It’s not the researcher’s fault if the libraries they use make breaking changes after a month; proof-of-concept code published with a paper is supposed to be static, and there’s often no incentive for the researcher to maintain it after publication.
At this point, venvs are the best workaround, but we can still wish for something better. As someone commented further up, being able to “import pytorch==2.0” and have multiple library versions coexist would go a long way.
Comments
It’s not the researcher’s fault if the libraries they use make breaking changes after a month; proof-of-concept code published with a paper is supposed to be static, and there’s often no incentive for the researcher to maintain it after publication.
At this point, venvs are the best workaround, but we can still wish for something better. As someone commented further up, being able to “import pytorch==2.0” and have multiple library versions coexist would go a long way.