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
ff41346c
提交
ff41346c
authored
6月 24, 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
fix bugs
1. remove the best order, cuz we won't use it now. 2. move the count_minimum_peak to the right place
上级
9a9a3b29
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
80 行增加
和
79 行删除
+80
-79
profiling.py
theano/compile/profiling.py
+80
-79
没有找到文件。
theano/compile/profiling.py
浏览文件 @
ff41346c
...
@@ -740,7 +740,8 @@ class ProfileStats(object):
...
@@ -740,7 +740,8 @@ class ProfileStats(object):
node_list
=
fgraph
.
nodes
node_list
=
fgraph
.
nodes
best_order
,
minimum_peak
=
count_minimum_peak
(
node_list
,
fgraph
)
_
,
minimum_peak
=
count_minimum_peak
(
node_list
,
fgraph
)
# for the best order, we dont use it now
max_minimum_peak
=
max
(
max_minimum_peak
,
minimum_peak
)
max_minimum_peak
=
max
(
max_minimum_peak
,
minimum_peak
)
...
@@ -865,6 +866,84 @@ class ProfileStats(object):
...
@@ -865,6 +866,84 @@ class ProfileStats(object):
self
.
optimizer_profile
[
0
]
.
print_profile
(
file
,
self
.
optimizer_profile
[
0
]
.
print_profile
(
file
,
self
.
optimizer_profile
[
1
])
self
.
optimizer_profile
[
1
])
def
count_minimum_peak
(
node_list
,
fgraph
):
mem_list
=
[]
current_mem
=
0
order_index
=
0
min_mem
=
0
order
=
[]
compute_map
=
fgraph
.
profile
.
compute_map
# compute_map use to check if a node is valid
node_mem
=
{}
for
node
in
self
.
apply_callcount
.
keys
():
sum_dense
=
0
for
out
in
node
.
outputs
:
sh
=
self
.
variable_shape
[
out
]
if
hasattr
(
out
.
type
,
'get_size'
):
v
=
out
.
type
.
get_size
(
sh
)
sum_dense
+=
v
node_mem
[
node
]
=
sum_dense
# node_mem use to calculate the node memory usage
def
check_node_state
(
node
):
"""
check if an Apply node is valid(has inputs but no outputs).
:param node: apply node
"""
inputs
=
node
.
inputs
outputs
=
node
.
outputs
deps
=
inputs
+
node
.
destroy_dependencies
computed_ins
=
all
(
compute_map
[
v
][
0
]
for
v
in
deps
)
computed_outs
=
all
(
compute_map
[
v
][
0
]
for
v
in
outputs
)
# check if there could be a compute_map
if
computed_ins
and
not
computed_outs
:
return
True
else
:
return
False
def
min_memory_generator
(
node_list
,
b
=
False
):
global
mem_list
,
current
,
order_index
,
min_mem
'''
enumerate all valid order( node with inputs in its compute_map)
compute the peak of all order and keep the order with the minimum peak.
return an order with minimum memory usage
:param node_list: a list of apply nodes
'''
for
i
in
range
(
len
(
node_list
)):
v
=
node_list
[
i
:
i
+
1
]
if
check_state
(
v
[
0
]):
if
len
(
node_list
)
==
1
:
yield
v
current_mem
+=
node_mem
[
v
[
0
]]
b
=
True
else
:
b
=
False
rest
=
node_list
[
:
i
]
+
node_list
[
i
+
1
:
]
for
p
in
count_min_memory_peak
(
rest
):
yield
v
+
p
current_mem
+=
node_mem
[
v
[
0
]]
if
b
:
mem_list
.
append
(
current_mem
)
if
not
min_mem
:
min_mem
=
current_mem
if
current_mem
<
min_mem
:
min_mem
=
current_mem
order_index
=
mem_list
.
index
(
current_mem
)
current_mem
=
0
gen
=
min_memory_generator
(
node_list
)
for
i
in
range
(
0
,
(
order_index
+
1
)):
order
=
gen
.
next
()
return
order
,
min_mem
if
0
:
# old code still to be ported from ProfileMode
if
0
:
# old code still to be ported from ProfileMode
def
long_print
(
self
,
file
=
sys
.
stderr
,
fct_name
=
None
,
message
=
None
,
def
long_print
(
self
,
file
=
sys
.
stderr
,
fct_name
=
None
,
message
=
None
,
...
@@ -1166,84 +1245,6 @@ if 0: # old code still to be ported from ProfileMode
...
@@ -1166,84 +1245,6 @@ if 0: # old code still to be ported from ProfileMode
n_ops_to_print
=
n_ops_to_print
,
print_apply
=
False
)
n_ops_to_print
=
n_ops_to_print
,
print_apply
=
False
)
def
count_minimum_peak
(
node_list
,
fgraph
):
mem_list
=
[]
current_mem
=
0
order_index
=
0
min_mem
=
0
order
=
[]
compute_map
=
fgraph
.
profile
.
compute_map
# compute_map use to check if a node is valid
node_mem
=
{}
for
node
in
self
.
apply_callcount
.
keys
():
sum_dense
=
0
for
out
in
node
.
outputs
:
sh
=
self
.
variable_shape
[
out
]
if
hasattr
(
out
.
type
,
'get_size'
):
v
=
out
.
type
.
get_size
(
sh
)
sum_dense
+=
v
node_mem
[
node
]
=
sum_dense
# node_mem use to calculate the node memory usage
def
check_node_state
(
node
):
"""
check if an Apply node is valid(has inputs but no outputs).
:param node: apply node
"""
inputs
=
node
.
inputs
outputs
=
node
.
outputs
deps
=
inputs
+
node
.
destroy_dependencies
computed_ins
=
all
(
compute_map
[
v
][
0
]
for
v
in
deps
)
computed_outs
=
all
(
compute_map
[
v
][
0
]
for
v
in
outputs
)
# check if there could be a compute_map
if
computed_ins
and
not
computed_outs
:
return
True
else
:
return
False
def
min_memory_generator
(
node_list
,
b
=
False
):
global
mem_list
,
current
,
order_index
,
min_mem
'''
enumerate all valid order( node with inputs in its compute_map)
compute the peak of all order and keep the order with the minimum peak.
return an order with minimum memory usage
:param node_list: a list of apply nodes
'''
for
i
in
range
(
len
(
node_list
)):
v
=
node_list
[
i
:
i
+
1
]
if
check_state
(
v
[
0
]):
if
len
(
node_list
)
==
1
:
yield
v
current_mem
+=
node_mem
[
v
[
0
]]
b
=
True
else
:
b
=
False
rest
=
node_list
[
:
i
]
+
node_list
[
i
+
1
:
]
for
p
in
count_min_memory_peak
(
rest
):
yield
v
+
p
current_mem
+=
node_mem
[
v
[
0
]]
if
b
:
mem_list
.
append
(
current_mem
)
if
not
min_mem
:
min_mem
=
current_mem
if
current_mem
<
min_mem
:
min_mem
=
current_mem
order_index
=
mem_list
.
index
(
current_mem
)
current_mem
=
0
gen
=
min_memory_generator
(
node_list
)
for
i
in
range
(
0
,
(
order_index
+
1
)):
order
=
gen
.
next
()
return
order
,
min_mem
class
ScanProfileStats
(
ProfileStats
):
class
ScanProfileStats
(
ProfileStats
):
...
...
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