Open racinmat opened 5 years ago
setup.py should depend on just tensorflow
, not tensorflow-gpu
unless you add the --gpu
option, so I'm not sure why that happened. Do you have suggestions for how we should change our setup.py to better support locally-installed tensorflow versions?
I compiled Tensorflow 1.12 from source on Windows, so I have Tensorflow qwith CUDA support, but in PIP, it is still called only tensorflow, not tensorflow-gpu. When installing Magenta by cloning it and calling
python setup.py develop
, it installed also the tensorflow-gpu, with older version, which completely broke my environment. I think it should detect if I have tensorflow with CUDA support even without the tensorflow-gpu package and not install it then.