如果存在软件包依赖项冲突，如何停止pip安装？

Question

I did pip install tensorflow-gpu==1.12.3 because my GPU can only support CUDA 9, while the latest tensorflow-gpu==1.14.0 requires a minimum of CUDA 10. One of the package dependencies is tensorboard .

I want to do pip install tensorflow . However, the latest version warns that a newer version of tensorboard must be installed, then automatically proceeded to uninstall the existing version and install the newer one. Is there a way to have pip halt or at least ask for confirmation instead of proceeding on its own in such a situation? I'm trying to figure out which is a compatible version of tensorflow that doesn't require newer versions of package dependencies shared with tensorflow-gpu .

Edit

ok just to clarify, this isn't a tensorflow-specific question, I'm just using it as an example because I just happen to be doing that right now.

To generalise: I installed an older version of package A, which has a dependency on an older version of package B. I want to install package C, which also has a dependency on package B. However, the latest version of C requires the latest version of B. When I try to install C, it automatically uninstalled the older B and installed the newest version. How do I stop it from doing that?

Right now my only workaround is to install a version of C, wait for it to automatically uninstall something, then interrupt the installation of C, and reinstall the version of the dependency package that was uninstalled. Then I step down to the next older version C, and repeat. Basically, a trial and error process.

Is there a more proper and elegant solution? I'm fine with it giving a warning that C requires a newer version of B and then terminating at that point, or prompting for a confirmation before proceeding to uninstall, instead of doing it automatically without asking.

I'm using PyCharm's venv, so alternatively I'm also fine if there's a way to revert the latest change to the environment ie uninstall C and whatever new dependency packages that's installed, reverting to the list of installed packages before I executed pip install C .

Hopefully that doesn't sound too convoluted. 😅

Answer 1

I have never done any work with tensorflow so I can't help you with that part, but I can suggest you using virtual environments when testing out stuff and not risking breaking system Python or upgrading packages you want at specific version.

python -m venv env to create virtual environment called env . source env/bin/activate to make the virtual environment active and deactivate to leave it.

Regarding the overwriting old packages, I haven't found options for promting installing new version nor an option to ignore requirements for a certain package. This is probably because developers can specify which version of dependencies are required and thus if pip finds outdated package, the package you wish to instal would probably not be usable with it.

However, you might try installing the newest version of tensorflow and let it install requirements it needs and then simply uninstall them using pip uninstall PACKAGE-NAME and install the version you wish with pip install PACKAGE-NAME==xxx .

Also, here is a guide that explains installation of tensorflow with CUDA 9 . It seems that they are using version 1.8 for CUDA 9.2 and 1.7 for CUDA 9.1.