Custom data types in plugins

Last modified: September 16, 2022



For common datatypes like Float or PointVector, the Gamera C++/Python argument passing interface provides an automatic wrapping mechanism that is described in Specifying arguments for plugin generation and dialog boxes.

Occasionally, these built in data types are not sufficient, and you need to pass custom types. A typical situation is that your return type is a tuple.

An example

Assume we need a plugin grey_stats that returns both the mean grey value and its variance on a greyscale image. As the return value is a tuple of two values, we cannot use one of the built in data types.

We therefore use the generic argument type Class in the Python wrapper for our plugin. Note that it is a good idea to give more information about the actual return value(s) in the documentation string:

class grey_stats(PluginFunction):
   """Returns mean and variance of the grey values as a tuple *(m,var)*"""
   self_type = ImageType([GREYSCALE])
   return_type = Class("m_var")

On the C++ side, the return type is PyObject* and can be constructed with any of the conversion functions from the Python C API, e.g. Py_BuildValue:

template<class T>
  PyObject* grey_stats(const T &img) {
  size_t i;
  double Ex = 0.0;  // E(X)
  double Ex2 = 0.0; // E(X^2)
  FloatVector* hist = histogram(img);
  for (i=0; i < hist->size(); i++) {
    Ex  += i * (*hist)[i];
    Ex2 += i*i * (*hist)[i];
  delete hist;
  return Py_BuildValue("ff", Ex, Ex2 - Ex*Ex);

In a Python script, the plugin can be called with

(m,v) = img.grey_stats()
# or:
m,v = img.grey_stats()

It is also possible to call the plugin from the C++ side, by simply converting the PyObject* return value back to C++ data types:

double m,v;
PyObject* obj;
obj = grey_stats(img);
PyArg_ParseTuple(obj, "ff", &m, &v);

The last line is necessary to avoid a memory leak. The Py_BuildValue function returns an "owned reference" (in Python lingo), which means that it has increased its reference count. To allow the PyObject to be deleted by the Python runtime library, its reference count must be decreased again. The actual deletion of the PyObject will then occur at some point in the future by the Python garbage collector.

Receiving a custom type in C++

When you have a PyObject* as an input argument of a plugin function, you can convert it to C++ data types with the conversion functions from the Python C API. These are

  • PyLong_AsLong and colleagues for ordinary scalar data types
  • PyArg_ParseTuple for tuples
  • PyList_GetItem for lists
  • PyObject_GetAttrString for properties of arbitrary classes

A fundamental problem with the conversion from PyObjects is that you can never be sure what actually is in the PyObject*. For instance, when you expect a list and apply PyList_GetItem, you cannot be sure that the PyObject* actually is a list. Nor can you safely assume that the list entries are of the data type you believe them to be. It is thus generally a good idea to check data types with PyList_Check and PyObject_TypeCheck, and to throw an exception when the wrong data type enters your function, e.g.

PyObject* myplugin(PyObject* list) {
    throw std::runtime_error("myplugin: Input argument is no list.");
  // ...

Returning a custom type from C++

To return a PyObject* from a plugin function, you can use the conversion functions from the Python C API. These are

  • Py_BuildValue for tuples or ordinary scalar values; as for all common scalar data types the Gamera argument wrapping can be used, calling Py_BuildValue for scalar types is usually only necessary for creating PyObjects for passing to returned lists or custom class properties.
  • PyList_New and PyList_SetItem for lists
  • PyInstance_New and PyObject_SetAttrString for arbitrary classes

PyObjects returned from a plugin function must always be "owned references". As both Py_Build_Value and PyList_New create owned references, this is usually automatically fulfilled. There is however one special case when you read a list entry from an input list with PyList_GetItem and write it to a return list with PyList_SetItem. As PyList_GetItem yields a "borrowed reference", its reference count must be increased if it is to be returned to Python. Here is an example that returns half of the input sequence and fills the rest with zeros:

PyObject* myplugin(PyObject* list) {
  size_t n,i;
  PyObject *retval, *entry;
  n = PyList_Size(list);
  retval = PyList_New(n);
  for (i=0; i<n/2; i++) {
    // PyList_GetItem returns a borrowed reference
    // => reference count must be manually increased (Py_SetItem does not do so)
    entry = PyList_GetItem(list, i);
    PyList_SetItem(retval, i, entry);
  for (i=n/2; i<n; i++) {
    // Py_BuildValue returns an "owned reference"
    // => no Py_INCREF necessary
    entry = Py_BuildValue("i", 0);
    PyList_SetItem(retval, i, entry);
  return retval;

It is important to note that, unlike PyList_SetItem, PyList_Append adds a reference to the added item, so that Py_INCREF must not be called in this case. In contrast, when you add an already owned reference to a list with PyList_Append, you must instead decrease its reference count with Py_DECREF:

PyObject* myplugin(PyObject* list) {
  size_t n,i;
  PyObject *retval, *entry;
  n = PyList_Size(list);
  retval = PyList_New(0);
  for (i=0; i<n/2; i++) {
    // PyList_Append already adds a reference to the added item
    PyList_Append(retval, PyList_GetItem(list, i));
  for (i=n/2; i<n; i++) {
    // Py_BuildValue returns an "owned reference" whose reference count
    // is again increased by PyList_Append => Py_DECREF necessary
    entry = Py_BuildValue("i", 0);
    PyList_Append(retval, entry);
  return retval;

Implementing a Python class in C++

Occasionally it can become necessary to write some methods of a custom Python class in C++. Let us assume, you want to implement the constructor of a custom class MyClass in C++.

The basic idea is to write a plugin function create_myclass that returns an object of type MyClass and to call this function in the Python constructor, i.e. in the __init__ method. This requires a Python wrapper of the form:

# the class definition
class MyClass:
  def __init__(self, arg1, arg2):
    mc = _myplugins.create_myclass(arg1, arg2)
    # copy over properties from mc to self
    self.property1 = mc.property1
    self.property2 = mc.property2
    # ...

# the plugin implementing the actual constructor of MyClass
class create_myclass(PluginFunction):
  self_type = None
  args = Args([Int("arg1"), Real("arg2")])
  return_type = Class('myclass', MyClass)

The C++ side of the create_myclass plugin will then be of the form (this primitive example simply passes the input arguments unaltered to the new class as properties, so it is not very useful, but nevertheless a nice demonstration of the main ideas):

PyObject* create_myclass(int arg1, double arg2) {
  // helper variable for temporary property storage
  PyObject* prop;
  // helper object for creating class instances (see below)
  // declared static so this is initialized only once
  static PyObject* my_class = NULL;

  // create a dictionary and store the properties therein
  // Note that PyDict_SetItemString (unlike PyList_SetItem) INCREFs the passed object
  // => the no longer needed reference returned by Py_BuildValue must be DECREFed
  PyObject* class_dict = PyDict_New();
  prop = Py_BuildValue("i", arg1);
  PyDict_SetItemString(class_dict, "property1", prop);
  prop = Py_BuildValue("f", arg2);
  PyDict_SetItemString(class_dict, "property2", prop);

  // create an instance of MyClass from the dictionary
  if (my_class == NULL) {
    my_class = PyClass_New(NULL, PyDict_New(), PyString_FromString("MyClass"));
  PyObject* ret = PyInstance_NewRaw(my_class, class_dict);
  return ret;