Add documentation about requirements and dependencies for installation on fresh conda env

[tlacombe]

Hello,

I tried installing this library by cloning the github repo and then running python setup.py install as indicated (on a fresh conda env) but unfortunately it did not work immediately.

• it seems tensorflow=2.2 cannot be installed with python = 3.9 yet (which is default on fresh conda env. I agree this is at the tensorflow level directly, but it may be worth mentioning in the readme of this library).
• If you install tensorflow=2.2 (either manually when creating the conda env, or running setup.py), you get absl-py=0.11 automatically installed, while tensorflow-metadata requires it to be <0.11, leading to an error when installing the library.
• tensorflow requires numpy>=1.19.2 to be installed. Having an older version of numpy leads to an error (numpy=1.19.1 was installed by default in my fresh env and it seems it was not updated when installing the library).
• It explicitly requires scipy=1.4.1 ; having scipy=1.5.2 (manually installed) led to an error when installing the library.
• Similarly, error: protobuf 4.0.0rc2 is installed but protobuf<4,>=3.7 is required by {'tensorflow-metadata'}

Code to reproduce the issues (note that of course it may depends on the way conda is set up for the user) :

conda create env --name robustness
conda activate robustness
pip setup.py install 

Trying iteratively, I managed to complete the installation (note that I did not try running the code yet) using: condra create env --name robustness python=3.8 setuptools numpy=1.19.2 protobuf=3.9 absl-py=0.10 (and then python setup.py install).

I hope this helps; if needed I can provide more logs.

[josipd]

Thanks for reporting and for the detailed analysis on the issue, I'll reproduce this on a fresh install and get back to you.