Main author here- hardest part was definitely parsing the paper and putting the pieces together, realising which components were properly specified and which were under specified. As of now I don’t actively maintain the repo but I do try to go through the issues and PRs. Open to major PRs as long as code stays neat and easy to understand.
https://www.reddit.com/r/MachineLearning/comments/8ctksk/did...
Would be happy to hear from the creators about what was the hardest part to implement, and future plans.