conda env genreation

[oeway]

Currently it seems there are some packages produced by get_conda_env are not necessary or unfit for different cases? Can we provide different flavors of it maybe?

from bioimageio.spec import ValidationSummary, save_bioimageio_package, load_model_description
from bioimageio.core import test_model

from bioimageio.spec.get_conda_env import get_conda_env
descr = load_model_description("emotional-cricket")
wf = descr.weights.pytorch_state_dict

assert wf is not None

conda_env_spec = get_conda_env(entry=wf, env_name="DROP")
print(conda_env_spec.model_dump(mode="json", exclude_defaults=True))

{'channels': ['pytorch', 'conda-forge', 'nodefaults'], 'dependencies': ['cpuonly', 'mkl ==2024.0.0', 'numpy <2', 'pip', 'pytorch==1.12.1', 'setuptools <70.0.0', 'torchaudio==0.12.1', 'torchvision==0.13.1']}

In the above example, I would maybe skip 'cpuonly' (because we will sometimes run it on gpu), 'numpy', 'pip' and 'setuptools' can be excluded as well.

[FynnBe]

• cpuonly yes, I suppose we should just change to gpu envs (that should be ale run on cpu as well). If that's not sufficient we need to provide two flavors.. let's see if we can avoid that.
• numpy is there for the pin <2 for pytorch < 2.3, for higher pytorch versions it will not be pinned
• setuptools is pinned for pytorch < 2.2 to avoid this issue

So for now, I'll just remove cpuonly (see #673 )