Array Backends

SGN-TS abstracts array operations behind an ArrayBackend interface. The default is NumpyBackend. When PyTorch is installed, TorchBackend enables GPU-accelerated operations. Both backends expose the same set of common operations (zeros, ones, cat, pad, matmul, etc.), so element code works identically regardless of which backend is active.

Using NumpyBackend

NumpyBackend is always available and is the default:

from sgnts.base import NumpyBackend

backend = NumpyBackend

z = backend.zeros((2, 1024))
o = backend.ones((2, 1024))
r = backend.arange(1024)

Using TorchBackend

Install PyTorch (pip install sgn-ts[torch]), then:

from sgnts.base import TorchBackend

backend = TorchBackend

z = backend.zeros((2, 1024))  # torch.Tensor

If PyTorch is not installed, importing TorchBackend still works but raises an error when any operation is called.

Setting the Backend in AdapterConfig

Elements that use AdapterConfig can specify which backend to use for internal buffer operations:

from sgnts.base import AdapterConfig, TorchBackend

config = AdapterConfig(backend=TorchBackend)

Converting Between Backends

Use the Converter transform to convert data between backends mid-pipeline:

from sgnts.transforms import Converter

# Convert from NumPy to PyTorch
conv = Converter(name="to_torch", backend=TorchBackend)

Creating a Custom Backend

Subclass ArrayBackend and implement all required class methods:

from sgnts.base import ArrayBackend

class MyBackend(ArrayBackend):
    @staticmethod
    def zeros(shape):
        ...

    @staticmethod
    def ones(shape):
        ...

    # ... implement all required methods