<a href="https://github.com/tencent-ailab/TLeague/blob/dev-open/tleague/learners/base_

why learner add api for response learner id ? about tleague HOT 2 CLOSED

tencent-ailab commented on August 17, 2024

why learner add api for response learner id ?

from tleague.

Comments (2)

pengsun commented on August 17, 2024 5

Hi, it serves for registering/discovering the learner task as defined here:

TLeague/tleague/utils/tl_types.py

Line 19 in 101ff85

class LearnerTask(object):

where all the learner tasks are maintained at League Manager.

To be detailed, when using Horovod for multi-GPU training, TLeague has M_L learners which always synchronize the NN parameters, with each learner associated to only one GPU card. Being consistent to the Horovod/mpi convention, we let rank 0 learner sponsor the task creating (by calling the _request_task):

TLeague/tleague/learners/pg_learner.py

Line 197 in 101ff85

if self.rank == 0:

while other rank i, i = 1, 2,...,M_L-1 learners query the existed learner task:

TLeague/tleague/learners/pg_learner.py

Line 215 in 101ff85

self.task = self._query_task()

Note all the M_L learners have an identical learner_id (which is the handle for the learner task) as passed in by the --learner_id arg:

TLeague/tleague/bin/run_pg_learner.py

Line 36 in 101ff85

flags.DEFINE_string("learner_id", "",

and the example here:
https://github.com/tencent-ailab/tleague_projpage/blob/c8df899d713a77dcc72aa6e2b9d892bcebc96cf8/pommerman/pommerman.yml.jinja2#L385