Adding a New Op

PREREQUISITES:

If you'd like to incorporate an operation that isn't covered by the existing library, you can create a custom Op. To incorporate your custom Op, you'll need to:

  • Register the new Op in a C++ file. The Op registration is independent of the implementation, and describes the semantics of how the Op is invoked. For example, it defines the Op name, and specifies its inputs and outputs.
  • Implement the Op in C++. This implementation is called a "kernel", and there can be multiple kernels for different architectures (e.g. CPUs, GPUs) or input / output types.
  • Optionally, create a Python wrapper. This wrapper is the public API to create the Op. A default wrapper is generated from the Op registration, which can be used directly or added to.
  • Optionally, write a function to compute gradients for the Op.
  • Optionally, write a function that describes the input and output shapes for the Op. This allows shape inference to work with your Op.
  • Test the Op, typically in Python. If you define gradients, you can verify them with the Python GradientChecker.

Define the Op's interface

You define the interface of an Op by registering it with the TensorFlow system. In the registration, you specify the name of your Op, its inputs (types and names) and outputs (types and names), as well as docstrings and any attrs the Op might require.

To see how this works, suppose you'd like to create an Op that takes a tensor of int32s and outputs a copy of the tensor, with all but the first element set to zero. Create file tensorflow/core/user_ops/zero_out.cc and add a call to the REGISTER_OP macro that defines the interface for such an Op:

#include "tensorflow/core/framework/op.h"

REGISTER_OP("ZeroOut")
    .Input("to_zero: int32")
    .Output("zeroed: int32");

This ZeroOut Op takes one tensor to_zero of 32-bit integers as input, and outputs a tensor zeroed of 32-bit integers.

A note on naming: The name of the Op should be unique and CamelCase. Names starting with an underscore (_) are reserved for internal use.

Implement the kernel for the Op

After you define the interface, provide one or more implementations of the Op. To create one of these kernels, create a class that extends OpKernel and overrides the Compute method. The Compute method provides one context argument of type OpKernelContext*, from which you can access useful things like the input and output tensors.

Important note: Instances of your OpKernel may be accessed concurrently. Your Compute method must be thread-safe. Guard any access to class members with a mutex (Or better yet, don't share state via class members! Consider using a ResourceMgr to keep track of Op state).

Add your kernel to the file you created above. The kernel might look something like this:

#include "tensorflow/core/framework/op_kernel.h"

using namespace tensorflow;

class ZeroOutOp : public OpKernel {
 public:
  explicit ZeroOutOp(OpKernelConstruction* context) : OpKernel(context) {}

  void Compute(OpKernelContext* context) override {
    // Grab the input tensor
    const Tensor& input_tensor = context->input(0);
    auto input = input_tensor.flat<int32>();

    // Create an output tensor
    Tensor* output_tensor = NULL;
    OP_REQUIRES_OK(context, context->allocate_output(0, input_tensor.shape(),
                                                     &output_tensor));
    auto output = output_tensor->flat<int32>();

    // Set all but the first element of the output tensor to 0.
    const int N = input.size();
    for (int i = 1; i < N; i++) {
      output(i) = 0;
    }

    // Preserve the first input value if possible.
    if (N > 0) output(0) = input(0);
  }
};

After implementing your kernel, you register it with the TensorFlow system. In the registration, you specify different constraints under which this kernel will run. For example, you might have one kernel made for CPUs, and a separate one for GPUs.

To do this for the ZeroOut op, add the following to zero_out.cc:

REGISTER_KERNEL_BUILDER(Name("ZeroOut").Device(DEVICE_CPU), ZeroOutOp);

Building the Op library

With TensorFlow binary installation

You should be able to compile zero_out.cc with a C++ compiler such as g++ or clang available on your system. The binary PIP package installs the header files and the library that you need to compile your Op in locations that are system specific. However, the TensorFlow python library provides the get_include function to get the header directory. Here is the output of this function on a Ubuntu machine.

$ python
>>> import tensorflow as tf
>>> tf.sysconfig.get_include()
'/usr/local/lib/python2.7/site-packages/tensorflow/include'

Assuming you have g++ installed, here is the sequence of commands you can use to compile your Op into a dynamic library.

TF_INC=$(python -c 'import tensorflow as tf; print(tf.sysconfig.get_include())')

g++ -std=c++11 -shared zero_out.cc -o zero_out.so -fPIC -I $TF_INC

On Mac OS X, the additional flag "-undefined dynamic_lookup" is required when building the .so file.

Note on gcc version 5: gcc5 uses the new C++ ABI. The binary pip packages available on the TensorFlow website are built with gcc4 that uses the older ABI. If you compile your op library with gcc5, add -D_GLIBCXX_USE_CXX11_ABI=0 to the command line to make the library compatible with the older abi.

With TensorFlow source installation

If you have TensorFlow sources installed, you can make use of TensorFlow's build system to compile your Op. Place a BUILD file with following Bazel build rule in the tensorflow/core/user_ops directory.

load("//tensorflow:tensorflow.bzl", "tf_custom_op_library")

tf_custom_op_library(
    name = "zero_out.so",
    srcs = ["zero_out.cc"],
)

Run the following command to build zero_out.so.

$ bazel build -c opt //tensorflow/core/user_ops:zero_out.so

Using the Op in Python

TensorFlow Python API provides the load_op_library function to load the dynamic library and register the Op with the TensorFlow framework. load_op_library returns a Python module, that contains the Python wrappers for the Op. Thus, once you have built the op, you can do the following to run it from Python :

import tensorflow as tf
zero_out_module = tf.load_op_library('zero_out.so')
with tf.Session(''):
  zero_out_module.zero_out([[1, 2], [3, 4]]).eval()

# Prints
array([[1, 0],
       [0, 0]], dtype=int32)

To make the Op available as a regular function import-able from a Python module, it maybe useful to have the load_op_library call in a Python source file as follows (see zero_out_op_1.py) :

import tensorflow as tf

_zero_out_module = tf.load_op_library('zero_out_op_kernel_1.so')
zero_out = _zero_out_module.zero_out

Verify it works

A good way to verify that you've successfully implemented your Op is to write a test for it. Create the file tensorflow/python/kernel_tests/zero_out_op_test.py with the contents:

import tensorflow as tf

class ZeroOutTest(tf.test.TestCase):
  def testZeroOut(self):
    zero_out_module = tf.load_op_library('zero_out.so')
    with self.test_session():
      result = zero_out_module.zero_out([5, 4, 3, 2, 1])
      self.assertAllEqual(result.eval(), [5, 0, 0, 0, 0])

Then run your test:

$ bazel test tensorflow/python:zero_out_op_test

Validation

The example above assumed that the Op applied to a tensor of any shape. What if it only applied to vectors? That means adding a check to the above OpKernel implementation.

  void Compute(OpKernelContext* context) override {
    // Grab the input tensor
    const Tensor& input_tensor = context->input(0);

    OP_REQUIRES(context, TensorShapeUtils::IsVector(input_tensor.shape()),
                errors::InvalidArgument("ZeroOut expects a 1-D vector."));
    // ...
  }

This asserts that the input is a vector, and returns having set the InvalidArgument status if it isn't. The OP_REQUIRES macro takes three arguments:

Alternatively, if you want to test whether a Status object returned from some function is an error, and if so return it, use OP_REQUIRES_OK. Both of these macros return from the function on error.

Op registration

Attrs

Ops can have attrs, whose values are set when the Op is added to a graph. These are used to configure the Op, and their values can be accessed both within the kernel implementation and in the types of inputs and outputs in the Op registration. Prefer using an input instead of an attr when possible, since inputs are more flexible. They can change every step, be set using a feed, etc. Attrs are used for things that can't be done with inputs: any configuration that affects the signature (number or type of inputs or outputs) or that can't change from step-to-step.

You define an attr when you register the Op, by specifying its name and type using the Attr method, which expects a spec of the form:

<name>: <attr-type-expr>

where <name> begins with a letter and can be composed of alphanumeric characters and underscores, and <attr-type-expr> is a type expression of the form described below

For example, if you'd like the ZeroOut Op to preserve a user-specified index, instead of only the 0th element, you can register the Op like so:

REGISTER_OP("ZeroOut")
    .Attr("preserve_index: int")
    .Input("to_zero: int32")
    .Output("zeroed: int32");

Your kernel can then access this attr in its constructor via the context parameter:

class ZeroOutOp : public OpKernel {
 public:
  explicit ZeroOutOp(OpKernelConstruction* context) : OpKernel(context) {
    // Get the index of the value to preserve
    OP_REQUIRES_OK(context,
                   context->GetAttr("preserve_index", &preserve_index_));
    // Check that preserve_index is positive
    OP_REQUIRES(context, preserve_index_ >= 0,
                errors::InvalidArgument("Need preserve_index >= 0, got ",
                                        preserve_index_));
  }
  void Compute(OpKernelContext* context) override {
    // ...
  }
 private:
  int preserve_index_;
};

which can then be used in the Compute method:

  void Compute(OpKernelContext* context) override {
    // ...

// Check that preserve_index is in range OP_REQUIRES(context, preserve_index_ < input.dimension(0), errors::InvalidArgument("preserve_index out of range"));
// Set all the elements of the output tensor to 0 const int N = input.size(); for (int i = 0; i < N; i++) { output_flat(i) = 0; }
// Preserve the requested input value output_flat(preserve_index_) = input(preserve_index_); }

To preserve backwards compatibility, you should specify a default value when adding an attr to an existing op:

REGISTER_OP("ZeroOut")
    .Attr("preserve_index: int = 0")
    .Input("to_zero: int32")
    .Output("zeroed: int32");

Attr types

The following types are supported in an attr:

  • string: Any sequence of bytes (not required to be UTF8).
  • int: A signed integer.
  • float: A floating point number.
  • bool: True or false.
  • type: One of the (non-ref) values of DataType.
  • shape: A TensorShapeProto.
  • tensor: A TensorProto.
  • list(<type>): A list of <type>, where <type> is one of the above types. Note that list(list(<type>)) is invalid.

See also: op_def_builder.cc:FinalizeAttr for a definitive list.

Default values & constraints

Attrs may have default values, and some types of attrs can have constraints. To define an attr with constraints, you can use the following <attr-type-expr>s:

  • {'<string1>', '<string2>'}: The value must be a string that has either the value <string1> or <string2>. The name of the type, string, is implied when you use this syntax. This emulates an enum:
REGISTER_OP("EnumExample")
    .Attr("e: {'apple', 'orange'}");
  • {<type1>, <type2>}: The value is of type type, and must be one of <type1> or <type2>, where <type1> and <type2> are supported tensor types. You don't specify that the type of the attr is type. This is implied when you have a list of types in {...}. For example, in this case the attr t is a type that must be an int32, a float, or a bool:
REGISTER_OP("RestrictedTypeExample")
    .Attr("t: {int32, float, bool}");
  • There are shortcuts for common type constraints:

    • numbertype: Type type restricted to the numeric (non-string and non-bool) types.
    • realnumbertype: Like numbertype without complex types.
    • quantizedtype: Like numbertype but just the quantized number types.

    The specific lists of types allowed by these are defined by the functions (like NumberTypes()) in tensorflow/core/framework/types.h. In this example the attr t must be one of the numeric types:

    c++ REGISTER_OP("NumberType") .Attr("t: numbertype");

    For this op:

    python tf.number_type(t=tf.int32) # Valid tf.number_type(t=tf.bool) # Invalid

  • int >= <n>: The value must be an int whose value is greater than or equal to <n>, where <n> is a natural number.

For example, the following Op registration specifies that the attr a must have a value that is at least 2:

REGISTER_OP("MinIntExample")
    .Attr("a: int >= 2");
  • list(<type>) >= <n>: A list of type <type> whose length is greater than or equal to <n>.

For example, the following Op registration specifies that the attr a is a list of types (either int32 or float), and that there must be at least 3 of them:

REGISTER_OP("TypeListExample")
    .Attr("a: list({int32, float}) >= 3");

To set a default value for an attr (making it optional in the generated code), add = <default> to the end, as in:

REGISTER_OP("AttrDefaultExample")
    .Attr("i: int = 0");

The supported syntax of the default value is what would be used in the proto representation of the resulting GraphDef definition.

Here are examples for how to specify a default for all types:

REGISTER_OP("AttrDefaultExampleForAllTypes")
   .Attr("s: string = 'foo'")
   .Attr("i: int = 0")
   .Attr("f: float = 1.0")
   .Attr("b: bool = true")
   .Attr("ty: type = DT_INT32")
   .Attr("sh: shape = { dim { size: 1 } dim { size: 2 } }")
   .Attr("te: tensor = { dtype: DT_INT32 int_val: 5 }")
   .Attr("l_empty: list(int) = []")
   .Attr("l_int: list(int) = [2, 3, 5, 7]");

Note in particular that the values of type type use the DT_* names for the types.

Polymorphism

Type Polymorphism

For ops that can take different types as input or produce different output types, you can specify an attr in an input or output type in the Op registration. Typically you would then register an OpKernel for each supported type.

For instance, if you'd like the ZeroOut Op to work on floats in addition to int32s, your Op registration might look like:

REGISTER_OP("ZeroOut")
    .Attr("T: {float, int32}")
    .Input("to_zero: T")
    .Output("zeroed: T");

Your Op registration now specifies that the input's type must be float, or int32, and that its output will be the same type, since both have type T.

A note on naming: Inputs, outputs, and attrs generally should be given snake_case names. The one exception is attrs that are used as the type of an input or in the type of an input. Those attrs can be inferred when the op is added to the graph and so don't appear in the op's function. For example, this last definition of ZeroOut will generate a Python function that looks like:

def zero_out(to_zero, name=None):
  """...
  Args:
    to_zero: A `Tensor`. Must be one of the following types:
        `float32`, `int32`.
    name: A name for the operation (optional).

  Returns:
    A `Tensor`. Has the same type as `to_zero`.
  """

If to_zero is passed an int32 tensor, then T is automatically set to int32 (well, actually DT_INT32). Those inferred attrs are given Capitalized or CamelCase names.

Compare this with an op that has a type attr that determines the output type:

REGISTER_OP("StringToNumber")
    .Input("string_tensor: string")
    .Output("output: out_type")
    .Attr("out_type: {float, int32}");
    .Doc(R"doc(
Converts each string in the input Tensor to the specified numeric type.
)doc");

In this case, the user has to specify the output type, as in the generated Python:

def string_to_number(string_tensor, out_type=None, name=None):
  """Converts each string in the input Tensor to the specified numeric type.

  Args:
    string_tensor: A `Tensor` of type `string`.
    out_type: An optional `tf.DType` from: `tf.float32, tf.int32`.
      Defaults to `tf.float32`.
    name: A name for the operation (optional).

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

#include "tensorflow/core/framework/op_kernel.h"
class ZeroOutInt32Op : public OpKernel { // as before };
class ZeroOutFloatOp : public OpKernel { public: explicit ZeroOutFloatOp(OpKernelConstruction* context) : OpKernel(context) {}
void Compute(OpKernelContext* context) override { // Grab the input tensor const Tensor& input_tensor = context->input(0); auto input = input_tensor.flat<float>();
// Create an output tensor Tensor* output = NULL; OP_REQUIRES_OK(context, context->allocate_output(0, input_tensor.shape(), &output)); auto output_flat = output->template flat<float>();
// Set all the elements of the output tensor to 0 const int N = input.size(); for (int i = 0; i < N; i++) { output_flat(i) = 0; }
// Preserve the first input value if (N > 0) output_flat(0) = input(0); } };
// Note that TypeConstraint<int32>("T") means that attr "T" (defined // in the Op registration above) must be "int32" to use this template // instantiation. REGISTER_KERNEL_BUILDER( Name("ZeroOut") .Device(DEVICE_CPU) .TypeConstraint<int32>("T"), ZeroOutOpInt32); REGISTER_KERNEL_BUILDER( Name("ZeroOut") .Device(DEVICE_CPU) .TypeConstraint<float>("T"), ZeroOutFloatOp);

To preserve backwards compatibility, you should specify a default value when adding an attr to an existing op:

REGISTER_OP("ZeroOut")
  .Attr("T: {float, int32} = DT_INT32")
  .Input("to_zero: T")
  .Output("zeroed: T")

Lets say you wanted to add more types, say double:

REGISTER_OP("ZeroOut")
    .Attr("T: {float, double, int32}")
    .Input("to_zero: T")
    .Output("zeroed: T");

Instead of writing another OpKernel with redundant code as above, often you will be able to use a C++ template instead. You will still have one kernel registration (REGISTER\_KERNEL\_BUILDER call) per overload.

template <typename T>
class ZeroOutOp : public OpKernel {
 public:
  explicit ZeroOutOp(OpKernelConstruction* context) : OpKernel(context) {}
void Compute(OpKernelContext* context) override { // Grab the input tensor const Tensor& input_tensor = context->input(0); auto input = input_tensor.flat<T>();
// Create an output tensor Tensor* output = NULL; OP_REQUIRES_OK(context, context->allocate_output(0, input_tensor.shape(), &output)); auto output_flat = output->template flat<T>();
// Set all the elements of the output tensor to 0 const int N = input.size(); for (int i = 0; i < N; i++) { output_flat(i) = 0; }
// Preserve the first input value if (N > 0) output_flat(0) = input(0); } };
// Note that TypeConstraint<int32>("T") means that attr "T" (defined // in the Op registration above) must be "int32" to use this template // instantiation. REGISTER_KERNEL_BUILDER( Name("ZeroOut") .Device(DEVICE_CPU) .TypeConstraint<int32>("T"), ZeroOutOp<int32>); REGISTER_KERNEL_BUILDER( Name("ZeroOut") .Device(DEVICE_CPU) .TypeConstraint<float>("T"), ZeroOutOp<float>); REGISTER_KERNEL_BUILDER( Name("ZeroOut") .Device(DEVICE_CPU) .TypeConstraint<double>("T"), ZeroOutOp<double>);

If you have more than a couple overloads, you can put the registration in a macro.

#include "tensorflow/core/framework/op_kernel.h"

#define REGISTER_KERNEL(type)                                       \
  REGISTER_KERNEL_BUILDER(                                          \
      Name("ZeroOut").Device(DEVICE_CPU).TypeConstraint<type>("T"), \
      ZeroOutOp<type>)

REGISTER_KERNEL(int32);
REGISTER_KERNEL(float);
REGISTER_KERNEL(double);

#undef REGISTER_KERNEL

Depending on the list of types you are registering the kernel for, you may be able to use a macro provided by tensorflow/core/framework/register_types.h:

#include "tensorflow/core/framework/op_kernel.h"
#include "tensorflow/core/framework/register_types.h"

REGISTER_OP("ZeroOut")
    .Attr("T: realnumbertype")
    .Input("to_zero: T")
    .Output("zeroed: T");

template <typename T>
class ZeroOutOp : public OpKernel { ... };

#define REGISTER_KERNEL(type)                                       \
  REGISTER_KERNEL_BUILDER(                                          \
      Name("ZeroOut").Device(DEVICE_CPU).TypeConstraint<type>("T"), \
      ZeroOutOp<type>)

TF_CALL_REAL_NUMBER_TYPES(REGISTER_KERNEL);

#undef REGISTER_KERNEL

List Inputs and Outputs

In addition to being able to accept or produce different types, ops can consume or produce a variable number of tensors.

In the next example, the attr T holds a list of types, and is used as the type of both the input in and the output out. The input and output are lists of tensors of that type (and the number and types of tensors in the output are the same as the input, since both have type T).

REGISTER_OP("PolymorphicListExample")
    .Attr("T: list(type)")
    .Input("in: T")
    .Output("out: T");

You can also place restrictions on what types can be specified in the list. In this next case, the input is a list of float and double tensors. The Op accepts, for example, input types (float, double, float) and in that case the output type would also be (float, double, float).

REGISTER_OP("ListTypeRestrictionExample")
    .Attr("T: list({float, double})")
    .Input("in: T")
    .Output("out: T");

If you want all the tensors in a list to be of the same type, you might do something like:

REGISTER_OP("IntListInputExample")
    .Attr("N: int")
    .Input("in: N * int32")
    .Output("out: int32");

This accepts a list of int32 tensors, and uses an int attr N to specify the length of the list.

This can be made type polymorphic as well. In the next example, the input is a list of tensors (with length "N") of the same (but unspecified) type ("T"), and the output is a single tensor of matching type:

REGISTER_OP("SameListInputExample")
    .Attr("N: int")
    .Attr("T: type")
    .Input("in: N * T")
    .Output("out: T");

By default, tensor lists have a minimum length of 1. You can change that default using a ">=" constraint on the corresponding attr. In this next example, the input is a list of at least 2 int32 tensors:

REGISTER_OP("MinLengthIntListExample")
    .Attr("N: int >= 2")
    .Input("in: N * int32")
    .Output("out: int32");

The same syntax works with "list(type)" attrs:

REGISTER_OP("MinimumLengthPolymorphicListExample")
    .Attr("T: list(type) >= 3")
    .Input("in: T")
    .Output("out: T");

Inputs and Outputs

To summarize the above, an Op registration can have multiple inputs and outputs:

REGISTER_OP("MultipleInsAndOuts")
    .Input("y: int32")
    .Input("z: float")
    .Output("a: string")
    .Output("b: int32");

Each input or output spec is of the form:

<name>: <io-type-expr>

where <name> begins with a letter and can be composed of alphanumeric characters and underscores. <io-type-expr> is one of the following type expressions:

  • <type>, where <type> is a supported input type (e.g. float, int32, string). This specifies a single tensor of the given type.

See the list of supported Tensor types.

REGISTER_OP("BuiltInTypesExample")
    .Input("integers: int32")
    .Input("complex_numbers: complex64");
  • <attr-type>, where <attr-type> is the name of an Attr with type type or list(type) (with a possible type restriction). This syntax allows for polymorphic ops.
REGISTER_OP("PolymorphicSingleInput")
    .Attr("T: type")
    .Input("in: T);

REGISTER_OP("RestrictedPolymorphicSingleInput")
    .Attr("T: {int32, int64}")
    .Input("in: T);

Referencing an attr of type list(type) allows you to accept a sequence of tensors.

REGISTER_OP("ArbitraryTensorSequenceExample")
    .Attr("T: list(type)")
    .Input("in: T")
    .Output("out: T");

REGISTER_OP("RestrictedTensorSequenceExample")
    .Attr("T: list({int32, int64})")
    .Input("in: T")
    .Output("out: T");

Note that the number and types of tensors in the output out is the same as in the input in, since both are of type T.

  • For a sequence of tensors with the same type: <number> * <type>, where <number> is the name of an Attr with type int. The <type> can either be a specific type like int32 or float, or the name of an attr with type type. As an example of the first, this Op accepts a list of int32 tensors:
REGISTER_OP("Int32SequenceExample")
    .Attr("NumTensors: int")
    .Input("in: NumTensors * int32")

Whereas this Op accepts a list of tensors of any type, as long as they are all the same:

REGISTER_OP("SameTypeSequenceExample")
    .Attr("NumTensors: int")
    .Attr("T: type")
    .Input("in: NumTensors * T")
  • For a reference to a tensor: Ref(<type>), where <type> is one of the previous types.

A note on naming: Any attr used in the type of an input will be inferred. By convention those inferred attrs use capital names (like T or N). Otherwise inputs, outputs, and attrs have names like function parameters (e.g. num_outputs). For more details, see the earlier note on naming.

For more details, see tensorflow/core/framework/op_def_builder.h.

Backwards compatibility

In general, changes to specifications must be backwards-compatible: changing the specification of an Op must not break prior serialized GraphDef protocol buffers constructed from older specfications. The details of GraphDef compatibility are described here.

There are several ways to preserve backwards-compatibility.

  1. Any new attrs added to an operation must have default values defined, and with that default value the Op must have the original behavior. To change an operation from not polymorphic to polymorphic, you must give a default value to the new type attr to preserve the original signature by default. For example, if your operation was:

c++ REGISTER_OP("MyGeneralUnaryOp") .Input("in: float") .Output("out: float");

you can make it polymorphic in a backwards-compatible way using:

c++ REGISTER_OP("MyGeneralUnaryOp") .Input("in: T") .Output("out: T") .Attr("T: numerictype = DT_FLOAT");

  1. You can safely make a constraint on an attr less restrictive. For example, you can change from {int32, int64} to {int32, int64, float} or type. Or you may change from {"apple", "orange"} to {"apple", "banana", "orange"} or string.

  2. You can change single inputs / outputs into list inputs / outputs, as long as the default for the list type matches the old signature.

  3. You can add a new list input / output, if it defaults to empty.

  4. Namespace any new Ops you create, by prefixing the Op names with something unique to your project. This avoids having your Op colliding with any Ops that might be included in future versions of TensorFlow.

  5. Plan ahead! Try to anticipate future uses for the Op. Some signature changes can't be done in a compatible way (for example, making a list of the same type into a list of varying types).

The full list of safe and unsafe changes can be found in tensorflow/core/framework/op_compatibility_test.cc. If you cannot make your change to an operation backwards compatible, then create a new operation with a new name with the new semantics.

Also note that while these changes can maintain GraphDef compatibility, the generated Python code may change in a way that isn't compatible with old callers. The Python API may be kept compatible by careful changes in a hand-written Python wrapper, by keeping the old signature except possibly adding new optional arguments to the end. Generally incompatible changes may only be made when TensorFlow's changes major versions, and must conform to the GraphDef version semantics.

GPU Support

You can implement different OpKernels and register one for CPU and another for GPU, just like you can register kernels for different types. There are several examples of kernels with GPU support in tensorflow/core/kernels/. Notice some kernels have a CPU version in a .cc file, a GPU version in a file ending in _gpu.cu.cc, and some code shared in common in a .h file.

For example, the pad op has everything but the GPU kernel in tensorflow/core/kernels/pad_op.cc. The GPU kernel is in tensorflow/core/kernels/pad_op_gpu.cu.cc, and the shared code is a templated class defined in tensorflow/core/kernels/pad_op.h. One thing to note, even when the GPU kernel version of pad is used, it still needs its "paddings" input in CPU memory. To mark that inputs or outputs are kept on the CPU, add a HostMemory() call to the kernel registration, e.g.:

#define REGISTER_GPU_KERNEL(T)                         \
  REGISTER_KERNEL_BUILDER(Name("Pad")                  \
                              .Device(DEVICE_GPU)      \
                              .TypeConstraint<T>("T")  \
                              .HostMemory("paddings"), \
                          PadOp<GPUDevice, T>)

Compiling the kernel for the GPU device

Look at cuda_op_kernel.cu.cc for an example that uses a CUDA kernel to implement an op. The tf_custom_op_library accepts a gpu_srcs argument in which the list of source files containing the CUDA kernels (*.cu.cc files) can be specified. For use with a binary installation of TensorFlow, the CUDA kernels have to be compiled with NVIDIA's nvcc compiler. Here is the sequence of commands you can use to compile the cuda_op_kernel.cu.cc and cuda_op_kernel.cc into a single dynamically loadable library:

nvcc -std=c++11 -c -o cuda_op_kernel.cu.o cuda_op_kernel.cu.cc \
-I $TF_INC -D GOOGLE_CUDA=1 -x cu -Xcompiler -fPIC

g++ -std=c++11 -shared -o cuda_op_kernel.so cuda_op_kernel.cc \
cuda_op_kernel.cu.o -I $TF_INC -fPIC -lcudart

cuda_op_kernel.so produced above can be loaded as usual in Python, using the tf.load_op_library function.

Implement the gradient in Python

Given a graph of ops, TensorFlow uses automatic differentiation (backpropagation) to add new ops representing gradients with respect to the existing ops (see Gradient Computation). To make automatic differentiation work for new ops, you must register a gradient function which computes gradients with respect to the ops' inputs given gradients with respect to the ops' outputs.

Mathematically, if an op computes \(y = f(x)\) the registered gradient op converts gradients \(\partial L/ \partial y\) of loss \(L\) with respect to \(y\) into gradients \(\partial L/ \partial x\) with respect to \(x\) via the chain rule:

$$\frac{\partial L}{\partial x} = \frac{\partial L}{\partial y} \frac{\partial y}{\partial x} = \frac{\partial L}{\partial y} \frac{\partial f}{\partial x}.$$

In the case of ZeroOut, only one entry in the input affects the output, so the gradient with respect to the input is a sparse "one hot" tensor. This is expressed as follows:

from tensorflow.python.framework import ops
from tensorflow.python.ops import array_ops
from tensorflow.python.ops import sparse_ops

@ops.RegisterGradient("ZeroOut")
def _zero_out_grad(op, grad):
  """The gradients for `zero_out`.

  Args:
    op: The `zero_out` `Operation` that we are differentiating, which we can use
      to find the inputs and outputs of the original op.
    grad: Gradient with respect to the output of the `zero_out` op.

  Returns:
    Gradients with respect to the input of `zero_out`.
  """
  to_zero = op.inputs[0]
  shape = array_ops.shape(to_zero)
  index = array_ops.zeros_like(shape)
  first_grad = array_ops.reshape(grad, [-1])[0]
  to_zero_grad = sparse_ops.sparse_to_dense(index, shape, first_grad, 0)
  return [to_zero_grad]  # List of one Tensor, since we have one input

Details about registering gradient functions with ops.RegisterGradient:

  • For an op with one output, the gradient function will take an Operation op and a Tensor grad and build new ops out of the tensors op.inputs[i], op.outputs[i], and grad. Information about any attrs can be found via op.get_attr.

  • If the op has multiple outputs, the gradient function will take op and grads, where grads is a list of gradients with respect to each output. The result of the gradient function must be a list of Tensor objects representing the gradients with respect to each input.

  • If there is no well-defined gradient for some input, such as for integer inputs used as indices, the corresponding returned gradient should be None. For example, for an op taking a floating point tensor x and an integer index i, the gradient function would return [x_grad, None].

  • If there is no meaningful gradient for the op at all, use ops.NoGradient("OpName") to disable automatic differentiation.

Note that at the time the gradient function is called, only the data flow graph of ops is available, not the tensor data itself. Thus, all computation must be performed using other tensorflow ops, to be run at graph execution time.

Implement a shape function in Python

The TensorFlow Python API has a feature called "shape inference" that provides information about the shapes of tensors without having to execute the graph. Shape inference is supported by "shape functions" that are registered for each op type, and perform two roles: asserting that the shapes of the inputs are compatible, and specifying the shapes for the outputs. A shape function is a Python function that takes an Operation as input, and returns a list of TensorShape objects (one per output of the op). To register a shape function, apply the tf.RegisterShape decorator to a shape function. For example, the ZeroOut op defined above would have a shape function like the following:

@tf.RegisterShape("ZeroOut")
def _zero_out_shape(op):
  """Shape function for the ZeroOut op.

  This is the unconstrained version of ZeroOut, which produces an output
  with the same shape as its input.
  """
  return [op.inputs[0].get_shape()]

A shape function can also constrain the shape of an input. For the version of ZeroOut with a vector shape constraint, the shape function would be as follows:

@tf.RegisterShape("ZeroOut")
def _zero_out_shape(op):
  """Shape function for the ZeroOut op.

  This is the constrained version of ZeroOut, which requires the input to
  have rank 1 (a vector).
  """
  input_shape = op.inputs[0].get_shape().with_rank(1)
  return [input_shape]

If your op is polymorphic with multiple inputs, use the properties of the operation to determine the number of shapes to check:

@tf.RegisterShape("IntListInputExample")
def _int_list_input_example_shape(op):
  """Shape function for the "IntListInputExample" op.

  All inputs and the output are matrices of the same size.
  """
  output_shape = tf.TensorShape(None)
  for input in op.inputs:
    output_shape = output_shape.merge_with(input.get_shape().with_rank(2))
  return [output_shape]

Since shape inference is an optional feature, and the shapes of tensors may vary dynamically, shape functions must be robust to incomplete shape information for any of the inputs. The merge_with method allows the caller to assert that two shapes are the same, even if either or both of them do not have complete information. Shape functions are defined for all of the standard Python ops, and provide many different usage examples.