Installation
============

Requirements
------------

- Python 3.12–3.13
- `PyTorch <https://pytorch.org/>`_ ≥ 2.0
- `NumPy <https://numpy.org/>`_ ≥ 2.0
- `zarr <https://zarr.readthedocs.io/>`_ ≥ 3.0
- `Cellpose <https://cellpose.readthedocs.io/>`_ ≥ 4.0
- `CAREamics <https://careamics.github.io/>`_ ≥ 0.0.9

Development install
-------------------

Clone the repository and install with `Poetry <https://python-poetry.org/>`_:

.. code-block:: bash

   git clone https://github.com/dv-bt/sphero-vem.git
   cd sphero-vem
   poetry install           # CPU and limited GPU acceleration
   poetry install -E cuda   # with CUDA 12.x GPU acceleration (Linux only)

GPU acceleration
----------------

PyTorch-based stages (registration, segmentation) support GPU execution via
PyTorch's native device management, including CUDA and MPS (Apple Silicon),
with no additional dependencies.

Array operation stages (denoising, shape analysis, spatial analysis) additionally
support CUDA acceleration via CuPy and CuCIM, installed with the ``cuda``
optional dependency group above. Automatic backend switching (NumPy ↔ CuPy,
scikit-image ↔ CuCIM) is handled at import time by ``utils.accelerator``.

Dataset and model weights
--------------------------

The annotated SBF-SEM dataset used to develop and benchmark this pipeline is
available at `BioImage Archive <https://doi.org/10.6019/S-BIAD3263>`_.
Fine-tuned Cellpose-SAM model weights are available on
`Zenodo <https://doi.org/10.5281/zenodo.19616546>`_.