def test_pslib_1(self):
"""Test cases for pslib."""
import paddle.fluid as fluid
from paddle.fluid.incubate.fleet.parameter_server.pslib import fleet
from paddle.fluid.incubate.fleet.parameter_server.pslib import PSLib
from paddle.fluid.incubate.fleet.base.role_maker import GeneralRoleMaker
try:
import netifaces
except:
print("warning: no netifaces, skip test_pslib_1")
return
os.environ["POD_IP"] = "127.0.0.1"
os.environ["PADDLE_PORT"] = "36001"
os.environ["TRAINING_ROLE"] = "TRAINER"
os.environ["PADDLE_TRAINER_ENDPOINTS"] = "127.0.0.1:36001"
os.environ["PADDLE_PSERVERS_IP_PORT_LIST"] = "127.0.0.1:36002"
os.environ["PADDLE_TRAINER_ID"] = "0"
role_maker = GeneralRoleMaker()
role_maker.generate_role()
place = fluid.CPUPlace()
exe = fluid.Executor(place)
fleet.init(role_maker)
train_program = fluid.Program()
startup_program = fluid.Program()
scope = fluid.Scope()
with fluid.program_guard(train_program, startup_program):
show = fluid.layers.data(name="show", shape=[-1, 1], \
dtype="float32", lod_level=1, append_batch_size=False)
fc = fluid.layers.fc(input=show, size=1, act=None)
label = fluid.layers.data(name="click", shape=[-1, 1], \
dtype="int64", lod_level=1, append_batch_size=False)
label_cast = fluid.layers.cast(label, dtype='float32')
cost = fluid.layers.log_loss(fc, label_cast)
try:
adam = fluid.optimizer.Adam(learning_rate=0.000005)
adam = fleet.distributed_optimizer(adam)
adam.minimize([cost], [scope])
fleet.run_server()
except:
print("do not support pslib test, skip")
return
fleet.clear_one_table(0)
from paddle.fluid.incubate.fleet.base.role_maker import \
MPISymetricRoleMaker
try:
role = MPISymetricRoleMaker()
role._all_reduce([1], [2])
except:
print("catch expected error of not inited")
try:
role = MPISymetricRoleMaker()
role._all_reduce([1], [2], "min")
except:
print("catch expected error of not inited")
try:
role = MPISymetricRoleMaker()
role._all_reduce([1], [2], "max")
except:
print("catch expected error of not inited")
try:
role = MPISymetricRoleMaker()
role._all_reduce([1], [2], "unknown")
except:
print("catch expected error of unknown type")
def do_training(self, args=None):
"""do training"""
avg_cost = self.net()
place = fluid.CPUPlace()
exe = fluid.Executor(place)
fleet.init()
# optimizer
optimizer = fluid.optimizer.Adam(learning_rate=0.001)
# 加入 fleet distributed_optimizer 加入分布式策略配置及多机优化
optimizer = fleet.distributed_optimizer(optimizer, strategy={"fleet_desc_file": "./thirdparty/pslib/fleet_desc.prototxt"})
optimizer.minimize(avg_cost)
train_info = []
# 启动server
if fleet.is_server():
fleet.init_server()
fleet.run_server()
# 启动worker
if fleet.is_worker():
train_data_path = 'thirdparty/data/dist_data/pslib/train_data'
train_data_files = []
for filename in os.listdir(train_data_path):
train_data_files.append(os.path.join(train_data_path, filename))
# fleet dataset
label = fluid.layers.data(name="click", shape=[-1, 1], dtype="int64", lod_level=1, append_batch_size=False)
data = fluid.layers.data(name="1", shape=[1], dtype="int64", lod_level=1)
dataset = fluid.DatasetFactory().create_dataset("InMemoryDataset")
dataset.set_use_var([label, data])
dataset.set_pipe_command("./python/bin/python ./thirdparty/pslib/dataset_generator.py")
dataset.set_batch_size(32)
dataset.set_thread(3)
dataset.set_filelist(train_data_files)
# 把数据读到内存
dataset.load_into_memory()
# 本地shuffle
dataset.local_shuffle()
# 初始化worker配置
fleet.init_worker()
exe.run(fluid.default_startup_program())
PASS_NUM = 1
for pass_id in range(PASS_NUM):
var_dict = {"loss": avg_cost}
global var_dict
class FetchVars(fluid.executor.FetchHandler):
def __init__(self, var_dict=None, period_secs=2):
super(FetchVars, self).__init__(var_dict, period_secs=2)
def handler(self, res_dict):
train_info.extend(res_dict["loss"])
print(train_info)
exe.train_from_dataset(
program=fluid.default_main_program(),
dataset=dataset,
fetch_handler=FetchVars(var_dict))
dataset.release_memory()
fleet.shrink_sparse_table()
fleet.shrink_dense_table(0.01, 11)
fleet.print_table_stat(0)
fleet.clear_one_table(0)
fleet.clear_model()
fleet.stop_worker()
return train_info