Some Install Feedback

[jkbhagatio]

Hi all, I recently installed on Ubuntu 19.10 and I have some feedback:

1: It looks like the yass readthedocs is deprecated and should probably be deleted, and the link to this in the github repository description should be removed.

2: Following the instructions in the github wiki, I have the following comments;

a) Using anaconda, I had issues sometimes with bash using python and pip when running set-up commands, E.g. bash would point to the default python and pip in usr/lib/ instead of my conda environment's python and pip even after properly initializing conda with bash, and activating the environment. I realized other people have had this issue: https://stackoverflow.com/questions/41060382/using-pip-to-install-packages-to-anaconda-environment, and the solution for me was just specifying the full path to python and pip when running these commands in the terminal. Maybe there could be a troubleshooting section on this wiki page or somewhere else in the wiki where this is mentioned?

b) the instructions currently ask the user to pip install -e the github repo before installing pytorch. This caused issues for me because pip would install some dependencies that conda would try to install when installing pytorch, leading to dependency conflicts.

I would swap the order of these instructions to run all conda installs first before running pip installs: this is also mentioned in Anaconda's best practices.

Additionally, if you are encouraging users to use anaconda, I would suggest streamlining a conda install by creating a yaml file that contains all required dependencies currently listed in setup.py (I don't think there is anything there that is only available on pypi: I think all requirements could also be installed via conda). This could make the process as simple as:

conda env create -f yass.yml
conda activate yass
conda develop <path/to/cloned/repo>

before the instructions for compiling the cuda code.

c) I got a cuda runtime error that gcc versions later than 8 are not supported for cuda toolkit 10.2. May want to mention this in install instructions.

Let me know what you think of these suggestions and if you want any help with this!

[calvinytong]

+1, another advantage of using conda is that you can manage and pin your cuda dependencies. This makes installation a lot easier because you can skip all the cuda configuration