AdamGleave
commented
9 months ago I see, having a submodule seems like the least-bad approach then.
On Fri, Dec 1, 2023, 12:19 Tom Tseng @.***> wrote:
@.**** commented on this pull request.
In .gitmodules https://github.com/AlignmentResearch/go_attack/pull/118#discussion_r1412539994 :
@@ -7,3 +7,6 @@ [submodule "engines/KataGo-raw"] path = engines/KataGo-raw url = @.***:lightvector/KataGo.git +[submodule "engines/KataGo-tensorflow"]
yeah, the tensorflow training/shuffling code has all been removed in KataGo v1.12. We need to either have this extra engines/KataGo-tensorflow submodule, or copy all the old tensorflow code back into KataGo-custom
— Reply to this email directly, view it on GitHub https://github.com/AlignmentResearch/go_attack/pull/118#discussion_r1412539994, or unsubscribe https://github.com/notifications/unsubscribe-auth/AALZ3I43VFTY6DDTBWPQMC3YHI3UJAVCNFSM6AAAAAAYEHUW7CVHI2DSMVQWIX3LMV43YUDVNRWFEZLROVSXG5CSMV3GSZLXHMYTONRQGUYTGNRWGM . You are receiving this because your review was requested.Message ID: @.***>
Update
KataGo-customto v1.12.4.Model training still uses TF so that we can finetune our TF cyclic adversary. If we want to finetune the new architectures using PyTorch model training, we'll need to further edit kubernetes/train.sh since some of the flags to train.py have changed (e.g., model kind is a flag -model-kind instead of a positional argument, and warmstarting is done with the -initial-checkpoint flag instead of done by copying files into a initial_weights directory).
kubernetes/shuffle_and_export.shwould need to switch to using the PyTorch code as well, since theshuffleddataformat is different between TF and PyTorch.