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
wmsofts
Paddle
提交
9d26f1a3
P
Paddle
项目概览
wmsofts
/
Paddle
与 Fork 源项目一致
Fork自
PaddlePaddle / Paddle
通知
1
Star
0
Fork
0
代码
文件
提交
分支
Tags
贡献者
分支图
Diff
Issue
0
列表
看板
标记
里程碑
合并请求
0
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
0
Issue
0
列表
看板
标记
里程碑
合并请求
0
合并请求
0
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
9d26f1a3
编写于
2月 15, 2018
作者:
Y
Yang Yang
浏览文件
操作
浏览文件
下载
电子邮件补丁
差异文件
callback to list of callbacks
上级
bea80b0d
变更
3
隐藏空白更改
内联
并排
Showing
3 changed file
with
22 addition
and
16 deletion
+22
-16
paddle/fluid/framework/executor.cc
paddle/fluid/framework/executor.cc
+1
-2
python/paddle/v2/fluid/backward.py
python/paddle/v2/fluid/backward.py
+20
-13
python/paddle/v2/fluid/optimizer.py
python/paddle/v2/fluid/optimizer.py
+1
-1
未找到文件。
paddle/fluid/framework/executor.cc
浏览文件 @
9d26f1a3
...
...
@@ -120,12 +120,11 @@ void Executor::Run(const ProgramDesc& pdesc, Scope* scope, int block_id,
for
(
auto
&
op_desc
:
block
.
AllOps
())
{
auto
op
=
paddle
::
framework
::
OpRegistry
::
CreateOp
(
*
op_desc
);
// VLOG(3)
<< op->DebugStringEx(local_scope);
VLOG
(
3
)
<<
place_
<<
" "
<<
op
->
DebugStringEx
(
local_scope
);
platform
::
DeviceContextPool
&
pool
=
platform
::
DeviceContextPool
::
Instance
();
platform
::
RecordEvent
record_event
(
op
->
Type
(),
pool
.
Get
(
place_
));
VLOG
(
3
)
<<
op
->
Type
();
op
->
Run
(
*
local_scope
,
place_
);
if
(
FLAGS_benchmark
)
{
VLOG
(
2
)
<<
"Memory used after operator "
+
op
->
Type
()
+
" running: "
...
...
python/paddle/v2/fluid/backward.py
浏览文件 @
9d26f1a3
...
...
@@ -234,7 +234,6 @@ def _callback_lookup_(op):
"ncclInit"
,
{
"parallel_scopes"
:
self
.
parallel_scopes_name
},
{
"Communicator"
:
[
'nccl_com__do_not_change_'
]},
{})
print
(
serialize_op_decs
(
op_desc
))
block
.
program
.
global_block
().
desc
.
append_op
().
copy_from
(
op_desc
)
self
.
has_inserted_nccl_init
=
True
...
...
@@ -285,9 +284,7 @@ def _append_backward_ops_(block,
val(str): corresponding forward variable name
callback(callable object): a callable object used to decorate new generated grad ops
"""
if
callbacks
is
None
:
callbacks
=
[]
else
:
if
callbacks
is
not
None
:
assert
(
isinstance
(
callbacks
,
list
))
for
cb
in
callbacks
:
if
not
hasattr
(
cb
,
'__call__'
):
...
...
@@ -302,12 +299,17 @@ def _append_backward_ops_(block,
if
op
.
has_attr
(
"sub_block"
):
sub_block
=
program
.
block
(
op
.
block_attr
(
"sub_block"
))
grad_sub_block
=
program
.
create_block
(
parent_idx
=
sub_block
.
idx
)
if
callbacks
is
None
:
callbacks
=
[
_callback_lookup_
(
op
)]
cb
=
_callback_lookup_
(
op
)
if
cb
is
not
None
:
if
callbacks
is
None
:
new_callbacks
=
[
cb
]
else
:
new_callbacks
=
callbacks
+
[
_callback_lookup_
(
op
)]
_append_backward_ops_
(
sub_block
,
sub_block
.
ops
,
grad_sub_block
,
no_grad_dict
,
grad_to_var
,
new_callbacks
)
else
:
callbacks
.
append
(
_callback_lookup_
(
op
))
_append_backward_ops_
(
sub_block
,
sub_block
.
ops
,
grad_sub_block
,
no_grad_dict
,
grad_to_var
,
callbacks
)
_append_backward_ops_
(
sub_block
,
sub_block
.
ops
,
grad_sub_block
,
no_grad_dict
,
grad_to_var
,
callbacks
)
grad_sub_block_list
.
append
(
grad_sub_block
.
desc
)
# Getting op's corresponding grad_op
...
...
@@ -327,8 +329,10 @@ def _append_backward_ops_(block,
new_op_desc
=
target_block
.
desc
.
append_op
()
new_op_desc
.
copy_from
(
op_desc
)
grad_to_var
[
"__current_op_desc__"
]
=
new_op_desc
for
cb
in
callbacks
:
cb
(
block
=
target_block
,
context
=
grad_to_var
)
if
callbacks
is
not
None
:
assert
(
isinstance
(
callbacks
,
list
))
for
cb
in
callbacks
:
cb
(
block
=
target_block
,
context
=
grad_to_var
)
def
_append_backward_vars_
(
block
,
start_op_idx
,
grad_to_var
,
grad_info_map
):
...
...
@@ -408,7 +412,8 @@ def _get_stop_gradients_(program):
return
no_grad_dict
def
append_backward
(
loss
,
parameter_list
=
None
,
no_grad_set
=
None
,
callback
=
None
):
def
append_backward
(
loss
,
parameter_list
=
None
,
no_grad_set
=
None
,
callbacks
=
None
):
"""
Append backward part to main_program
...
...
@@ -424,6 +429,8 @@ def append_backward(loss, parameter_list=None, no_grad_set=None, callback=None):
(list[(Variable,Variable)]): list of (parameter, gradient) pair.
"""
assert
isinstance
(
loss
,
framework
.
Variable
)
if
callbacks
is
not
None
:
isinstance
(
callbacks
,
list
)
program
=
loss
.
block
.
program
if
no_grad_set
is
None
:
...
...
@@ -451,7 +458,7 @@ def append_backward(loss, parameter_list=None, no_grad_set=None, callback=None):
no_grad_dict
[
0
].
update
(
map
(
_append_grad_suffix_
,
block_no_grad_set
))
_append_backward_ops_
(
root_block
,
op_path
,
root_block
,
no_grad_dict
,
grad_to_var
,
callback
)
grad_to_var
,
callback
s
)
# Because calc_gradient may be called multiple times,
# we need rename the internal gradient variables so that they have
...
...
python/paddle/v2/fluid/optimizer.py
浏览文件 @
9d26f1a3
...
...
@@ -225,7 +225,7 @@ class Optimizer(object):
`create_optimization_pass()` into one.
"""
params_grads
=
append_backward
(
loss
,
parameter_list
,
no_grad_set
,
error_clip_callback
)
[
error_clip_callback
]
)
params_grads
=
append_gradient_clip_ops
(
params_grads
)
...
...
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