ArrayBackend

class odl.ArrayBackend(impl: str, array_namespace: ModuleType, available_dtypes: dict[str, object], array_type: type, array_constructor: Callable, from_dlpack: Callable, make_contiguous: Callable, identifier_of_dtype: Callable, available_devices: list[str], to_cpu: Callable, to_numpy: Callable, to_device: Callable)

Bases: object

Class to implement the array backend associated to each TensorSpace Implementations.

Attributes

implstr

The implementation of the backend, e.g ‘numpy’

array_namespaceModuleType

The actual namespace of the backend, e.g np

available_dtypesdict

A dictionnary mapping a Number/str datatype to the corresponding backend-specific datatype, e.g {float:np.float64, ‘float64’, np.float64, …}

array_typetype

The type of the array once implemented by the backend, e.g np.ndarray

array_constructorCallable

The function the backend uses to create an array, e.g np.asarray

from_dlpackCallable

Stand-in for the from_dlpack method of the Python Array API. We would rather use that directly, but there are multiple inconsistencies of its behaviour particular in different PyTorch versions, so we need to wrap it as a workaround.

make_contiguousCallable

The function the backend uses to make an array contiguous, e.g np.ascontiguousasarray

identifier_of_dtypeCallable

The function used to get a string representation of a backend-specific dtype

available_deviceslist[str]

List of devices accepted by the backend

to_cpuCallable

Function to copy an array to the CPU

to_numpy: Callable

Function to create a Numpy version of an array

__init__(impl: str, array_namespace: ModuleType, available_dtypes: dict[str, object], array_type: type, array_constructor: Callable, from_dlpack: Callable, make_contiguous: Callable, identifier_of_dtype: Callable, available_devices: list[str], to_cpu: Callable, to_numpy: Callable, to_device: Callable) → None