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
ede4b230
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
ede4b230
编写于
9月 29, 2018
作者:
T
tensor-tang
提交者:
GitHub
9月 29, 2018
浏览文件
操作
浏览文件
下载
差异文件
Merge pull request #13553 from jczaja/prv-fused_embedding_fc_lstm_op
Adding fused_embedding_fc_lstm op
上级
618b3297
e202f33a
变更
10
展开全部
隐藏空白更改
内联
并排
Showing
10 changed file
with
987 addition
and
9 deletion
+987
-9
paddle/fluid/framework/ir/CMakeLists.txt
paddle/fluid/framework/ir/CMakeLists.txt
+1
-0
paddle/fluid/framework/ir/embedding_fc_lstm_fuse_pass.cc
paddle/fluid/framework/ir/embedding_fc_lstm_fuse_pass.cc
+243
-0
paddle/fluid/framework/ir/embedding_fc_lstm_fuse_pass.h
paddle/fluid/framework/ir/embedding_fc_lstm_fuse_pass.h
+40
-0
paddle/fluid/framework/ir/graph_pattern_detector.cc
paddle/fluid/framework/ir/graph_pattern_detector.cc
+18
-0
paddle/fluid/framework/ir/graph_pattern_detector.h
paddle/fluid/framework/ir/graph_pattern_detector.h
+17
-0
paddle/fluid/inference/analysis/analyzer.h
paddle/fluid/inference/analysis/analyzer.h
+9
-8
paddle/fluid/inference/api/paddle_inference_api.h
paddle/fluid/inference/api/paddle_inference_api.h
+1
-1
paddle/fluid/inference/tests/api/analyzer_text_classification_tester.cc
...nference/tests/api/analyzer_text_classification_tester.cc
+13
-0
paddle/fluid/operators/fused_embedding_fc_lstm_op.cc
paddle/fluid/operators/fused_embedding_fc_lstm_op.cc
+604
-0
paddle/fluid/operators/fused_embedding_fc_lstm_op.h
paddle/fluid/operators/fused_embedding_fc_lstm_op.h
+41
-0
未找到文件。
paddle/fluid/framework/ir/CMakeLists.txt
浏览文件 @
ede4b230
...
...
@@ -34,6 +34,7 @@ endif ()
pass_library
(
attention_lstm_fuse_pass inference
)
pass_library
(
infer_clean_graph_pass inference
)
pass_library
(
fc_lstm_fuse_pass inference
)
pass_library
(
embedding_fc_lstm_fuse_pass inference
)
pass_library
(
fc_gru_fuse_pass inference
)
pass_library
(
seq_concat_fc_fuse_pass inference
)
...
...
paddle/fluid/framework/ir/embedding_fc_lstm_fuse_pass.cc
0 → 100644
浏览文件 @
ede4b230
// 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.
// See the License for the specific language governing permissions and
// limitations under the License.
#include "paddle/fluid/framework/ir/embedding_fc_lstm_fuse_pass.h"
#include <algorithm>
#include <string>
#include "paddle/fluid/framework/lod_tensor.h"
#include "paddle/fluid/operators/math/blas.h"
#include "paddle/fluid/operators/math/cpu_vec.h"
#include "paddle/fluid/operators/math/fc_compute.h"
#include "paddle/fluid/platform/cpu_info.h"
namespace
paddle
{
namespace
framework
{
namespace
ir
{
static
int
BuildFusion
(
Graph
*
graph
,
const
std
::
string
&
name_scope
,
Scope
*
scope
,
bool
with_fc_bias
)
{
GraphPatternDetector
gpd
;
auto
*
pattern
=
gpd
.
mutable_pattern
();
// Build pattern
PDNode
*
x
=
pattern
->
NewNode
(
patterns
::
PDNodeName
(
name_scope
,
"x"
))
->
assert_is_op_input
(
"lookup_table"
)
->
assert_var_not_persistable
();
patterns
::
Embedding
embedding_pattern
(
pattern
,
name_scope
);
// TODO(jczaja): Intermediate can only be for val that are not used anywhere
// but lookup table output may go into other LSTM (for reverse
// direction)
auto
*
embedding_out
=
embedding_pattern
(
x
);
patterns
::
FC
fc_pattern
(
pattern
,
name_scope
);
// fc_out is a tmp var, will be removed after fuse, so marked as intermediate.
auto
*
fc_out
=
fc_pattern
(
embedding_out
,
with_fc_bias
)
->
AsIntermediate
();
patterns
::
LSTM
lstm_pattern
(
pattern
,
name_scope
);
lstm_pattern
(
fc_out
);
// Create New OpDesc
auto
embedding_lstm_creator
=
[
&
](
Node
*
embedding
,
Node
*
W
,
Node
*
lstm
,
Node
*
input
,
Node
*
weight_x
,
Node
*
weight_h
,
Node
*
bias
,
Node
*
hidden
,
Node
*
cell
,
Node
*
xx
,
Node
*
fc_bias
)
{
OpDesc
op_desc
;
op_desc
.
SetType
(
"fused_embedding_fc_lstm"
);
#define SET_IN(Key, node__) op_desc.SetInput(#Key, {node__->Name()});
SET_IN
(
Ids
,
input
);
SET_IN
(
WeightH
,
weight_h
);
// Neet to have this passed as We need Wc data for peephole connections
SET_IN
(
Bias
,
bias
);
#undef SET_IN
// Multiply embeddings with Weights
PADDLE_ENFORCE
(
scope
);
const
std
::
string
&
embeddings
=
patterns
::
UniqueKey
(
"Embeddings"
);
auto
*
embeddings_var
=
scope
->
Var
(
embeddings
);
PADDLE_ENFORCE
(
embeddings_var
);
auto
*
embeddings_tensor
=
embeddings_var
->
GetMutable
<
framework
::
LoDTensor
>
();
// Get WeightX size: [single_embedding, fc_size]
// and embedding size: [dict_size, single_embedding]
// and create new size of embeddings eg. [dict_size , hidden_size]
auto
*
embedding_var
=
scope
->
FindVar
(
W
->
Name
());
PADDLE_ENFORCE
(
embedding_var
);
const
auto
&
embedding_tensor
=
embedding_var
->
Get
<
framework
::
LoDTensor
>
();
const
auto
&
weightx_tensor
=
scope
->
FindVar
(
weight_x
->
Name
())
->
Get
<
framework
::
LoDTensor
>
();
embeddings_tensor
->
Resize
(
{
embedding_tensor
.
dims
()[
0
],
weightx_tensor
.
dims
()[
1
]});
// Multiplie embeddings via WeightsX and add bias
auto
embedding_data
=
embedding_tensor
.
data
<
float
>
();
auto
weightx_data
=
weightx_tensor
.
data
<
float
>
();
auto
embeddings_data
=
embeddings_tensor
->
mutable_data
<
float
>
(
platform
::
CPUPlace
());
// Adding biases to GEMM result to be
auto
*
lstm_bias_var
=
scope
->
FindVar
(
bias
->
Name
());
PADDLE_ENFORCE
(
lstm_bias_var
);
const
auto
&
lstm_bias_tensor
=
lstm_bias_var
->
Get
<
framework
::
LoDTensor
>
();
auto
alpha
=
1.0
f
;
auto
beta
=
1.0
f
;
int
m
=
embedding_tensor
.
dims
()[
0
];
int
n
=
weightx_tensor
.
dims
()[
1
];
int
k
=
embedding_tensor
.
dims
()[
1
];
// Copy only gate biases values (only actual bias data, not peephole
// weights)
std
::
vector
<
float
>
combined_biases
;
combined_biases
.
reserve
(
n
);
std
::
copy_n
(
lstm_bias_tensor
.
data
<
float
>
(),
n
,
std
::
back_inserter
(
combined_biases
));
if
(
with_fc_bias
)
{
// Add FC-bias with LSTM-bias (into GEMM result to be)
auto
*
fc_bias_var
=
scope
->
FindVar
(
fc_bias
->
Name
());
const
auto
&
fc_bias_tensor
=
fc_bias_var
->
Get
<
framework
::
LoDTensor
>
();
for
(
int
i
=
0
;
i
<
fc_bias_tensor
.
numel
();
i
++
)
{
combined_biases
[
i
]
+=
fc_bias_tensor
.
data
<
float
>
()[
i
];
}
}
// broadcast biases
std
::
vector
<
float
>
ones
(
m
,
1.0
f
);
paddle
::
operators
::
math
::
CBlas
<
float
>::
GEMM
(
CblasRowMajor
,
CblasNoTrans
,
CblasNoTrans
,
m
,
n
,
1
,
alpha
,
&
ones
[
0
],
1
,
&
combined_biases
[
0
],
n
,
0.0
f
,
embeddings_data
,
n
);
// Wx*embeddings + biases
paddle
::
operators
::
math
::
CBlas
<
float
>::
GEMM
(
CblasRowMajor
,
CblasNoTrans
,
CblasNoTrans
,
m
,
n
,
k
,
alpha
,
embedding_data
,
k
,
weightx_data
,
n
,
beta
,
embeddings_data
,
n
);
op_desc
.
SetInput
(
"Embeddings"
,
{
embeddings
});
// Create temp variables.
const
std
::
string
BatchedInput
=
patterns
::
UniqueKey
(
"BatchedInput"
);
const
std
::
string
BatchedCellPreAct
=
patterns
::
UniqueKey
(
"BatchedCellPreAct"
);
const
std
::
string
BatchedGate
=
patterns
::
UniqueKey
(
"BatchedGate"
);
scope
->
Var
(
BatchedInput
)
->
GetMutable
<
framework
::
LoDTensor
>
();
scope
->
Var
(
BatchedCellPreAct
)
->
GetMutable
<
framework
::
LoDTensor
>
();
scope
->
Var
(
BatchedGate
)
->
GetMutable
<
framework
::
LoDTensor
>
();
op_desc
.
SetInput
(
"H0"
,
{});
op_desc
.
SetInput
(
"C0"
,
{});
op_desc
.
SetOutput
(
"Hidden"
,
{
hidden
->
Name
()});
op_desc
.
SetOutput
(
"Cell"
,
{
cell
->
Name
()});
op_desc
.
SetOutput
(
"XX"
,
{
xx
->
Name
()});
op_desc
.
SetOutput
(
"BatchedGate"
,
{
BatchedGate
});
op_desc
.
SetOutput
(
"BatchCellPreAct"
,
{
BatchedCellPreAct
});
op_desc
.
SetOutput
(
"BatchedInput"
,
{
BatchedInput
});
op_desc
.
SetAttr
(
"is_reverse"
,
lstm
->
Op
()
->
GetAttr
(
"is_reverse"
));
op_desc
.
SetAttr
(
"use_peepholes"
,
lstm
->
Op
()
->
GetAttr
(
"use_peepholes"
));
// TODO(TJ): get from attr
op_desc
.
SetAttr
(
"use_seq"
,
true
);
PADDLE_ENFORCE
(
graph
->
Has
(
kParamScopeAttr
));
auto
*
scope
=
graph
->
Get
<
Scope
*>
(
kParamScopeAttr
);
#define OP_SET_OUT(x) \
const std::string x = patterns::UniqueKey(#x); \
op_desc.SetOutput(#x, {x}); \
scope->Var(x)->GetMutable<LoDTensor>()
OP_SET_OUT
(
BatchedCell
);
OP_SET_OUT
(
BatchedHidden
);
OP_SET_OUT
(
ReorderedH0
);
OP_SET_OUT
(
ReorderedC0
);
#undef OP_SET_OUT
auto
*
op
=
graph
->
CreateOpNode
(
&
op_desc
);
IR_NODE_LINK_TO
(
input
,
op
);
IR_NODE_LINK_TO
(
weight_x
,
op
);
IR_NODE_LINK_TO
(
weight_h
,
op
);
IR_NODE_LINK_TO
(
bias
,
op
);
IR_NODE_LINK_TO
(
op
,
hidden
);
return
op
;
};
int
fusion_count
{
0
};
auto
handler
=
[
&
](
const
GraphPatternDetector
::
subgraph_t
&
subgraph
,
Graph
*
g
)
{
GET_IR_NODE_FROM_SUBGRAPH
(
lstm
,
lstm
,
lstm_pattern
);
GET_IR_NODE_FROM_SUBGRAPH
(
Weight
,
Weight
,
lstm_pattern
);
GET_IR_NODE_FROM_SUBGRAPH
(
Bias
,
Bias
,
lstm_pattern
);
GET_IR_NODE_FROM_SUBGRAPH
(
Cell
,
Cell
,
lstm_pattern
);
GET_IR_NODE_FROM_SUBGRAPH
(
Hidden
,
Hidden
,
lstm_pattern
);
GET_IR_NODE_FROM_SUBGRAPH
(
lookup_table
,
lookup_table
,
embedding_pattern
);
GET_IR_NODE_FROM_SUBGRAPH
(
W
,
W
,
embedding_pattern
);
GET_IR_NODE_FROM_SUBGRAPH
(
w
,
w
,
fc_pattern
);
GET_IR_NODE_FROM_SUBGRAPH
(
mul
,
mul
,
fc_pattern
);
// TODO(jczaja): Add support for is_sparse / is_distributed
auto
is_sparse
=
boost
::
get
<
bool
>
(
lookup_table
->
Op
()
->
GetAttr
(
"is_sparse"
));
auto
is_distributed
=
boost
::
get
<
bool
>
(
lookup_table
->
Op
()
->
GetAttr
(
"is_distributed"
));
if
(
is_sparse
==
true
||
is_distributed
==
true
)
{
return
;
}
if
(
with_fc_bias
)
{
GET_IR_NODE_FROM_SUBGRAPH
(
fc_out
,
Out
,
fc_pattern
);
GET_IR_NODE_FROM_SUBGRAPH
(
fc_bias
,
bias
,
fc_pattern
);
GET_IR_NODE_FROM_SUBGRAPH
(
elementwise_add
,
elementwise_add
,
fc_pattern
);
embedding_lstm_creator
(
lookup_table
,
W
,
lstm
,
subgraph
.
at
(
x
),
w
,
Weight
,
Bias
,
Hidden
,
Cell
,
fc_out
,
fc_bias
);
// Remove unneeded nodes.
// TODO(jczaja): Proper removing of lookup table
std
::
unordered_set
<
const
Node
*>
marked_nodes
(
//{lookup_table, mul, lstm, elementwise_add, fc_bias, W});
{
mul
,
lstm
,
elementwise_add
,
fc_bias
});
GraphSafeRemoveNodes
(
graph
,
marked_nodes
);
}
else
{
GET_IR_NODE_FROM_SUBGRAPH
(
fc_out
,
mul_out
,
fc_pattern
);
embedding_lstm_creator
(
lookup_table
,
W
,
lstm
,
subgraph
.
at
(
x
),
w
,
Weight
,
Bias
,
Hidden
,
Cell
,
fc_out
,
nullptr
);
// Remove unneeded nodes.
// TODO(jczaja): Proper removing of lookup table
// std::unordered_set<const Node*> marked_nodes({lookup_table, W, mul,
// lstm});
std
::
unordered_set
<
const
Node
*>
marked_nodes
({
mul
,
lstm
});
GraphSafeRemoveNodes
(
graph
,
marked_nodes
);
}
++
fusion_count
;
};
gpd
(
graph
,
handler
);
return
fusion_count
;
}
std
::
unique_ptr
<
ir
::
Graph
>
EmbeddingFCLSTMFusePass
::
ApplyImpl
(
std
::
unique_ptr
<
ir
::
Graph
>
graph
)
const
{
FusePassBase
::
Init
(
name_scope_
,
graph
.
get
());
int
fusion_count
=
BuildFusion
(
graph
.
get
(),
name_scope_
,
param_scope
(),
true
/*with_fc_bias*/
);
AddStatis
(
fusion_count
);
return
graph
;
}
}
// namespace ir
}
// namespace framework
}
// namespace paddle
REGISTER_PASS
(
embedding_fc_lstm_fuse_pass
,
paddle
::
framework
::
ir
::
EmbeddingFCLSTMFusePass
);
paddle/fluid/framework/ir/embedding_fc_lstm_fuse_pass.h
0 → 100644
浏览文件 @
ede4b230
// 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.
// See the License for the specific language governing permissions and
// limitations under the License.
#pragma once
#include "paddle/fluid/framework/ir/fuse_pass_base.h"
#include "paddle/fluid/framework/ir/graph.h"
#include "paddle/fluid/framework/ir/graph_pattern_detector.h"
namespace
paddle
{
namespace
framework
{
namespace
ir
{
// Fusing of Embedding , FC and LSTM op
// Just FC without bias
class
EmbeddingFCLSTMFusePass
:
public
FusePassBase
{
public:
virtual
~
EmbeddingFCLSTMFusePass
()
{}
protected:
std
::
unique_ptr
<
ir
::
Graph
>
ApplyImpl
(
std
::
unique_ptr
<
ir
::
Graph
>
graph
)
const
;
const
std
::
string
name_scope_
{
"embedding_fc_lstm_fuse"
};
};
}
// namespace ir
}
// namespace framework
}
// namespace paddle
paddle/fluid/framework/ir/graph_pattern_detector.cc
浏览文件 @
ede4b230
...
...
@@ -692,6 +692,24 @@ PDNode *patterns::FC::operator()(paddle::framework::ir::PDNode *x,
}
}
PDNode
*
patterns
::
Embedding
::
operator
()(
PDNode
*
x
)
{
x
->
assert_is_op_input
(
"lookup_table"
,
"Ids"
);
auto
*
lookup_table_op
=
pattern
->
NewNode
(
lookup_table_repr
())
->
assert_is_op
(
"lookup_table"
);
#define NEW_NODE(arg__, io__) \
auto *arg__ = pattern->NewNode(arg__##_repr()) \
->assert_is_op_##io__("lookup_table", #arg__);
NEW_NODE
(
W
,
input
);
NEW_NODE
(
Out
,
output
);
#undef NEW_NODE
lookup_table_op
->
LinksFrom
({
x
,
W
});
lookup_table_op
->
LinksTo
({
Out
});
return
Out
;
}
PDNode
*
patterns
::
LSTM
::
operator
()(
PDNode
*
x
)
{
x
->
assert_is_op_input
(
"lstm"
,
"Input"
);
auto
*
lstm_op
=
pattern
->
NewNode
(
lstm_repr
())
->
assert_is_op
(
"lstm"
);
...
...
paddle/fluid/framework/ir/graph_pattern_detector.h
浏览文件 @
ede4b230
...
...
@@ -418,6 +418,23 @@ struct FC : public PatternBase {
PATTERN_DECL_NODE
(
Out
);
};
// Embedding
struct
Embedding
:
public
PatternBase
{
Embedding
(
PDPattern
*
pattern
,
const
std
::
string
&
name_scope
)
:
PatternBase
(
pattern
,
name_scope
,
"embedding"
)
{}
PDNode
*
operator
()(
PDNode
*
x
);
// declare operator node's name
PATTERN_DECL_NODE
(
lookup_table
);
// Inputs
//
PATTERN_DECL_NODE
(
Ids
);
PATTERN_DECL_NODE
(
W
);
// embeddings
// Outputs
PATTERN_DECL_NODE
(
Out
);
};
struct
LSTM
:
public
PatternBase
{
LSTM
(
PDPattern
*
pattern
,
const
std
::
string
&
name_scope
)
:
PatternBase
(
pattern
,
name_scope
,
"lstm"
)
{}
...
...
paddle/fluid/inference/analysis/analyzer.h
浏览文件 @
ede4b230
...
...
@@ -64,14 +64,15 @@ class Analyzer : public OrderedRegistry<PassManager> {
// larger fusion.
const
std
::
vector
<
std
::
string
>
all_ir_passes_
{{
// Manual update the passes here.
"infer_clean_graph_pass"
,
//
"attention_lstm_fuse_pass"
,
//
"fc_lstm_fuse_pass"
,
//
"mul_lstm_fuse_pass"
,
//
"fc_gru_fuse_pass"
,
//
"mul_gru_fuse_pass"
,
//
"seq_concat_fc_fuse_pass"
,
//
"fc_fuse_pass"
,
//
"infer_clean_graph_pass"
,
//
"attention_lstm_fuse_pass"
,
//
"embedding_fc_lstm_fuse_pass"
,
//
"fc_lstm_fuse_pass"
,
//
"mul_lstm_fuse_pass"
,
//
"fc_gru_fuse_pass"
,
//
"mul_gru_fuse_pass"
,
//
"seq_concat_fc_fuse_pass"
,
//
"fc_fuse_pass"
,
//
#ifdef PADDLE_WITH_MKLDNN
"conv_relu_mkldnn_fuse_pass"
,
//
#endif
...
...
paddle/fluid/inference/api/paddle_inference_api.h
浏览文件 @
ede4b230
...
...
@@ -263,7 +263,7 @@ struct AnalysisConfig : public NativeConfig {
bool
enable_ir_optim
=
true
;
// Manually determine the IR passes to run.
IrPassMode
ir_mode
{
IrPassMode
::
kExclude
};
std
::
vector
<
std
::
string
>
ir_passes
;
std
::
vector
<
std
::
string
>
ir_passes
{
"embedding_fc_lstm_fuse_pass"
}
;
// NOT stable yet.
bool
use_feed_fetch_ops
{
true
};
...
...
paddle/fluid/inference/tests/api/analyzer_text_classification_tester.cc
浏览文件 @
ede4b230
...
...
@@ -104,5 +104,18 @@ TEST(Analyzer_Text_Classification, compare) {
CompareNativeAndAnalysis
(
cfg
,
input_slots_all
);
}
TEST
(
Analyzer_Text_Classification
,
compare_against_embedding_fc_lstm_fused
)
{
AnalysisConfig
cfg
;
SetConfig
(
&
cfg
);
// Enable embedding_fc_lstm_fuse_pass (disabled by default)
auto
it
=
std
::
find
(
cfg
.
ir_passes
.
begin
(),
cfg
.
ir_passes
.
end
(),
"embedding_fc_lstm_fuse_pass"
);
if
(
it
!=
cfg
.
ir_passes
.
end
())
cfg
.
ir_passes
.
erase
(
it
);
std
::
vector
<
std
::
vector
<
PaddleTensor
>>
input_slots_all
;
SetInput
(
&
input_slots_all
);
CompareNativeAndAnalysis
(
cfg
,
input_slots_all
);
}
}
// namespace inference
}
// namespace paddle
paddle/fluid/operators/fused_embedding_fc_lstm_op.cc
0 → 100644
浏览文件 @
ede4b230
此差异已折叠。
点击以展开。
paddle/fluid/operators/fused_embedding_fc_lstm_op.h
0 → 100644
浏览文件 @
ede4b230
/* Copyright (c) 2016 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.
See the License for the specific language governing permissions and
limitations under the License. */
#pragma once
#include "paddle/fluid/framework/op_registry.h"
namespace
paddle
{
namespace
operators
{
using
LoDTensor
=
framework
::
LoDTensor
;
using
Tensor
=
framework
::
Tensor
;
class
FusedEmbeddingFCLSTMOp
:
public
framework
::
OperatorWithKernel
{
public:
using
framework
::
OperatorWithKernel
::
OperatorWithKernel
;
void
InferShape
(
framework
::
InferShapeContext
*
ctx
)
const
override
;
protected:
framework
::
OpKernelType
GetExpectedKernelType
(
const
framework
::
ExecutionContext
&
ctx
)
const
override
;
};
class
FusedEmbeddingFCLSTMOpMaker
:
public
framework
::
OpProtoAndCheckerMaker
{
public:
void
Make
()
override
;
};
}
// namespace operators
}
// namespace paddle
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