def run_trainer(self, args):
"""run trainer"""
from paddle.fluid.incubate.fleet.parameter_server.distribute_transpiler import fleet
import paddle.fluid as fluid
from paddle.fluid.transpiler.ps_dispatcher import RoundRobin
from paddle.fluid.transpiler.ps_dispatcher import HashName
fluid.default_startup_program().random_seed = 1
fluid.default_main_program().random_seed = 1
if args.role.upper() != "TRAINER":
raise ValueError("args role must be TRAINER")
role = role_maker.UserDefinedRoleMaker(
current_id=args.current_id,
role=role_maker.Role.WORKER,
worker_num=args.trainers,
server_endpoints=args.endpoints.split(","))
fleet.init(role)
strategy = DistributeTranspilerConfig()
strategy.sync_mode = args.run_params["sync_mode"]
strategy.async_mode = args.run_params["async_mode"]
strategy.mode = "pserver"
strategy.slice_var_up = args.run_params['slice_var_up']
strategy.enable_dc_asgd = args.run_params['enable_dc_asgd']
if args.run_params['split_method']:
strategy.split_method = HashName
strategy.split_method = RoundRobin
strategy.wait_port = args.run_params['wait_port']
strategy.runtime_split_send_recv = args.run_params['runtime_split_send_recv']
strategy.use_hierarchical_allreduce = args.run_params['use_hierarchical_allreduce']
# strategy.hierarchical_allreduce_exter_nranks = args.run_params['hierarchical_allreduce_exter_nranks']
# strategy.hierarchical_allreduce_inter_nranks = args.run_params['hierarchical_allreduce_inter_nranks']
strategy.geo_sgd_mode = args.run_params['geo_sgd']
strategy.geo_sgd_need_push_nums = args.run_params['push_nums']
avg_cost = self.net()
optimizer = fluid.optimizer.SGD(LEARNING_RATE)
optimizer = fleet.distributed_optimizer(optimizer, strategy)
optimizer.minimize(avg_cost)
losses = self.do_training(fleet, args)
losses = "" if not losses else losses
print(losses)
def test_communicator_init_and_start(self):
role = role_maker.UserDefinedRoleMaker(
current_id=0,
role=role_maker.Role.WORKER,
worker_num=2,
server_endpoints=["127.0.0.1:6001", "127.0.0.1:6002"])
fleet.init(role)
avg_cost = self.net()
optimizer = fluid.optimizer.SGD(0.01)
strategy = DistributeTranspilerConfig()
strategy.sync_mode = True
strategy.wait_port = False
optimizer = fleet.distributed_optimizer(optimizer, strategy)
optimizer.minimize(avg_cost)
comm = Communicator(fleet.main_program)
comm.start()
time.sleep(10)
comm.stop()
def test_communicator_async(self):
role = role_maker.UserDefinedRoleMaker(
current_id=0,
role=role_maker.Role.WORKER,
worker_num=2,
server_endpoints=["127.0.0.1:6001", "127.0.0.1:6002"])
fleet.init(role)
avg_cost = self.net()
optimizer = fluid.optimizer.SGD(0.01)
strategy = DistributeTranspilerConfig()
strategy.sync_mode = False
strategy.runtime_split_send_recv = True
strategy.wait_port = False
optimizer = fleet.distributed_optimizer(optimizer, strategy)
optimizer.minimize(avg_cost)
fleet.init_worker()
time.sleep(10)
fleet.stop_worker()