How to train and test ilql model?

[timercrack]

I can train the bert model use bert.phase1, and test it without problem.
I don't know much about the ilql model. It looks like an advanced version of lql. What is the purpuse of extract_policy.py in ilql directory and how to use it?
What step to do to train and test the iql model? I have prepared iql training data and reward plugin but don't know how to start.

[Cryolite]

Please use train.py for training. extract_policy.py implements the AWR algorithm, which is used in the original IQL paper. However, extract_policy.py is for extracting a policy model from a trained Q-function model. If you simply want to make predictions, you can use the Q-function model directly, and I think there is no need to run extract_policy.py.

[timercrack]

Thank you for your answser, I have another question, I saw your get_reward implementation is based on game_rank and game_score, and reward is set to 0.0 when it is called not at the end of a game because it does not have such value.
In this way, will other data be useless except for the data at the end of the game?
Can I use the round rank and round delta score instead when game is not end?