Compiling Python classes with @jitclass


This is a early version of jitclass support. Not all compiling features are exposed or implemented, yet.

Numba supports code generation for classes via the numba.experimental.jitclass() decorator. A class can be marked for optimization using this decorator along with a specification of the types of each field. We call the resulting class object a jitclass. All methods of a jitclass are compiled into nopython functions. The data of a jitclass instance is allocated on the heap as a C-compatible structure so that any compiled functions can have direct access to the underlying data, bypassing the interpreter.

Basic usage

Here’s an example of a jitclass:

import numpy as np
from numba import int32, float32    # import the types
from numba.experimental import jitclass

spec = [
    ('value', int32),               # a simple scalar field
    ('array', float32[:]),          # an array field

class Bag(object):
    def __init__(self, value):
        self.value = value
        self.array = np.zeros(value, dtype=np.float32)

    def size(self):
        return self.array.size

    def increment(self, val):
        for i in range(self.size):
            self.array[i] += val
        return self.array

    def add(x, y):
        return x + y

n = 21
mybag = Bag(n)

In the above example, a spec is provided as a list of 2-tuples. The tuples contain the name of the field and the Numba type of the field. Alternatively, user can use a dictionary (an OrderedDict preferably for stable field ordering), which maps field names to types.

The definition of the class requires at least a __init__ method for initializing each defined fields. Uninitialized fields contains garbage data. Methods and properties (getters and setters only) can be defined. They will be automatically compiled.

Inferred class member types from type annotations with as_numba_type

Fields of a jitclass can also be inferred from Python type annotations.

from typing import List
from numba.experimental import jitclass
from numba.typed import List as NumbaList

class Counter:
    value: int

    def __init__(self):
        self.value = 0

    def get(self) -> int:
        ret = self.value
        self.value += 1
        return ret

class ListLoopIterator:
    counter: Counter
    items: List[float]

    def __init__(self, items: List[float]):
        self.items = items
        self.counter = Counter()

    def get(self) -> float:
        idx = self.counter.get() % len(self.items)
        return self.items[idx]

items = NumbaList([3.14, 2.718, 0.123, -4.])
loop_itr = ListLoopIterator(items)

Any type annotations on the class will be used to extend the spec if that field is not already present. The Numba type corresponding to the given Python type is inferred using as_numba_type. For example, if we have the class

@jitclass([("w", int32), ("y", float64[:])])
class Foo:
    w: int
    x: float
    y: np.ndarray
    z: SomeOtherType

    def __init__(self, w: int, x: float, y: np.ndarray, z: SomeOtherType):

then the full spec used for Foo will be:

  • "w": int32 (specified in the spec)

  • "x": float64 (added from type annotation)

  • "y": array(float64, 1d, A) (specified in the spec)

  • "z": numba.as_numba_type(SomeOtherType) (added from type annotation)

Here SomeOtherType could be any supported Python type (e.g. bool, typing.Dict[int, typing.Tuple[float, float]], or another jitclass).

Note that only type annotations on the class will be used to infer spec elements. Method type annotations (e.g. those of __init__ above) are ignored.

Numba requires knowing the dtype and rank of NumPy arrays, which cannot currently be expressed with type annotations. Because of this, NumPy arrays need to be included in the spec explicitly.

Specifying numba.typed containers as class members explicitly

The following patterns demonstrate how to specify a numba.typed.Dict or numba.typed.List explicitly as part of the spec passed to jitclass.

First, using explicit Numba types and explicit construction.

from numba import types, typed
from numba.experimental import jitclass

# key and value types
kv_ty = (types.int64, types.unicode_type)

# A container class with:
# * member 'd' holding a typed dictionary of int64 -> unicode string (kv_ty)
# * member 'l' holding a typed list of float64
@jitclass([('d', types.DictType(*kv_ty)),
           ('l', types.ListType(types.float64))])
class ContainerHolder(object):
    def __init__(self):
        # initialize the containers
        self.d = typed.Dict.empty(*kv_ty)
        self.l = typed.List.empty_list(types.float64)

container = ContainerHolder()
container.d[1] = "apple"
container.d[2] = "orange"
print(container.d) # {1: apple, 2: orange}
print(container.l) # [123.0, 456.0]

Another useful pattern is to use the numba.typed container attribute _numba_type_ to find the type of a container, this can be accessed directly from an instance of the container in the Python interpreter. The same information can be obtained by calling numba.typeof() on the instance. For example:

from numba import typed, typeof
from numba.experimental import jitclass

d = typed.Dict()
d[1] = "apple"
d[2] = "orange"
l = typed.List()

@jitclass([('d', typeof(d)), ('l', typeof(l))])
class ContainerInstHolder(object):
    def __init__(self, dict_inst, list_inst):
        self.d = dict_inst
        self.l = list_inst

container = ContainerInstHolder(d, l)
print(container.d) # {1: apple, 2: orange}
print(container.l) # [123.0, 456.0]

It is worth noting that the instance of the container in a jitclass must be initialized before use, for example, this will cause an invalid memory access as self.d is written to without d being initialized as a type.Dict instance of the type specified.

from numba import types
from numba.experimental import jitclass

dict_ty = types.DictType(types.int64, types.unicode_type)

@jitclass([('d', dict_ty)])
class NotInitialisingContainer(object):
    def __init__(self):
        self.d[10] = "apple" # this is invalid, `d` is not initialized

NotInitialisingContainer() # segmentation fault/memory access violation

Support operations

The following operations of jitclasses work in both the interpreter and Numba compiled functions:

  • calling the jitclass class object to construct a new instance (e.g. mybag = Bag(123));

  • read/write access to attributes and properties (e.g. mybag.value);

  • calling methods (e.g. mybag.increment(3));

  • calling static methods as instance attributes (e.g. mybag.add(1, 1));

  • calling static methods as class attributes (e.g. Bag.add(1, 2));

  • using select dunder methods (e.g. __add__ with mybag + otherbag);

Using jitclasses in Numba compiled function is more efficient. Short methods can be inlined (at the discretion of LLVM inliner). Attributes access are simply reading from a C structure. Using jitclasses from the interpreter has the same overhead of calling any Numba compiled function from the interpreter. Arguments and return values must be unboxed or boxed between Python objects and native representation. Values encapsulated by a jitclass does not get boxed into Python object when the jitclass instance is handed to the interpreter. It is during attribute access to the field values that they are boxed. Calling static methods as class attributes is only supported outside of the class definition (i.e. code cannot call Bag.add() from within another method of Bag).

Supported dunder methods

The following dunder methods may be defined for jitclasses:

  • __abs__

  • __bool__

  • __complex__

  • __contains__

  • __float__

  • __getitem__

  • __hash__

  • __index__

  • __int__

  • __len__

  • __setitem__

  • __str__

  • __eq__

  • __ne__

  • __ge__

  • __gt__

  • __le__

  • __lt__

  • __add__

  • __floordiv__

  • __lshift__

  • __matmul__

  • __mod__

  • __mul__

  • __neg__

  • __pos__

  • __pow__

  • __rshift__

  • __sub__

  • __truediv__

  • __and__

  • __or__

  • __xor__

  • __iadd__

  • __ifloordiv__

  • __ilshift__

  • __imatmul__

  • __imod__

  • __imul__

  • __ipow__

  • __irshift__

  • __isub__

  • __itruediv__

  • __iand__

  • __ior__

  • __ixor__

  • __radd__

  • __rfloordiv__

  • __rlshift__

  • __rmatmul__

  • __rmod__

  • __rmul__

  • __rpow__

  • __rrshift__

  • __rsub__

  • __rtruediv__

  • __rand__

  • __ror__

  • __rxor__

Refer to the Python Data Model documentation for descriptions of these methods.


  • A jitclass class object is treated as a function (the constructor) inside a Numba compiled function.

  • isinstance() only works in the interpreter.

  • Manipulating jitclass instances in the interpreter is not optimized, yet.

  • Support for jitclasses are available on CPU only. (Note: Support for GPU devices is planned for a future release.)

The decorator: @jitclass

numba.experimental.jitclass(cls_or_spec=None, spec=None)

A function for creating a jitclass. Can be used as a decorator or function.

Different use cases will cause different arguments to be set.

If specified, spec gives the types of class fields. It must be a dictionary or sequence. With a dictionary, use collections.OrderedDict for stable ordering. With a sequence, it must contain 2-tuples of (fieldname, fieldtype).

Any class annotations for field names not listed in spec will be added. For class annotation x: T we will append ("x", as_numba_type(T)) to the spec if x is not already a key in spec.

If used as a decorator, returns a callable that takes a class object and
returns a compiled version.
If used as a function, returns the compiled class (an instance of


  1. cls_or_spec = None, spec = None

>>> @jitclass()
... class Foo:
...     ...
  1. cls_or_spec = None, spec = spec

>>> @jitclass(spec=spec)
... class Foo:
...     ...
  1. cls_or_spec = Foo, spec = None

>>> @jitclass
... class Foo:
...     ...

4) cls_or_spec = spec, spec = None In this case we update cls_or_spec, spec = None, cls_or_spec.

>>> @jitclass(spec)
... class Foo:
...     ...
  1. cls_or_spec = Foo, spec = spec

>>> JitFoo = jitclass(Foo, spec)