""Represents a value produced by an `Operation`.

A `Tensor` is a symbolic handle to one of the outputs of an

`Operation`. It does not hold the values of that operation's output,

but instead provides a means of computing those values in a

TensorFlow [`Session`](../../api_docs/python/client.md#Session).

This class has two primary purposes:

1. A `Tensor` can be passed as an input to another `Operation`.

This builds a dataflow connection between operations, which

enables TensorFlow to execute an entire `Graph` that represents a

large, multi-step computation.

2. After the graph has been launched in a session, the value of the

`Tensor` can be computed by passing it to

[`Session.run()`](../../api_docs/python/client.md#Session.run).

`t.eval()` is a shortcut for calling

`tf.get_default_session().run(t)`.

In the following example, `c`, `d`, and `e` are symbolic `Tensor`

objects, whereas `result` is a numpy array that stores a concrete

value:

```python

# Build a dataflow graph.

c = tf.constant([[1.0, 2.0], [3.0, 4.0]])

d = tf.constant([[1.0, 1.0], [0.0, 1.0]])

e = tf.matmul(c, d)

# Construct a `Session` to execut the graph.

sess = tf.Session()

# Execute the graph and store the value that `e` represents in `result`.

result = sess.run(e)

```

@@dtype

@@name

@@value_index

@@graph

@@op

@@consumers

@@eval

@@get_shape

@@set_shape

"""

# List of Python operators that we allow to override.

OVERLOADABLE_OPERATORS = {

# Binary.

"__add__",

"__radd__",

"__sub__",

"__rsub__",

"__mul__",

"__rmul__",

"__div__",

"__rdiv__",

"__truediv__",

"__rtruediv__",

"__floordiv__",

"__rfloordiv__",

"__mod__",

"__rmod__",

"__lt__",

"__le__",

"__gt__",

"__ge__",

"__and__",

"__rand__",

"__or__",

"__ror__",

"__xor__",

"__rxor__",

"__getitem__",

# Unary.

"__invert__",

"__neg__",

"__abs__"

}

def __init__(self, op, value_index, dtype):

"""Creates a new `Tensor`.

Args:

op: An `Operation`. `Operation` that computes this tensor.

value_index: An `int`. Index of the operation's endpoint that produces

this tensor.

dtype: A `DType`. Type of elements stored in this tensor.

Raises:

TypeError: If the op is not an `Operation`.

"""

if not isinstance(op, Operation):

raise TypeError("op needs to be an Operation: %s" % op)

self._op = op

self._value_index = value_index

self._dtype = dtypes.as_dtype(dtype)

self._shape = tensor_shape.unknown_shape()

# List of operations that use this Tensor as input. We maintain this list

# to easily navigate a computation graph.

self._consumers = []

@property

def op(self):

"""The `Operation` that produces this tensor as an output."""

return self._op

@property

def dtype(self):

"""The `DType` of elements in this tensor."""

return self._dtype

@property

def graph(self):

"""The `Graph` that contains this tensor."""

return self._op.graph

@property

def name(self):

"""The string name of this tensor."""

if not self._op.name:

raise ValueError("Operation was not named: %s" % self._op)

return "%s:%d" % (self._op.name, self._value_index)

@property

def device(self):

"""The name of the device on which this tensor will be produced, or None."""

return self._op.device

def _shape_as_list(self):

if self._shape.ndims is not None:

return [dim.value for dim in self._shape.dims]

else:

return None

def get_shape(self):

"""Returns the `TensorShape` that represents the shape of this tensor.

The shape is computed using shape inference functions that are

registered for each `Operation` type using `tf.RegisterShape`.

See [`TensorShape`](../../api_docs/python/framework.md#TensorShape) for more

details of what a shape represents.

The inferred shape of a tensor is used to provide shape

information without having to launch the graph in a session. This

can be used for debugging, and providing early error messages. For

example:

```python

c = tf.constant([[1.0, 2.0, 3.0], [4.0, 5.0, 6.0]])

print c.get_shape()

==> TensorShape([Dimension(2), Dimension(3)])

d = tf.constant([[1.0, 0.0], [0.0, 1.0], [1.0, 0.0], [0.0, 1.0]])

print d.get_shape()

==> TensorShape([Dimension(4), Dimension(2)])

# Raises a ValueError, because `c` and `d` do not have compatible

# inner dimensions.

e = tf.matmul(c, d)

f = tf.matmul(c, d, transpose_a=True, transpose_b=True)

print f.get_shape()

==> TensorShape([Dimension(3), Dimension(4)])

```

In some cases, the inferred shape may have unknown dimensions. If

the caller has additional information about the values of these

dimensions, `Tensor.set_shape()` can be used to augment the

inferred shape.

Returns:

A `TensorShape` representing the shape of this tensor.

"""

return self._shape

def set_shape(self, shape):

"""Updates the shape of this tensor.

This method can be called multiple times, and will merge the given

`shape` with the current shape of this tensor. It can be used to

provide additional information about the shape of this tensor that

cannot be inferred from the graph alone. For example, this can be used

to provide additional information about the shapes of images:

```python

_, image_data = tf.TFRecordReader(...).read(...)

image = tf.image.decode_png(image_data, channels=3)

# The height and width dimensions of `image` are data dependent, and

# cannot be computed without executing the op.

print image.get_shape()

==> TensorShape([Dimension(None), Dimension(None), Dimension(3)])

# We know that each image in this dataset is 28 x 28 pixels.

image.set_shape([28, 28, 3])

print image.get_shape()

==> TensorShape([Dimension(28), Dimension(28), Dimension(3)])

```

Args:

shape: A `TensorShape` representing the shape of this tensor.

Raises:

ValueError: If `shape` is not compatible with the current shape of

this tensor.

"""

self._shape = self._shape.merge_with(shape)

@property

def value_index(self):

"""The index of this tensor in the outputs of its `Operation`."""

return self._value_index

def consumers(self):

"""Returns a list of `Operation`s that consume this tensor.

Returns:

A list of `Operation`s.

"""

return self._consumers

def _add_consumer(self, consumer):

"""Add a consumer to this tensor.

Args:

consumer: an Operation.

Raises:

TypeError: if the consumer is not an Operation.

"""

if not isinstance(consumer, Operation):

raise TypeError("Consumer must be an Operation: %s" % consumer)

self._consumers.append(consumer)

def _as_node_def_input(self):

"""Return a value to use for the NodeDef "input" attribute.

The returned string can be used in a NodeDef "input" attribute

to indicate that the NodeDef uses this Tensor as input.

Raises:

ValueError: if this Tensor's Operation does not have a name.

Returns:

a string.

"""

if not self._op.name:

raise ValueError("Operation was not named: %s" % self._op)

if self._value_index == 0:

return self._op.name

else:

return "%s:%d" % (self._op.name, self._value_index)

def __str__(self):

return "Tensor(\"%s\"%s%s%s)" % (

self.name,

(", shape=%s" % self.get_shape())

if self.get_shape().ndims is not None else "",

(", dtype=%s" % self._dtype.name) if self._dtype else "",

(", device=%s" % self.device) if self.device else "")

def __hash__(self):

# Necessary to support Python's collection membership operators

return id(self)

def __eq__(self, other):

# Necessary to support Python's collection membership operators

return id(self) == id(other)

# NOTE(mrry): This enables the Tensor's overloaded "right" binary

# operators to run when the left operand is an ndarray, because it

# accords the Tensor class higher priority than an ndarray, or a

# numpy matrix.

# TODO(mrry): Convert this to using numpy's __numpy_ufunc__

# mechanism, which allows more control over how Tensors interact

# with ndarrays.

__array_priority__ = 100

@staticmethod

def _override_operator(operator, func):

"""Overrides (string) operator on Tensors to call func.

Args:

operator: the string name of the operator to override.

func: the function that replaces the overriden operator.

Raises:

ValueError: If operator has already been overwritten,

or if operator is not allowed to be overwritten.

"""

existing = getattr(Tensor, operator, None)

if existing is not None:

# Check to see if this is a default method-wrapper or slot wrapper which

# will be true for the comparison operators.

if not isinstance(existing, type(object.__lt__)):

raise ValueError("operator %s cannot be overwritten again." % operator)

if operator not in Tensor.OVERLOADABLE_OPERATORS:

raise ValueError("Overriding %s is disallowed" % operator)

setattr(Tensor, operator, func)

def __iter__(self):

"""Dummy method to prevent iteration. Do not call.

NOTE(mrry): If we register __getitem__ as an overloaded operator,

Python will valiantly attempt to iterate over the Tensor from 0 to

infinity. Declaring this method prevents this unintended

behavior.

Raises:

TypeError: when invoked.

"""

raise TypeError("'Tensor' object is not iterable")

def eval(self, feed_dict=None, session=None):

"""Evaluates this tensor in a `Session`.

Calling this method will execute all preceding operations that

produce the inputs needed for the operation that produces this

tensor.

*N.B.* Before invoking `Tensor.eval()`, its graph must have been

launched in a session, and either a default session must be

available, or `session` must be specified explicitly.

Args:

feed_dict: A dictionary that maps `Tensor` objects to feed values.

See [`Session.run()`](../../api_docs/python/client.md#Session.run) for a

description of the valid feed values.

session: (Optional.) The `Session` to be used to evaluate this tensor. If

none, the default session will be used.

Returns:

A numpy array corresponding to the value of this tensor.

"""

return _eval_using_default_session(self, feed_dict, self.graph, session)

def _TensorTensorConversionFunction(t, dtype=None, name=None):

_ = name

if dtype and not dtype.is_compatible_with(t.dtype):

raise ValueError(

"Tensor conversion requested dtype %s for Tensor with dtype %s: %r"

% (dtype.name, t.dtype.name, str(t)))

return t

_tensor_conversion_func_registry = {

0: [(Tensor, _TensorTensorConversionFunction)]}

def convert_to_tensor(value, dtype=None, name=None):

"""Converts the given `value` to a `Tensor`.

This function converts Python objects of various types to `Tensor`

objects. It accepts `Tensor` objects, numpy arrays, Python lists,

and Python scalars. For example:

```python

import numpy as np

array = np.random.rand((32, 100, 100))

def my_func(arg):

arg = tf.convert_to_tensor(arg, dtype=tf.float32)

return tf.matmul(arg, arg) + arg

# The following calls are equivalent.

value_1 = my_func(tf.constant([[1.0, 2.0], [3.0, 4.0]]))

value_2 = my_func([[1.0, 2.0], [3.0, 4.0]])

value_3 = my_func(np.array([[1.0, 2.0], [3.0, 4.0]], dtype=np.float32))

```

This function can be useful when composing a new operation in Python

(such as `my_func` in the example above). All standard Python op

constructors apply this function to each of their Tensor-valued

inputs, which allows those ops to accept numpy arrays, Python lists,

and scalars in addition to `Tensor` objects.

Args:

value: An object whose type has a registered `Tensor` conversion function.

dtype: Optional element type for the returned tensor. If missing, the

type is inferred from the type of `value`.

name: Optional name to use if a new `Tensor` is created.

Returns:

A `Tensor` based on `value`.

Raises:

TypeError: If no conversion function is registered for `value`.

RuntimeError: If a registered conversion function returns an invalid value.

"""

error_prefix = "" if name is None else "%s: " % name

if dtype is not None:

dtype = dtypes.as_dtype(dtype)

for _, funcs_at_priority in sorted(_tensor_conversion_func_registry.items()):

for base_type, conversion_func in funcs_at_priority:

if isinstance(value, base_type):

ret = conversion_func(value, dtype=dtype, name=name)

if not isinstance(ret, Tensor):

raise RuntimeError(

"%sConversion function %r for type %s returned non-Tensor: %r"

% (error_prefix, conversion_func, base_type, ret))

if dtype and not dtype.is_compatible_with(ret.dtype):

raise RuntimeError(

"%sConversion function %r for type %s returned incompatible "

"dtype: requested = %s, actual = %s"

% (error_prefix, conversion_func, base_type,

dtype.name, ret.dtype.name))

return ret

raise TypeError("%sCannot convert %r with type %s to Tensor: "

"no conversion function registered."

% (error_prefix, value, type(value)))

def convert_to_tensor_or_indexed_slices(value, dtype=None, name=None):

"""Converts the given object to a `Tensor` or an `IndexedSlices`.

If `value` is an `IndexedSlices` it is returned

unmodified. Otherwise, it is converted to a `Tensor` using

`convert_to_tensor()`.

Args:

value: An `IndexedSlices` or an object that can be consumed by

`convert_to_tensor()`.

dtype: (Optional.) The required `DType` of the returned `Tensor` or

`IndexedSlices`.

name: (Optional.) A name to use if a new `Tensor` is created.

Returns:

An `Tensor` or an `IndexedSlices` based on `value`.

Raises:

ValueError: If `dtype` does not match the element type of `value`.

"""

if isinstance(value, IndexedSlices):

if dtype and not dtypes.as_dtype(dtype).is_compatible_with(value.dtype):

raise ValueError(

"Tensor conversion requested dtype %s for Tensor with dtype %s: %r"

% (dtypes.as_dtype(dtype).name, value.dtype.name, str(value)))

return value

else:

return convert_to_tensor(value, dtype, name)

def convert_n_to_tensor_or_indexed_slices(values, dtype=None, name=None):

"""Converts `values` to a list of `Tensor` or `IndexedSlices` objects.

Args:

values: A list of `None`, `IndexedSlices`, or objects that can be consumed

by `convert_to_tensor()`.

dtype: (Optional.) The required `DType` of the returned `Tensor`

`IndexedSlices`.

name: (Optional.) A name prefix to used when a new `Tensor` is

created, in which case element `i` will be given the name `name

+ '_' + i`.

Returns:

A list of `Tensor` and/or `IndexedSlices` objects.

Raises:

TypeError: If no conversion function is registered for an element in

`values`.

RuntimeError: If a registered conversion function returns an invalid

value.

"""

if not isinstance(values, collections.Sequence):

raise TypeError("values must be a list.")

ret = []

for i, value in enumerate(values):

if value is None:

ret.append(value)

else:

n = None if name is None else "%s_%d" % (name, i)

ret.append(

convert_to_tensor_or_indexed_slices(value, dtype=dtype, name=n))

return ret

def register_tensor_conversion_function(base_type, conversion_func,

priority=100):

"""Registers a function for converting objects of base_type to Tensor.

The conversion function must have the following signature:

def conversion_func(value, dtype=None, name=None):

# ...

It must return a Tensor with the given dtype if specified. If the

conversion function creates a new Tensor, it should use the given

name if specified. All exceptions will be propagated to the caller.

NOTE: The conversion functions will execute in order of priority,

followed by order of registration. To ensure that a conversion

function F runs before another conversion function G, ensure that

F is registered with a smaller priority than G.

Args:

base_type: The base type or tuple of base types for all objects that

`conversion_func` accepts.

conversion_func: A function that converts instances of base_type to Tensor.

priority: Optional integer that indicates the priority for applying this

conversion function. Conversion functions with smaller priority values

run earlier than conversion functions with larger priority values.

Defaults to 100.

Raises:

TypeError: If the arguments do not have the appropriate type.

"""

if not (isinstance(base_type, type) or

(isinstance(base_type, tuple)

and all(isinstance(x, type) for x in base_type))):

raise TypeError("base_type must be a type or a tuple of types.")

if not callable(conversion_func):

raise TypeError("conversion_func must be callable.")

try:

funcs_at_priority = _tensor_conversion_func_registry[priority]

except KeyError:

funcs_at_priority = []

_tensor_conversion_func_registry[priority] = funcs_at_priority

funcs_at_priority.append((base_type, conversion_func))

class IndexedSlices(object):

"""A sparse representation of a set of tensor slices at given indices.

This class is a simple wrapper for a pair of `Tensor` objects:

* `values`: A `Tensor` of any dtype with shape `[D0, D1, ..., Dn]`.

* `indices`: A 1-D integer `Tensor` with shape `[D0]`.

An `IndexedSlices` is typically used to represent a subset of a larger

tensor `dense` of shape `[LARGE0, D1, .. , DN]` where `LARGE0 >> D0`.

The values in `indices` are the indices in the first dimension of

the slices that have been extracted from the larger tensor.

The dense tensor `dense` represented by an `IndexedSlices` `slices` has

```python

dense[slices.indices[i], :, :, :, ...] = slices.values[i, :, :, :, ...]

```

The `IndexedSlices` class is used principally in the definition of

gradients for operations that have sparse gradients

(e.g. [`tf.gather`](../../api_docs/python/array_ops.md#gather)).

Contrast this representation with

[`SparseTensor`](../../api_docs/python/sparse_ops.md#SparseTensor),

which uses multi-dimensional indices and scalar values.

@@__init__

@@values

@@indices

@@dense_shape

@@name

@@dtype

@@device

@@op

"""

def __init__(self, values, indices, dense_shape=None):

"""Creates an `IndexedSlices`."""

self._values = values

self._indices = indices

self._dense_shape = dense_shape

@property

def values(self):

"""A `Tensor` containing the values of the slices."""

return self._values

@property

def indices(self):

"""A 1-D `Tensor` containing the indices of the slices."""

return self._indices

@property

def dense_shape(self):

"""A 1-D `Tensor` containing the shape of the corresponding dense tensor."""

return self._dense_shape

@property

def name(self):

"""The name of this `IndexedSlices`."""

return self.values.name

@property

def device(self):

"""The name of the device on which `values` will be produced, or `None`."""

return self.values.device

@property

def op(self):

"""The `Operation` that produces `values` as an output."""

return self.values.op

@property

def dtype(self):

"""The `DType` of elements in this tensor."""

return self.values.dtype

def __str__(self):

return "IndexedSlices(indices=%s, values=%s)" % (

self._indices, self._values)

def assert_same_graph(items, expected_graph=None):

"""Asserts all items are from the same graph.

Args:

items: List of graph items (e.g., Variable, Tensor, SparseTensor,

Operation, or IndexedSlices).

expected_graph: Expected graph. If not specified, assert all tensors are

from the same graph.

Returns:

items, for chaining.

Raises:

ValueError: If any graphs do not match.

"""

for item in items:

if not expected_graph:

expected_graph = item.graph

elif expected_graph != item.graph:

raise ValueError("Items must be from the same graph.")

return items

class SparseTensor(object):

"""Represents a sparse tensor.

Tensorflow represents a sparse tensor as three separate dense tensors:

`indices`, `values`, and `dense_shape`. In Python, the three tensors are

collected into a `SparseTensor` class for ease of use. If you have separate

`indices`, `values`, and `dense_shape` tensors, wrap them in a `SparseTensor`

object before passing to the Ops below.

Concretely, the sparse tensor `SparseTensor(values, indices, dense_shape)` is

* `indices`: A 2-D int64 tensor of shape `[N, ndims]`.

* `values`: A 1-D tensor of any type and shape `[N]`.

* `dense_shape`: A 1-D int64 tensor of shape `[ndims]`.

where `N` and `ndims` are the number of values, and number of dimensions in

the `SparseTensor` respectively.

The corresponding dense tensor satisfies

```python

dense.shape = dense_shape

dense[tuple(indices[i])] = values[i]

```

By convention, `indices` should be sorted in row-major order (or equivalently

lexigraphic order on the tuples `indices[i]`). This is not enforced when

`SparseTensor` objects are constructed, but most Ops assume correct ordering.

If the ordering is wrong, it can be fixed by calling `sparse_reorder` on the

misordered `SparseTensor`.

Example: The sparse tensor

```python

SparseTensor(values=[1, 2], indices=[[0, 0], [1, 2]], shape=[3, 4])

```

represents the dense tensor

```python

[[1, 0, 0, 0]

[0, 0, 2, 0]

[0, 0, 0, 0]]

```

@@__init__

@@indices

@@values

@@dtype

@@shape

@@graph

"""

def __init__(self, indices, values, shape):

"""Creates a `SparseTensor`.

Args:

indices: A 2-D int64 tensor of shape `[N, ndims]`.

values: A 1-D tensor of any type and shape `[N]`.

dense_shape: A 1-D int64 tensor of shape `[ndims]`.

Returns:

A `SparseTensor`

"""

with op_scope([indices, values, shape], None, "SparseTensor"):

indices = convert_to_tensor(indices, name="indices")

values = convert_to_tensor(values, name="values")

shape = convert_to_tensor(shape, name="shape")

self._indices = indices

self._values = values

self._shape = shape

indices_shape = indices.get_shape().with_rank(2)

values_shape = values.get_shape().with_rank(1)

shape_shape = shape.get_shape().with_rank(1)

# Assert number of rows in indices match the number of elements in values.

indices_shape[0].merge_with(values_shape[0])

# Assert number of columns in indices matches the number of elements in

# shape.

indices_shape[1].merge_with(shape_shape[0])

@property

def indices(self):

"""The indices of non-zero values in the represented dense tensor.

Returns:

A 2-D Tensor of int64 with shape `[N, ndims]`, where `N` is the

number of non-zero values in the tensor, and `ndims` is the rank.

"""

return self._indices

@property

def values(self):

"""The non-zero values in the represented dense tensor.

Returns:

A 1-D Tensor of any data type.

"""

return self._values

@property

def dtype(self):

"""The `DType` of elements in this tensor."""

return self._values.dtype

@property

def shape(self):

"""A 1-D Tensor of int64 representing the shape of the dense tensor."""

return self._shape

@property

def graph(self):

"""The `Graph` that contains the index, value, and shape tensors."""

return self._indices.graph

def __str__(self):

return "SparseTensor(indices=%s, values=%s, shape=%s)" % (

self._indices, self._values, self._shape)

SparseTensorValue = collections.namedtuple("SparseTensorValue",

["indices", "values", "shape"])

def _device_string(dev_spec):

if isinstance(dev_spec, pydev.Device):

return dev_spec.to_string()

else:

return dev_spec

def _NodeDef(op_type, name, device=None, attrs=None):

"""Create a NodeDef proto.

Args:

op_type: Value for the "op" attribute of the NodeDef proto.

name: Value for the "name" attribute of the NodeDef proto.

device: string, device, or function from NodeDef to string.

Value for the "device" attribute of the NodeDef proto.

attrs: optional list for the "attr" attribute of the NodeDef proto.

Returns:

A graph_pb2.NodeDef protocol buffer.

"""

node_def = graph_pb2.NodeDef()

node_def.op = compat.as_bytes(op_type)

node_def.name = compat.as_bytes(name)

if attrs is not None:

for k, v in six.iteritems(attrs):

node_def.attr[k].CopyFrom(v)

if device is not None:

if callable(device):

node_def.device = device(node_def)

else:

node_def.device = _device_string(device)

return node_def

# Copied from core/framework/node_def_util.cc

# TODO(mrry,josh11b): Consolidate this validation in C++ code.

_VALID_OP_NAME_REGEX = re.compile("[A-Za-z0-9.][A-Za-z0-9_.\\-/]*")

class Operation(object):

"""Represents a graph node that performs computation on tensors.

An `Operation` is a node in a TensorFlow `Graph` that takes zero or

more `Tensor` objects as input, and produces zero or more `Tensor`

objects as output. Objects of type `Operation` are created by

calling a Python op constructor (such as

[`tf.matmul()`](../../api_docs/python/math_ops.md#matmul))

or [`Graph.create_op()`](../../api_docs/python/framework.md#Graph.create_op).

For example `c = tf.matmul(a, b)` creates an `Operation` of type

"MatMul" that takes tensors `a` and `b` as input, and produces `c`

as output.

After the graph has been launched in a session, an `Operation` can

be executed by passing it to

[`Session.run()`](../../api_docs/python/client.md#Session.run).

`op.run()` is a shortcut for calling `tf.get_default_session().run(op)`.

@@name

@@type

@@inputs

@@control_inputs

@@outputs

@@device

@@graph

@@run

@@get_attr

@@traceback

"""

def __init__(self, node_def, g, inputs=None, output_types=None,

control_inputs=None, input_types=None, original_op=None,

op_def=None):

"""Creates an `Operation`.

NOTE: This constructor validates the name of the `Operation` (passed

as `node_def.name`). Valid `Operation` names match the following

regular expression:

[A-Za-z0-9.][A-Za-z0-9_.\\-/]*

Args:

node_def: `graph_pb2.NodeDef`. `NodeDef` for the `Operation`.

Used for attributes of `graph_pb2.NodeDef`, typically `name`,

`op`, and `device`. The `input` attribute is irrelevant here

as it will be computed when generating the model.

g: `Graph`. The parent graph.

inputs: list of `Tensor` objects. The inputs to this `Operation`.

output_types: list of `DType` objects. List of the types of the

`Tensors` computed by this operation. The length of this list indicates

the number of output endpoints of the `Operation`.

control_inputs: list of operations or tensors from which to have a

control dependency.

input_types: List of `DType` objects representing the

types of the tensors accepted by the `Operation`. By default

uses `[x.dtype.base_dtype for x in inputs]`. Operations that expect

reference-typed inputs must specify these explicitly.

original_op: Optional. Used to associate the new `Operation` with an

existing `Operation` (for example, a replica with the op that was

replicated).

op_def: Optional. The `op_def_pb2.OpDef` proto that describes the

op type that this `Operation` represents.

Raises:

TypeError: if control inputs are not Operations or Tensors,

or if `node_def` is not a `NodeDef`,

or if `g` is not a `Graph`,

or if `inputs` are not tensors,

or if `inputs` and `input_types` are incompatible.

ValueError: if the `node_def` name is not valid.

"""

if not isinstance(node_def, graph_pb2.NodeDef):

raise TypeError("node_def needs to be a NodeDef: %s" % node_def)

if node_def.ByteSize() >= (1 << 31) or node_def.ByteSize() < 0:

raise ValueError(

"Cannot create an Operation with a NodeDef larger than 2GB.")

if not _VALID_OP_NAME_REGEX.match(node_def.name):

raise ValueError("'%s' is not a valid node name" % node_def.name)

if not isinstance(g, Graph):

raise TypeError("g needs to be a Graph: %s" % g)

self._node_def = copy.deepcopy(node_def)

self._graph = g

if inputs is None:

inputs = []

self._inputs = inputs

for a in self._inputs:

if not isinstance(a, Tensor):

raise TypeError("input needs to be a Tensor: %s" % a)

# Mark that we consume the inputs.

a._add_consumer(self) # pylint: disable=protected-access

if output_types is None:

output_types = []

self._output_types = output_types

self._outputs = [Tensor(self, i, output_type)

for i, output_type in enumerate(output_types)]