An up to date, complete and unified install document

The number one source of apparent bugs and issues plaguing users is version mismatches and user errors during install.

I have tried clean installs of many different guides and instructionals. None of them follow the same steps and none of them completly work. By combining them and manually research errors and reccomended versions for various prerequisites i eventually got it working. Thpugh any attempt to get tensorboard working has been unsuccessful. my most promising lead was that it needed tensorflow installed. Which now results in an error when trying to run training and didn’t solve the tensorboard issue.

The official docs on github seem to be the least coherent, most out of date and possibly even wrong in some places. It also splits itself between a seperate page that poorly describes setting up an enviroment and reccomends different versions of packages compared to the main page.
Theres no clear reccomended versions i could find. the version reccomended int he official docs are incompatible with some of the prequisite modules auto installed when installing the mlagents python package into an environment. This conclusion is based on the errors i got and the facte that they were resolved by using a different python version

Code monkeys tutorials are missing entire steps and use comands and such that no longer work or possibly even exist. Which of course isn’t his fault, presumably they still work for that old version of mlagents.

Their are a handful of others scattered about the internet with variations of all these issues.

The best was in the hummingbird tutorial. But there were still a couple of tweaks i ahd to make myself. installing missing packages and updating others. which the tutorial doesn’t cover at all.

I might attempt a write up myself if i can ever reliably figure it out. But im new to all things python and somone more experienced would obviously be better.
Some key requirements, atleats in my opnion, would include:

  • Anaconda - many, possibly most devs using ml agents will not be familiar with python. unity is a c# focussed engine that attracts novices and relies heavily on GUIs. Equivelant python commands should also be mentioned alongside. But the target audience of this system is unity devs.
  • optional steps such as download the source code and examples needs to be on a different page or seperate section entirley to the main install. having them randomly inserted between other steps makes the page harder to read, confusing and increases the risk of user error.
  • A table of versions is absolutely the bare minimum. without this any tutorial is useless, as users guess combinations of different versions of packages through trial and error as not only i did, but just about every other person who found a solution in forums seems to have.
  • The ideal would be a guided or automatic installer. Given the modularity of python and that the install proccess is basically a bunch command lines. this shouldn’t be hard at all. But perhaps theres a good reason for it? again im not well versed in python, so this might be off base.
2 Likes

the pytorch version referenced: ml-agents/docs/Installation.md at develop · Unity-Technologies/ml-agents · GitHub is not listed in anaconda, likely because it isnt an lts version

Yeah, I try to download pytorch 1.7.1 like is recommended and get “ERROR: Could not find a version that satisfies the requirement torch~=1.7.1” Seems like the supported pytorch version is ancient.