Closed FrugoFruit90 closed 2 years ago
This is just the code for the original (KDD 2020) paper. The personalize models paper has not been published yet so we don't have a code release.
can I get a personalized model if I use training data that is from one particular player?
In theory yes, but we needed about 10 million games to get good performance when starting from scratch. So I don't think there's anyone with enough games for that to work.
isn't it possible to use a pre-trained model and then fine-tune it with a small set of games from a particular user?
Yes, we're working on the code release for that right now: https://github.com/CSSLab/maia-individual
Would love to use that feature too.
Is there a script that would do pretty much the same as train_maia.py, but starting with a given set of weights?
This achieves the highest move-matching accuracy in the paper but I didn't find it mentioned in the readme.