league.league

base_league

BaseLeague

class ding.league.base_league.BaseLeague(cfg: EasyDict)

Overview:

League, proposed by Google Deepmind AlphaStar. Can manage multiple players in one league.

Interface:

get_job_info, judge_snapshot, update_active_player, finish_job, save_checkpoint

Note

In __init__ method, league would also initialized players as well(in _init_players method).

finish_job(job_info: dict) → None

Overview:

Finish current job. Update shared payoff to record the game results.

Arguments:

• job_info (dict): A dict containing job result information.

get_job_info(player_id: str | None = None, eval_flag: bool = False) → dict

Overview:

Get info dict of the job which is to be launched to an active player.

Arguments:

• player_id (str): The active player’s id.

• eval_flag (bool): Whether this is an evaluation job.

Returns:

• job_info (dict): Job info.

ReturnsKeys:

• necessary: launch_player (the active player)

judge_snapshot(player_id: str, force: bool = False) → bool

Overview:

Judge whether a player is trained enough for snapshot. If yes, call player’s snapshot, create a historical player(prepare the checkpoint and add it to the shared payoff), then mutate it, and return True. Otherwise, return False.

Arguments:

• player_id (ActivePlayer): The active player’s id.

Returns:

• snapshot_or_not (dict): Whether the active player is snapshotted.

static save_checkpoint(src_checkpoint, dst_checkpoint) → None

Overview:

Copy a checkpoint from path src_checkpoint to path dst_checkpoint.

Arguments:

• src_checkpoint (str): Source checkpoint’s path, e.g. s3://alphastar_fake_data/ckpt.pth

• dst_checkpoint (str): Destination checkpoint’s path, e.g. s3://alphastar_fake_data/ckpt.pth

update_active_player(player_info: dict) → None

Overview:

Update an active player’s info.

Arguments:

• player_info (dict): Info dict of the player which is to be updated.

ArgumentsKeys:

• necessary: player_id, train_iteration

create_league

Overview:

Given the key (league_type), create a new league instance if in league_mapping’s values, or raise an KeyError. In other words, a derived league must first register then call create_league to get the instance object.

Arguments:

• cfg (EasyDict): league config, necessary keys: [league.import_module, league.learner_type]

Returns:

• league (BaseLeague): the created new league, should be an instance of one of league_mapping’s values

league.one_vs_one_league

one_vs_one_league

OneVsOneLeague

class ding.league.one_vs_one_league.OneVsOneLeague(cfg: EasyDict)

Overview:

One vs One battle game league. Decide which two players will play against each other.

Interface:

__init__, run, close, finish_job, update_active_player

finish_job(job_info: dict) → None

Overview:

Finish current job. Update shared payoff to record the game results.

Arguments:

• job_info (dict): A dict containing job result information.

get_job_info(player_id: str | None = None, eval_flag: bool = False) → dict

Overview:

Get info dict of the job which is to be launched to an active player.

Arguments:

• player_id (str): The active player’s id.

• eval_flag (bool): Whether this is an evaluation job.

Returns:

• job_info (dict): Job info.

ReturnsKeys:

• necessary: launch_player (the active player)

judge_snapshot(player_id: str, force: bool = False) → bool

Overview:

Judge whether a player is trained enough for snapshot. If yes, call player’s snapshot, create a historical player(prepare the checkpoint and add it to the shared payoff), then mutate it, and return True. Otherwise, return False.

Arguments:

• player_id (ActivePlayer): The active player’s id.

Returns:

• snapshot_or_not (dict): Whether the active player is snapshotted.

static save_checkpoint(src_checkpoint, dst_checkpoint) → None

Overview:

Copy a checkpoint from path src_checkpoint to path dst_checkpoint.

Arguments:

• src_checkpoint (str): Source checkpoint’s path, e.g. s3://alphastar_fake_data/ckpt.pth

• dst_checkpoint (str): Destination checkpoint’s path, e.g. s3://alphastar_fake_data/ckpt.pth

update_active_player(player_info: dict) → None

Overview:

Update an active player’s info.

Arguments:

• player_info (dict): Info dict of the player which is to be updated.

ArgumentsKeys:

• necessary: player_id, train_iteration