DirectML support

[Teranis]

See #1169 for details

Adds functions to implement torch-directml. This library uses Directx12 to accelerate PyTorch, which is slower than CUDA but has the benefit of being compatible with any Directx12 device on Windows, including AMD GPUs.

Also includes a working example!

[mrariden]

@Teranis This looks good! Can you also add some documentation in the docs/installation.rst file?

[Teranis]

Hi, I modified the cellpose code so that it tries to use DirectML if CUDA and MPS are not available. I tried to follow the existing logic for it. I also included installation instructions (which are now a lot easier than before.

[codecov]

Codecov Report

Attention: Patch coverage is 40.00000% with 24 lines in your changes missing coverage. Please review.

Project coverage is 47.08%. Comparing base (682a1de) to head (e08b47f).
Report is 46 commits behind head on main.

Files with missing lines	Patch %	Lines
cellpose/core.py	23.80%	16 Missing ⚠️
cellpose/models.py	58.82%	7 Missing ⚠️
cellpose/dynamics.py	50.00%	1 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #1181      +/-   ##
==========================================
- Coverage   47.13%   47.08%   -0.05%     
==========================================
  Files          16       16              
  Lines        3711     3761      +50     
==========================================
+ Hits         1749     1771      +22     
- Misses       1962     1990      +28     

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[Teranis]

Hi @mrariden ,
I have added a unit test for core.assign_device. I'm not sure how I would realistically cover the other functionality. I would run everything, but with a different fixture as the model:

@pytest.fixture()
def cellposemodel_fixture_24layer_directml():
    """ This is functionally identical to CellposeModel but uses mock class """
    gpu = True
    with patch('torch.cuda.is_available', return_value=False), \
            patch('torch.backends.mps.is_available', return_value=False):
        model = MockCellposeModel(24, gpu=gpu)
    yield model

Then, I would compare the model output to the ground truth. I'm confused about how to do these last steps or why you strictly stick to CLI.

Kind regards,
Timon Stegmaier

[Teranis]

Hi @mrariden ,
I was wondering if there is an update, or if I need to do anything else.
Kind regards,
Timon Stegmaier