So tests with Python 3.9+ are failing with:

skorch/tests/conftest.py:163: in
import wandb
/opt/hostedtoolcache/Python/3.11.9/x64/lib/python3.11/site-packages/wandb/init.py:27: in
from wandb import sdk as wandb_sdk
/opt/hostedtoolcache/Python/3.11.9/x64/lib/python3.11/site-packages/wandb/sdk/init.py:25: in
from .artifacts.artifact import Artifact
/opt/hostedtoolcache/Python/3.11.9/x64/lib/python3.11/site-packages/wandb/sdk/artifacts/artifact.py:45: in
from wandb import data_types, env, util
/opt/hostedtoolcache/Python/3.11.9/x64/lib/python3.11/site-packages/wandb/data_types.py:32: in
from .sdk.data_types import _dtypes
/opt/hostedtoolcache/Python/3.11.9/x64/lib/python3.11/site-packages/wandb/sdk/data_types/dtypes.py:393: in
NumberType.types.append(np.float)
/opt/hostedtoolcache/Python/3.11.9/x64/lib/python3.11/site-packages/numpy/init.py:397: in getattr
raise AttributeError(
E AttributeError: np.float_ was removed in the NumPy 2.0 release. Use np.float64 instead.

This comes from wandb, which doesn't work with numpy 2.0: https://github.com/wandb/wandb/issues => 7810. There isn't much we can do from the skorch side, but this should only wandb users, who should pin to numpy v1.26.

I'll keep this PR open for now and will re-trigger once wandb releases a fix. If that takes too long, we can pin numpy for our CI.

Update: wandb has released a fix. Now encountering a different error on Windows and MacOS:

RuntimeError: Numpy is not available

PyTorch team has been made aware of the issue: https://github.com/pytorch/pytorch/issues => 128860

Edit: According to this comment, the numpy 2.0 fix for PyTorch will not be backported to PyTorch versions below 2.3. Since we want to test those versions, though, there isn't really a choice but to use numpy < v2.0. As soon as we drop official support for versions below 2.3, we can remove the pin.

The numpy issues still seem to persist, as can for instance be seen here: https://github.com/skorch-dev/skorch/actions/runs/10682583044/job/29744698675?pr=1064. All runs using either Python > 3.8 and PyTorch < 2.3 fail, as they'll use numpy > 2.0, which is incompatible with the respective PyTorch versions.

The numpy issue was gone intermittently. I think that was because one of the dependencies (possibly accelerate) had pinned it. Now that the pin is gone, the issues resurface. But either way, we should not depend on another dependency to do the pinning for us.

ping @ottonemo @thomasjpfan