worker.learner.comm

learner_comm

BaseCommLearner

class ding.worker.learner.comm.base_comm_learner.BaseCommLearner(cfg: EasyDict)

Overview:

Abstract baseclass for CommLearner.

Interfaces:

__init__, send_policy, get_data, send_learn_info， start, close

Property:

hooks4call

__init__(cfg: EasyDict) → None

Overview:

Initialization method.

Arguments:

• cfg (EasyDict): Config dict

close() → None

Overview:

Close comm learner.

abstract get_data(batch_size: int) → list

Overview:

Get batched meta data from coordinator. Will be registered in base learner.

Arguments:

• batch_size (int): Batch size.

Returns:

• stepdata (list): A list of training data, each element is one trajectory.

abstract send_learn_info(learn_info: dict) → None

Overview:

Send learn info to coordinator. Will be registered in base learner.

Arguments:

• learn_info (dict): Learn info in dict type.

abstract send_policy(state_dict: dict) → None

Overview:

Save learner’s policy in corresponding path. Will be registered in base learner.

Arguments:

• state_dict (dict): State dict of the runtime policy.

start() → None

Overview:

Start comm learner.

FlaskFileSystemLearner

class ding.worker.learner.comm.flask_fs_learner.FlaskFileSystemLearner(cfg: EasyDict)

Overview:

An implementation of CommLearner, using flask and the file system.

Interfaces:

__init__, send_policy, get_data, send_learn_info, start, close

Property:

hooks4call

__init__(cfg: EasyDict) → None

Overview:

Init method.

Arguments:

• cfg (EasyDict): Config dict.

close() → None

Overview:

Join learner thread and close learner if still running. Then close learner slave and comm learner itself.

get_data(batch_size: int) → List[Callable]

Overview:

Get a list of data loading function, which can be implemented by dataloader to read data from files.

Arguments:

• batch_size (int): Batch size.

Returns:

• data (List[Callable]): A list of callable data loading function.

send_learn_info(learn_info: dict) → None

Overview:

Store learn info dict in queue, which will be retrieved by callback function “deal_with_learner_learn” in learner slave, then will be sent to coordinator.

Arguments:

• learn_info (dict): Learn info in dict type. Keys are like ‘learner_step’, ‘priority_info’ ‘finished_task’, etc. You can refer to learn_info``(``worker/learner/base_learner.py) for details.

send_policy(state_dict: dict) → None

Overview:

Save learner’s policy in corresponding path, called by SendPolicyHook.

Arguments:

• state_dict (dict): State dict of the policy.

start() → None

Overview:

Start comm learner itself and the learner slave.

create_comm_learner

Overview:

Given the key(comm_learner_name), create a new comm learner instance if in comm_map’s values, or raise an KeyError. In other words, a derived comm learner must first register, then can call create_comm_learner to get the instance.

Arguments:

• cfg (dict): Learner config. Necessary keys: [import_names, comm_learner_type].

Returns:

• learner (BaseCommLearner): The created new comm learner, should be an instance of one of comm_map’s values.