arXiv:1902.04522 paper discussion

[alreadydone]

Kudos for the release of ELFv2 and a beautifully written paper! As people study it I am sure many will come up with questions, and in fact I have two right now, so I just open this thread to invite answers:

1. I think https://facebook.ai/developers/tools/elf-opengo is supposed to contain the auxiliary datasets (the ladder set is what is interesting to me), but currently it shows 404.

2. There are some numbers missing when I viewed the pdf and when I looked at the TeX source I found they're not there:

   We initially started with the synchronous approach before switching to the asynchronous approach. Switching offered two benefits: (1) Both selfplay generation and training realized a drastic speedup. The asynchronous approach achieves over 5x \ytnote{not sure this is accurate} the selfplay throughput of the synchronous approach on our hardware setup. (2) The ratio of selfplay games to training minibatches increased by \ytnote{need to fix this: TODO}, thus helping to prevent overfitting.

Are these numbers available now, just you forgot to include them?

[jma127]

We commend you on your watchful GitHub eye 😄

• Minutes ago, we officially announced our V2 work on ELF OpenGo in a Facebook AI Research blog post. Please give it a read! All links should be working now.
• Good catch on that missing value -- the paper will be updated during tonight's Arxiv announcement (roughly 5pm Pacific) with that value.

[KvanTTT]

The link still does not work :(