add lars to fleet meta optimizer (#25884)
JZ-LIANG
5 years ago
committed by
GitHub
parent
8d2896f1fe
commit
8ebffc78c9
@ -0,0 +1,83 @@
|
||||
# Copyright (c) 2020 PaddlePaddle Authors. All Rights Reserved.
|
||||
#
|
||||
# Licensed under the Apache License, Version 2.0 (the "License");
|
||||
# you may not use this file except in compliance with the License.
|
||||
# You may obtain a copy of the License at
|
||||
#
|
||||
# http://www.apache.org/licenses/LICENSE-2.0
|
||||
#
|
||||
# Unless required by applicable law or agreed to in writing, software
|
||||
# distributed under the License is distributed on an "AS IS" BASIS,
|
||||
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
||||
# See the License for the specific language governing permissions and
|
||||
|
||||
from paddle.fluid.optimizer import Momentum, LarsMomentumOptimizer
|
||||
from .meta_optimizer_base import MetaOptimizerBase
|
||||
import logging
|
||||
|
||||
__all__ = ["LarsOptimizer"]
|
||||
|
||||
|
||||
class LarsOptimizer(MetaOptimizerBase):
|
||||
def __init__(self, optimizer):
|
||||
super(LarsOptimizer, self).__init__(optimizer)
|
||||
self.inner_opt = optimizer
|
||||
self.lars_opt = None
|
||||
# we do not allow meta optimizer to be inner optimizer currently
|
||||
self.meta_optimizers_white_list = []
|
||||
|
||||
def _set_basic_info(self, loss, role_maker, user_defined_optimizer,
|
||||
user_defined_strategy):
|
||||
super(LarsOptimizer, self)._set_basic_info(
|
||||
loss, role_maker, user_defined_optimizer, user_defined_strategy)
|
||||
|
||||
opt = self.inner_opt
|
||||
if not isinstance(opt, Momentum):
|
||||
return
|
||||
|
||||
configs = self.user_defined_strategy.lars_configs
|
||||
|
||||
self.lars_opt = LarsMomentumOptimizer(
|
||||
learning_rate=opt._learning_rate,
|
||||
momentum=opt._momentum,
|
||||
lars_coeff=configs['lars_coeff'],
|
||||
lars_weight_decay=configs['lars_weight_decay'],
|
||||
parameter_list=opt._parameter_list,
|
||||
regularization=opt.regularization,
|
||||
grad_clip=opt._grad_clip,
|
||||
name=opt._name)
|
||||
|
||||
def _can_apply(self):
|
||||
if self.user_defined_strategy.lars:
|
||||
if not isinstance(self.inner_opt, Momentum):
|
||||
logging.warn(
|
||||
"lars need the inner optimizer to be Momentum optimizer.")
|
||||
return False
|
||||
return True
|
||||
return False
|
||||
|
||||
def _disable_strategy(self, dist_strategy):
|
||||
dist_strategy.lars = False
|
||||
dist_strategy.lars_configs = {
|
||||
'lars_coeff': 0.001,
|
||||
'lars_weight_decay': 0.0005,
|
||||
}
|
||||
|
||||
def backward(self,
|
||||
loss,
|
||||
startup_program=None,
|
||||
parameter_list=None,
|
||||
no_grad_set=None,
|
||||
callbacks=None):
|
||||
return self.lars_opt.backward(loss, startup_program, parameter_list,
|
||||
no_grad_set, callbacks)
|
||||
|
||||
def minimize_impl(self,
|
||||
loss,
|
||||
startup_program=None,
|
||||
parameter_list=None,
|
||||
no_grad_set=None):
|
||||
optimize_ops, params_grads = \
|
||||
self.lars_opt.minimize(loss, startup_program,
|
||||
parameter_list, no_grad_set)
|
||||
return optimize_ops, params_grads
|
||||
@ -0,0 +1,73 @@
|
||||
# Copyright (c) 2020 PaddlePaddle Authors. All Rights Reserved.
|
||||
#
|
||||
# Licensed under the Apache License, Version 2.0 (the "License");
|
||||
# you may not use this file except in compliance with the License.
|
||||
# You may obtain a copy of the License at
|
||||
#
|
||||
# http://www.apache.org/licenses/LICENSE-2.0
|
||||
#
|
||||
# Unless required by applicable law or agreed to in writing, software
|
||||
# distributed under the License is distributed on an "AS IS" BASIS,
|
||||
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
||||
# See the License for the specific language governing permissions and
|
||||
# limitations under the License.
|
||||
|
||||
import unittest
|
||||
import paddle
|
||||
import os
|
||||
import paddle.fleet as fleet
|
||||
import paddle.fluid.incubate.fleet.base.role_maker as role_maker
|
||||
|
||||
|
||||
class TestFleetLarsMetaOptimizer(unittest.TestCase):
|
||||
def setUp(self):
|
||||
os.environ["POD_IP"] = "127.0.0.1"
|
||||
os.environ["PADDLE_TRAINER_ENDPOINTS"] = "127.0.0.1:36001"
|
||||
os.environ["PADDLE_TRAINERS_NUM"] = "2"
|
||||
os.environ["PADDLE_PSERVERS_IP_PORT_LIST"] = \
|
||||
"127.0.0.1:36001,127.0.0.2:36001"
|
||||
|
||||
def net(self):
|
||||
role = role_maker.PaddleCloudRoleMaker(is_collective=True)
|
||||
fleet.init(role)
|
||||
input_x = paddle.fluid.layers.data(
|
||||
name="x", shape=[32], dtype='float32')
|
||||
input_y = paddle.fluid.layers.data(name="y", shape=[1], dtype='int64')
|
||||
|
||||
fc_1 = paddle.fluid.layers.fc(input=input_x, size=64, act='tanh')
|
||||
fc_2 = paddle.fluid.layers.fc(input=fc_1, size=256, act='tanh')
|
||||
prediction = paddle.fluid.layers.fc(input=[fc_2], size=2, act='softmax')
|
||||
cost = paddle.fluid.layers.cross_entropy(
|
||||
input=prediction, label=input_y)
|
||||
avg_cost = paddle.fluid.layers.mean(x=cost)
|
||||
|
||||
strategy = paddle.fleet.DistributedStrategy()
|
||||
strategy.lars = True
|
||||
strategy.lars_configs = {
|
||||
"lars_coeff": 0.001,
|
||||
"lars_weight_decay": 0.0005,
|
||||
}
|
||||
|
||||
return avg_cost, strategy
|
||||
|
||||
def test_lars_optimizer(self):
|
||||
avg_cost, strategy = self.net()
|
||||
optimizer = paddle.optimizer.Momentum(learning_rate=0.01, momentum=0.9)
|
||||
optimizer = fleet.distributed_optimizer(optimizer, strategy=strategy)
|
||||
optimizer.minimize(avg_cost)
|
||||
|
||||
ops = [op.type for op in avg_cost.block.ops]
|
||||
self.assertIn('lars_momentum', ops)
|
||||
|
||||
def test_lars_not_apply_with_adam(self):
|
||||
avg_cost, strategy = self.net()
|
||||
optimizer = paddle.optimizer.Adam(learning_rate=0.01)
|
||||
optimizer = fleet.distributed_optimizer(optimizer, strategy=strategy)
|
||||
optimizer.minimize(avg_cost)
|
||||
|
||||
ops = [op.type for op in avg_cost.block.ops]
|
||||
self.assertNotIn('lars_momentum', ops)
|
||||
|
||||
|
||||
if __name__ == "__main__":
|
||||
unittest.main()
|
||||
Loading…
Reference in new issue