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
c26f2b21
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 >>
未验证
提交
c26f2b21
编写于
10月 15, 2018
作者:
T
Tao Luo
提交者:
GitHub
10月 15, 2018
浏览文件
操作
浏览文件
下载
差异文件
Merge pull request #13813 from sfraczek/sfraczek/conv-bn-fuse-pass-full-eigen
rewrite conv_bn fuse pass fully to eigen
上级
fededdda
50c5e9b0
变更
1
隐藏空白更改
内联
并排
Showing
1 changed file
with
35 addition
and
104 deletion
+35
-104
paddle/fluid/framework/ir/conv_bn_fuse_pass.cc
paddle/fluid/framework/ir/conv_bn_fuse_pass.cc
+35
-104
未找到文件。
paddle/fluid/framework/ir/conv_bn_fuse_pass.cc
浏览文件 @
c26f2b21
...
...
@@ -44,89 +44,6 @@ namespace ir {
GET_IR_NODE_FROM_SUBGRAPH(bn_saved_mean, bn_saved_mean, pattern_name); \
GET_IR_NODE_FROM_SUBGRAPH(bn_saved_variance, bn_saved_variance, pattern_name)
template
<
typename
UnaryOperation
>
LoDTensor
tensor_apply
(
const
LoDTensor
&
vec
,
UnaryOperation
f
)
{
LoDTensor
vec_y
;
vec_y
.
Resize
(
vec
.
dims
());
const
float
*
x
=
vec
.
data
<
float
>
();
float
*
y
=
vec_y
.
mutable_data
<
float
>
(
platform
::
CPUPlace
());
for
(
int64_t
i
=
0
;
i
<
vec
.
numel
();
i
++
)
{
y
[
i
]
=
f
(
x
[
i
]);
}
return
vec_y
;
}
void
tensor_apply_inplace
(
LoDTensor
*
vec
,
float
(
*
f
)(
float
))
{
float
*
data
=
vec
->
mutable_data
<
float
>
(
platform
::
CPUPlace
());
for
(
int64_t
i
=
0
;
i
<
vec
->
numel
();
i
++
)
{
data
[
i
]
=
f
(
data
[
i
]);
}
}
template
<
typename
BinaryOperation
>
LoDTensor
tensor_apply_eltwise
(
const
LoDTensor
&
vec_a
,
const
LoDTensor
&
vec_b
,
BinaryOperation
f
)
{
PADDLE_ENFORCE_EQ
(
vec_a
.
dims
(),
vec_b
.
dims
());
LoDTensor
vec_y
;
vec_y
.
Resize
(
vec_a
.
dims
());
const
float
*
a
=
vec_a
.
data
<
float
>
();
const
float
*
b
=
vec_b
.
data
<
float
>
();
float
*
y
=
vec_y
.
mutable_data
<
float
>
(
platform
::
CPUPlace
());
for
(
int64_t
i
=
0
;
i
<
vec_a
.
numel
();
i
++
)
{
y
[
i
]
=
f
(
a
[
i
],
b
[
i
]);
}
return
vec_y
;
}
template
<
typename
BinaryOperation
>
LoDTensor
tensor_apply_eltwise_broadcast
(
const
LoDTensor
&
vec_a
,
const
LoDTensor
&
vec_b
,
BinaryOperation
f
)
{
PADDLE_ENFORCE_EQ
(
vec_a
.
dims
().
size
(),
2
);
PADDLE_ENFORCE_EQ
(
vec_b
.
dims
().
size
(),
2
);
PADDLE_ENFORCE_EQ
(
vec_a
.
dims
()[
0
],
vec_b
.
dims
()[
0
]);
PADDLE_ENFORCE_EQ
(
vec_b
.
dims
()[
1
],
1
);
LoDTensor
vec_y
;
vec_y
.
Resize
(
vec_a
.
dims
());
const
float
*
a
=
vec_a
.
data
<
float
>
();
const
float
*
b
=
vec_b
.
data
<
float
>
();
float
*
y
=
vec_y
.
mutable_data
<
float
>
(
platform
::
CPUPlace
());
size_t
a_height
=
vec_a
.
dims
()[
0
];
size_t
a_width
=
vec_a
.
dims
()[
1
];
for
(
size_t
h
=
0
;
h
<
a_height
;
h
++
)
{
for
(
size_t
w
=
0
;
w
<
a_width
;
++
w
)
{
*
(
y
++
)
=
f
(
*
(
a
++
),
b
[
h
]);
}
}
return
vec_y
;
}
// reshape to two dimensions {A, B * C * ...}
void
make_tensor_2d
(
LoDTensor
*
tensor_to_reshape
)
{
auto
dims_count
=
tensor_to_reshape
->
dims
().
size
();
PADDLE_ENFORCE_GT
(
dims_count
,
0
);
int
size2
=
1
;
for
(
int
i
=
1
;
i
<
dims_count
;
i
++
)
{
size2
*=
tensor_to_reshape
->
dims
()[
i
];
}
tensor_to_reshape
->
Resize
(
make_ddim
({
tensor_to_reshape
->
dims
()[
0
],
size2
}));
}
void
recompute_conv_weights
(
LoDTensor
*
weights
,
LoDTensor
*
tmp
)
{
// remember the weights tensor shape {A, B, C, ...}
auto
weights_shape
=
weights
->
dims
();
// reduce the weights to 2d {A, B * C * ...}
make_tensor_2d
(
weights
);
// make tmp tensor 2d by adding 1 as second dim {A, 1}
make_tensor_2d
(
tmp
);
*
weights
=
tensor_apply_eltwise_broadcast
(
*
weights
,
*
tmp
,
std
::
multiplies
<
float
>
());
// reshape weights to the original dims {A, B, C, ...}
weights
->
Resize
(
weights_shape
);
}
void
recompute_bias_and_weights
(
const
Scope
*
scope
,
ir
::
Node
*
conv_weight
,
//
const
ir
::
Node
&
bn_scale
,
//
...
...
@@ -135,6 +52,13 @@ void recompute_bias_and_weights(const Scope* scope,
const
ir
::
Node
&
bn_variance
,
//
LoDTensor
*
eltwise_y_in_tensor
,
//
float
epsilon
)
{
using
EigenVectorArrayMap
=
Eigen
::
Map
<
Eigen
::
Array
<
float
,
Eigen
::
Dynamic
,
1
>>
;
using
ConstEigenVectorArrayMap
=
Eigen
::
Map
<
const
Eigen
::
Array
<
float
,
Eigen
::
Dynamic
,
1
>>
;
using
EigenMatrixArrayMap
=
Eigen
::
Map
<
Eigen
::
Array
<
float
,
Eigen
::
Dynamic
,
Eigen
::
Dynamic
,
Eigen
::
RowMajor
>>
;
// Re-compute bias of conv2d from BN
PADDLE_ENFORCE_EQ
(
eltwise_y_in_tensor
->
dims
(),
bn_bias_tensor
.
dims
());
...
...
@@ -143,31 +67,38 @@ void recompute_bias_and_weights(const Scope* scope,
scope
->
FindVar
(
bn_variance
.
Name
())
->
GetMutable
<
LoDTensor
>
();
auto
*
mean_tensor
=
scope
->
FindVar
(
bn_mean
.
Name
())
->
GetMutable
<
LoDTensor
>
();
auto
std_tensor
=
LoDTensor
();
std_tensor
.
Resize
(
bn_bias_tensor
.
dims
());
std_tensor
=
tensor_apply
(
*
variance_tensor
,
[
&
](
float
x
)
{
return
x
+
epsilon
;
});
ConstEigenVectorArrayMap
scale_array
(
scale_tensor
->
data
<
float
>
(),
scale_tensor
->
numel
(),
1
);
EigenVectorArrayMap
variance_array
(
variance_tensor
->
mutable_data
<
float
>
(
platform
::
CPUPlace
()),
variance_tensor
->
numel
(),
1
);
ConstEigenVectorArrayMap
mean_array
(
mean_tensor
->
data
<
float
>
(),
mean_tensor
->
numel
(),
1
);
ConstEigenVectorArrayMap
bn_bias_array
(
bn_bias_tensor
.
data
<
float
>
(),
bn_bias_tensor
.
numel
(),
1
);
using
EigenVectorArrayMap
=
Eigen
::
Map
<
Eigen
::
Array
<
float
,
Eigen
::
Dynamic
,
1
>>
;
// variance will not be used anymore, so make it std_array and then tmp_array
variance_array
+=
epsilon
;
variance_array
=
variance_array
.
sqrt
();
variance_array
=
scale_array
/
variance_array
;
EigenVectorArrayMap
eltwise_y_in_array
(
eltwise_y_in_tensor
->
mutable_data
<
float
>
(
platform
::
CPUPlace
()),
eltwise_y_in_tensor
->
numel
(),
1
);
EigenVectorArrayMap
std_vec
(
std_tensor
.
mutable_data
<
float
>
(
platform
::
CPUPlace
()),
std_tensor
.
numel
(),
1
);
std_vec
=
std_vec
.
sqrt
();
auto
tmp_tensor
=
tensor_apply_eltwise
(
*
scale_tensor
,
std_tensor
,
std
::
divides
<
float
>
());
auto
tensor_minus
=
tensor_apply_eltwise
(
*
eltwise_y_in_tensor
,
*
mean_tensor
,
std
::
minus
<
float
>
());
auto
tensor_mul
=
tensor_apply_eltwise
(
tensor_minus
,
tmp_tensor
,
std
::
multiplies
<
float
>
());
*
eltwise_y_in_tensor
=
tensor_apply_eltwise
(
tensor_mul
,
bn_bias_tensor
,
std
::
plus
<
float
>
());
eltwise_y_in_array
=
((
eltwise_y_in_array
-
mean_array
)
*
variance_array
)
+
bn_bias_array
;
// Re-compute weight of conv2d from BN
auto
*
current_param
=
scope
->
FindVar
(
conv_weight
->
Name
())
->
GetMutable
<
LoDTensor
>
();
recompute_conv_weights
(
current_param
,
&
tmp_tensor
);
auto
*
weights
=
scope
->
FindVar
(
conv_weight
->
Name
())
->
GetMutable
<
LoDTensor
>
();
auto
weights_shape
=
weights
->
dims
();
auto
weights_shape_2d
=
flatten_to_2d
(
weights_shape
,
1
);
EigenMatrixArrayMap
weights_array_2d
(
weights
->
mutable_data
<
float
>
(
platform
::
CPUPlace
()),
weights_shape_2d
[
0
],
weights_shape_2d
[
1
]);
weights_array_2d
.
colwise
()
*=
variance_array
;
}
std
::
unique_ptr
<
ir
::
Graph
>
ConvBNFusePass
::
ApplyImpl
(
...
...
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