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
PaddleDetection
提交
98aebc46
P
PaddleDetection
项目概览
PaddlePaddle
/
PaddleDetection
大约 1 年 前同步成功
通知
694
Star
11112
Fork
2696
代码
文件
提交
分支
Tags
贡献者
分支图
Diff
Issue
184
列表
看板
标记
里程碑
合并请求
40
Wiki
0
Wiki
分析
仓库
DevOps
项目成员
Pages
P
PaddleDetection
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
184
Issue
184
列表
看板
标记
里程碑
合并请求
40
合并请求
40
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
98aebc46
编写于
3月 28, 2022
作者:
S
shangliang Xu
提交者:
GitHub
3月 28, 2022
浏览文件
操作
浏览文件
下载
电子邮件补丁
差异文件
refine coslr, add last_plateau_epochs (#5401)
上级
504e89ba
变更
1
隐藏空白更改
内联
并排
Showing
1 changed file
with
36 addition
and
10 deletion
+36
-10
ppdet/optimizer.py
ppdet/optimizer.py
+36
-10
未找到文件。
ppdet/optimizer.py
浏览文件 @
98aebc46
...
...
@@ -41,12 +41,21 @@ class CosineDecay(object):
max_epochs (int): max epochs for the training process.
if you commbine cosine decay with warmup, it is recommended that
the max_iters is much larger than the warmup iter
use_warmup (bool): whether to use warmup. Default: True.
min_lr_ratio (float): minimum learning rate ratio. Default: 0.
last_plateau_epochs (int): use minimum learning rate in
the last few epochs. Default: 0.
"""
def
__init__
(
self
,
max_epochs
=
1000
,
use_warmup
=
True
,
eta_min
=
0.
):
def
__init__
(
self
,
max_epochs
=
1000
,
use_warmup
=
True
,
min_lr_ratio
=
0.
,
last_plateau_epochs
=
0
):
self
.
max_epochs
=
max_epochs
self
.
use_warmup
=
use_warmup
self
.
eta_min
=
eta_min
self
.
min_lr_ratio
=
min_lr_ratio
self
.
last_plateau_epochs
=
last_plateau_epochs
def
__call__
(
self
,
base_lr
=
None
,
...
...
@@ -56,21 +65,38 @@ class CosineDecay(object):
assert
base_lr
is
not
None
,
"either base LR or values should be provided"
max_iters
=
self
.
max_epochs
*
int
(
step_per_epoch
)
last_plateau_iters
=
self
.
last_plateau_epochs
*
int
(
step_per_epoch
)
min_lr
=
base_lr
*
self
.
min_lr_ratio
if
boundary
is
not
None
and
value
is
not
None
and
self
.
use_warmup
:
# use warmup
warmup_iters
=
len
(
boundary
)
for
i
in
range
(
int
(
boundary
[
-
1
]),
max_iters
):
boundary
.
append
(
i
)
decayed_lr
=
base_lr
*
0.5
*
(
math
.
cos
(
(
i
-
warmup_iters
)
*
math
.
pi
/
(
max_iters
-
warmup_iters
))
+
1
)
decayed_lr
=
decayed_lr
if
decayed_lr
>
self
.
eta_min
else
self
.
eta_min
value
.
append
(
decayed_lr
)
if
i
<
max_iters
-
last_plateau_iters
:
decayed_lr
=
min_lr
+
(
base_lr
-
min_lr
)
*
0.5
*
(
math
.
cos
(
(
i
-
warmup_iters
)
*
math
.
pi
/
(
max_iters
-
warmup_iters
-
last_plateau_iters
))
+
1
)
value
.
append
(
decayed_lr
)
else
:
value
.
append
(
min_lr
)
return
optimizer
.
lr
.
PiecewiseDecay
(
boundary
,
value
)
elif
last_plateau_iters
>
0
:
# not use warmup, but set `last_plateau_epochs` > 0
boundary
=
[]
value
=
[]
for
i
in
range
(
max_iters
):
if
i
<
max_iters
-
last_plateau_iters
:
decayed_lr
=
min_lr
+
(
base_lr
-
min_lr
)
*
0.5
*
(
math
.
cos
(
i
*
math
.
pi
/
(
max_iters
-
last_plateau_iters
))
+
1
)
value
.
append
(
decayed_lr
)
else
:
value
.
append
(
min_lr
)
if
i
>
0
:
boundary
.
append
(
i
)
return
optimizer
.
lr
.
PiecewiseDecay
(
boundary
,
value
)
return
optimizer
.
lr
.
CosineAnnealingDecay
(
base_lr
,
T_max
=
max_iters
,
eta_min
=
self
.
eta_min
)
base_lr
,
T_max
=
max_iters
,
eta_min
=
min_lr
)
@
serializable
...
...
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