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
夏有凉风,冬有雪
Paddle
提交
b6b7ab63
P
Paddle
项目概览
夏有凉风,冬有雪
/
Paddle
与 Fork 源项目一致
Fork自
PaddlePaddle / Paddle
通知
1
Star
1
Fork
0
代码
文件
提交
分支
Tags
贡献者
分支图
Diff
Issue
1
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
1
Issue
1
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
b6b7ab63
编写于
11月 21, 2017
作者:
G
guosheng
浏览文件
操作
浏览文件
下载
电子邮件补丁
差异文件
Fix calculations in gru_unit_op to be consistent with gru_op
上级
f191c820
变更
2
显示空白变更内容
内联
并排
Showing
2 changed file
with
46 addition
and
39 deletion
+46
-39
paddle/operators/gru_unit_op.h
paddle/operators/gru_unit_op.h
+38
-32
python/paddle/v2/fluid/tests/test_gru_unit_op.py
python/paddle/v2/fluid/tests/test_gru_unit_op.py
+8
-7
未找到文件。
paddle/operators/gru_unit_op.h
浏览文件 @
b6b7ab63
...
...
@@ -146,35 +146,27 @@ class GRUUnitGradKernel : public framework::OpKernel<T> {
auto
*
weight_grad
=
context
.
Output
<
Tensor
>
(
framework
::
GradVarName
(
"Weight"
));
auto
*
bias_grad
=
context
.
Output
<
Tensor
>
(
framework
::
GradVarName
(
"Bias"
));
input_grad
->
mutable_data
<
T
>
(
context
.
GetPlace
());
hidden_prev_grad
->
mutable_data
<
T
>
(
context
.
GetPlace
());
weight_grad
->
mutable_data
<
T
>
(
context
.
GetPlace
());
Tensor
gate_grad
;
gate_grad
.
mutable_data
<
T
>
(
input
->
dims
(),
context
.
GetPlace
());
Tensor
reset_hidden_prev_grad
;
reset_hidden_prev_grad
.
mutable_data
<
T
>
(
reset_hidden_prev
->
dims
(),
context
.
GetPlace
());
int
batch_size
=
input
->
dims
()[
0
];
int
frame_size
=
hidden_prev
->
dims
()[
1
];
const
T
*
hidden_prev_data
=
hidden_prev
->
data
<
T
>
();
T
*
hidden_prev_grad_data
=
hidden_prev_grad
->
data
<
T
>
();
const
T
*
weight_data
=
weight
->
data
<
T
>
();
T
*
weight_grad_data
=
weight_grad
->
data
<
T
>
();
T
*
gate_grad_data
=
gate_grad
.
data
<
T
>
(
);
T
*
gate_grad_data
=
gate_grad
.
mutable_data
<
T
>
(
input
->
dims
(),
context
.
GetPlace
()
);
const
T
*
reset_hidden_prev_data
=
reset_hidden_prev
->
data
<
T
>
();
T
*
reset_hidden_prev_grad_data
=
reset_hidden_prev_grad
.
data
<
T
>
();
T
*
reset_hidden_prev_grad_data
=
reset_hidden_prev_grad
.
mutable_data
<
T
>
(
reset_hidden_prev
->
dims
(),
context
.
GetPlace
());
auto
h_p
=
EigenMatrix
<
T
>::
From
(
*
hidden_prev
);
auto
g
=
EigenMatrix
<
T
>::
From
(
*
gate
);
auto
d_h
=
EigenMatrix
<
T
>::
From
(
*
hidden_grad
);
auto
d_x
=
EigenMatrix
<
T
>::
From
(
*
input_grad
);
auto
d_h_p
=
EigenMatrix
<
T
>::
From
(
*
hidden_prev_grad
);
auto
d_g
=
EigenMatrix
<
T
>::
From
(
gate_grad
);
auto
d_r_h_p
=
EigenMatrix
<
T
>::
From
(
reset_hidden_prev_grad
);
auto
place
=
context
.
GetEigenDevice
<
Place
>
();
int
batch_size
=
input
->
dims
()[
0
];
int
frame_size
=
hidden_prev
->
dims
()[
1
];
Eigen
::
array
<
int
,
2
>
extents
({{
batch_size
,
frame_size
}});
Eigen
::
array
<
int
,
2
>
u_offsets
({{
0
,
0
}});
auto
u
=
g
.
slice
(
u_offsets
,
extents
);
// update gate
...
...
@@ -195,28 +187,42 @@ class GRUUnitGradKernel : public framework::OpKernel<T> {
gate_grad_data
+
frame_size
*
2
,
frame_size
*
3
,
weight_data
+
frame_size
*
frame_size
*
2
,
frame_size
,
0
,
reset_hidden_prev_grad_data
,
frame_size
);
// backward for unactivated reset gate
ActGradCompute
(
context
.
Attr
<
int
>
(
"gate_activation"
),
place
,
r
,
r
,
d_g
.
slice
(
r_offsets
,
extents
),
d_r_h_p
*
h_p
);
// backward for weight
if
(
weight_grad
)
{
T
*
weight_grad_data
=
weight_grad
->
mutable_data
<
T
>
(
context
.
GetPlace
());
// backward for state_weight
math
::
gemm
<
Place
,
T
>
(
context
.
device_context
(),
true
,
false
,
frame_size
,
frame_size
,
batch_size
,
1
,
reset_hidden_prev_data
,
frame_size
,
gate_grad_data
+
frame_size
*
2
,
frame_size
*
3
,
0
,
weight_grad_data
+
frame_size
*
frame_size
*
2
,
frame_size
);
// backward for unactivated reset gate
ActGradCompute
(
context
.
Attr
<
int
>
(
"gate_activation"
),
place
,
r
,
r
,
d_g
.
slice
(
r_offsets
,
extents
),
d_r_h_p
*
h_p
);
// backward for update_gate_weight and reset_gate_weight
math
::
gemm
<
Place
,
T
>
(
context
.
device_context
(),
true
,
false
,
frame_size
,
frame_size
*
2
,
batch_size
,
1
,
hidden_prev_data
,
frame_size
,
gate_grad_data
,
frame_size
*
3
,
0
,
weight_grad_data
,
frame_size
*
2
);
}
// backward for hidden_prev
if
(
hidden_prev_grad
)
{
T
*
hidden_prev_grad_data
=
hidden_prev_grad
->
mutable_data
<
T
>
(
context
.
GetPlace
());
auto
d_h_p
=
EigenMatrix
<
T
>::
From
(
*
hidden_prev_grad
);
d_h_p
.
device
(
place
)
=
d_r_h_p
*
r
+
d_h
*
(
u
.
constant
(
T
(
1
))
-
u
);
math
::
gemm
<
Place
,
T
>
(
context
.
device_context
(),
false
,
true
,
batch_size
,
frame_size
,
frame_size
*
2
,
1
,
gate_grad_data
,
frame_size
*
3
,
weight_data
,
frame_size
*
2
,
1
,
hidden_prev_grad_data
,
frame_size
);
}
// backward for input
if
(
input_grad
)
{
input_grad
->
mutable_data
<
T
>
(
context
.
GetPlace
());
auto
d_x
=
EigenMatrix
<
T
>::
From
(
*
input_grad
);
d_x
.
device
(
place
)
=
d_g
;
}
// backward for bias
if
(
bias_grad
)
{
bias_grad
->
mutable_data
<
T
>
(
context
.
GetPlace
());
...
...
python/paddle/v2/fluid/tests/test_gru_unit_op.py
浏览文件 @
b6b7ab63
...
...
@@ -28,8 +28,8 @@ def relu(x):
class
TestGRUUnitOp
(
OpTest
):
batch_size
=
3
frame_size
=
5
batch_size
=
5
frame_size
=
10
activate
=
{
GRUActivationType
.
identity
:
identity
,
GRUActivationType
.
sigmoid
:
sigmoid
,
...
...
@@ -92,9 +92,7 @@ class TestGRUUnitOp(OpTest):
self
.
check_output
()
def
test_check_grad
(
self
):
self
.
check_grad
(
[
'Input'
,
'HiddenPrev'
,
'Weight'
],
[
'Hidden'
],
max_relative_error
=
0.007
)
self
.
check_grad
([
'Input'
,
'HiddenPrev'
,
'Weight'
],
[
'Hidden'
])
class
TestGRUUnitOpWithBias
(
TestGRUUnitOp
):
...
...
@@ -110,9 +108,12 @@ class TestGRUUnitOpWithBias(TestGRUUnitOp):
}
def
test_check_grad
(
self
):
self
.
check_grad
([
'Input'
,
'HiddenPrev'
,
'Weight'
,
'Bias'
],
[
'Hidden'
])
def
test_check_grad_ingore_input
(
self
):
self
.
check_grad
(
[
'
Input'
,
'
HiddenPrev'
,
'Weight'
,
'Bias'
],
[
'Hidden'
],
max_relative_error
=
0.007
)
[
'HiddenPrev'
,
'Weight'
,
'Bias'
],
[
'Hidden'
],
no_grad_set
=
set
(
'Input'
)
)
if
__name__
==
'__main__'
:
...
...
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