Support Numpy 2.0

[MalteEbner]

Error using lightly with numpy 2.0

Numpy 2.0.0 was released on June 16th 2024 and does not work with lightly.

Using numpy==2.0.0 and lightly==1.5.6 and python==3.10.8:

(.venv) user@device repo % python -c "import lightly"                           
Traceback (most recent call last):
  File "<string>", line 1, in <module>
  File "/Users/malteebnerlightly/Documents/GitHub/lightspeed/.venv/lib/python3.10/site-packages/lightly/__init__.py", line 88, in <module>
    from lightly.api import _version_checking
  File "/Users/malteebnerlightly/Documents/GitHub/lightspeed/.venv/lib/python3.10/site-packages/lightly/api/__init__.py", line 7, in <module>
    from lightly.api.api_workflow_client import ApiWorkflowClient
  File "/Users/malteebnerlightly/Documents/GitHub/lightspeed/.venv/lib/python3.10/site-packages/lightly/api/api_workflow_client.py", line 21, in <module>
    from lightly.api.api_workflow_selection import _SelectionMixin
  File "/Users/malteebnerlightly/Documents/GitHub/lightspeed/.venv/lib/python3.10/site-packages/lightly/api/api_workflow_selection.py", line 30, in <module>
    class _SelectionMixin:
  File "/Users/malteebnerlightly/Documents/GitHub/lightspeed/.venv/lib/python3.10/site-packages/lightly/api/api_workflow_selection.py", line 32, in _SelectionMixin
    self, al_scores: Dict[str, NDArray[np.float_]], query_tag_id: str
  File "/Users/malteebnerlightly/Documents/GitHub/lightspeed/.venv/lib/python3.10/site-packages/numpy/__init__.py", line 397, in __getattr__
    raise AttributeError(
AttributeError: `np.float_` was removed in the NumPy 2.0 release. Use `np.float64` instead.. Did you mean: 'float16'?

Reproduce

Run in a new venv, such that numpy 2.0 is installed:

pip install lightly 
python -c "import lightly"  

Workaround

Just install an older numpy version, e.g. using pip install numpy==1.26.

[philippmwirth]

It looks like torchvision 0.19 will be compatible with numpy 2.0 (source: https://github.com/pytorch/vision/issues/8460#issuecomment-2145217340). Let's require numpy 1.x for now.

[adamjstewart]

Why can't we support both numpy 1 and 2 as soon as possible? Why do we need to wait for other dependencies? Seems like it's a simple fix.

[philippmwirth]

Hey @adamjstewart, thanks for the feedback. We can already migrate the library to NumPy 2.0 as in #1561 but the problem of broken upstream libraries remains. In particular, torchvision 0.18.0 is only 98% compatible (see the tracker issue https://github.com/numpy/numpy/issues/26191) and it appears that we are affected by the remaining two percent: Applying color jitter results in the following error (failing GitHub action):

    transformed: Tensor = self.transform(image)
/opt/hostedtoolcache/Python/3.10.14/x64/lib/python3.10/site-packages/torchvision/transforms/transforms.py:95: in __call__
    img = t(img)
/opt/hostedtoolcache/Python/3.10.14/x64/lib/python3.10/site-packages/torch/nn/modules/module.py:1532: in _wrapped_call_impl
    return self._call_impl(*args, **kwargs)
/opt/hostedtoolcache/Python/3.10.14/x64/lib/python3.10/site-packages/torch/nn/modules/module.py:1541: in _call_impl
    return forward_call(*args, **kwargs)
/opt/hostedtoolcache/Python/3.10.14/x64/lib/python3.10/site-packages/torchvision/transforms/transforms.py:540: in forward
    img = t(img)
/opt/hostedtoolcache/Python/3.10.14/x64/lib/python3.10/site-packages/torch/nn/modules/module.py:1532: in _wrapped_call_impl
    return self._call_impl(*args, **kwargs)
/opt/hostedtoolcache/Python/3.10.14/x64/lib/python3.10/site-packages/torch/nn/modules/module.py:1541: in _call_impl
    return forward_call(*args, **kwargs)
/opt/hostedtoolcache/Python/3.10.14/x64/lib/python3.10/site-packages/torchvision/transforms/transforms.py:1280: in forward
    img = F.adjust_hue(img, hue_factor)
/opt/hostedtoolcache/Python/3.10.14/x64/lib/python3.10/site-packages/torchvision/transforms/functional.py:959: in adjust_hue
    return F_pil.adjust_hue(img, hue_factor)
_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ 

img = <PIL.Image.Image image mode=RGB size=100x100 at 0x7FAC861DADA0>
hue_factor = -0.020240534096956253

    @torch.jit.unused
    def adjust_hue(img: Image.Image, hue_factor: float) -> Image.Image:
        if not (-0.5 <= hue_factor <= 0.5):
            raise ValueError(f"hue_factor ({hue_factor}) is not in [-0.5, 0.5].")

        if not _is_pil_image(img):
            raise TypeError(f"img should be PIL Image. Got {type(img)}")

        input_mode = img.mode
        if input_mode in {"L", "1", "I", "F"}:
            return img

        h, s, v = img.convert("HSV").split()

        np_h = np.array(h, dtype=np.uint8)
        # uint8 addition take cares of rotation across boundaries
        with np.errstate(over="ignore"):
>           np_h += np.uint8(hue_factor * 255)
E           OverflowError: Python integer -5 out of bounds for uint8

/opt/hostedtoolcache/Python/3.10.14/x64/lib/python3.10/site-packages/torchvision/transforms/_functional_pil.py:114: OverflowError

[adamjstewart]

It's not possible for your tests to pass without waiting for torchvision, but it is possible for the next release to be compatible with numpy 2 (futureproof). This way, as soon as torchvision makes a new release, the last month of lightly releases will already immediately support numpy 2.

[philippmwirth]

We appreciate your input. However, relaxing the numpy version requirement again will lead to more broken installations, which we want to avoid. We've chosen to keep numpy constrained to <2.0 for the time being.

[MalteEbner]

For the moment, there is no torchvision version compatible with numpy 2 thus we restricted the requirements to not use numpy 2: https://github.com/lightly-ai/lightly/pull/1562

Once a compatible torchvision version is released, we can have a look again.