If I want to start model training from scratch

[junbon1120]

Hi, I had a difficulty when I tried to start training a model using the Mortal source code.
When I try to run "train.py", it loads the trained model from the config file for the "test_player" and starts training the model.
I would like a clue as to what to do if I don't have a trained model for "test_player".

[Equim-chan]

Currently the code doesn't handle bootstrapping yet, but I think you could just temporarily comment out the load-from-file part in TestPlayer, which will keep the params randomized, until you have a model that can be used as a baseline.

[Equim-chan]

Mortal only recognizes mjai format which encodes to newline-separated JSON. GZip is optional.

[junbon1120]

Do you use libriichi to convert Tenhou log data to mjai format?

[Equim-chan]

No. Mortal and libriichi only recognize mjai.

[junbon1120]

Thanks, I will try to convert Tenhou log data to mjai format using mjai-reviewer code!

[junbon1120]

Hai
I was able to convert to mjai format using the mjai-reviewer code. Thanks for your comment.
I am aware that I need to run "train_grp.py" and "train.py" to learn Mortal, is this understanding correct?
Do I need to run any other Code?

[jacy]

what is the project for if can't train from scratch?

[phu54321]

How much training time is spent on the current online model?

[phu54321]

Since this whole project is under AGPL, should I distribute the weights if I plan to create an web app based on this? Thanks

[Equim-chan]

You don't have to, since weights are not a part of the source code.