Skip to content
体验新版
项目
组织
正在加载...
登录
切换导航
打开侧边栏
PaddlePaddle
Paddle
提交
c70f5920
P
Paddle
项目概览
PaddlePaddle
/
Paddle
12 个月 前同步成功
通知
2292
Star
20931
Fork
5422
代码
文件
提交
分支
Tags
贡献者
分支图
Diff
Issue
1423
列表
看板
标记
里程碑
合并请求
543
Wiki
0
Wiki
分析
仓库
DevOps
项目成员
Pages
P
Paddle
项目概览
项目概览
详情
发布
仓库
仓库
文件
提交
分支
标签
贡献者
分支图
比较
Issue
1,423
Issue
1,423
列表
看板
标记
里程碑
合并请求
543
合并请求
543
Pages
分析
分析
仓库分析
DevOps
Wiki
0
Wiki
成员
成员
收起侧边栏
关闭侧边栏
动态
分支图
创建新Issue
提交
Issue看板
体验新版 GitCode,发现更多精彩内容 >>
未验证
提交
c70f5920
编写于
7月 30, 2020
作者:
M
mapingshuo
提交者:
GitHub
7月 30, 2020
浏览文件
操作
浏览文件
下载
电子邮件补丁
差异文件
add gradient Merge optimizer to meta (#25763)
* add gradient Merge optimizer to meta, test=develop
上级
caa90a65
变更
7
显示空白变更内容
内联
并排
Showing
7 changed file
with
154 addition
and
2 deletion
+154
-2
python/paddle/fleet/base/distributed_strategy.py
python/paddle/fleet/base/distributed_strategy.py
+29
-0
python/paddle/fleet/base/meta_optimizer_factory.py
python/paddle/fleet/base/meta_optimizer_factory.py
+6
-1
python/paddle/fleet/meta_optimizers/__init__.py
python/paddle/fleet/meta_optimizers/__init__.py
+5
-1
python/paddle/fleet/meta_optimizers/gradient_merge_optimizer.py
.../paddle/fleet/meta_optimizers/gradient_merge_optimizer.py
+53
-0
python/paddle/fluid/optimizer.py
python/paddle/fluid/optimizer.py
+6
-0
python/paddle/fluid/tests/unittests/CMakeLists.txt
python/paddle/fluid/tests/unittests/CMakeLists.txt
+2
-0
python/paddle/fluid/tests/unittests/test_fleet_gradient_merge_meta_optimizer.py
...sts/unittests/test_fleet_gradient_merge_meta_optimizer.py
+53
-0
未找到文件。
python/paddle/fleet/base/distributed_strategy.py
浏览文件 @
c70f5920
...
...
@@ -521,6 +521,23 @@ class DistributedStrategy(object):
@
property
def
gradient_merge
(
self
):
"""
Gradient Merge, also called as Gradient Accumulation,
is a strategy for large batch training. With this strategy,
model parameter will not be updated until user-defined steps.
For each step, the forward network and the backward network
will run to calculate the gradient of model parameters.
For every k step, the optimization network will run,
applying a specific optimization method (such as SGD, Adam)
to model parameters.
Examples:
.. code-block:: python
import paddle.fleet as fleet
strategy = fleet.DistributedStrategy()
strategy.gradient_merge = True
strategy.gradient_merge_configs = {"k_steps": 4, "avg": True}
"""
return
self
.
strategy
.
gradient_merge
@
gradient_merge
.
setter
...
...
@@ -532,6 +549,18 @@ class DistributedStrategy(object):
@
property
def
gradient_merge_configs
(
self
):
"""
the key-value configs of distribute_strategy
Keys:
k_steps (int): the update period of the parameters
avg (bool): whether to average the gradients of each mini-batch,
the default value is `True`
Example:
import paddle.fleet as fleet
strategy = fleet.DistributedStrategy()
strategy.gradient_merge = True
strategy.gradient_merge_configs = {"k_steps": 4, "avg": True}
"""
return
get_msg_dict
(
self
.
strategy
.
gradient_merge_configs
)
@
gradient_merge_configs
.
setter
...
...
python/paddle/fleet/base/meta_optimizer_factory.py
浏览文件 @
c70f5920
...
...
@@ -13,11 +13,16 @@
# limitations under the License.
from
..meta_optimizers
import
RecomputeOptimizer
from
..meta_optimizers
import
GradientMergeOptimizer
from
..meta_optimizers
import
GraphExecutionOptimizer
__all__
=
[
"MetaOptimizerFactory"
]
meta_optimizer_names
=
[
"RecomputeOptimizer"
,
"GraphExecutionOptimizer"
]
meta_optimizer_names
=
[
"RecomputeOptimizer"
,
"GradientMergeOptimizer"
,
"GraphExecutionOptimizer"
,
]
class
MetaOptimizerFactory
(
object
):
...
...
python/paddle/fleet/meta_optimizers/__init__.py
浏览文件 @
c70f5920
...
...
@@ -12,6 +12,10 @@
# See the License for the specific language governing permissions and
from
.recompute_optimizer
import
RecomputeOptimizer
from
.gradient_merge_optimizer
import
GradientMergeOptimizer
from
.graph_execution_optimizer
import
GraphExecutionOptimizer
__all__
=
[
'RecomputeOptimizer'
]
__all__
=
[
'RecomputeOptimizer'
,
'GradientMergeOptimizer'
,
]
python/paddle/fleet/meta_optimizers/gradient_merge_optimizer.py
0 → 100644
浏览文件 @
c70f5920
# Copyright (c) 2019 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
GradientMergeOptimizer
as
GM
from
.meta_optimizer_base
import
MetaOptimizerBase
__all__
=
[
"GradientMergeOptimizer"
]
class
GradientMergeOptimizer
(
MetaOptimizerBase
):
def
__init__
(
self
,
optimizer
):
super
(
GradientMergeOptimizer
,
self
).
__init__
(
optimizer
)
self
.
inner_opt
=
optimizer
self
.
wrapped_opt
=
GM
(
optimizer
)
self
.
meta_optimizers_white_list
=
[]
def
_set_basic_info
(
self
,
loss
,
role_maker
,
user_defined_optimizer
,
user_defined_strategy
):
super
(
GradientMergeOptimizer
,
self
).
_set_basic_info
(
loss
,
role_maker
,
user_defined_optimizer
,
user_defined_strategy
)
self
.
wrapped_opt
.
_set_k_steps
(
self
.
user_defined_strategy
.
gradient_merge_configs
[
"k_steps"
])
self
.
wrapped_opt
.
_set_avg
(
self
.
user_defined_strategy
.
gradient_merge_configs
[
"avg"
])
def
_can_apply
(
self
):
can_apply
=
(
self
.
user_defined_strategy
.
gradient_merge
==
True
)
and
\
self
.
user_defined_strategy
.
gradient_merge_configs
[
"k_steps"
]
>
1
return
can_apply
def
_disable_strategy
(
self
,
dist_strategy
):
dist_strategy
.
gradient_merge
=
False
dist_strategy
.
gradient_merge_configs
=
{
"k_steps"
:
1
,
"avg"
:
True
}
def
minimize_impl
(
self
,
loss
,
startup_program
=
None
,
parameter_list
=
None
,
no_grad_set
=
None
):
optimize_ops
,
params_grads
=
\
self
.
wrapped_opt
.
minimize
(
loss
,
startup_program
,
parameter_list
,
no_grad_set
)
return
optimize_ops
,
params_grads
python/paddle/fluid/optimizer.py
浏览文件 @
c70f5920
...
...
@@ -5017,6 +5017,12 @@ class GradientMergeOptimizer(object):
self
.
type
=
"gradient_merge"
self
.
avg
=
avg
def
_set_k_steps
(
self
,
k_steps
):
self
.
k_steps
=
k_steps
def
_set_avg
(
self
,
avg
):
self
.
avg
=
avg
def
minimize
(
self
,
loss
,
startup_program
=
None
,
...
...
python/paddle/fluid/tests/unittests/CMakeLists.txt
浏览文件 @
c70f5920
...
...
@@ -32,6 +32,7 @@ list(APPEND MIXED_DIST_TEST_OPS test_fleet_checkpoint)
list
(
APPEND MIXED_DIST_TEST_OPS test_collective_optimizer
)
list
(
APPEND MIXED_DIST_TEST_OPS test_fleet_base
)
list
(
APPEND MIXED_DIST_TEST_OPS test_fleet_meta_optimizer
)
list
(
APPEND MIXED_DIST_TEST_OPS test_fleet_gradient_merge_meta_optimizer
)
list
(
APPEND MIXED_DIST_TEST_OPS test_fleet_private_function
)
foreach
(
TEST_OP
${
MIXED_DIST_TEST_OPS
}
)
list
(
REMOVE_ITEM TEST_OPS
${
TEST_OP
}
)
...
...
@@ -364,6 +365,7 @@ if(WITH_DISTRIBUTE)
if
(
NOT APPLE
)
py_test_modules
(
test_fleet_base MODULES test_fleet_base ENVS
${
dist_ENVS
}
)
py_test_modules
(
test_fleet_meta_optimizer MODULES test_fleet_meta_optimizer ENVS
${
dist_ENVS
}
)
py_test_modules
(
test_fleet_gradient_merge_meta_optimizer MODULES test_fleet_gradient_merge_meta_optimizer ENVS
${
dist_ENVS
}
)
py_test_modules
(
test_fleet_private_function MODULES test_fleet_private_function ENVS
${
dist_ENVS
}
)
endif
(
NOT APPLE
)
if
(
WITH_DGC
)
...
...
python/paddle/fluid/tests/unittests/test_fleet_gradient_merge_meta_optimizer.py
0 → 100644
浏览文件 @
c70f5920
# 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
TestFleetGradientMergeMetaOptimizer
(
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
test_gradient_merge_optimizer
(
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
=
64
,
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
.
gradient_merge
=
True
strategy
.
gradient_merge_configs
=
{
"k_steps"
:
2
,
"avg"
:
True
}
optimizer
=
paddle
.
optimizer
.
SGD
(
learning_rate
=
0.01
)
optimizer
=
fleet
.
distributed_optimizer
(
optimizer
,
strategy
=
strategy
)
optimizer
.
minimize
(
avg_cost
)
if
__name__
==
"__main__"
:
unittest
.
main
()
编辑
预览
Markdown
is supported
0%
请重试
或
添加新附件
.
添加附件
取消
You are about to add
0
people
to the discussion. Proceed with caution.
先完成此消息的编辑!
取消
想要评论请
注册
或
登录