Closed ed1d1a8d closed 1 year ago
Check out this pull request on
See visual diffs & provide feedback on Jupyter Notebooks.
Powered by ReviewNB
Name | Link |
---|---|
Latest commit | 27d0fc4a23c4eebf7ee073906e0d0586ff21d064 |
Latest deploy log | https://app.netlify.com/sites/goattack/deploys/63b758f2abd5af0009923293 |
Thanks for working on this Tony, great to have a more rigorous eval of how much compute we used! Left a number of minor comments / clarifying questions. Please request re-review once addressed.
LGTM apart from the question about ratio of rows:moves not being 2, and the # of visits used during training. Both of those could change the numbers to a non-trivial degree, but if it looks like resolving them will end up being more of a research problem we might want to merge this PR early and just open another ticket to address it.
So I talked with lightvector and he gave a lot of suggested improvements to our estimation procedure. In particular, he recommended we benchmark the actual selfplay/victimplay process to determine the ratio between data rows and total number of moves played (this might explain our variation from 1.48 to 2.22).
Doing the improved estimation factoring in lightvector's suggestions will be fairly involved, so I made a new issue for it here: https://github.com/HumanCompatibleAI/KataGoVisualizer/issues/41. I agree that we can merge this PR and address this issue later.
Fixed all the other small things and re-requesting review.
The number of FLOPs used by KataGo is estimated in
notebooks/iclr2022/estimate-flops-katago.ipynb
. We perform the estimate by:ptflops
andthop
libraries. We use the newly written pytorch version of KataGo for this (added as a submodule).The number of FLOPs used to train our adversary is estimated in
notebooks/iclr2022/estimate-flops-adv.ipynb
. We perform this estimate by:We assume the selfplay/victimplay dominates the total compute.
Updated compute estimates: