Multiple install-issues under windows and linux

[rnbwdsh]

After hours on tinkering around on (arch) linux and not getting it to run with miniconda and optirun (running code on an intel-nvidia-laptop-gpu) I gave up after 3 hours and tried to get it running on windows.

So far I've got multiple remarks:

• For me following your instructions, I end up with tensorflow 1.10, not 1.1 as you state in the readme.txt Same with keras 2.07 vs keras 2.06 -> Typo?
• On windows it's just activate chemvae without source
• You have a jupyter-example, but you don't require jupyter in the environment.yml, so even if it's installed natively it doesn't work
• Also the train_vae NEEDS a gpu-device, yet you don't install tensorflow-gpu in the environment.yml (but in the requirements.txt, which adds to the confusion) - i found two ways that also fix the requirements for cuda+cudnn
  • This is fixed by removing tensorflow fully from the requirements and after installing, running conda install tensorflow-gpu -> this leads to theano as default backend -> must change C:\Users\<username>\Anaconda3\envs\chemvae\etc\conda\activate.d\keras_activate.bat to ackend=tensorflow.
  • Alternative solution: Changing it to tensorflow-gpu and running pip install --upgrade tensorflow-gpu==1.10 after install - otherwise it's not properly recognizing my gpu-device. This also requires editing the keras_activate.bat.

How to reproduce: Install the visual studio 2015 build tools on a fresh install of windows 10, then run the install script and then the python -m chemvae.train_vae