Alternative dependency recommendations

[ajleite]

Thank you for this very exciting work! I am looking forward to putting this meta-learning algorithm through its paces on a wide variety of environments.

I wanted to share my experience getting this program set up here in case it might be helpful for anybody else.

On my Ubuntu 20.10 system, for which no gcc versions below 7 are available, it was nearly impossible to compile the plasma tensorflow ops for the tf 1.13 binaries because tf 1.13 and 1.14 are compiled with gcc 4.8, which is now relatively ancient. So I wanted to use tf 1.15 instead.

However, ray 0.7.7 doesn't play well with tf 1.15 -- ironically, ray 0.7.6 performs better. (ref: ray-project/ray#8729)

I was able to get your code running by using the following alternative versions:
tensorflow-gpu==1.15.5
ray==0.7.6

(Also -- I needed to compile Python from source to get Python 3.7, which is the last version that supported tf 1.x, running on Ubuntu 20.04. That wasn't unexpected or challenging for me, but I imagine some of our less technically-inclined peers, particularly in my home discipline of cognitive science, might have issues with it or at least want to be pointed in that direction. In order to get tensorboard working, I needed to install libsqlite-dev in addition to the required libraries before building Python.)

Mr. Kirsch, I suspect these steps will only get harder for people over time. Even though docker has a number of disadvantages in terms of integrating with the rest of an external researcher's particular toolkit, you might want to consider making a Dockerfile for this project. (I don't understand docker fully but I will link you to a position paper here: https://arxiv.org/abs/1410.0846.)