CPU and GPU segmentation
========================

The Omnipose wrapper calls ``cellpose_omni.core.use_gpu()`` to select GPU use.
When a usable CUDA GPU is unavailable, the normal workflow uses the CPU.
Other analysis stages use NumPy, SciPy, image processing, and plotting libraries;
the segmentation GPU setting does not move the entire pipeline onto a GPU.

Check the notebook environment
------------------------------

.. code-block:: python

   import sys
   import torch
   print(sys.executable)
   print(torch.__version__)
   print("CUDA available:", torch.cuda.is_available())

The wrapper prints ``GPU activated?`` when initialized. A first segmentation
run can require a model download. Reusing masks does not require a segmentation
model download.

Optional acceleration
---------------------

Start with the CPU installation, then choose a compatible
PyTorch build for the actual NVIDIA driver and GPU. Consult the official
`PyTorch installer <https://pytorch.org/get-started/locally/>`_ and
`Omnipose GPU guidance <https://omnipose.readthedocs.io/installation.html>`_.

An optional Windows setup uses Python 3.10, an RTX A6000, PyTorch
``2.7.1+cu118`` and torchvision ``0.22.1+cu118``. Select versions compatible
with your own GPU and driver.
Restart the notebook kernel after changing PyTorch.

If CUDA is unavailable, check the interpreter, installed PyTorch build, and
driver compatibility. A ``False`` CUDA check by itself does not prevent CPU
segmentation. See :doc:`../workflow/segmentation` for the actual wrapper settings.