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
极客李华
Paddle
提交
217db273
P
Paddle
项目概览
极客李华
/
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
217db273
编写于
3月 05, 2019
作者:
D
dengkaipeng
浏览文件
操作
浏览文件
下载
电子邮件补丁
差异文件
add mkldnn support. test=develop
上级
6cb66721
变更
5
显示空白变更内容
内联
并排
Showing
5 changed file
with
111 addition
and
48 deletion
+111
-48
paddle/fluid/operators/mkldnn/softmax_mkldnn_op.cc
paddle/fluid/operators/mkldnn/softmax_mkldnn_op.cc
+93
-35
paddle/fluid/operators/softmax_cudnn_op.cu.cc
paddle/fluid/operators/softmax_cudnn_op.cu.cc
+0
-1
paddle/fluid/operators/softmax_op.cc
paddle/fluid/operators/softmax_op.cc
+6
-5
python/paddle/fluid/layers/nn.py
python/paddle/fluid/layers/nn.py
+11
-6
python/paddle/fluid/tests/unittests/test_layers.py
python/paddle/fluid/tests/unittests/test_layers.py
+1
-1
未找到文件。
paddle/fluid/operators/mkldnn/softmax_mkldnn_op.cc
浏览文件 @
217db273
...
...
@@ -110,28 +110,51 @@ class SoftmaxMKLDNNKernel : public paddle::framework::OpKernel<T> {
"It must use CPUPlace."
);
auto
&
dev_ctx
=
ctx
.
template
device_context
<
MKLDNNDeviceContext
>();
auto
mkldnn_engine
=
dev_ctx
.
GetEngine
();
const
Tensor
*
input
=
ctx
.
Input
<
Tensor
>
(
"X"
);
Tensor
*
outp
ut
=
ctx
.
Output
<
Tensor
>
(
"Out"
);
const
Tensor
*
X
=
ctx
.
Input
<
Tensor
>
(
"X"
);
Tensor
*
O
ut
=
ctx
.
Output
<
Tensor
>
(
"Out"
);
PADDLE_ENFORCE_EQ
(
input
->
dims
(),
outp
ut
->
dims
(),
X
->
dims
(),
O
ut
->
dims
(),
"The shape of softmax's input and output must be identical."
);
const
int
axis
=
ctx
.
Attr
<
int
>
(
"axis"
);
int
rank
=
X
->
dims
().
size
();
// make sure 'output' holds memory, which will be shared by
// 'flattened_output' later.
output
->
mutable_data
<
T
>
(
ctx
.
GetPlace
());
Out
->
mutable_data
<
T
>
(
ctx
.
GetPlace
());
std
::
vector
<
int
>
perm
,
shape
;
CalcTransPermAndShapeByAxis
(
*
X
,
axis
,
&
perm
,
&
shape
);
Tensor
X_2d
,
Out_2d
;
Tensor
X_trans
,
Out_trans
;
if
(
axis
!=
-
1
&&
axis
!=
rank
-
1
)
{
X_trans
.
mutable_data
<
T
>
(
framework
::
make_ddim
(
shape
),
ctx
.
GetPlace
());
Out_trans
.
mutable_data
<
T
>
(
framework
::
make_ddim
(
shape
),
ctx
.
GetPlace
());
TransCompute
<
MKLDNNDeviceContext
,
T
>
(
rank
,
dev_ctx
,
*
X
,
&
X_trans
,
perm
);
TransCompute
<
MKLDNNDeviceContext
,
T
>
(
rank
,
dev_ctx
,
*
Out
,
&
Out_trans
,
perm
);
X_2d
=
framework
::
ReshapeToMatrix
(
X_trans
,
rank
-
1
);
Out_2d
=
framework
::
ReshapeToMatrix
(
Out_trans
,
rank
-
1
);
}
else
{
X_2d
=
framework
::
ReshapeToMatrix
(
*
X
,
rank
-
1
);
Out_2d
=
framework
::
ReshapeToMatrix
(
*
Out
,
rank
-
1
);
}
// flatten input and output to 2-D matrixs
auto
dims
=
input
->
dims
();
// input and output share the same shape
auto
flattened_dims
=
framework
::
flatten_to_2d
(
dims
,
dims
.
size
()
-
1
);
framework
::
Tensor
flattened_input
;
framework
::
Tensor
flattened_output
;
flattened_input
.
ShareDataWith
(
*
input
).
Resize
(
flattened_dims
);
flattened_output
.
ShareDataWith
(
*
output
).
Resize
(
flattened_dims
);
const
T
*
input_data
=
flattened_input
.
data
<
T
>
();
T
*
output_data
=
flattened_output
.
mutable_data
<
T
>
(
ctx
.
GetPlace
());
std
::
vector
<
int
>
src_tz
=
paddle
::
framework
::
vectorize2int
(
flattened_dims
);
// auto dims = input->dims(); // input and output share the same shape
// auto flattened_dims = framework::flatten_to_2d(dims, dims.size() - 1);
// framework::Tensor flattened_input;
// framework::Tensor flattened_output;
// flattened_input.ShareDataWith(*input).Resize(flattened_dims);
// flattened_output.ShareDataWith(*output).Resize(flattened_dims);
// const T* input_data = flattened_input.data<T>();
// T* output_data = flattened_output.mutable_data<T>(ctx.GetPlace());
const
T
*
input_data
=
X_2d
.
data
<
T
>
();
T
*
output_data
=
Out_2d
.
mutable_data
<
T
>
(
ctx
.
GetPlace
());
// std::vector<int> src_tz = paddle::framework::vectorize2int(flattened_dims);
std
::
vector
<
int
>
src_tz
=
paddle
::
framework
::
vectorize2int
(
X_2d
.
dims
());
std
::
vector
<
int
>
dst_tz
=
src_tz
;
// Same memory descriptor to be used for input and output
memory
::
dims
softmax_tz
=
{
src_tz
[
0
],
src_tz
[
1
]};
...
...
@@ -178,6 +201,10 @@ class SoftmaxMKLDNNKernel : public paddle::framework::OpKernel<T> {
output_data
[
i
]
<
threshold
?
threshold
:
output_data
[
i
];
}
}
if
(
axis
!=
-
1
&&
axis
!=
rank
-
1
)
{
TransCompute
<
MKLDNNDeviceContext
,
T
>
(
rank
,
dev_ctx
,
Out_trans
,
Out
,
perm
);
}
}
};
...
...
@@ -190,33 +217,60 @@ class SoftmaxMKLDNNGradKernel : public paddle::framework::OpKernel<T> {
auto
&
dev_ctx
=
ctx
.
template
device_context
<
MKLDNNDeviceContext
>();
auto
mkldnn_engine
=
dev_ctx
.
GetEngine
();
const
Tensor
*
outp
ut
=
ctx
.
Input
<
Tensor
>
(
"Out"
);
auto
*
d
o
ut
=
ctx
.
template
Input
<
Tensor
>(
framework
::
GradVarName
(
"Out"
));
auto
*
d
x
=
const
Tensor
*
O
ut
=
ctx
.
Input
<
Tensor
>
(
"Out"
);
auto
*
d
O
ut
=
ctx
.
template
Input
<
Tensor
>(
framework
::
GradVarName
(
"Out"
));
auto
*
d
X
=
ctx
.
template
Output
<
framework
::
Tensor
>(
framework
::
GradVarName
(
"X"
));
PADDLE_ENFORCE_EQ
(
d
out
->
dims
(),
dx
->
dims
(),
d
Out
->
dims
(),
dX
->
dims
(),
"The shape of softmax_grad's input and output must be identical."
);
const
int
axis
=
ctx
.
Attr
<
int
>
(
"axis"
);
int
rank
=
Out
->
dims
().
size
();
// make sure 'dx' holds memory, which will be shared by 'flattened_dx'
// later.
dx
->
template
mutable_data
<
T
>(
ctx
.
GetPlace
());
auto
dims
=
dout
->
dims
();
// input and output share the same shape
auto
flattened_dims
=
framework
::
flatten_to_2d
(
dims
,
dims
.
size
()
-
1
);
framework
::
Tensor
flattened_output
;
framework
::
Tensor
flattened_dout
;
framework
::
Tensor
flattened_dx
;
flattened_output
.
ShareDataWith
(
*
output
).
Resize
(
flattened_dims
);
flattened_dout
.
ShareDataWith
(
*
dout
).
Resize
(
flattened_dims
);
flattened_dx
.
ShareDataWith
(
*
dx
).
Resize
(
flattened_dims
);
const
T
*
dst_data
=
flattened_output
.
data
<
T
>
();
const
T
*
diff_dst_ptr
=
flattened_dout
.
template
data
<
T
>();
T
*
diff_src_ptr
=
flattened_dx
.
template
mutable_data
<
T
>(
ctx
.
GetPlace
());
std
::
vector
<
int
>
dst_tz
=
paddle
::
framework
::
vectorize2int
(
flattened_dims
);
dX
->
template
mutable_data
<
T
>(
ctx
.
GetPlace
());
std
::
vector
<
int
>
perm
,
shape
;
CalcTransPermAndShapeByAxis
(
*
dX
,
axis
,
&
perm
,
&
shape
);
Tensor
dX_2d
,
Out_2d
,
dOut_2d
;
Tensor
dX_trans
,
Out_trans
,
dOut_trans
;
if
(
axis
!=
-
1
&&
axis
!=
rank
-
1
)
{
dX_trans
.
mutable_data
<
T
>
(
framework
::
make_ddim
(
shape
),
ctx
.
GetPlace
());
Out_trans
.
mutable_data
<
T
>
(
framework
::
make_ddim
(
shape
),
ctx
.
GetPlace
());
dOut_trans
.
mutable_data
<
T
>
(
framework
::
make_ddim
(
shape
),
ctx
.
GetPlace
());
TransCompute
<
MKLDNNDeviceContext
,
T
>
(
rank
,
dev_ctx
,
*
dX
,
&
dX_trans
,
perm
);
TransCompute
<
MKLDNNDeviceContext
,
T
>
(
rank
,
dev_ctx
,
*
Out
,
&
Out_trans
,
perm
);
TransCompute
<
MKLDNNDeviceContext
,
T
>
(
rank
,
dev_ctx
,
*
dOut
,
&
dOut_trans
,
perm
);
dX_2d
=
framework
::
ReshapeToMatrix
(
dX_trans
,
rank
-
1
);
Out_2d
=
framework
::
ReshapeToMatrix
(
Out_trans
,
rank
-
1
);
dOut_2d
=
framework
::
ReshapeToMatrix
(
dOut_trans
,
rank
-
1
);
}
else
{
dX_2d
=
framework
::
ReshapeToMatrix
(
*
dX
,
rank
-
1
);
Out_2d
=
framework
::
ReshapeToMatrix
(
*
Out
,
rank
-
1
);
dOut_2d
=
framework
::
ReshapeToMatrix
(
*
dOut
,
rank
-
1
);
}
// auto dims = dout->dims(); // input and output share the same shape
// auto flattened_dims = framework::flatten_to_2d(dims, dims.size() - 1);
// framework::Tensor flattened_output;
// framework::Tensor flattened_dout;
// framework::Tensor flattened_dx;
// flattened_output.ShareDataWith(*output).Resize(flattened_dims);
// flattened_dout.ShareDataWith(*dout).Resize(flattened_dims);
// flattened_dx.ShareDataWith(*dx).Resize(flattened_dims);
// const T* dst_data = flattened_output.data<T>();
// const T* diff_dst_ptr = flattened_dout.template data<T>();
// T* diff_src_ptr = flattened_dx.template mutable_data<T>(ctx.GetPlace());
const
T
*
dst_data
=
Out_2d
.
data
<
T
>
();
const
T
*
diff_dst_ptr
=
dOut_2d
.
template
data
<
T
>();
T
*
diff_src_ptr
=
dX_2d
.
template
mutable_data
<
T
>(
ctx
.
GetPlace
());
std
::
vector
<
int
>
dst_tz
=
paddle
::
framework
::
vectorize2int
(
Out_2d
.
dims
());
std
::
vector
<
int
>
src_tz
(
dst_tz
);
// Same memory descriptor to be used for input and output
...
...
@@ -261,6 +315,10 @@ class SoftmaxMKLDNNGradKernel : public paddle::framework::OpKernel<T> {
std
::
vector
<
primitive
>
pipeline
{
*
softmax_bwd_p
};
stream
(
stream
::
kind
::
eager
).
submit
(
pipeline
).
wait
();
if
(
axis
!=
-
1
&&
axis
!=
rank
-
1
)
{
TransCompute
<
MKLDNNDeviceContext
,
T
>
(
rank
,
dev_ctx
,
dX_trans
,
dX
,
perm
);
}
}
};
}
// namespace operators
...
...
paddle/fluid/operators/softmax_cudnn_op.cu.cc
浏览文件 @
217db273
...
...
@@ -28,7 +28,6 @@ class SoftmaxCUDNNKernel : public framework::OpKernel<T> {
auto
&
dev_ctx
=
context
.
template
device_context
<
platform
::
CUDADeviceContext
>();
auto
*
X
=
context
.
Input
<
Tensor
>
(
"X"
);
auto
*
Out
=
context
.
Output
<
Tensor
>
(
"Out"
);
// auto dims = X->dims();
const
int
axis
=
context
.
Attr
<
int
>
(
"axis"
);
int
rank
=
X
->
dims
().
size
();
...
...
paddle/fluid/operators/softmax_op.cc
浏览文件 @
217db273
...
...
@@ -85,10 +85,10 @@ class SoftmaxOpMaker : public framework::OpProtoAndCheckerMaker {
void
Make
()
override
{
AddInput
(
"X"
,
"The input tensor of softmax, "
"whose
last
dimension is the input_feature_dimensions."
);
"whose
:attr:`axis`
dimension is the input_feature_dimensions."
);
AddOutput
(
"Out"
,
"The normalized values with the same shape as X."
);
AddAttr
<
int
>
(
"axis"
,
"The dimension of Input(x) to perform softmax,"
"The dimension
index
of Input(x) to perform softmax,"
"default -1 for last dimension"
)
.
SetDefault
(
-
1
);
AddAttr
<
bool
>
(
...
...
@@ -115,12 +115,13 @@ Softmax Operator.
The input of the softmax operator is a tensor of any rank. The output tensor
has the same shape as the input.
The input tensor will first be logically flattened to a 2-D matrix. The matrix's
second dimension(row length) is as same as the last dimension of the input
The :attr:`axis` th dimension of the input tensor will be permuted to the last.
Then the input tensor will be logically flattened to a 2-D matrix. The matrix's
second dimension(row length) is as same as the :attr:`axis` dimension of the input
tensor, and the first dimension(column length) is the product of all other
dimensions of the input tensor. For each row of the matrix, the softmax operator
squashes the K-dimensional(K is the width of the matrix, which is also the size
of the input tensor's
last
dimension) vector of arbitrary real values to a
of the input tensor's
:attr:`axis`
dimension) vector of arbitrary real values to a
K-dimensional vector of real values in the range [0, 1] that add up to 1.
It computes the exponential of the given dimension and the sum of exponential
values of all the other dimensions in the K-dimensional vector input.
...
...
python/paddle/fluid/layers/nn.py
浏览文件 @
217db273
...
...
@@ -1819,17 +1819,18 @@ def sequence_softmax(input, use_cudnn=False, name=None):
return
softmax_out
def
softmax
(
input
,
use_cudnn
=
False
,
name
=
None
):
def
softmax
(
input
,
use_cudnn
=
False
,
name
=
None
,
axis
=-
1
):
"""
The input of the softmax operator is a tensor of any rank. The output tensor
has the same shape as the input.
The input tensor will first be logically flattened to a 2-D matrix. The matrix's
second dimension(row length) is as same as the last dimension of the input
The :attr:`axis` th dimension of the input tensor will be permuted to the last.
Then the input tensor will be logically flattened to a 2-D matrix. The matrix's
second dimension(row length) is as same as the :attr:`axis` th dimension of the input
tensor, and the first dimension(column length) is the product of all other
dimensions of the input tensor. For each row of the matrix, the softmax operator
squashes the K-dimensional(K is the width of the matrix, which is also the size
of the input tensor's
last
dimension) vector of arbitrary real values to a
of the input tensor's
:attr:`axis` th
dimension) vector of arbitrary real values to a
K-dimensional vector of real values in the range [0, 1] that add up to 1.
It computes the exponential of the given dimension and the sum of exponential
...
...
@@ -1851,6 +1852,7 @@ def softmax(input, use_cudnn=False, name=None):
False by default. Default: False
name (str|None): A name for this layer(optional). If set None, the layer
will be named automatically. Default: None.
axis (int): The index of dimension to perform softmax calculation. Default: -1.
Returns:
Variable: output of softmax
...
...
@@ -1860,7 +1862,7 @@ def softmax(input, use_cudnn=False, name=None):
.. code-block:: python
fc = fluid.layers.fc(input=x, size=10)
softmax = fluid.layers.softmax(input=fc)
softmax = fluid.layers.softmax(input=fc
, axis=1
)
"""
helper
=
LayerHelper
(
'softmax'
,
**
locals
())
...
...
@@ -1870,7 +1872,10 @@ def softmax(input, use_cudnn=False, name=None):
type
=
"softmax"
,
inputs
=
{
"X"
:
input
},
outputs
=
{
"Out"
:
softmax_out
},
attrs
=
{
"use_cudnn"
:
use_cudnn
})
attrs
=
{
"axis"
:
axis
,
"use_cudnn"
:
use_cudnn
})
return
softmax_out
...
...
python/paddle/fluid/tests/unittests/test_layers.py
浏览文件 @
217db273
...
...
@@ -513,7 +513,7 @@ class TestBook(unittest.TestCase):
with
program_guard
(
program
):
data
=
layers
.
data
(
name
=
'data'
,
shape
=
[
10
],
dtype
=
'float32'
)
hid
=
layers
.
fc
(
input
=
data
,
size
=
20
)
self
.
assertIsNotNone
(
layers
.
softmax
(
hid
))
self
.
assertIsNotNone
(
layers
.
softmax
(
hid
,
axis
=
1
))
print
(
str
(
program
))
def
test_space_to_depth
(
self
):
...
...
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