Advanced topics

For brevity, the rest of this chapter assumes that the following two lines are present:

#include <pybind11/pybind11.h>

namespace py = pybind11;

Exporting constants and mutable objects

To expose a C++ constant, use the attr function to register it in a module as shown below. The int_ class is one of many small wrapper objects defined in pybind11/pytypes.h. General objects (including integers) can also be converted using the function cast.

PYBIND11_PLUGIN(example) {
    py::module m("example", "pybind11 example plugin");
    m.attr("MY_CONSTANT") = py::int_(123);
    m.attr("MY_CONSTANT_2") = py::cast(new MyObject());
}

Operator overloading

Suppose that we’re given the following Vector2 class with a vector addition and scalar multiplication operation, all implemented using overloaded operators in C++.

class Vector2 {
public:
    Vector2(float x, float y) : x(x), y(y) { }

    Vector2 operator+(const Vector2 &v) const { return Vector2(x + v.x, y + v.y); }
    Vector2 operator*(float value) const { return Vector2(x * value, y * value); }
    Vector2& operator+=(const Vector2 &v) { x += v.x; y += v.y; return *this; }
    Vector2& operator*=(float v) { x *= v; y *= v; return *this; }

    friend Vector2 operator*(float f, const Vector2 &v) {
        return Vector2(f * v.x, f * v.y);
    }

    std::string toString() const {
        return "[" + std::to_string(x) + ", " + std::to_string(y) + "]";
    }
private:
    float x, y;
};

The following snippet shows how the above operators can be conveniently exposed to Python.

#include <pybind11/operators.h>

PYBIND11_PLUGIN(example) {
    py::module m("example", "pybind11 example plugin");

    py::class_<Vector2>(m, "Vector2")
        .def(py::init<float, float>())
        .def(py::self + py::self)
        .def(py::self += py::self)
        .def(py::self *= float())
        .def(float() * py::self)
        .def("__repr__", &Vector2::toString);

    return m.ptr();
}

Note that a line like

.def(py::self * float())

is really just short hand notation for

.def("__mul__", [](const Vector2 &a, float b) {
    return a * b;
})

This can be useful for exposing additional operators that don’t exist on the C++ side, or to perform other types of customization.

Note

To use the more convenient py::self notation, the additional header file pybind11/operators.h must be included.

See also

The file example/example3.cpp contains a complete example that demonstrates how to work with overloaded operators in more detail.

Callbacks and passing anonymous functions

The C++11 standard brought lambda functions and the generic polymorphic function wrapper std::function<> to the C++ programming language, which enable powerful new ways of working with functions. Lambda functions come in two flavors: stateless lambda function resemble classic function pointers that link to an anonymous piece of code, while stateful lambda functions additionally depend on captured variables that are stored in an anonymous lambda closure object.

Here is a simple example of a C++ function that takes an arbitrary function (stateful or stateless) with signature int -> int as an argument and runs it with the value 10.

int func_arg(const std::function<int(int)> &f) {
    return f(10);
}

The example below is more involved: it takes a function of signature int -> int and returns another function of the same kind. The return value is a stateful lambda function, which stores the value f in the capture object and adds 1 to its return value upon execution.

std::function<int(int)> func_ret(const std::function<int(int)> &f) {
    return [f](int i) {
        return f(i) + 1;
    };
}

After including the extra header file pybind11/functional.h, it is almost trivial to generate binding code for both of these functions.

#include <pybind11/functional.h>

PYBIND11_PLUGIN(example) {
    py::module m("example", "pybind11 example plugin");

    m.def("func_arg", &func_arg);
    m.def("func_ret", &func_ret);

    return m.ptr();
}

The following interactive session shows how to call them from Python.

$ python
>>> import example
>>> def square(i):
...     return i * i
...
>>> example.func_arg(square)
100L
>>> square_plus_1 = example.func_ret(square)
>>> square_plus_1(4)
17L
>>>

Note

This functionality is very useful when generating bindings for callbacks in C++ libraries (e.g. a graphical user interface library).

The file example/example5.cpp contains a complete example that demonstrates how to work with callbacks and anonymous functions in more detail.

Warning

Keep in mind that passing a function from C++ to Python (or vice versa) will instantiate a piece of wrapper code that translates function invocations between the two languages. Copying the same function back and forth between Python and C++ many times in a row will cause these wrappers to accumulate, which can decrease performance.

Overriding virtual functions in Python

Suppose that a C++ class or interface has a virtual function that we’d like to to override from within Python (we’ll focus on the class Animal; Dog is given as a specific example of how one would do this with traditional C++ code).

class Animal {
public:
    virtual ~Animal() { }
    virtual std::string go(int n_times) = 0;
};

class Dog : public Animal {
public:
    std::string go(int n_times) {
        std::string result;
        for (int i=0; i<n_times; ++i)
            result += "woof! ";
        return result;
    }
};

Let’s also suppose that we are given a plain function which calls the function go() on an arbitrary Animal instance.

std::string call_go(Animal *animal) {
    return animal->go(3);
}

Normally, the binding code for these classes would look as follows:

PYBIND11_PLUGIN(example) {
    py::module m("example", "pybind11 example plugin");

    py::class_<Animal> animal(m, "Animal");
    animal
        .def("go", &Animal::go);

    py::class_<Dog>(m, "Dog", animal)
        .def(py::init<>());

    m.def("call_go", &call_go);

    return m.ptr();
}

However, these bindings are impossible to extend: Animal is not constructible, and we clearly require some kind of “trampoline” that redirects virtual calls back to Python.

Defining a new type of Animal from within Python is possible but requires a helper class that is defined as follows:

class PyAnimal : public Animal {
public:
    /* Inherit the constructors */
    using Animal::Animal;

    /* Trampoline (need one for each virtual function) */
    std::string go(int n_times) {
        PYBIND11_OVERLOAD_PURE(
            std::string, /* Return type */
            Animal,      /* Parent class */
            go,          /* Name of function */
            n_times      /* Argument(s) */
        );
    }
};

The macro PYBIND11_OVERLOAD_PURE() should be used for pure virtual functions, and PYBIND11_OVERLOAD() should be used for functions which have a default implementation.

There are also two alternate macros PYBIND11_OVERLOAD_PURE_NAME() and PYBIND11_OVERLOAD_NAME() which take a string-valued name argument after the Name of the function slot. This is useful when the C++ and Python versions of the function have different names, e.g. operator() vs __call__.

The binding code also needs a few minor adaptations (highlighted):

PYBIND11_PLUGIN(example) {
    py::module m("example", "pybind11 example plugin");

    py::class_<PyAnimal> animal(m, "Animal");
    animal
        .alias<Animal>()
        .def(py::init<>())
        .def("go", &Animal::go);

    py::class_<Dog>(m, "Dog", animal)
        .def(py::init<>());

    m.def("call_go", &call_go);

    return m.ptr();
}

Importantly, the trampoline helper class is used as the template argument to class_, and a call to class_::alias() informs the binding generator that this is merely an alias for the underlying type Animal. Following this, we are able to define a constructor as usual.

The Python session below shows how to override Animal::go and invoke it via a virtual method call.

>>> from example import *
>>> d = Dog()
>>> call_go(d)
u'woof! woof! woof! '
>>> class Cat(Animal):
...     def go(self, n_times):
...             return "meow! " * n_times
...
>>> c = Cat()
>>> call_go(c)
u'meow! meow! meow! '

Please take a look at the General notes regarding convenience macros before using this feature.

See also

The file example/example12.cpp contains a complete example that demonstrates how to override virtual functions using pybind11 in more detail.

General notes regarding convenience macros

pybind11 provides a few convenience macros such as PYBIND11_MAKE_OPAQUE() and PYBIND11_DECLARE_HOLDER_TYPE(), and PYBIND11_OVERLOAD_*. Since these are “just” macros that are evaluated in the preprocessor (which has no concept of types), they will get confused by commas in a template argument such as PYBIND11_OVERLOAD(MyReturnValue<T1, T2>, myFunc). In this case, the preprocessor assumes that the comma indicates the beginnning of the next parameter. Use a typedef to bind the template to another name and use it in the macro to avoid this problem.

Global Interpreter Lock (GIL)

The classes gil_scoped_release and gil_scoped_acquire can be used to acquire and release the global interpreter lock in the body of a C++ function call. In this way, long-running C++ code can be parallelized using multiple Python threads. Taking the previous section as an example, this could be realized as follows (important changes highlighted):

class PyAnimal : public Animal {
public:
    /* Inherit the constructors */
    using Animal::Animal;

    /* Trampoline (need one for each virtual function) */
    std::string go(int n_times) {
        /* Acquire GIL before calling Python code */
        py::gil_scoped_acquire acquire;

        PYBIND11_OVERLOAD_PURE(
            std::string, /* Return type */
            Animal,      /* Parent class */
            go,          /* Name of function */
            n_times      /* Argument(s) */
        );
    }
};

PYBIND11_PLUGIN(example) {
    py::module m("example", "pybind11 example plugin");

    py::class_<PyAnimal> animal(m, "Animal");
    animal
        .alias<Animal>()
        .def(py::init<>())
        .def("go", &Animal::go);

    py::class_<Dog>(m, "Dog", animal)
        .def(py::init<>());

    m.def("call_go", [](Animal *animal) -> std::string {
        /* Release GIL before calling into (potentially long-running) C++ code */
        py::gil_scoped_release release;
        return call_go(animal);
    });

    return m.ptr();
}

Passing STL data structures

When including the additional header file pybind11/stl.h, conversions between std::vector<>, std::list<>, std::set<>, and std::map<> and the Python list, set and dict data structures are automatically enabled. The types std::pair<> and std::tuple<> are already supported out of the box with just the core pybind11/pybind11.h header.

Note

Arbitrary nesting of any of these types is supported.

See also

The file example/example2.cpp contains a complete example that demonstrates how to pass STL data types in more detail.

Binding sequence data types, iterators, the slicing protocol, etc.

Please refer to the supplemental example for details.

See also

The file example/example6.cpp contains a complete example that shows how to bind a sequence data type, including length queries (__len__), iterators (__iter__), the slicing protocol and other kinds of useful operations.

Return value policies

Python and C++ use wildly different ways of managing the memory and lifetime of objects managed by them. This can lead to issues when creating bindings for functions that return a non-trivial type. Just by looking at the type information, it is not clear whether Python should take charge of the returned value and eventually free its resources, or if this is handled on the C++ side. For this reason, pybind11 provides a several return value policy annotations that can be passed to the module::def() and class_::def() functions. The default policy is return_value_policy::automatic.

Return value policy	Description
return_value_policy::automatic	This is the default return value policy, which falls back to the policy return_value_policy::take_ownership when the return value is a pointer. Otherwise, it uses return_value::move or return_value::copy for rvalue and lvalue references, respectively. See below for a description of what all of these different policies do.
return_value_policy::automatic_reference	As above, but use policy return_value_policy::reference when the return value is a pointer. You probably won’t need to use this.
return_value_policy::take_ownership	Reference an existing object (i.e. do not create a new copy) and take ownership. Python will call the destructor and delete operator when the object’s reference count reaches zero. Undefined behavior ensues when the C++ side does the same..
return_value_policy::copy	Create a new copy of the returned object, which will be owned by Python. This policy is comparably safe because the lifetimes of the two instances are decoupled.
return_value_policy::move	Use std::move to move the return value contents into a new instance that will be owned by Python. This policy is comparably safe because the lifetimes of the two instances (move source and destination) are decoupled.
return_value_policy::reference	Reference an existing object, but do not take ownership. The C++ side is responsible for managing the object’s lifetime and deallocating it when it is no longer used. Warning: undefined behavior will ensue when the C++ side deletes an object that is still referenced and used by Python.
return_value_policy::reference_internal	This policy only applies to methods and properties. It references the object without taking ownership similar to the above return_value_policy::reference policy. In contrast to that policy, the function or property’s implicit this argument (called the parent) is considered to be the the owner of the return value (the child). pybind11 then couples the lifetime of the parent to the child via a reference relationship that ensures that the parent cannot be garbage collected while Python is still using the child. More advanced variations of this scheme are also possible using combinations of return_value_policy::reference and the keep_alive call policy described next.

The following example snippet shows a use case of the return_value_policy::reference_internal policy.

class Example {
public:
    Internal &get_internal() { return internal; }
private:
    Internal internal;
};

PYBIND11_PLUGIN(example) {
    py::module m("example", "pybind11 example plugin");

    py::class_<Example>(m, "Example")
        .def(py::init<>())
        .def("get_internal", &Example::get_internal, "Return the internal data",
                             py::return_value_policy::reference_internal);

    return m.ptr();
}

Warning

Code with invalid call policies might access unitialized memory or free data structures multiple times, which can lead to hard-to-debug non-determinism and segmentation faults, hence it is worth spending the time to understand all the different options in the table above.

Note

The next section on Additional call policies discusses call policies that can be specified in addition to a return value policy from the list above. Call policies indicate reference relationships that can involve both return values and parameters of functions.

Note

As an alternative to elaborate call policies and lifetime management logic, consider using smart pointers (see the section on Smart pointers for details). Smart pointers can tell whether an object is still referenced from C++ or Python, which generally eliminates the kinds of inconsistencies that can lead to crashes or undefined behavior. For functions returning smart pointers, it is not necessary to specify a return value policy.

Additional call policies

In addition to the above return value policies, further call policies can be specified to indicate dependencies between parameters. There is currently just one policy named keep_alive<Nurse, Patient>, which indicates that the argument with index Patient should be kept alive at least until the argument with index Nurse is freed by the garbage collector; argument indices start at one, while zero refers to the return value. For methods, index one refers to the implicit this pointer, while regular arguments begin at index two. Arbitrarily many call policies can be specified.

Consider the following example: the binding code for a list append operation that ties the lifetime of the newly added element to the underlying container might be declared as follows:

py::class_<List>(m, "List")
    .def("append", &List::append, py::keep_alive<1, 2>());

Note

keep_alive is analogous to the with_custodian_and_ward (if Nurse, Patient != 0) and with_custodian_and_ward_postcall (if Nurse/Patient == 0) policies from Boost.Python.

See also

The file example/example13.cpp contains a complete example that demonstrates using keep_alive in more detail.

Implicit type conversions

Suppose that instances of two types A and B are used in a project, and that an A can easily be converted into an instance of type B (examples of this could be a fixed and an arbitrary precision number type).

py::class_<A>(m, "A")
    /// ... members ...

py::class_<B>(m, "B")
    .def(py::init<A>())
    /// ... members ...

m.def("func",
    [](const B &) { /* .... */ }
);

To invoke the function func using a variable a containing an A instance, we’d have to write func(B(a)) in Python. On the other hand, C++ will automatically apply an implicit type conversion, which makes it possible to directly write func(a).

In this situation (i.e. where B has a constructor that converts from A), the following statement enables similar implicit conversions on the Python side:

py::implicitly_convertible<A, B>();

Unique pointers

Given a class Example with Python bindings, it’s possible to return instances wrapped in C++11 unique pointers, like so

std::unique_ptr<Example> create_example() { return std::unique_ptr<Example>(new Example()); }

m.def("create_example", &create_example);

In other words, there is nothing special that needs to be done. While returning unique pointers in this way is allowed, it is illegal to use them as function arguments. For instance, the following function signature cannot be processed by pybind11.

void do_something_with_example(std::unique_ptr<Example> ex) { ... }

The above signature would imply that Python needs to give up ownership of an object that is passed to this function, which is generally not possible (for instance, the object might be referenced elsewhere).

Smart pointers

This section explains how to pass values that are wrapped in “smart” pointer types with internal reference counting. For the simpler C++11 unique pointers, refer to the previous section.

The binding generator for classes, class_, takes an optional second template type, which denotes a special holder type that is used to manage references to the object. When wrapping a type named Type, the default value of this template parameter is std::unique_ptr<Type>, which means that the object is deallocated when Python’s reference count goes to zero.

It is possible to switch to other types of reference counting wrappers or smart pointers, which is useful in codebases that rely on them. For instance, the following snippet causes std::shared_ptr to be used instead.

py::class_<Example, std::shared_ptr<Example> /* <- holder type */> obj(m, "Example");

Note that any particular class can only be associated with a single holder type.

To enable transparent conversions for functions that take shared pointers as an argument or that return them, a macro invocation similar to the following must be declared at the top level before any binding code:

PYBIND11_DECLARE_HOLDER_TYPE(T, std::shared_ptr<T>);

Note

The first argument of PYBIND11_DECLARE_HOLDER_TYPE() should be a placeholder name that is used as a template parameter of the second argument. Thus, feel free to use any identifier, but use it consistently on both sides; also, don’t use the name of a type that already exists in your codebase.

One potential stumbling block when using holder types is that they need to be applied consistently. Can you guess what’s broken about the following binding code?

class Child { };

class Parent {
public:
   Parent() : child(std::make_shared<Child>()) { }
   Child *get_child() { return child.get(); }  /* Hint: ** DON'T DO THIS ** */
private:
    std::shared_ptr<Child> child;
};

PYBIND11_PLUGIN(example) {
    py::module m("example");

    py::class_<Child, std::shared_ptr<Child>>(m, "Child");

    py::class_<Parent, std::shared_ptr<Parent>>(m, "Parent")
       .def(py::init<>())
       .def("get_child", &Parent::get_child);

    return m.ptr();
}

The following Python code will cause undefined behavior (and likely a segmentation fault).

from example import Parent
print(Parent().get_child())

The problem is that Parent::get_child() returns a pointer to an instance of Child, but the fact that this instance is already managed by std::shared_ptr<...> is lost when passing raw pointers. In this case, pybind11 will create a second independent std::shared_ptr<...> that also claims ownership of the pointer. In the end, the object will be freed twice since these shared pointers have no way of knowing about each other.

There are two ways to resolve this issue:

1. For types that are managed by a smart pointer class, never use raw pointers in function arguments or return values. In other words: always consistently wrap pointers into their designated holder types (such as std::shared_ptr<...>). In this case, the signature of get_child() should be modified as follows:

std::shared_ptr<Child> get_child() { return child; }

2. Adjust the definition of Child by specifying std::enable_shared_from_this<T> (see cppreference for details) as a base class. This adds a small bit of information to Child that allows pybind11 to realize that there is already an existing std::shared_ptr<...> and communicate with it. In this case, the declaration of Child should look as follows:

class Child : public std::enable_shared_from_this<Child> { };

Please take a look at the General notes regarding convenience macros before using this feature.

See also

The file example/example8.cpp contains a complete example that demonstrates how to work with custom reference-counting holder types in more detail.

Custom constructors

The syntax for binding constructors was previously introduced, but it only works when a constructor with the given parameters actually exists on the C++ side. To extend this to more general cases, let’s take a look at what actually happens under the hood: the following statement

py::class_<Example>(m, "Example")
    .def(py::init<int>());

is short hand notation for

py::class_<Example>(m, "Example")
    .def("__init__",
        [](Example &instance, int arg) {
            new (&instance) Example(arg);
        }
    );

In other words, init() creates an anonymous function that invokes an in-place constructor. Memory allocation etc. is already take care of beforehand within pybind11.

Catching and throwing exceptions

When C++ code invoked from Python throws an std::exception, it is automatically converted into a Python Exception. pybind11 defines multiple special exception classes that will map to different types of Python exceptions:

C++ exception type	Python exception type
std::exception	RuntimeError
std::bad_alloc	MemoryError
std::domain_error	ValueError
std::invalid_argument	ValueError
std::length_error	ValueError
std::out_of_range	ValueError
std::range_error	ValueError
pybind11::stop_iteration	StopIteration (used to implement custom iterators)
pybind11::index_error	IndexError (used to indicate out of bounds accesses in __getitem__, __setitem__, etc.)
pybind11::value_error	ValueError (used to indicate wrong value passed in container.remove(...)
pybind11::error_already_set	Indicates that the Python exception flag has already been initialized

When a Python function invoked from C++ throws an exception, it is converted into a C++ exception of type error_already_set whose string payload contains a textual summary.

There is also a special exception cast_error that is thrown by handle::call() when the input arguments cannot be converted to Python objects.

Treating STL data structures as opaque objects

pybind11 heavily relies on a template matching mechanism to convert parameters and return values that are constructed from STL data types such as vectors, linked lists, hash tables, etc. This even works in a recursive manner, for instance to deal with lists of hash maps of pairs of elementary and custom types, etc.

However, a fundamental limitation of this approach is that internal conversions between Python and C++ types involve a copy operation that prevents pass-by-reference semantics. What does this mean?

Suppose we bind the following function

void append_1(std::vector<int> &v) {
   v.push_back(1);
}

and call it from Python, the following happens:

>>> v = [5, 6]
>>> append_1(v)
>>> print(v)
[5, 6]

As you can see, when passing STL data structures by reference, modifications are not propagated back the Python side. A similar situation arises when exposing STL data structures using the def_readwrite or def_readonly functions:

/* ... definition ... */

class MyClass {
    std::vector<int> contents;
};

/* ... binding code ... */

py::class_<MyClass>(m, "MyClass")
    .def(py::init<>)
    .def_readwrite("contents", &MyClass::contents);

In this case, properties can be read and written in their entirety. However, an append operaton involving such a list type has no effect:

>>> m = MyClass()
>>> m.contents = [5, 6]
>>> print(m.contents)
[5, 6]
>>> m.contents.append(7)
>>> print(m.contents)
[5, 6]

To deal with both of the above situations, pybind11 provides a macro named PYBIND11_MAKE_OPAQUE(T) that disables the template-based conversion machinery of types, thus rendering them opaque. The contents of opaque objects are never inspected or extracted, hence they can be passed by reference. For instance, to turn std::vector<int> into an opaque type, add the declaration

PYBIND11_MAKE_OPAQUE(std::vector<int>);

before any binding code (e.g. invocations to class_::def(), etc.). This macro must be specified at the top level, since instantiates a partial template overload. If your binding code consists of multiple compilation units, it must be present in every file preceding any usage of std::vector<int>. Opaque types must also have a corresponding class_ declaration to associate them with a name in Python, and to define a set of available operations:

py::class_<std::vector<int>>(m, "IntVector")
    .def(py::init<>())
    .def("clear", &std::vector<int>::clear)
    .def("pop_back", &std::vector<int>::pop_back)
    .def("__len__", [](const std::vector<int> &v) { return v.size(); })
    .def("__iter__", [](std::vector<int> &v) {
       return py::make_iterator(v.begin(), v.end());
    }, py::keep_alive<0, 1>()) /* Keep vector alive while iterator is used */
    // ....

Please take a look at the General notes regarding convenience macros before using this feature.

See also

The file example/example14.cpp contains a complete example that demonstrates how to create and expose opaque types using pybind11 in more detail.

Transparent conversion of dense and sparse Eigen data types

Eigen [1] is C++ header-based library for dense and sparse linear algebra. Due to its popularity and widespread adoption, pybind11 provides transparent conversion support between Eigen and Scientific Python linear algebra data types.

Specifically, when including the optional header file pybind11/eigen.h, pybind11 will automatically and transparently convert

1. Static and dynamic Eigen dense vectors and matrices to instances of numpy.ndarray (and vice versa).

1. Eigen sparse vectors and matrices to instances of scipy.sparse.csr_matrix/scipy.sparse.csc_matrix (and vice versa).

This makes it possible to bind most kinds of functions that rely on these types. One major caveat are functions that take Eigen matrices by reference and modify them somehow, in which case the information won’t be propagated to the caller.

/* The Python bindings of this function won't replicate
   the intended effect of modifying the function argument */
void scale_by_2(Eigen::Vector3f &v) {
   v *= 2;
}

To see why this is, refer to the section on Treating STL data structures as opaque objects (although that section specifically covers STL data types, the underlying issue is the same). The next two sections discuss an efficient alternative for exposing the underlying native Eigen types as opaque objects in a way that still integrates with NumPy and SciPy.

[1]	http://eigen.tuxfamily.org

See also

The file example/eigen.cpp contains a complete example that shows how to pass Eigen sparse and dense data types in more detail.

Buffer protocol

Python supports an extremely general and convenient approach for exchanging data between plugin libraries. Types can expose a buffer view [2], which provides fast direct access to the raw internal data representation. Suppose we want to bind the following simplistic Matrix class:

class Matrix {
public:
    Matrix(size_t rows, size_t cols) : m_rows(rows), m_cols(cols) {
        m_data = new float[rows*cols];
    }
    float *data() { return m_data; }
    size_t rows() const { return m_rows; }
    size_t cols() const { return m_cols; }
private:
    size_t m_rows, m_cols;
    float *m_data;
};

The following binding code exposes the Matrix contents as a buffer object, making it possible to cast Matrices into NumPy arrays. It is even possible to completely avoid copy operations with Python expressions like np.array(matrix_instance, copy = False).

py::class_<Matrix>(m, "Matrix")
   .def_buffer([](Matrix &m) -> py::buffer_info {
        return py::buffer_info(
            m.data(),                            /* Pointer to buffer */
            sizeof(float),                       /* Size of one scalar */
            py::format_descriptor<float>::value, /* Python struct-style format descriptor */
            2,                                   /* Number of dimensions */
            { m.rows(), m.cols() },              /* Buffer dimensions */
            { sizeof(float) * m.rows(),          /* Strides (in bytes) for each index */
              sizeof(float) }
        );
    });

The snippet above binds a lambda function, which can create py::buffer_info description records on demand describing a given matrix. The contents of py::buffer_info mirror the Python buffer protocol specification.

struct buffer_info {
    void *ptr;
    size_t itemsize;
    std::string format;
    int ndim;
    std::vector<size_t> shape;
    std::vector<size_t> strides;
};

To create a C++ function that can take a Python buffer object as an argument, simply use the type py::buffer as one of its arguments. Buffers can exist in a great variety of configurations, hence some safety checks are usually necessary in the function body. Below, you can see an basic example on how to define a custom constructor for the Eigen double precision matrix (Eigen::MatrixXd) type, which supports initialization from compatible buffer objects (e.g. a NumPy matrix).

/* Bind MatrixXd (or some other Eigen type) to Python */
typedef Eigen::MatrixXd Matrix;

typedef Matrix::Scalar Scalar;
constexpr bool rowMajor = Matrix::Flags & Eigen::RowMajorBit;

py::class_<Matrix>(m, "Matrix")
    .def("__init__", [](Matrix &m, py::buffer b) {
        typedef Eigen::Stride<Eigen::Dynamic, Eigen::Dynamic> Strides;

        /* Request a buffer descriptor from Python */
        py::buffer_info info = b.request();

        /* Some sanity checks ... */
        if (info.format != py::format_descriptor<Scalar>::value)
            throw std::runtime_error("Incompatible format: expected a double array!");

        if (info.ndim != 2)
            throw std::runtime_error("Incompatible buffer dimension!");

        auto strides = Strides(
            info.strides[rowMajor ? 0 : 1] / sizeof(Scalar),
            info.strides[rowMajor ? 1 : 0] / sizeof(Scalar));

        auto map = Eigen::Map<Matrix, 0, Strides>(
            static_cat<Scalar *>(info.ptr), info.shape[0], info.shape[1], strides);

        new (&m) Matrix(map);
    });

For reference, the def_buffer() call for this Eigen data type should look as follows:

.def_buffer([](Matrix &m) -> py::buffer_info {
    return py::buffer_info(
        m.data(),                /* Pointer to buffer */
        sizeof(Scalar),          /* Size of one scalar */
        /* Python struct-style format descriptor */
        py::format_descriptor<Scalar>::value,
        /* Number of dimensions */
        2,
        /* Buffer dimensions */
        { (size_t) m.rows(),
          (size_t) m.cols() },
        /* Strides (in bytes) for each index */
        { sizeof(Scalar) * (rowMajor ? m.cols() : 1),
          sizeof(Scalar) * (rowMajor ? 1 : m.rows()) }
    );
 })

For a much easier approach of binding Eigen types (although with some limitations), refer to the section on Transparent conversion of dense and sparse Eigen data types.

See also

The file example/example7.cpp contains a complete example that demonstrates using the buffer protocol with pybind11 in more detail.

[2]	http://docs.python.org/3/c-api/buffer.html

NumPy support

By exchanging py::buffer with py::array in the above snippet, we can restrict the function so that it only accepts NumPy arrays (rather than any type of Python object satisfying the buffer protocol).

In many situations, we want to define a function which only accepts a NumPy array of a certain data type. This is possible via the py::array_t<T> template. For instance, the following function requires the argument to be a NumPy array containing double precision values.

void f(py::array_t<double> array);

When it is invoked with a different type (e.g. an integer or a list of integers), the binding code will attempt to cast the input into a NumPy array of the requested type. Note that this feature requires the :file:pybind11/numpy.h header to be included.

Data in NumPy arrays is not guaranteed to packed in a dense manner; furthermore, entries can be separated by arbitrary column and row strides. Sometimes, it can be useful to require a function to only accept dense arrays using either the C (row-major) or Fortran (column-major) ordering. This can be accomplished via a second template argument with values py::array::c_style or py::array::f_style.

void f(py::array_t<double, py::array::c_style | py::array::forcecast> array);

The py::array::forcecast argument is the default value of the second template paramenter, and it ensures that non-conforming arguments are converted into an array satisfying the specified requirements instead of trying the next function overload.

Vectorizing functions

Suppose we want to bind a function with the following signature to Python so that it can process arbitrary NumPy array arguments (vectors, matrices, general N-D arrays) in addition to its normal arguments:

double my_func(int x, float y, double z);

After including the pybind11/numpy.h header, this is extremely simple:

m.def("vectorized_func", py::vectorize(my_func));

Invoking the function like below causes 4 calls to be made to my_func with each of the array elements. The significant advantage of this compared to solutions like numpy.vectorize() is that the loop over the elements runs entirely on the C++ side and can be crunched down into a tight, optimized loop by the compiler. The result is returned as a NumPy array of type numpy.dtype.float64.

>>> x = np.array([[1, 3],[5, 7]])
>>> y = np.array([[2, 4],[6, 8]])
>>> z = 3
>>> result = vectorized_func(x, y, z)

The scalar argument z is transparently replicated 4 times. The input arrays x and y are automatically converted into the right types (they are of type numpy.dtype.int64 but need to be numpy.dtype.int32 and numpy.dtype.float32, respectively)

Sometimes we might want to explicitly exclude an argument from the vectorization because it makes little sense to wrap it in a NumPy array. For instance, suppose the function signature was

double my_func(int x, float y, my_custom_type *z);

This can be done with a stateful Lambda closure:

// Vectorize a lambda function with a capture object (e.g. to exclude some arguments from the vectorization)
m.def("vectorized_func",
    [](py::array_t<int> x, py::array_t<float> y, my_custom_type *z) {
        auto stateful_closure = [z](int x, float y) { return my_func(x, y, z); };
        return py::vectorize(stateful_closure)(x, y);
    }
);

In cases where the computation is too complicated to be reduced to vectorize, it will be necessary to create and access the buffer contents manually. The following snippet contains a complete example that shows how this works (the code is somewhat contrived, since it could have been done more simply using vectorize).

#include <pybind11/pybind11.h>
#include <pybind11/numpy.h>

namespace py = pybind11;

py::array_t<double> add_arrays(py::array_t<double> input1, py::array_t<double> input2) {
    auto buf1 = input1.request(), buf2 = input2.request();

    if (buf1.ndim != 1 || buf2.ndim != 1)
        throw std::runtime_error("Number of dimensions must be one");

    if (buf1.shape[0] != buf2.shape[0])
        throw std::runtime_error("Input shapes must match");

    auto result = py::array(py::buffer_info(
        nullptr,            /* Pointer to data (nullptr -> ask NumPy to allocate!) */
        sizeof(double),     /* Size of one item */
        py::format_descriptor<double>::value(), /* Buffer format */
        buf1.ndim,          /* How many dimensions? */
        { buf1.shape[0] },  /* Number of elements for each dimension */
        { sizeof(double) }  /* Strides for each dimension */
    ));

    auto buf3 = result.request();

    double *ptr1 = (double *) buf1.ptr,
           *ptr2 = (double *) buf2.ptr,
           *ptr3 = (double *) buf3.ptr;

    for (size_t idx = 0; idx < buf1.shape[0]; idx++)
        ptr3[idx] = ptr1[idx] + ptr2[idx];

    return result;
}

PYBIND11_PLUGIN(test) {
    py::module m("test");
    m.def("add_arrays", &add_arrays, "Add two NumPy arrays");
    return m.ptr();
}

See also

The file example/example10.cpp contains a complete example that demonstrates using vectorize() in more detail.

Functions taking Python objects as arguments

pybind11 exposes all major Python types using thin C++ wrapper classes. These wrapper classes can also be used as parameters of functions in bindings, which makes it possible to directly work with native Python types on the C++ side. For instance, the following statement iterates over a Python dict:

void print_dict(py::dict dict) {
    /* Easily interact with Python types */
    for (auto item : dict)
        std::cout << "key=" << item.first << ", "
                  << "value=" << item.second << std::endl;
}

Available types include handle, object, bool_, int_, float_, str, bytes, tuple, list, dict, slice, none, capsule, iterable, iterator, function, buffer, array, and array_t.

In this kind of mixed code, it is often necessary to convert arbitrary C++ types to Python, which can be done using cast():

MyClass *cls = ..;
py::object obj = py::cast(cls);

The reverse direction uses the following syntax:

py::object obj = ...;
MyClass *cls = obj.cast<MyClass *>();

When conversion fails, both directions throw the exception cast_error. It is also possible to call python functions via operator().

py::function f = <...>;
py::object result_py = f(1234, "hello", some_instance);
MyClass &result = result_py.cast<MyClass>();

The special f(*args) and f(*args, **kwargs) syntax is also supported to supply arbitrary argument and keyword lists, although these cannot be mixed with other parameters.

py::function f = <...>;
py::tuple args = py::make_tuple(1234);
py::dict kwargs;
kwargs["y"] = py::cast(5678);
py::object result = f(*args, **kwargs);

See also

The file example/example2.cpp contains a complete example that demonstrates passing native Python types in more detail. The file example/example11.cpp discusses usage of args and kwargs.

Default arguments revisited

The section on Default arguments previously discussed basic usage of default arguments using pybind11. One noteworthy aspect of their implementation is that default arguments are converted to Python objects right at declaration time. Consider the following example:

py::class_<MyClass>("MyClass")
    .def("myFunction", py::arg("arg") = SomeType(123));

In this case, pybind11 must already be set up to deal with values of the type SomeType (via a prior instantiation of py::class_<SomeType>), or an exception will be thrown.

Another aspect worth highlighting is that the “preview” of the default argument in the function signature is generated using the object’s __repr__ method. If not available, the signature may not be very helpful, e.g.:

FUNCTIONS
...
|  myFunction(...)
|      Signature : (MyClass, arg : SomeType = <SomeType object at 0x101b7b080>) -> NoneType
...

The first way of addressing this is by defining SomeType.__repr__. Alternatively, it is possible to specify the human-readable preview of the default argument manually using the arg_t notation:

py::class_<MyClass>("MyClass")
    .def("myFunction", py::arg_t<SomeType>("arg", SomeType(123), "SomeType(123)"));

Sometimes it may be necessary to pass a null pointer value as a default argument. In this case, remember to cast it to the underlying type in question, like so:

py::class_<MyClass>("MyClass")
    .def("myFunction", py::arg("arg") = (SomeType *) nullptr);

Binding functions that accept arbitrary numbers of arguments and keywords arguments

Python provides a useful mechanism to define functions that accept arbitrary numbers of arguments and keyword arguments:

def generic(*args, **kwargs):
    # .. do something with args and kwargs

Such functions can also be created using pybind11:

void generic(py::args args, py::kwargs kwargs) {
    /// .. do something with args
    if (kwargs)
        /// .. do something with kwargs
}

/// Binding code
m.def("generic", &generic);

(See example/example11.cpp). The class py::args derives from py::list and py::kwargs derives from py::dict Note that the kwargs argument is invalid if no keyword arguments were actually provided. Please refer to the other examples for details on how to iterate over these, and on how to cast their entries into C++ objects.

Partitioning code over multiple extension modules

It’s straightforward to split binding code over multiple extension modules, while referencing types that are declared elsewhere. Everything “just” works without any special precautions. One exception to this rule occurs when extending a type declared in another extension module. Recall the basic example from Section Inheritance.

py::class_<Pet> pet(m, "Pet");
pet.def(py::init<const std::string &>())
   .def_readwrite("name", &Pet::name);

py::class_<Dog>(m, "Dog", pet /* <- specify parent */)
    .def(py::init<const std::string &>())
    .def("bark", &Dog::bark);

Suppose now that Pet bindings are defined in a module named basic, whereas the Dog bindings are defined somewhere else. The challenge is of course that the variable pet is not available anymore though it is needed to indicate the inheritance relationship to the constructor of class_<Dog>. However, it can be acquired as follows:

py::object pet = (py::object) py::module::import("basic").attr("Pet");

py::class_<Dog>(m, "Dog", pet)
    .def(py::init<const std::string &>())
    .def("bark", &Dog::bark);

Alternatively, we can rely on the base tag, which performs an automated lookup of the corresponding Python type. However, this also requires invoking the import function once to ensure that the pybind11 binding code of the module basic has been executed.

py::module::import("basic");

py::class_<Dog>(m, "Dog", py::base<Pet>())
    .def(py::init<const std::string &>())
    .def("bark", &Dog::bark);

Naturally, both methods will fail when there are cyclic dependencies.

Note that compiling code which has its default symbol visibility set to hidden (e.g. via the command line flag -fvisibility=hidden on GCC/Clang) can interfere with the ability to access types defined in another extension module. Workarounds include changing the global symbol visibility (not recommended, because it will lead unnecessarily large binaries) or manually exporting types that are accessed by multiple extension modules:

#ifdef _WIN32
#  define EXPORT_TYPE __declspec(dllexport)
#else
#  define EXPORT_TYPE __attribute__ ((visibility("default")))
#endif

class EXPORT_TYPE Dog : public Animal {
    ...
};

Pickling support

Python’s pickle module provides a powerful facility to serialize and de-serialize a Python object graph into a binary data stream. To pickle and unpickle C++ classes using pybind11, two additional functions must be provided. Suppose the class in question has the following signature:

class Pickleable {
public:
    Pickleable(const std::string &value) : m_value(value) { }
    const std::string &value() const { return m_value; }

    void setExtra(int extra) { m_extra = extra; }
    int extra() const { return m_extra; }
private:
    std::string m_value;
    int m_extra = 0;
};

The binding code including the requisite __setstate__ and __getstate__ methods [3] looks as follows:

py::class_<Pickleable>(m, "Pickleable")
    .def(py::init<std::string>())
    .def("value", &Pickleable::value)
    .def("extra", &Pickleable::extra)
    .def("setExtra", &Pickleable::setExtra)
    .def("__getstate__", [](const Pickleable &p) {
        /* Return a tuple that fully encodes the state of the object */
        return py::make_tuple(p.value(), p.extra());
    })
    .def("__setstate__", [](Pickleable &p, py::tuple t) {
        if (t.size() != 2)
            throw std::runtime_error("Invalid state!");

        /* Invoke the in-place constructor. Note that this is needed even
           when the object just has a trivial default constructor */
        new (&p) Pickleable(t[0].cast<std::string>());

        /* Assign any additional state */
        p.setExtra(t[1].cast<int>());
    });

An instance can now be pickled as follows:

try:
    import cPickle as pickle  # Use cPickle on Python 2.7
except ImportError:
    import pickle

p = Pickleable("test_value")
p.setExtra(15)
data = pickle.dumps(p, 2)

Note that only the cPickle module is supported on Python 2.7. The second argument to dumps is also crucial: it selects the pickle protocol version 2, since the older version 1 is not supported. Newer versions are also fine—for instance, specify -1 to always use the latest available version. Beware: failure to follow these instructions will cause important pybind11 memory allocation routines to be skipped during unpickling, which will likely lead to memory corruption and/or segmentation faults.

See also

The file example/example15.cpp contains a complete example that demonstrates how to pickle and unpickle types using pybind11 in more detail.

[3]	http://docs.python.org/3/library/pickle.html#pickling-class-instances

Generating documentation using Sphinx

Sphinx [4] has the ability to inspect the signatures and documentation strings in pybind11-based extension modules to automatically generate beautiful documentation in a variety formats. The python_example repository [5] contains a simple example repository which uses this approach.

There are two potential gotchas when using this approach: first, make sure that the resulting strings do not contain any TAB characters, which break the docstring parsing routines. You may want to use C++11 raw string literals, which are convenient for multi-line comments. Conveniently, any excess indentation will be automatically be removed by Sphinx. However, for this to work, it is important that all lines are indented consistently, i.e.:

// ok
m.def("foo", &foo, R"mydelimiter(
    The foo function

    Parameters
    ----------
)mydelimiter");

// *not ok*
m.def("foo", &foo, R"mydelimiter(The foo function

    Parameters
    ----------
)mydelimiter");

[4]	http://www.sphinx-doc.org

[5]	http://github.com/pybind/python_example