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
兽拳
Paddle
提交
424dd2fc
P
Paddle
项目概览
兽拳
/
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 >>
未验证
提交
424dd2fc
编写于
4月 05, 2018
作者:
Q
Qiao Longfei
提交者:
GitHub
4月 05, 2018
浏览文件
操作
浏览文件
下载
差异文件
Merge pull request #9597 from jacquesqiao/sgd-support-update-selected-rows
Sgd support update selected rows
上级
7bf82f82
ff4208e6
变更
10
隐藏空白更改
内联
并排
Showing
10 changed file
with
201 addition
and
83 deletion
+201
-83
paddle/fluid/framework/operator.cc
paddle/fluid/framework/operator.cc
+11
-0
paddle/fluid/framework/operator.h
paddle/fluid/framework/operator.h
+2
-0
paddle/fluid/framework/selected_rows.cc
paddle/fluid/framework/selected_rows.cc
+5
-1
paddle/fluid/framework/selected_rows.h
paddle/fluid/framework/selected_rows.h
+13
-1
paddle/fluid/operators/lookup_table_op.cc
paddle/fluid/operators/lookup_table_op.cc
+4
-18
paddle/fluid/operators/lookup_table_op.h
paddle/fluid/operators/lookup_table_op.h
+8
-10
paddle/fluid/operators/sgd_op.cc
paddle/fluid/operators/sgd_op.cc
+7
-6
paddle/fluid/operators/sgd_op.h
paddle/fluid/operators/sgd_op.h
+80
-43
python/paddle/fluid/tests/unittests/test_lookup_table_op.py
python/paddle/fluid/tests/unittests/test_lookup_table_op.py
+4
-4
python/paddle/fluid/tests/unittests/test_sgd_op.py
python/paddle/fluid/tests/unittests/test_sgd_op.py
+67
-0
未找到文件。
paddle/fluid/framework/operator.cc
浏览文件 @
424dd2fc
...
...
@@ -35,6 +35,17 @@ std::vector<std::tuple<platform::Place, LibraryType>> kKernelPriority = {
std
::
make_tuple
(
platform
::
CPUPlace
(),
LibraryType
::
kPlain
),
};
proto
::
VarType
::
Type
GetDataTypeOfVar
(
const
Variable
*
var
)
{
if
(
var
->
IsType
<
framework
::
LoDTensor
>
())
{
return
framework
::
ToDataType
(
var
->
Get
<
framework
::
LoDTensor
>
().
type
());
}
else
if
(
var
->
IsType
<
framework
::
SelectedRows
>
())
{
return
framework
::
ToDataType
(
var
->
Get
<
framework
::
SelectedRows
>
().
value
().
type
());
}
else
{
PADDLE_THROW
(
"Var should be LoDTensor or SelectedRows"
);
}
}
static
DDim
GetDims
(
const
Scope
&
scope
,
const
std
::
string
&
name
)
{
Variable
*
var
=
scope
.
FindVar
(
name
);
if
(
var
==
nullptr
)
{
...
...
paddle/fluid/framework/operator.h
浏览文件 @
424dd2fc
...
...
@@ -61,6 +61,8 @@ inline std::string GradVarName(const std::string& var_name) {
return
var_name
+
kGradVarSuffix
;
}
proto
::
VarType
::
Type
GetDataTypeOfVar
(
const
Variable
*
var
);
class
OperatorBase
;
class
ExecutionContext
;
...
...
paddle/fluid/framework/selected_rows.cc
浏览文件 @
424dd2fc
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved.
/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
...
...
@@ -13,6 +16,7 @@ limitations under the License. */
namespace
paddle
{
namespace
framework
{
void
SerializeToStream
(
std
::
ostream
&
os
,
const
SelectedRows
&
selected_rows
,
const
platform
::
DeviceContext
&
dev_ctx
)
{
{
// the 1st field, uint32_t version
...
...
paddle/fluid/framework/selected_rows.h
浏览文件 @
424dd2fc
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved.
/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
...
...
@@ -47,6 +50,15 @@ class SelectedRows {
void
set_rows
(
const
Vector
<
int64_t
>&
rows
)
{
rows_
=
rows
;
}
/**
* get the index of id in rows
*/
int64_t
index
(
int64_t
id
)
const
{
auto
it
=
std
::
find
(
rows_
.
begin
(),
rows_
.
end
(),
id
);
PADDLE_ENFORCE
(
it
!=
rows_
.
end
(),
"id should be in rows"
);
return
static_cast
<
int64_t
>
(
std
::
distance
(
rows_
.
begin
(),
it
));
}
DDim
GetCompleteDims
()
const
{
std
::
vector
<
int64_t
>
dims
=
vectorize
(
value_
->
dims
());
dims
[
0
]
=
height_
;
...
...
paddle/fluid/operators/lookup_table_op.cc
浏览文件 @
424dd2fc
...
...
@@ -18,22 +18,6 @@ limitations under the License. */
namespace
paddle
{
namespace
operators
{
static
inline
framework
::
OpKernelType
ExpectedKernelType
(
const
framework
::
ExecutionContext
&
ctx
)
{
auto
*
table_var
=
ctx
.
InputVar
(
"W"
);
if
(
table_var
->
IsType
<
LoDTensor
>
())
{
return
framework
::
OpKernelType
(
framework
::
ToDataType
(
table_var
->
Get
<
LoDTensor
>
().
type
()),
ctx
.
device_context
());
}
else
if
(
table_var
->
IsType
<
SelectedRows
>
())
{
return
framework
::
OpKernelType
(
framework
::
ToDataType
(
table_var
->
Get
<
SelectedRows
>
().
value
().
type
()),
ctx
.
device_context
());
}
else
{
PADDLE_THROW
(
"W should be LoDTensor or SelectedRows"
);
}
}
class
LookupTableOp
:
public
framework
::
OperatorWithKernel
{
public:
using
framework
::
OperatorWithKernel
::
OperatorWithKernel
;
...
...
@@ -67,7 +51,8 @@ class LookupTableOp : public framework::OperatorWithKernel {
protected:
framework
::
OpKernelType
GetExpectedKernelType
(
const
framework
::
ExecutionContext
&
ctx
)
const
override
{
return
ExpectedKernelType
(
ctx
);
auto
data_type
=
framework
::
GetDataTypeOfVar
(
ctx
.
InputVar
(
"W"
));
return
framework
::
OpKernelType
(
data_type
,
ctx
.
device_context
());
}
};
...
...
@@ -138,7 +123,8 @@ class LookupTableOpGrad : public framework::OperatorWithKernel {
protected:
framework
::
OpKernelType
GetExpectedKernelType
(
const
framework
::
ExecutionContext
&
ctx
)
const
override
{
return
ExpectedKernelType
(
ctx
);
auto
data_type
=
framework
::
GetDataTypeOfVar
(
ctx
.
InputVar
(
"W"
));
return
framework
::
OpKernelType
(
data_type
,
ctx
.
device_context
());
}
};
...
...
paddle/fluid/operators/lookup_table_op.h
浏览文件 @
424dd2fc
...
...
@@ -30,13 +30,7 @@ using LoDTensor = framework::LoDTensor;
using
SelectedRows
=
framework
::
SelectedRows
;
using
DDim
=
framework
::
DDim
;
static
constexpr
int64_t
kNoPadding
=
-
1
;
inline
size_t
getIndex
(
const
std
::
vector
<
int64_t
>
&
rows
,
int64_t
value
)
{
auto
it
=
std
::
find
(
rows
.
begin
(),
rows
.
end
(),
value
);
PADDLE_ENFORCE
(
it
!=
rows
.
end
(),
"id should be in rows"
);
return
static_cast
<
size_t
>
(
std
::
distance
(
rows
.
begin
(),
it
));
}
constexpr
int64_t
kNoPadding
=
-
1
;
template
<
typename
T
>
class
LookupTableKernel
:
public
framework
::
OpKernel
<
T
>
{
...
...
@@ -55,7 +49,9 @@ class LookupTableKernel : public framework::OpKernel<T> {
auto
*
table_t
=
context
.
Input
<
SelectedRows
>
(
"W"
);
table_dim
=
table_t
->
value
().
dims
();
}
else
{
PADDLE_THROW
(
"table only support LoDTensor and SelectedRows"
);
PADDLE_THROW
(
"The parameter W of a LookupTable "
"must be either LoDTensor or SelectedRows"
);
}
int64_t
*
ids
;
...
...
@@ -107,7 +103,7 @@ class LookupTableKernel : public framework::OpKernel<T> {
memset
(
output
+
i
*
row_width
,
0
,
row_width
*
sizeof
(
T
));
}
else
{
PADDLE_ENFORCE_GE
(
ids
[
i
],
0
);
auto
id_index
=
getIndex
(
table_t
.
rows
(),
ids
[
i
]);
auto
id_index
=
table_t
.
index
(
ids
[
i
]);
memcpy
(
output
+
i
*
row_width
,
table
+
id_index
*
row_width
,
row_width
*
sizeof
(
T
));
}
...
...
@@ -128,7 +124,9 @@ class LookupTableGradKernel : public framework::OpKernel<T> {
auto
*
table_t
=
context
.
Input
<
SelectedRows
>
(
"W"
);
table_dim
=
table_t
->
value
().
dims
();
}
else
{
PADDLE_THROW
(
"table only support LoDTensor and SelectedRows"
);
PADDLE_THROW
(
"The parameter W of a LookupTable "
"must be either LoDTensor or SelectedRows"
);
}
bool
is_sparse
=
context
.
Attr
<
bool
>
(
"is_sparse"
);
...
...
paddle/fluid/operators/sgd_op.cc
浏览文件 @
424dd2fc
...
...
@@ -43,9 +43,8 @@ class SGDOp : public framework::OperatorWithKernel {
protected:
framework
::
OpKernelType
GetExpectedKernelType
(
const
framework
::
ExecutionContext
&
ctx
)
const
override
{
return
framework
::
OpKernelType
(
framework
::
ToDataType
(
ctx
.
Input
<
framework
::
LoDTensor
>
(
"Param"
)
->
type
()),
ctx
.
GetPlace
());
auto
data_type
=
framework
::
GetDataTypeOfVar
(
ctx
.
InputVar
(
"Param"
));
return
framework
::
OpKernelType
(
data_type
,
ctx
.
device_context
());
}
};
...
...
@@ -53,10 +52,12 @@ class SGDOpMaker : public framework::OpProtoAndCheckerMaker {
public:
SGDOpMaker
(
OpProto
*
proto
,
OpAttrChecker
*
op_checker
)
:
OpProtoAndCheckerMaker
(
proto
,
op_checker
)
{
AddInput
(
"Param"
,
"(Tensor) Input parameter"
);
AddInput
(
"Param"
,
"(Tensor
or SelectedRows
) Input parameter"
);
AddInput
(
"LearningRate"
,
"(Tensor) Learning rate of SGD"
);
AddInput
(
"Grad"
,
"(Tensor) Input gradient"
);
AddOutput
(
"ParamOut"
,
"(Tensor) Output parameter"
);
AddInput
(
"Grad"
,
"(Tensor or SelectedRows) Input gradient"
);
AddOutput
(
"ParamOut"
,
"(Tensor or SelectedRows, same with Param) "
"Output parameter, should share the same memory with Param"
);
AddComment
(
R"DOC(
SGD operator
...
...
paddle/fluid/operators/sgd_op.h
浏览文件 @
424dd2fc
...
...
@@ -23,60 +23,97 @@ namespace operators {
template
<
typename
T
>
class
SGDOpKernel
:
public
framework
::
OpKernel
<
T
>
{
public:
void
Compute
(
const
framework
::
ExecutionContext
&
ctx
)
const
override
{
auto
*
param
=
ctx
.
Input
<
framework
::
Tensor
>
(
"Param"
);
auto
*
param_out
=
ctx
.
Output
<
framework
::
Tensor
>
(
"ParamOut"
);
auto
*
learning_rate
=
ctx
.
Input
<
framework
::
Tensor
>
(
"LearningRate"
);
auto
*
grad_var
=
ctx
.
InputVar
(
"Grad"
);
// Actually, all tensors are LoDTensor except SelectedRows.
if
(
grad_var
->
IsType
<
framework
::
LoDTensor
>
())
{
param_out
->
mutable_data
<
T
>
(
ctx
.
GetPlace
());
auto
*
grad
=
ctx
.
Input
<
framework
::
Tensor
>
(
"Grad"
);
auto
p
=
framework
::
EigenVector
<
T
>::
Flatten
(
*
param
);
auto
g
=
framework
::
EigenVector
<
T
>::
Flatten
(
*
grad
);
auto
o
=
framework
::
EigenVector
<
T
>::
Flatten
(
*
param_out
);
auto
*
lr
=
learning_rate
->
data
<
T
>
();
o
=
p
-
lr
[
0
]
*
g
;
}
else
if
(
grad_var
->
IsType
<
framework
::
SelectedRows
>
())
{
// TODO(qijun): In Sparse SGD operator, in-place update is enforced.
// This manual optimization brings difficulty to track data dependency.
// It's better to find a more elegant solution.
PADDLE_ENFORCE_EQ
(
param
,
param_out
);
auto
*
grad
=
ctx
.
Input
<
framework
::
SelectedRows
>
(
"Grad"
);
void
Compute
(
const
framework
::
ExecutionContext
&
ctx
)
const
override
{
const
auto
*
learning_rate
=
ctx
.
Input
<
framework
::
Tensor
>
(
"LearningRate"
);
const
auto
*
param_var
=
ctx
.
InputVar
(
"Param"
);
const
auto
*
grad_var
=
ctx
.
InputVar
(
"Grad"
);
if
(
param_var
->
IsType
<
framework
::
LoDTensor
>
())
{
const
auto
*
param
=
ctx
.
Input
<
framework
::
Tensor
>
(
"Param"
);
auto
*
param_out
=
ctx
.
Output
<
framework
::
Tensor
>
(
"ParamOut"
);
// Actually, all tensors are LoDTensor except SelectedRows.
if
(
grad_var
->
IsType
<
framework
::
LoDTensor
>
())
{
param_out
->
mutable_data
<
T
>
(
ctx
.
GetPlace
());
const
auto
*
grad
=
ctx
.
Input
<
framework
::
Tensor
>
(
"Grad"
);
auto
p
=
framework
::
EigenVector
<
T
>::
Flatten
(
*
param
);
auto
g
=
framework
::
EigenVector
<
T
>::
Flatten
(
*
grad
);
auto
o
=
framework
::
EigenVector
<
T
>::
Flatten
(
*
param_out
);
auto
*
lr
=
learning_rate
->
data
<
T
>
();
o
=
p
-
lr
[
0
]
*
g
;
}
else
if
(
grad_var
->
IsType
<
framework
::
SelectedRows
>
())
{
// TODO(qijun): In Sparse SGD operator, in-place update is enforced.
// This manual optimization brings difficulty to track data dependency.
// It's better to find a more elegant solution.
PADDLE_ENFORCE_EQ
(
param
,
param_out
);
const
auto
*
grad
=
ctx
.
Input
<
framework
::
SelectedRows
>
(
"Grad"
);
// for distributed training, a sparse var may be empty,
// just skip updating.
if
(
grad
->
rows
().
size
()
==
0
)
{
return
;
}
auto
grad_height
=
grad
->
height
();
auto
out_dims
=
param_out
->
dims
();
PADDLE_ENFORCE_EQ
(
grad_height
,
out_dims
[
0
]);
auto
&
grad_value
=
grad
->
value
();
auto
&
grad_rows
=
grad
->
rows
();
size_t
grad_row_numel
=
grad_value
.
numel
()
/
grad_rows
.
size
();
PADDLE_ENFORCE_EQ
(
grad_row_numel
,
param_out
->
numel
()
/
grad_height
);
auto
*
grad_data
=
grad_value
.
data
<
T
>
();
auto
*
out_data
=
param_out
->
data
<
T
>
();
auto
*
lr
=
learning_rate
->
data
<
T
>
();
for
(
size_t
i
=
0
;
i
<
grad_rows
.
size
();
i
++
)
{
PADDLE_ENFORCE
(
grad_rows
[
i
]
<
grad_height
,
"Input rows index should less than height"
);
for
(
int64_t
j
=
0
;
j
<
grad_row_numel
;
j
++
)
{
out_data
[
grad_rows
[
i
]
*
grad_row_numel
+
j
]
-=
lr
[
0
]
*
grad_data
[
i
*
grad_row_numel
+
j
];
}
}
}
else
{
PADDLE_THROW
(
"Unsupported Variable Type of Grad"
);
}
}
else
if
(
param_var
->
IsType
<
framework
::
SelectedRows
>
())
{
PADDLE_ENFORCE
(
grad_var
->
IsType
<
framework
::
SelectedRows
>
(),
"when param "
"is SelectedRows, gradient should also be SelectedRows"
);
const
auto
&
param
=
param_var
->
Get
<
framework
::
SelectedRows
>
();
auto
*
param_out
=
ctx
.
Output
<
framework
::
SelectedRows
>
(
"ParamOut"
);
const
auto
&
grad
=
grad_var
->
Get
<
framework
::
SelectedRows
>
();
// for distributed training, a sparse var may be empty,
// just skip updating.
if
(
grad
->
rows
().
size
()
==
0
)
{
if
(
grad
.
rows
().
size
()
==
0
)
{
return
;
}
auto
in_height
=
grad
->
height
();
auto
out_dims
=
param_out
->
dims
();
PADDLE_ENFORCE_EQ
(
in_height
,
out_dims
[
0
]);
auto
&
in_value
=
grad
->
value
();
auto
&
in_rows
=
grad
->
rows
();
size_t
param_row_width
=
param
.
value
().
numel
()
/
param
.
rows
().
size
();
size_t
grad_row_width
=
grad
.
value
().
numel
()
/
grad
.
rows
().
size
();
PADDLE_ENFORCE_EQ
(
param_row_width
,
grad_row_width
,
"param_row should have the same size with grad_row"
);
int64_t
in_row_numel
=
in_value
.
numel
()
/
in_rows
.
size
();
PADDLE_ENFORCE_EQ
(
in_row_numel
,
param_out
->
numel
()
/
in_height
);
auto
*
in_data
=
in_value
.
data
<
T
>
();
auto
*
out_data
=
param_out
->
data
<
T
>
();
auto
*
lr
=
learning_rate
->
data
<
T
>
();
for
(
size_t
i
=
0
;
i
<
in_rows
.
size
();
i
++
)
{
PADDLE_ENFORCE
(
in_rows
[
i
]
<
in_height
,
const
auto
*
lr
=
learning_rate
->
data
<
T
>
();
const
auto
*
grad_data
=
grad
.
value
().
data
<
T
>
();
auto
*
out_data
=
param_out
->
mutable_value
()
->
data
<
T
>
();
for
(
size_t
i
=
0
;
i
<
grad
.
rows
().
size
();
i
++
)
{
PADDLE_ENFORCE
(
grad
.
rows
()[
i
]
<
grad
.
height
(),
"Input rows index should less than height"
);
for
(
int64_t
j
=
0
;
j
<
in_row_numel
;
j
++
)
{
out_data
[
in_rows
[
i
]
*
in_row_numel
+
j
]
-=
lr
[
0
]
*
in_data
[
i
*
in_row_numel
+
j
];
int64_t
id_index
=
param
.
index
(
grad
.
rows
()[
i
]);
for
(
int64_t
j
=
0
;
j
<
grad_row_width
;
j
++
)
{
out_data
[
id_index
*
grad_row_width
+
j
]
-=
lr
[
0
]
*
grad_data
[
i
*
grad_row_width
+
j
];
}
}
}
else
{
PADDLE_THROW
(
"Unsupported Variable Type of
Grad
"
);
PADDLE_THROW
(
"Unsupported Variable Type of
Parameter
"
);
}
}
};
...
...
python/paddle/fluid/tests/unittests/test_lookup_table_op.py
浏览文件 @
424dd2fc
...
...
@@ -115,18 +115,18 @@ class TestLookupTableWIsSelectedRows(OpTest):
w_array
=
np
.
ones
((
len
(
rows
),
row_numel
)).
astype
(
"float32"
)
for
i
in
range
(
len
(
rows
)):
w_array
[
i
]
*=
i
ids
_tensor
=
w_selected_rows
.
get_tensor
()
ids
_tensor
.
set
(
w_array
,
place
)
w
_tensor
=
w_selected_rows
.
get_tensor
()
w
_tensor
.
set
(
w_array
,
place
)
# create Out Variable
O
ut_tensor
=
scope
.
var
(
'Out'
).
get_tensor
()
o
ut_tensor
=
scope
.
var
(
'Out'
).
get_tensor
()
# create and run lookup_table operator
lookup_table
=
Operator
(
"lookup_table"
,
W
=
'W'
,
Ids
=
'Ids'
,
Out
=
'Out'
)
lookup_table
.
run
(
scope
,
place
)
# get result from Out
result_array
=
np
.
array
(
O
ut_tensor
)
result_array
=
np
.
array
(
o
ut_tensor
)
# all(): return True if all elements of the iterable are true (or if the iterable is empty)
for
idx
,
row
in
enumerate
(
ids_array
):
assert
(
row
[
0
]
==
result_array
[
idx
]).
all
()
...
...
python/paddle/fluid/tests/unittests/test_sgd_op.py
浏览文件 @
424dd2fc
...
...
@@ -97,5 +97,72 @@ class TestSparseSGDOp(unittest.TestCase):
self
.
check_with_place
(
place
)
class
TestSGDOpOptimizeSelectedRows
(
unittest
.
TestCase
):
def
check_with_place
(
self
,
place
):
scope
=
core
.
Scope
()
row_width
=
12
# create and initialize Grad Variable
grad_height
=
10
grad_rows
=
[
0
,
4
,
7
]
grad_selected_rows
=
scope
.
var
(
'Grad'
).
get_selected_rows
()
grad_selected_rows
.
set_height
(
grad_height
)
grad_selected_rows
.
set_rows
(
grad_rows
)
grad_array
=
np
.
ones
((
len
(
grad_rows
),
row_width
)).
astype
(
"float32"
)
grad_array
[
0
,
0
]
=
2.0
grad_array
[
2
,
8
]
=
4.0
grad_tensor
=
grad_selected_rows
.
get_tensor
()
grad_tensor
.
set
(
grad_array
,
place
)
# create and initialize Param Variable
# create and initialize W Variable
param_rows
=
[
0
,
1
,
2
,
3
,
4
,
5
,
6
,
7
]
# init Param
w_selected_rows
=
scope
.
var
(
'Param'
).
get_selected_rows
()
w_selected_rows
.
set_height
(
len
(
param_rows
))
w_selected_rows
.
set_rows
(
param_rows
)
w_array
=
np
.
ones
((
len
(
param_rows
),
row_width
)).
astype
(
"float32"
)
for
i
in
range
(
len
(
param_rows
)):
w_array
[
i
]
*=
i
w_tensor
=
w_selected_rows
.
get_tensor
()
w_tensor
.
set
(
w_array
,
place
)
w_before_optimize
=
np
.
array
(
w_tensor
)
# create and initialize LeraningRate Variable
lr_value
=
0.1
lr
=
scope
.
var
(
'LearningRate'
).
get_tensor
()
lr_array
=
np
.
full
((
1
),
lr_value
).
astype
(
"float32"
)
lr
.
set
(
lr_array
,
place
)
# optimize with Python
w_after_optimize
=
np
.
copy
(
w_before_optimize
)
for
index
,
id
in
enumerate
(
grad_rows
):
w_after_optimize
[
id
]
=
w_before_optimize
[
id
]
-
lr_value
*
grad_array
[
index
]
# create and run sgd operator
sgd_op
=
Operator
(
"sgd"
,
Param
=
'Param'
,
Grad
=
'Grad'
,
ParamOut
=
'Param'
,
LearningRate
=
'LearningRate'
)
sgd_op
.
run
(
scope
,
place
)
# get and compare result
result_array
=
np
.
array
(
w_tensor
)
assert
(
result_array
==
w_after_optimize
).
all
()
def
test_sparse_parameter_sgd
(
self
):
places
=
[
core
.
CPUPlace
()]
# do not support GPU kernel currently
for
place
in
places
:
self
.
check_with_place
(
place
)
if
__name__
==
"__main__"
:
unittest
.
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