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
f2265d9f
P
Paddle
项目概览
PaddlePaddle
/
Paddle
大约 1 年 前同步成功
通知
2298
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
未验证
提交
f2265d9f
编写于
3月 16, 2020
作者:
W
WangXi
提交者:
GitHub
3月 16, 2020
浏览文件
操作
浏览文件
下载
电子邮件补丁
差异文件
Fix problem use recompute and dgc same time (#23010)
上级
8fb8b952
变更
2
隐藏空白更改
内联
并排
Showing
2 changed file
with
23 addition
and
3 deletion
+23
-3
python/paddle/fluid/optimizer.py
python/paddle/fluid/optimizer.py
+3
-0
python/paddle/fluid/tests/unittests/test_dgc_optimizer.py
python/paddle/fluid/tests/unittests/test_dgc_optimizer.py
+20
-3
未找到文件。
python/paddle/fluid/optimizer.py
浏览文件 @
f2265d9f
...
...
@@ -3926,6 +3926,9 @@ class RecomputeOptimizer(Optimizer):
parameter_list
,
no_grad_set
,
checkpoints
=
self
.
_checkpoints
)
# Note: since we can't use all_reduce_op now,
# dgc_op should be the last op of one grad.
self
.
_optimizer
.
_append_dgc_ops
(
params_grads
)
return
params_grads
def
apply_optimize
(
self
,
loss
,
startup_program
,
params_grads
):
...
...
python/paddle/fluid/tests/unittests/test_dgc_optimizer.py
浏览文件 @
f2265d9f
...
...
@@ -35,7 +35,8 @@ class TestDGCMomentumOptimizer(unittest.TestCase):
def
check_dgc_momentum_optimizer
(
self
,
dims
=
[
5
,
10
,
8
],
name
=
"momentum"
,
regularization
=
None
):
regularization
=
None
,
use_recompute
=
False
):
init_program
=
framework
.
Program
()
program
=
framework
.
Program
()
block
=
program
.
global_block
()
...
...
@@ -72,6 +73,13 @@ class TestDGCMomentumOptimizer(unittest.TestCase):
local_grad_clip_norm
=
1.0
,
num_trainers
=
2
,
regularization
=
regularization
)
if
use_recompute
:
dgc_momentum_optimizer
=
optimizer
.
RecomputeOptimizer
(
dgc_momentum_optimizer
)
dgc_momentum_optimizer
.
get_accumulators
=
dgc_momentum_optimizer
.
_optimizer
.
get_accumulators
dgc_momentum_optimizer
.
get_velocity_str
=
dgc_momentum_optimizer
.
_optimizer
.
get_velocity_str
mean_out
=
block
.
create_var
(
dtype
=
"float32"
,
shape
=
[
1
],
lod_level
=
0
,
name
=
"mean.out"
)
block
.
append_op
(
...
...
@@ -112,8 +120,9 @@ class TestDGCMomentumOptimizer(unittest.TestCase):
self
.
assertAlmostEqual
(
op
.
attr
(
'regular_coeff'
),
coeff
)
print
(
"dgc regular_coeff="
+
str
(
coeff
))
with
open
(
"test_dgc_optimizer_"
+
name
+
".log"
,
"w"
)
as
f
:
program_to_code
(
program
,
fout
=
f
)
# for local test debug
#with open("test_dgc_optimizer_" + name + str(use_recompute) + ".log", "w") as f:
# program_to_code(program, fout=f)
def
test_momentum_without_dgc
(
self
):
self
.
check_dgc_momentum_optimizer
(
...
...
@@ -130,6 +139,14 @@ class TestDGCMomentumOptimizer(unittest.TestCase):
self
.
check_dgc_momentum_optimizer
(
dims
=
[
16
,
1024
,
8
],
name
=
"dgc_momentum"
)
def
test_momentum_with_dgc_recompute
(
self
):
# 16 * 1024 = 16384, use dgc momentum
self
.
check_dgc_momentum_optimizer
(
dims
=
[
16
,
1024
,
8
],
name
=
"dgc_momentum"
,
regularization
=
regularizer
.
L2Decay
(
1e-4
),
use_recompute
=
True
)
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