Closed jkterry1 closed 2 years ago
I'm happy to specify a higher version but I don't think I agree with this, random versions of numpy can cause serious speed regressions out of nowhere.
I've used numpy for quite about a decade now and I'm the current maintainer of both gym and pettingzoo and have never seen this before, do you like have a link for that I could see?
Regardless though, no other major python package I'm aware of specifies a precise version of numpy in their requirements, which is inherently going to cause installation issues, and different versions of numpy definitely don't have changes in what the algorithms do or major breaking API changes, which is the real point of what requirements specify anyways.
As an example of a regression that has bitten me before: https://github.com/numpy/numpy/issues/14281
Sure, but specifying to avoid these performance issues along still isn't the standard for requirements in my experience and still results in issues with installation alongside other packages if you're installing via the specified requirements. Most packages depend on ranges of pytorch and tensorflow the version of which can have absolutely profound impacts on performance.
If there's a particular package this is causing incompatibility with, I'm happy to upgrade the version number!
Right now requirements.txt specifies an exact version of numpy. There's no reasonable way this is necessary and causes conflicts when installing in systems