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
5eea5db9
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 >>
提交
5eea5db9
编写于
5月 29, 2018
作者:
T
tangwei12
浏览文件
操作
浏览文件
下载
电子邮件补丁
差异文件
optimized checkpoint and save_model
上级
514b2427
变更
3
隐藏空白更改
内联
并排
Showing
3 changed file
with
58 addition
and
44 deletion
+58
-44
python/paddle/fluid/__init__.py
python/paddle/fluid/__init__.py
+1
-0
python/paddle/fluid/io.py
python/paddle/fluid/io.py
+28
-33
python/paddle/fluid/trainer.py
python/paddle/fluid/trainer.py
+29
-11
未找到文件。
python/paddle/fluid/__init__.py
浏览文件 @
5eea5db9
...
...
@@ -26,6 +26,7 @@ from trainer import BeginEpochEvent
from
trainer
import
EndEpochEvent
from
trainer
import
BeginStepEvent
from
trainer
import
EndStepEvent
from
trainer
import
CheckpointConfig
import
inferencer
from
inferencer
import
Inferencer
...
...
python/paddle/fluid/io.py
浏览文件 @
5eea5db9
...
...
@@ -491,7 +491,6 @@ CHECKPOINT_SEPARATOR = "_"
def
save_checkpoint
(
executor
,
checkpoint_dir
=
None
,
max_num_checkpoints
=
3
,
save_interval_secs
=
600
,
main_program
=
None
):
"""
Save Checkpoint will save persistable LodTensor variables from main_program in checkpoint directory,
...
...
@@ -511,15 +510,10 @@ def save_checkpoint(executor,
if
not
os
.
path
.
isdir
(
checkpoint_dir
):
os
.
makedirs
(
checkpoint_dir
)
serial
=
_get_lastest_checkpoint_dir
(
checkpoint_dir
)
if
serial
>=
0
and
not
_interval_secs_exceed
(
_get_serial_dir
(
serial
,
checkpoint_dir
),
save_interval_secs
):
return
serial
+=
1
cur_dir
=
_get_serial_dir
(
serial
,
checkpoint_dir
)
serial
=
_get_lastest_checkpoint_dir
(
checkpoint_dir
)
+
1
cur_dir
=
_get_serial_dir
(
checkpoint_dir
,
serial
)
load
_persist_vars_without_grad
(
executor
,
cur_dir
,
main_program
)
save
_persist_vars_without_grad
(
executor
,
cur_dir
,
main_program
)
_write_success
(
cur_dir
)
_lru_delete
(
checkpoint_dir
,
max_num_checkpoints
)
...
...
@@ -542,7 +536,7 @@ def load_checkpoint(executor, checkpoint_dir=None, main_program=None):
if
serial
<
0
:
return
cur_dir
=
_get_serial_dir
(
serial
,
checkpoint_dir
)
cur_dir
=
_get_serial_dir
(
checkpoint_dir
,
serial
)
load_persist_vars_without_grad
(
executor
,
cur_dir
,
main_program
)
...
...
@@ -559,11 +553,6 @@ def clean_checkpoint(checkpoint_dir, delete_dir=False):
os
.
rmdir
(
checkpoint_dir
)
def
_get_serial_dir
(
serial
,
checkpoint_dir
):
serial_folder
=
CHECKPOINT_PREFIX
+
CHECKPOINT_SEPARATOR
+
str
(
serial
)
return
os
.
path
.
join
(
checkpoint_dir
,
serial_folder
)
def
_is_checkpoint_var
(
var
):
"""
the checkpoint will not save or load all the variables.
...
...
@@ -582,29 +571,37 @@ def _is_checkpoint_var(var):
return
var
.
persistable
def
_interval_secs_exceed
(
dirname
,
save_interval_secs
):
dir_time
=
os
.
path
.
getmtime
(
dirname
)
if
save_interval_secs
>
(
time
.
time
()
-
dir_time
):
return
False
return
True
def
_get_dir_serial
(
dirname
):
_
,
serial
=
dirname
.
split
(
CHECKPOINT_SEPARATOR
)
serial_num
=
-
1
try
:
serial_num
=
int
(
serial
)
except
ValueError
:
serial_num
=
-
1
return
serial_num
def
_get_serial_dir
(
dirname
,
serial
):
serial_folder
=
CHECKPOINT_PREFIX
+
CHECKPOINT_SEPARATOR
+
str
(
serial
)
return
os
.
path
.
join
(
dirname
,
serial_folder
)
def
_lru_delete
(
dirname
,
max_num_checkpoints
=
3
):
dirs
=
os
.
listdir
(
dirname
)
serial
s
=
[]
serial
_map
=
{}
for
serial
in
dirs
:
try
:
serials
.
append
(
int
(
serial
))
except
ValueError
:
continue
serial_num
=
_get_dir_serial
(
serial
)
serial_map
[
serial_num
]
=
serial
if
len
(
serial
s
)
<=
max_num_checkpoints
:
if
len
(
serial
_map
.
keys
()
)
<=
max_num_checkpoints
:
return
serials
=
serial_map
.
keys
()
serials
.
sort
(
reverse
=
True
)
serials
=
serials
[
max_num_checkpoints
:]
for
serial
in
serials
:
cur_dir
=
os
.
path
.
join
(
dirname
,
str
(
serial
)
)
cur_dir
=
_get_serial_dir
(
dirname
,
serial
)
shutil
.
rmtree
(
cur_dir
)
...
...
@@ -633,20 +630,18 @@ def _get_lastest_checkpoint_dir(checkpoint_dir):
"""
is _SUCCESS in this dir
"""
_
,
serial
=
cur_dir
.
split
(
CHECKPOINT_SEPARATOR
)
try
:
int
(
serial
)
except
ValueError
:
serial
=
_get_dir_serial
(
cur_dir
)
if
serial
==
-
1
:
return
-
1
if
not
os
.
path
.
isdir
(
os
.
path
.
join
(
checkpoint_dir
,
cur_dir
)):
return
-
1
success_path
=
os
.
path
.
join
(
_get_serial_dir
(
serial
,
checkpoint_dir
),
SUCCESS_MARK_FILENAME
)
_get_serial_dir
(
checkpoint_dir
,
serial
),
SUCCESS_MARK_FILENAME
)
if
os
.
path
.
isfile
(
success_path
):
return
int
(
serial
)
return
serial
if
not
os
.
path
.
isdir
(
checkpoint_dir
):
return
-
1
...
...
python/paddle/fluid/trainer.py
浏览文件 @
5eea5db9
...
...
@@ -60,11 +60,24 @@ class CheckpointConfig(object):
def
__init__
(
self
,
checkpoint_dir
=
None
,
max_num_checkpoints
=
3
,
save_interval_secs
=
600
):
epoch_interval
=
1
,
step_interval
=
10
):
if
checkpoint_dir
is
None
:
self
.
checkpoint_dir
=
os
.
getcwd
()
else
:
self
.
checkpoint_dir
=
checkpoint_dir
self
.
max_num_checkpoints
=
max_num_checkpoints
self
.
save_interval_secs
=
save_interval_secs
if
epoch_interval
<
1
:
self
.
epoch_interval
=
1
else
:
self
.
epoch_interval
=
epoch_interval
if
step_interval
<
1
:
self
.
step_interval
=
10
else
:
self
.
step_interval
=
step_interval
def
check_and_get_place
(
place
):
...
...
@@ -290,14 +303,6 @@ class Trainer(object):
exe
=
executor
.
Executor
(
self
.
place
)
io
.
save_persistables
(
exe
,
dirname
=
param_path
)
def
_save_checkpoint
(
self
):
if
self
.
checkpoint
and
self
.
chief
:
exe
=
executor
.
Executor
(
self
.
place
)
io
.
save_checkpoint
(
exe
,
self
.
checkpoint
.
checkpoint_dir
,
self
.
checkpoint
.
max_num_checkpoints
,
self
.
checkpoint
.
save_interval_secs
,
self
.
train_program
)
@
contextlib
.
contextmanager
def
_prog_and_scope_guard
(
self
):
with
framework
.
program_guard
(
...
...
@@ -343,8 +348,9 @@ class Trainer(object):
])
else
:
metrics
=
exe
.
run
(
feed
=
data
,
fetch_list
=
[])
event_handler
(
EndStepEvent
(
epoch_id
,
step_id
,
metrics
))
self
.
_save_checkpoint
()
self
.
_save_checkpoint
(
epoch_id
,
step_id
)
event_handler
(
EndEpochEvent
(
epoch_id
))
def
_test_by_executor
(
self
,
reader
,
feed_order
,
fetch_list
):
...
...
@@ -384,6 +390,18 @@ class Trainer(object):
loss_name
=
self
.
train_func_outputs
[
0
].
name
)
return
self
.
_get_parallel_executor
()
def
_save_checkpoint
(
self
,
epoch_id
,
step_id
):
if
not
self
.
checkpoint
or
not
self
.
chief
:
return
if
epoch_id
%
self
.
checkpoint
.
epoch_interval
==
0
and
step_id
%
self
.
checkpoint
.
step_interval
==
0
:
exe
=
executor
.
Executor
(
self
.
place
)
io
.
save_checkpoint
(
executor
=
exe
,
checkpoint_dir
=
self
.
checkpoint
.
checkpoint_dir
,
max_num_checkpoints
=
self
.
checkpoint
.
max_num_checkpoints
,
main_program
=
self
.
train_program
)
def
build_feed_var_list
(
program
,
feed_order
):
if
not
isinstance
(
program
,
framework
.
Program
):
...
...
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