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
db6e7b56
提交
db6e7b56
authored
5月 07, 2015
作者:
--global
浏览文件
操作
浏览文件
下载
电子邮件补丁
差异文件
Modify ScanSaveMem to keep a buffer big enough for the memory reuse feature
上级
4ae75bc3
隐藏空白字符变更
内嵌
并排
正在显示
2 个修改的文件
包含
27 行增加
和
3 行删除
+27
-3
scan_op.py
theano/scan_module/scan_op.py
+1
-1
scan_opt.py
theano/scan_module/scan_opt.py
+26
-2
没有找到文件。
theano/scan_module/scan_op.py
浏览文件 @
db6e7b56
...
...
@@ -675,7 +675,7 @@ class Scan(PureOp):
self
.
n_sit_sot
+
self
.
n_nit_sot
)
wrapped_inputs
=
[
Param
(
x
,
borrow
=
False
)
for
x
in
self
.
inputs
]
wrapped_outputs
=
[
Out
(
x
,
borrow
=
(
x
not
in
self
.
inputs
)
)
for
x
in
wrapped_outputs
=
[
Out
(
x
,
borrow
=
True
)
for
x
in
self
.
outputs
[:
slices
]]
wrapped_outputs
+=
self
.
outputs
[
slices
:]
profile
=
None
...
...
theano/scan_module/scan_opt.py
浏览文件 @
db6e7b56
...
...
@@ -1228,8 +1228,32 @@ class ScanSaveMem(gof.Optimizer):
if
start
==
0
or
store_steps
[
i
]
==
0
:
store_steps
[
i
]
=
0
else
:
pval
=
select_max
(
nw_steps
-
start
+
init_l
[
i
],
init_l
[
i
])
# The "+ 1" is because if the memory pre-allocation
# mechanism used to in the Scan op to reduce overhead.
# To prevent aliasing between the inputs and outputs
# of recurrent states, it requires that the buffer be
# large enough to that, the new state and the oldest
# tap needed don't occupy the sample place in the
# circular buffer. For now, this only needs to be done
# for mitsots and sitsots (because mitmots are not
# currently supported by the mechanism) and only if
# the inner function has more then one output
# (otherwise, there is no risk of aliasing because
# once the output is computed, the oldest tap can
# safely be overwritten).
first_mitsot_idx
=
node
.
op
.
n_mit_mot
last_sitsot_idx
=
(
node
.
op
.
n_mit_mot
+
node
.
op
.
n_mit_sot
+
node
.
op
.
n_sit_sot
-
1
)
if
(
i
>=
first_mitsot_idx
and
i
<=
last_sitsot_idx
and
len
(
node
.
op
.
outputs
)
>
1
):
pval
=
select_max
(
nw_steps
-
start
+
init_l
[
i
],
init_l
[
i
]
+
1
)
else
:
pval
=
select_max
(
nw_steps
-
start
+
init_l
[
i
],
init_l
[
i
])
if
store_steps
[
i
]
!=
-
1
:
pval
=
select_max
(
pval
,
store_steps
[
i
])
...
...
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