Skip to content

fiery.diffeo.backends

Interchangeable backends for the sampling operations.

Every layer and function in this package delegates warping (pull), splatting (push, count) and sampled gradients (grad) to a backend module. Three are available:

  • torch uses torch.nn.functional.grid_sample. It is fast, but it does not implement all the boundary conditions used by the metrics and it approximates splatting with a small-deformation model.
  • interpol uses fiery-interpol, which implements every operator and boundary condition consistently in TorchScript. This is the default backend.
  • jitfields uses jitfields, which implements the same operators in C++/CUDA. It needs cupy and cppyy, so it is an optional dependency.

Pick one per layer with the backend= argument, or for a whole block with the backend context manager.

Attributes

default_backend module-attribute 

default_backend = interpol

Backend used by every layer and function that is not given an explicit backend= argument. Defaults to interpol.

Functions:

backend 

backend(bck)

Temporarily set the default backend.

Parameters:

Name Type Description Default
bck module

Backend to use inside the block. Must be one of the modules under fiery.diffeo.backends.

 required

Yields:

Name Type Description
bck module

The backend that was set.

Examples:

from fiery.diffeo.layers import BCH, Exp
from fiery.diffeo.backends import backend, jitfields

with backend(jitfields):
    layer1 = Exp()
    layer2 = BCH()