arraybridge

ArrayBridge provides explicit conversion and shared lifecycle utilities for NumPy, CuPy, PyTorch, TensorFlow, JAX, and pyclesperanto arrays.

Core dependencies are NumPy and metaclass-registry. Other frameworks are optional.

Quick start

import numpy as np

from arraybridge import convert_memory, detect_memory_type

value = np.arange(6).reshape(2, 3)
assert detect_memory_type(value) == "numpy"

copy = convert_memory(
    value,
    source_type="numpy",
    target_type="numpy",
    gpu_id=0,
)

convert_memory requires the declared source type, target type, and device id. The source and target MemoryType declarations perform conversion directly. The device id is required even for CPU conversions so call sites have one stable signature. A GPU target must declare that identifier as available; ArrayBridge does not silently place the value on the CPU.

Declarative decorators

from arraybridge import numpy

@numpy
def normalize(image):
    return image / max(float(image.max()), 1.0)

The framework decorators attach input_memory_type, output_memory_type, and execution_memory_type metadata, provide dtype/slice runtime parameters, and add framework-specific stream/OOM handling where supported. They do not convert inputs or outputs between frameworks and do not accept a gpu_id argument. A host runtime must call convert_memory at the boundary it plans and scope execution using the execution declaration.

Stack utilities

import numpy as np

from arraybridge import stack_slices, unstack_slices

slices = [np.zeros((8, 8)), np.ones((8, 8))]
stack = stack_slices(slices, memory_type="numpy", gpu_id=0)
restored = unstack_slices(stack, memory_type="numpy", gpu_id=0)

stack_slices requires non-empty 2D inputs. unstack_slices requires a 3D array. Both validate shape and use explicit target memory/device declarations.

Installation

pip install arraybridge
pip install "arraybridge[torch]"
pip install "arraybridge[cupy]"

Documentation: https://arraybridge.readthedocs.io

Native array geometry

MemoryType.reshape(array, shape) and MemoryType.broadcast_to(array, shape) operate in the array's framework and device. MemoryType.ones_like(reference, shape=..., dtype=...) allocates on the reference's device; omitting shape retains its shape. Reshape and broadcast preserve views where the backend can represent them. Broadcasting is not a promise of a writable, independent buffer.

These operations never silently project pixels to NumPy. Pyclesperanto cannot provide native changed-shape reshape or broadcasting, so those requests raise NotImplementedError. Its native ones allocation supports one to three dimensions and its supported integer/float32 dtypes; boolean allocation is explicitly unsupported. Existing conversion and disk boundaries remain separate.