Skip to content
项目
群组
代码片段
帮助
当前项目
正在载入...
登录 / 注册
切换导航面板
P
pytensor
项目
项目
详情
活动
周期分析
仓库
仓库
文件
提交
分支
标签
贡献者
图表
比较
统计图
议题
0
议题
0
列表
看板
标记
里程碑
合并请求
0
合并请求
0
CI / CD
CI / CD
流水线
作业
日程
统计图
Wiki
Wiki
代码片段
代码片段
成员
成员
折叠边栏
关闭边栏
活动
图像
聊天
创建新问题
作业
提交
问题看板
Open sidebar
testgroup
pytensor
Commits
7c7a5f6d
提交
7c7a5f6d
authored
8月 02, 2014
作者:
Roy Xue
浏览文件
操作
浏览文件
下载
电子邮件补丁
差异文件
Updates
上级
695db885
隐藏空白字符变更
内嵌
并排
正在显示
1 个修改的文件
包含
43 行增加
和
49 行删除
+43
-49
profiling.py
theano/compile/profiling.py
+43
-49
没有找到文件。
theano/compile/profiling.py
浏览文件 @
7c7a5f6d
...
@@ -25,7 +25,6 @@ import numpy
...
@@ -25,7 +25,6 @@ import numpy
import
theano
import
theano
from
theano.gof
import
graph
from
theano.gof
import
graph
from
theano.gof
import
vm
from
theano.configparser
import
AddConfigVar
,
BoolParam
,
IntParam
from
theano.configparser
import
AddConfigVar
,
BoolParam
,
IntParam
...
@@ -663,7 +662,6 @@ class ProfileStats(object):
...
@@ -663,7 +662,6 @@ class ProfileStats(object):
node_memory_saved_by_view
=
0
node_memory_saved_by_view
=
0
node_memory_saved_by_inplace
=
0
node_memory_saved_by_inplace
=
0
node_idx
=
0
node_idx
=
0
for
node
in
order
:
for
node
in
order
:
val
=
nodes_mem
[
node
]
val
=
nodes_mem
[
node
]
dmap
=
getattr
(
node
.
op
,
'destroy_map'
,
None
)
dmap
=
getattr
(
node
.
op
,
'destroy_map'
,
None
)
...
@@ -696,12 +694,19 @@ class ProfileStats(object):
...
@@ -696,12 +694,19 @@ class ProfileStats(object):
max_minimum_peak
=
0
max_minimum_peak
=
0
def
count_minimum_peak
(
node_list
,
fgraph
,
nodes_mem
):
def
count_minimum_peak
(
node_list
,
fgraph
,
nodes_mem
):
global
maybe_executed
,
mem_count
,
mem_bound
global
maybe_executed
,
mem_count
,
mem_bound
,
last_user
,
freed
order
=
[]
order
=
[]
min_order
=
[]
min_order
=
[]
node_list
=
list
(
node_list
)
node_list
=
list
(
node_list
)
min_mem
=
sys
.
maxint
min_mem
=
sys
.
maxint
current_mem
=
0
current_mem
=
0
mem_count
=
0
mem_bound
=
numpy
.
inf
computed
=
set
()
last_user
=
{}
for
node
in
node_list
:
for
ins
in
node
.
inputs
:
last_user
[
ins
]
=
0
def
check_node_state
(
node
):
def
check_node_state
(
node
):
"""
"""
...
@@ -733,89 +738,78 @@ class ProfileStats(object):
...
@@ -733,89 +738,78 @@ class ProfileStats(object):
if
c
!=
"output"
and
check_node_state
(
c
):
if
c
!=
"output"
and
check_node_state
(
c
):
executables_nodes
.
add
(
c
)
executables_nodes
.
add
(
c
)
mem_count
=
0
mem_bound
=
numpy
.
inf
# dependencies = {}
dependencies
=
fgraph
.
profile
.
dependencies
print
dependencies
compute
=
set
()
last_user
=
{}
for
node
in
node_list
:
for
node
in
node_list
:
for
ins
in
node
.
inputs
:
for
ins
in
node
.
inputs
:
last_user
[
ins
]
=
node
last_user
[
ins
]
+=
1
for
outs
in
node
.
outputs
:
for
outs
in
node
.
outputs
:
computed
.
add
(
outs
)
computed
.
add
(
outs
)
# for node in node_list[0].inputs[0]:
# dependencies[node] = []
# if val.owner and val.clients:
# ls = []
# for c in val.clients:
# if c[0] is not 'output':
# ls += c[0].outputs
# dependencies[val] += ls
def
min_memory_generator
(
executables_nodes
):
def
min_memory_generator
(
executables_nodes
):
global
mem_count
,
mem_bound
global
mem_count
,
mem_bound
,
last_user
,
freed
for
node
in
executables_nodes
:
for
node
in
executables_nodes
:
new_exec_nodes
=
executables_nodes
.
copy
()
new_exec_nodes
=
executables_nodes
.
copy
()
new_exec_nodes
.
remove
(
node
)
new_exec_nodes
.
remove
(
node
)
mem_created
=
0
mem_created
=
0
mem_freed
=
0
for
var
in
node
.
outputs
:
compute_map
[
var
][
0
]
=
1
idx
=
0
idx
=
0
dmap
=
getattr
(
node
.
op
,
'destroy_map'
,
None
)
dmap
=
getattr
(
node
.
op
,
'destroy_map'
,
None
)
vmap
=
getattr
(
node
.
op
,
'view_map'
,
None
)
vmap
=
getattr
(
node
.
op
,
'view_map'
,
None
)
mem_created
=
sum
(
nodes_mem
[
node
])
# for i in nodes_mem[node]:
# if (dmap and idx in dmap) or (vmap and idx in vmap):
# continue
# elif not isinstance(i, str):
# mem_created += i
# idx += 1
mem_count
+=
mem_created
# Add memory created
mem_freed
=
0
for
ins
in
node
.
inputs
:
for
ins
in
node
.
inputs
:
if
(
ins
in
computed
)
and
(
ins
not
in
fgraph
.
outputs
)
and
(
node
==
last_user
[
ins
]):
last_user
[
ins
]
-=
1
mem_freed
+=
var_mem
[
ins
]
# for val in node.outputs:
# add mem_create
# if (dependencies[val] and val.owner and val not in fgraph.outputs):
for
i
in
nodes_mem
[
node
]:
# # print dependencies[val]
if
(
dmap
and
idx
in
dmap
)
or
(
vmap
and
idx
in
vmap
):
# # print compute_map
continue
# for v in dependencies[val]:
elif
not
isinstance
(
i
,
str
):
# if compute_map[v] == 1:
mem_created
+=
i
# print "yes"
idx
+=
1
# mem_freed += var_mem[val]
#add mem_freed, this part is not working well
for
ins
in
node
.
inputs
:
if
(
ins
in
computed
)
and
(
ins
not
in
fgraph
.
outputs
)
and
(
last_user
[
ins
]
==
0
):
if
not
isinstance
(
var_mem
[
ins
],
str
):
mem_freed
+=
var_mem
[
ins
]
mem_count
+=
mem_created
mem_count
-=
mem_freed
mem_count
-=
mem_freed
# Reduce memory freed
# check if cut path now
if
mem_count
>
mem_bound
:
if
mem_count
>
mem_bound
:
mem_count
-=
mem_created
mem_count
-=
mem_created
mem_count
+=
mem_freed
mem_count
+=
mem_freed
for
ins
in
node
.
inputs
:
last_user
[
ins
]
+=
1
continue
continue
for
var
in
node
.
outputs
:
compute_map
[
var
][
0
]
=
1
for
var
in
node
.
outputs
:
for
var
in
node
.
outputs
:
for
c
,
_
in
var
.
clients
:
for
c
,
_
in
var
.
clients
:
if
c
!=
"output"
and
check_node_state
(
c
):
if
c
!=
"output"
and
check_node_state
(
c
):
new_exec_nodes
.
add
(
c
)
new_exec_nodes
.
add
(
c
)
if
not
new_exec_nodes
:
if
not
new_exec_nodes
:
yield
[
node
]
yield
[
node
]
#update mem_bound
if
mem_count
<
mem_bound
:
if
mem_count
<
mem_bound
:
# update the mem_bound
mem_bound
=
mem_count
mem_bound
=
mem_count
else
:
else
:
for
p
in
min_memory_generator
(
new_exec_nodes
):
for
p
in
min_memory_generator
(
new_exec_nodes
):
yield
[
node
]
+
p
yield
[
node
]
+
p
# resetting part
mem_count
-=
mem_created
mem_count
-=
mem_created
mem_count
+=
mem_freed
mem_count
+=
mem_freed
# reset
for
ins
in
node
.
inputs
:
last_user
[
ins
]
+=
1
for
var
in
node
.
outputs
:
for
var
in
node
.
outputs
:
compute_map
[
var
][
0
]
=
0
compute_map
[
var
][
0
]
=
0
for
order
in
min_memory_generator
(
executables_nodes
):
for
order
in
min_memory_generator
(
executables_nodes
):
continue
continue
...
...
编写
预览
Markdown
格式
0%
重试
或
添加新文件
添加附件
取消
您添加了
0
人
到此讨论。请谨慎行事。
请先完成此评论的编辑!
取消
请
注册
或者
登录
后发表评论