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
GoAI
Paddle
提交
91e8aada
P
Paddle
项目概览
GoAI
/
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
91e8aada
编写于
8月 08, 2017
作者:
D
dzhwinter
提交者:
GitHub
8月 08, 2017
浏览文件
操作
浏览文件
下载
差异文件
Merge branch 'develop' into type_alias
上级
aa5090de
e31a469e
变更
35
隐藏空白更改
内联
并排
Showing
35 changed file
with
1276 addition
and
110 deletion
+1276
-110
.travis.yml
.travis.yml
+1
-1
doc/api/v2/config/layer.rst
doc/api/v2/config/layer.rst
+10
-0
doc/templates/conf.py.cn.in
doc/templates/conf.py.cn.in
+3
-7
doc/templates/conf.py.en.in
doc/templates/conf.py.en.in
+3
-7
paddle/framework/CMakeLists.txt
paddle/framework/CMakeLists.txt
+8
-7
paddle/framework/op_registry.h
paddle/framework/op_registry.h
+0
-6
paddle/framework/operator.h
paddle/framework/operator.h
+9
-1
paddle/framework/pybind.cc
paddle/framework/pybind.cc
+21
-2
paddle/gserver/layers/KmaxSeqScoreLayer.cpp
paddle/gserver/layers/KmaxSeqScoreLayer.cpp
+117
-0
paddle/gserver/layers/SubNestedSequenceLayer.cpp
paddle/gserver/layers/SubNestedSequenceLayer.cpp
+176
-0
paddle/gserver/tests/CMakeLists.txt
paddle/gserver/tests/CMakeLists.txt
+10
-0
paddle/gserver/tests/test_KmaxSeqScore.cpp
paddle/gserver/tests/test_KmaxSeqScore.cpp
+160
-0
paddle/gserver/tests/test_LayerGrad.cpp
paddle/gserver/tests/test_LayerGrad.cpp
+78
-0
paddle/operators/CMakeLists.txt
paddle/operators/CMakeLists.txt
+2
-0
paddle/operators/cross_entropy_op.cc
paddle/operators/cross_entropy_op.cc
+4
-3
paddle/operators/net_op.h
paddle/operators/net_op.h
+9
-0
paddle/operators/uniform_random_op.cc
paddle/operators/uniform_random_op.cc
+84
-0
paddle/operators/uniform_random_op.cu
paddle/operators/uniform_random_op.cu
+70
-0
paddle/parameter/Argument.cpp
paddle/parameter/Argument.cpp
+20
-0
paddle/parameter/Argument.h
paddle/parameter/Argument.h
+24
-0
paddle/scripts/travis/build_doc.sh
paddle/scripts/travis/build_doc.sh
+2
-7
proto/CMakeLists.txt
proto/CMakeLists.txt
+1
-1
python/paddle/trainer/config_parser.py
python/paddle/trainer/config_parser.py
+35
-0
python/paddle/trainer_config_helpers/layers.py
python/paddle/trainer_config_helpers/layers.py
+92
-1
python/paddle/trainer_config_helpers/tests/configs/file_list.sh
.../paddle/trainer_config_helpers/tests/configs/file_list.sh
+2
-1
python/paddle/trainer_config_helpers/tests/configs/protostr/test_kmax_seq_socre_layer.protostr
...tests/configs/protostr/test_kmax_seq_socre_layer.protostr
+66
-0
python/paddle/trainer_config_helpers/tests/configs/protostr/test_seq_select_layers.protostr
...rs/tests/configs/protostr/test_seq_select_layers.protostr
+37
-0
python/paddle/trainer_config_helpers/tests/configs/test_kmax_seq_socre_layer.py
...config_helpers/tests/configs/test_kmax_seq_socre_layer.py
+11
-0
python/paddle/trainer_config_helpers/tests/configs/test_seq_select_layers.py
...er_config_helpers/tests/configs/test_seq_select_layers.py
+11
-0
python/paddle/v2/framework/tests/CMakeLists.txt
python/paddle/v2/framework/tests/CMakeLists.txt
+2
-0
python/paddle/v2/framework/tests/gradient_checker.py
python/paddle/v2/framework/tests/gradient_checker.py
+149
-3
python/paddle/v2/framework/tests/op_test_util.py
python/paddle/v2/framework/tests/op_test_util.py
+4
-3
python/paddle/v2/framework/tests/test_cross_entropy_op.py
python/paddle/v2/framework/tests/test_cross_entropy_op.py
+14
-2
python/paddle/v2/framework/tests/test_softmax_op.py
python/paddle/v2/framework/tests/test_softmax_op.py
+6
-58
python/paddle/v2/framework/tests/test_uniform_random_op.py
python/paddle/v2/framework/tests/test_uniform_random_op.py
+35
-0
未找到文件。
.travis.yml
浏览文件 @
91e8aada
...
...
@@ -38,7 +38,7 @@ before_install:
# Paddle is using protobuf 3.1 currently. Protobuf 3.2 breaks the compatibility. So we specify the python
# protobuf version.
-
pip install numpy wheel 'protobuf==3.1' sphinx==1.5.6 recommonmark sphinx-rtd-theme==0.1.9 virtualenv pre-commit requests==2.9.2 LinkChecker
-
pip install rarfile
-
pip install rarfile
nltk==3.2.2 scipy==0.19.0 recordio matplotlib Pillow
-
curl https://glide.sh/get | bash
-
eval "$(GIMME_GO_VERSION=1.8.3 gimme)"
-
go get -u github.com/alecthomas/gometalinter
...
...
doc/api/v2/config/layer.rst
浏览文件 @
91e8aada
...
...
@@ -257,6 +257,16 @@ seq_concat
.. autoclass:: paddle.v2.layer.seq_concat
:noindex:
kmax_sequence_score
-------------------
.. autoclass:: paddle.v2.layer.kmax_sequence_score
:noindex:
sub_nested_seq
--------------
.. autoclass:: paddle.v2.layer.sub_nested_seq
:noindex:
Reshaping Layers
================
...
...
doc/templates/conf.py.cn.in
浏览文件 @
91e8aada
...
...
@@ -13,15 +13,11 @@
# serve to show the default.
import sys
import os, subprocess
sys.path.insert(0, os.path.abspath('@PROJ_ROOT@/python'))
import shlex
from recommonmark import parser, transform
try:
import py_paddle
import paddle
import paddle.v2
except ImportError:
print("Must install paddle python package before generating documentation")
sys.exit(1)
import paddle
import paddle.v2
MarkdownParser = parser.CommonMarkParser
AutoStructify = transform.AutoStructify
...
...
doc/templates/conf.py.en.in
浏览文件 @
91e8aada
...
...
@@ -13,15 +13,11 @@
# serve to show the default.
import sys
import os, subprocess
sys.path.insert(0, os.path.abspath('@PROJ_ROOT@/python'))
import shlex
from recommonmark import parser, transform
try:
import py_paddle
import paddle
import paddle.v2
except ImportError:
print("Must install paddle python package before generating documentation")
sys.exit(1)
import paddle
import paddle.v2
MarkdownParser = parser.CommonMarkParser
...
...
paddle/framework/CMakeLists.txt
浏览文件 @
91e8aada
...
...
@@ -40,11 +40,12 @@ if(WITH_PYTHON)
cc_library
(
paddle_pybind SHARED
SRCS pybind.cc
DEPS pybind python backward
fc_op
sgd_op
add_op
mean_op
cross_entropy_op
fill_zeros_like_op
recurrent_op
)
fc_op
sgd_op
add_op
mean_op
cross_entropy_op
recurrent_op
uniform_random_op
fill_zeros_like_op
)
endif
(
WITH_PYTHON
)
paddle/framework/op_registry.h
浏览文件 @
91e8aada
...
...
@@ -260,12 +260,6 @@ class OpRegistry {
return
CreateOp
(
op_desc
.
type
(),
inputs
,
outputs
,
attrs
);
}
static
bool
SupportGPU
(
const
std
::
string
&
op_type
)
{
OperatorWithKernel
::
OpKernelKey
key
;
key
.
place_
=
platform
::
GPUPlace
();
return
OperatorWithKernel
::
AllOpKernels
().
at
(
op_type
).
count
(
key
)
!=
0
;
}
static
std
::
shared_ptr
<
OperatorBase
>
CreateGradOp
(
const
OperatorBase
&
op
)
{
PADDLE_ENFORCE
(
!
op
.
IsNetOp
(),
"Use framework::Backward to get backward ops"
);
...
...
paddle/framework/operator.h
浏览文件 @
91e8aada
...
...
@@ -88,6 +88,8 @@ class OperatorBase {
virtual
bool
IsNetOp
()
const
{
return
false
;
}
virtual
bool
SupportGPU
()
const
{
return
false
;
}
/// rename inputs outputs name
void
Rename
(
const
std
::
string
&
old_name
,
const
std
::
string
&
new_name
);
...
...
@@ -308,7 +310,7 @@ class OperatorWithKernel : public OperatorBase {
using
OpKernelMap
=
std
::
unordered_map
<
OpKernelKey
,
std
::
unique_ptr
<
OpKernel
>
,
OpKernelHash
>
;
void
InferShape
(
const
Scope
&
scope
)
const
{
void
InferShape
(
const
Scope
&
scope
)
const
override
{
InferShape
(
InferShapeContext
(
this
,
scope
));
}
...
...
@@ -324,6 +326,12 @@ class OperatorWithKernel : public OperatorBase {
return
g_all_op_kernels
;
}
bool
SupportGPU
()
const
override
{
OperatorWithKernel
::
OpKernelKey
key
;
key
.
place_
=
platform
::
GPUPlace
();
return
OperatorWithKernel
::
AllOpKernels
().
at
(
type_
).
count
(
key
)
!=
0
;
}
protected:
virtual
void
InferShape
(
const
InferShapeContext
&
ctx
)
const
=
0
;
};
...
...
paddle/framework/pybind.cc
浏览文件 @
91e8aada
...
...
@@ -39,6 +39,7 @@ USE_OP(softmax);
USE_OP
(
rowwise_add
);
USE_OP
(
fill_zeros_like
);
USE_OP_WITHOUT_KERNEL
(
recurrent_op
);
USE_OP
(
uniform_random
);
namespace
paddle
{
namespace
framework
{
...
...
@@ -56,6 +57,26 @@ void ExposeOperator(ClassType &m) {
[](
const
typename
ClassType
::
type
&
op
)
->
std
::
vector
<
std
::
string
>
{
return
op
.
outputs_
;
})
.
def
(
"inputs"
,
[](
const
typename
ClassType
::
type
&
op
)
->
std
::
vector
<
std
::
string
>
{
return
op
.
inputs_
;
})
.
def
(
"support_gpu"
,
&
ClassType
::
type
::
SupportGPU
)
.
def
(
"temp_outputs"
,
[](
const
typename
ClassType
::
type
&
op
)
->
std
::
vector
<
std
::
string
>
{
auto
iter
=
op
.
attrs_
.
find
(
"temporary_index"
);
std
::
vector
<
std
::
string
>
ret
;
if
(
iter
==
op
.
attrs_
.
end
())
{
return
ret
;
}
else
{
auto
tmp_idx
=
boost
::
get
<
std
::
vector
<
int
>>
(
iter
->
second
);
ret
.
reserve
(
tmp_idx
.
size
());
for
(
auto
&
index
:
tmp_idx
)
{
ret
.
push_back
(
op
.
outputs_
.
at
(
index
));
}
return
ret
;
}
})
.
def
(
"__str__"
,
&
ClassType
::
type
::
DebugString
);
}
...
...
@@ -201,8 +222,6 @@ All parameter, weight, gradient are variables in Paddle.
return
OpRegistry
::
CreateOp
(
desc
);
});
operator_base
.
def_static
(
"support_gpu"
,
&
OpRegistry
::
SupportGPU
);
operator_base
.
def
(
"backward"
,
[](
const
OperatorBase
&
forwardOp
,
const
std
::
unordered_set
<
std
::
string
>
&
no_grad_vars
)
{
...
...
paddle/gserver/layers/KmaxSeqScoreLayer.cpp
0 → 100644
浏览文件 @
91e8aada
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
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 "Layer.h"
namespace
paddle
{
class
KmaxSeqScoreLayer
:
public
Layer
{
private:
MatrixPtr
scores_
;
size_t
beamSize_
;
void
kmaxScorePerSeq
(
const
real
*
score
,
real
*
sortedRes
,
const
ICpuGpuVectorPtr
seqStartPos
);
public:
explicit
KmaxSeqScoreLayer
(
const
LayerConfig
&
config
)
:
Layer
(
config
)
{}
bool
init
(
const
LayerMap
&
layerMap
,
const
ParameterMap
&
parameterMap
)
override
;
void
forward
(
PassType
passType
)
override
;
void
backward
(
const
UpdateCallback
&
callback
=
nullptr
)
override
;
};
REGISTER_LAYER
(
kmax_seq_score
,
KmaxSeqScoreLayer
);
bool
KmaxSeqScoreLayer
::
init
(
const
LayerMap
&
layerMap
,
const
ParameterMap
&
parameterMap
)
{
bool
ret
=
Layer
::
init
(
layerMap
,
parameterMap
);
CHECK_EQ
(
1U
,
inputLayers_
.
size
());
beamSize_
=
config_
.
beam_size
();
CHECK_GE
(
beamSize_
,
1U
);
setNeedSequenceInfo
(
false
);
setNeedGradient
(
false
);
return
ret
;
}
void
KmaxSeqScoreLayer
::
kmaxScorePerSeq
(
const
real
*
scores
,
real
*
sortedIds
,
const
ICpuGpuVectorPtr
seqStartPos
)
{
int
*
starts
=
seqStartPos
->
getMutableData
(
false
);
std
::
vector
<
real
>
indices
;
for
(
size_t
i
=
0
;
i
<
seqStartPos
->
getSize
()
-
1
;
++
i
)
{
int
seqLen
=
starts
[
i
+
1
]
-
starts
[
i
];
int
k
=
std
::
min
(
static_cast
<
int
>
(
beamSize_
),
seqLen
);
indices
.
resize
(
seqLen
,
0
);
std
::
iota
(
begin
(
indices
),
end
(
indices
),
0.
);
std
::
vector
<
real
>
tmpScore
(
scores
+
starts
[
i
],
scores
+
starts
[
i
+
1
]);
std
::
partial_sort
(
begin
(
indices
),
begin
(
indices
)
+
k
,
end
(
indices
),
[
&
](
size_t
a
,
size_t
b
)
{
return
tmpScore
[
a
]
>
tmpScore
[
b
];
});
memcpy
(
sortedIds
+
(
i
*
beamSize_
),
indices
.
data
(),
k
*
sizeof
(
real
));
}
}
void
KmaxSeqScoreLayer
::
forward
(
PassType
passType
)
{
Layer
::
forward
(
passType
);
const
Argument
&
input
=
getInput
(
0
);
const
MatrixPtr
inputScore
=
getInputValue
(
0
);
CHECK
(
input
.
hasSeq
()
||
input
.
hasSubseq
())
<<
"input of "
<<
getName
()
<<
" must be a sequence or a nested sequence."
;
CHECK_EQ
(
input
.
value
->
getWidth
(),
1UL
)
<<
"input of "
<<
getName
()
<<
" is score over a sequence or a nested sequence, so its width "
<<
" must be 1."
;
if
(
useGpu_
)
{
// this Layer runs only in CPU, if the model is runing on GPU,
// then copy the input to this layer from GPU to CPU.
Matrix
::
resizeOrCreate
(
scores_
,
inputScore
->
getHeight
(),
1
,
false
/* trans */
,
false
/* useGpu */
);
scores_
->
copyFrom
(
*
inputScore
);
}
else
{
scores_
=
inputScore
;
}
Matrix
::
resizeOrCreate
(
output_
.
value
,
input
.
hasSubseq
()
?
input
.
getNumSubSequences
()
:
input
.
getNumSequences
(),
beamSize_
,
false
,
false
);
output_
.
value
->
one
();
output_
.
value
->
mulScalar
(
-
1.
);
kmaxScorePerSeq
(
scores_
->
getData
(),
output_
.
value
->
getData
(),
input
.
hasSubseq
()
?
input
.
subSequenceStartPositions
:
input
.
sequenceStartPositions
);
}
void
KmaxSeqScoreLayer
::
backward
(
const
UpdateCallback
&
callback
)
{}
}
// namespace paddle
paddle/gserver/layers/SubNestedSequenceLayer.cpp
0 → 100644
浏览文件 @
91e8aada
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
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 "Layer.h"
#include "paddle/math/Matrix.h"
#include "paddle/math/Vector.h"
#include "paddle/utils/Logging.h"
#include "paddle/utils/Stat.h"
namespace
paddle
{
class
SubNestedSequenceLayer
:
public
Layer
{
public:
explicit
SubNestedSequenceLayer
(
const
LayerConfig
&
config
)
:
Layer
(
config
)
{}
bool
init
(
const
LayerMap
&
layerMap
,
const
ParameterMap
&
parameterMap
)
override
;
void
forward
(
PassType
passType
)
override
;
void
backward
(
const
UpdateCallback
&
callback
=
nullptr
)
override
;
private:
/*
* This functions generates the indices of rows in a batch according to the
* indices of selected sub-sequence in each sequence.
*
* Examples:
* selectedIndices:
* [
* [0, 1, -1],
* [0, 1, 2],
* [0, -1, -1],
* [0, 2, 3],
* ]
* inputSeqInfo:
* [
* [0,3,4],
* [4,5,7,10,15],
* [15,20],
* [20,22,23,25,28]
* ]
*
* ths output is saved to private member rowIndice_;
* [0,1,2,3,4,5,6,7,8,9,10,11,12,13,14,15,
* 16,17,18,19,20,21,22,23,24,25,26,27]
*/
void
calSelectedCols
(
const
MatrixPtr
selectedIndices
,
const
std
::
vector
<
std
::
vector
<
int
>>&
inputSeqInfo
);
// if the second input of this layer is on GPU memory, copy it to CPU memory.
MatrixPtr
selIdsCpu_
;
// reorganized sequenceStartPositions and subSequenceStartPositions
// into a 2d vector to facilitate the sequence selection process.
std
::
vector
<
std
::
vector
<
int
>>
inputSeqInfoVec_
;
// the final selected row indices in a batch,
// rowIdx_ and selectedRows_ actually share a same memory.
IVectorPtr
rowIndice_
;
std
::
vector
<
int
>
selectedRows_
;
};
REGISTER_LAYER
(
sub_nested_seq
,
SubNestedSequenceLayer
);
bool
SubNestedSequenceLayer
::
init
(
const
LayerMap
&
layerMap
,
const
ParameterMap
&
parameterMap
)
{
/* Initialize the basic parent class */
Layer
::
init
(
layerMap
,
parameterMap
);
CHECK_EQ
(
2U
,
inputLayers_
.
size
());
setNeedSequenceInfo
(
false
);
return
true
;
}
void
SubNestedSequenceLayer
::
calSelectedCols
(
const
MatrixPtr
selectedIndices
,
const
std
::
vector
<
std
::
vector
<
int
>>&
inputSeqInfo
)
{
selectedRows_
.
clear
();
std
::
vector
<
int
>
outSeqStartInfo
(
1
,
0
);
std
::
vector
<
int
>
outSubSeqStartInfo
(
1
,
0
);
size_t
seqNum
=
selectedIndices
->
getHeight
();
size_t
beamSize
=
selectedIndices
->
getWidth
();
for
(
size_t
i
=
0
;
i
<
seqNum
;
++
i
)
{
for
(
size_t
j
=
0
;
j
<
beamSize
;
++
j
)
{
if
(
selectedIndices
->
getElement
(
i
,
j
)
==
-
1.
)
break
;
int
selSubSeqIdx
=
selectedIndices
->
getElement
(
i
,
j
);
CHECK_GT
(
inputSeqInfoVec_
[
i
].
size
()
-
1
,
selSubSeqIdx
);
size_t
subSeqLen
=
inputSeqInfoVec_
[
i
][
selSubSeqIdx
+
1
]
-
inputSeqInfoVec_
[
i
][
selSubSeqIdx
];
for
(
size_t
k
=
0
;
k
<
subSeqLen
;
++
k
)
selectedRows_
.
push_back
(
inputSeqInfoVec_
[
i
][
selSubSeqIdx
]
+
k
);
outSubSeqStartInfo
.
push_back
(
outSubSeqStartInfo
.
back
()
+
subSeqLen
);
}
outSeqStartInfo
.
push_back
(
outSubSeqStartInfo
.
back
());
}
if
(
useGpu_
)
{
rowIndice_
=
IVector
::
create
(
selectedRows_
.
size
(),
useGpu_
);
rowIndice_
->
copyFrom
(
selectedRows_
.
data
(),
selectedRows_
.
size
());
}
else
{
rowIndice_
=
IVector
::
create
(
selectedRows_
.
data
(),
selectedRows_
.
size
(),
useGpu_
);
}
// create the sequence information for the output.
ICpuGpuVector
::
resizeOrCreate
(
output_
.
sequenceStartPositions
,
outSeqStartInfo
.
size
(),
false
);
output_
.
sequenceStartPositions
->
copyFrom
(
outSeqStartInfo
.
data
(),
outSeqStartInfo
.
size
(),
false
);
ICpuGpuVector
::
resizeOrCreate
(
output_
.
subSequenceStartPositions
,
outSubSeqStartInfo
.
size
(),
false
);
output_
.
subSequenceStartPositions
->
copyFrom
(
outSubSeqStartInfo
.
data
(),
outSubSeqStartInfo
.
size
(),
false
);
}
void
SubNestedSequenceLayer
::
forward
(
PassType
passType
)
{
Layer
::
forward
(
passType
);
const
Argument
&
inputSeq
=
getInput
(
0
);
CHECK
(
inputSeq
.
hasSubseq
())
<<
"The first input of SubNestSequence layer "
<<
"must be a nested sequence."
;
const
MatrixPtr
selectedIndices
=
getInputValue
(
1
);
CHECK_EQ
(
inputSeq
.
getNumSequences
(),
selectedIndices
->
getHeight
());
if
(
dynamic_cast
<
GpuMatrix
*>
(
selectedIndices
.
get
()))
{
/*
* Currently, the second input for this layer is generated by
* kmax_sequence_score_layer whose output is always stored on CPU,
* or a data_layer which canbe on GPU.
*
* If the second input is on GPU, copy it to CPU memory, because this
* input always uses very few memory, and operations related to it are
* all logic control, not computations.
*/
Matrix
::
resizeOrCreate
(
selIdsCpu_
,
selectedIndices
->
getHeight
(),
selectedIndices
->
getWidth
(),
false
/* trans */
,
false
/* useGpu */
);
selIdsCpu_
->
copyFrom
(
*
selectedIndices
);
}
else
{
selIdsCpu_
=
selectedIndices
;
}
Argument
::
reorganizeSeqInfo
(
inputSeq
.
sequenceStartPositions
,
inputSeq
.
subSequenceStartPositions
,
inputSeqInfoVec_
);
calSelectedCols
(
selIdsCpu_
,
inputSeqInfoVec_
);
resetOutput
(
selectedRows_
.
size
(),
getSize
());
getOutputValue
()
->
selectRows
(
*
getInputValue
(
0
),
*
rowIndice_
);
}
void
SubNestedSequenceLayer
::
backward
(
const
UpdateCallback
&
callback
)
{
MatrixPtr
inputSeqGrad
=
getInputGrad
(
0
);
MatrixPtr
outputGrad
=
getOutputGrad
();
if
(
inputSeqGrad
)
outputGrad
->
addToRows
(
*
inputSeqGrad
,
*
rowIndice_
);
}
}
// namespace paddle
paddle/gserver/tests/CMakeLists.txt
浏览文件 @
91e8aada
...
...
@@ -66,6 +66,16 @@ add_unittest_without_exec(test_BatchNorm
add_test
(
NAME test_BatchNorm
COMMAND test_BatchNorm
)
################# test_KmaxSeqScore #######################
add_unittest_without_exec
(
test_KmaxSeqScore
test_KmaxSeqScore.cpp
LayerGradUtil.cpp
)
add_test
(
NAME test_KmaxSeqScore
COMMAND test_KmaxSeqScore
)
################## test_Evaluator #######################
add_unittest
(
test_Evaluator
test_Evaluator.cpp
)
...
...
paddle/gserver/tests/test_KmaxSeqScore.cpp
0 → 100644
浏览文件 @
91e8aada
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
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 <gtest/gtest.h>
#include <algorithm>
#include <string>
#include <vector>
#include "ModelConfig.pb.h"
#include "paddle/gserver/layers/DataLayer.h"
#include "paddle/trainer/Trainer.h"
#include "paddle/utils/GlobalConstants.h"
#include "LayerGradUtil.h"
#include "paddle/testing/TestUtil.h"
using
namespace
paddle
;
// NOLINT
using
namespace
std
;
// NOLINT
DECLARE_bool
(
use_gpu
);
DECLARE_int32
(
gpu_id
);
DECLARE_bool
(
thread_local_rand_use_global_seed
);
vector
<
int
>
randSampling
(
int
range
,
int
n
)
{
CHECK_GE
(
range
,
n
);
vector
<
int
>
num
(
range
);
iota
(
begin
(
num
),
end
(
num
),
0
);
if
(
range
==
n
)
return
num
;
random_shuffle
(
begin
(
num
),
end
(
num
));
num
.
resize
(
n
);
return
num
;
}
void
genRandomSeqInfo
(
vector
<
int
>&
seqStartPosition
,
vector
<
int
>&
subSeqStartPosition
)
{
const
int
maxSeqNum
=
100
;
// generate random start position information
int
seqNum
=
1
+
(
rand
()
%
maxSeqNum
);
seqStartPosition
.
resize
(
seqNum
+
1
,
0
);
subSeqStartPosition
.
resize
(
1
,
0
);
for
(
int
i
=
0
;
i
<
seqNum
;
++
i
)
{
int
subSeqLen
=
1
+
(
rand
()
%
maxSeqNum
);
for
(
int
j
=
0
;
j
<
subSeqLen
;
++
j
)
subSeqStartPosition
.
push_back
(
subSeqStartPosition
.
back
()
+
subSeqLen
);
seqStartPosition
[
i
+
1
]
=
subSeqStartPosition
.
back
();
}
}
void
genRandomGroundTruth
(
real
*
values
,
vector
<
vector
<
int
>>&
groundTruth
,
vector
<
int
>&
startPos
,
size_t
beamSize
)
{
groundTruth
.
resize
(
startPos
.
size
()
-
1
,
vector
<
int
>
(
beamSize
,
-
1
));
for
(
size_t
i
=
0
;
i
<
startPos
.
size
()
-
1
;
++
i
)
{
int
seqLen
=
startPos
[
i
+
1
]
-
startPos
[
i
];
vector
<
int
>
pos
=
randSampling
(
seqLen
,
min
(
static_cast
<
int
>
(
beamSize
),
seqLen
));
for
(
size_t
j
=
0
;
j
<
pos
.
size
();
++
j
)
{
groundTruth
[
i
][
j
]
=
pos
[
j
];
values
[
startPos
[
i
]
+
pos
[
j
]]
=
1.
;
}
}
}
void
checkLayerOut
(
vector
<
vector
<
int
>>
groundTruth
,
real
*
layerOut
,
size_t
beamSize
)
{
for
(
size_t
i
=
0
;
i
<
groundTruth
.
size
();
++
i
)
{
int
begPos
=
i
*
beamSize
;
vector
<
real
>
tmp
(
layerOut
+
begPos
,
layerOut
+
begPos
+
beamSize
);
sort
(
begin
(
tmp
),
end
(
tmp
));
sort
(
begin
(
groundTruth
[
i
]),
end
(
groundTruth
[
i
]));
for
(
size_t
j
=
0
;
j
<
beamSize
;
++
j
)
CHECK_EQ
(
tmp
[
j
],
groundTruth
[
i
][
j
]);
}
}
TEST
(
Layer
,
kmaxSeqScoreLayer
)
{
const
size_t
maxBeamSize
=
100
;
int
beamSize
=
1
+
(
rand
()
%
maxBeamSize
);
vector
<
int
>
seqStartPosition
;
vector
<
int
>
subSeqStartPosition
;
genRandomSeqInfo
(
seqStartPosition
,
subSeqStartPosition
);
MatrixPtr
inValue
=
Matrix
::
create
(
subSeqStartPosition
.
back
(),
1
,
false
,
false
);
for
(
auto
hasSubseq
:
{
false
,
true
})
{
vector
<
vector
<
int
>>
groundTruth
;
inValue
->
randomizeUniform
();
genRandomGroundTruth
(
inValue
->
getData
(),
groundTruth
,
hasSubseq
?
subSeqStartPosition
:
seqStartPosition
,
beamSize
);
for
(
auto
useGpu
:
{
false
,
true
})
{
TestConfig
config
;
config
.
layerConfig
.
set_type
(
"kmax_seq_score"
);
config
.
layerConfig
.
set_beam_size
(
beamSize
);
if
(
hasSubseq
)
{
config
.
inputDefs
.
push_back
({
INPUT_SELF_DEFINE_DATA
,
"scores"
,
inValue
,
seqStartPosition
,
subSeqStartPosition
});
}
else
{
config
.
inputDefs
.
push_back
(
{
INPUT_SELF_DEFINE_DATA
,
"scores"
,
inValue
,
seqStartPosition
});
}
config
.
layerConfig
.
add_inputs
();
// data layer initialize
std
::
vector
<
DataLayerPtr
>
dataLayers
;
LayerMap
layerMap
;
vector
<
Argument
>
datas
;
initDataLayer
(
config
,
&
dataLayers
,
&
datas
,
&
layerMap
,
"kmax_seq_score"
,
100
/* actually this parameter is unused in self-defined input*/
,
false
,
useGpu
);
// test layer initialize
std
::
vector
<
ParameterPtr
>
parameters
;
LayerPtr
kmaxSeqScoreLayer
;
FLAGS_use_gpu
=
useGpu
;
initTestLayer
(
config
,
&
layerMap
,
&
parameters
,
&
kmaxSeqScoreLayer
);
kmaxSeqScoreLayer
->
forward
(
PASS_TRAIN
);
const
MatrixPtr
outValue
=
kmaxSeqScoreLayer
->
getOutputValue
();
CHECK_EQ
(
outValue
->
getHeight
(),
hasSubseq
?
subSeqStartPosition
.
size
()
-
1
:
seqStartPosition
.
size
()
-
1
);
CHECK_EQ
(
outValue
->
getWidth
(),
beamSize
);
checkLayerOut
(
groundTruth
,
outValue
->
getData
(),
beamSize
);
}
}
}
int
main
(
int
argc
,
char
**
argv
)
{
testing
::
InitGoogleTest
(
&
argc
,
argv
);
initMain
(
argc
,
argv
);
FLAGS_thread_local_rand_use_global_seed
=
true
;
srand
((
size_t
)(
time
(
NULL
)));
return
RUN_ALL_TESTS
();
}
paddle/gserver/tests/test_LayerGrad.cpp
浏览文件 @
91e8aada
...
...
@@ -1899,6 +1899,84 @@ TEST(Layer, CropLayer) {
}
}
vector
<
real
>
randSampling
(
real
range
,
int
n
)
{
CHECK_GE
(
range
,
n
);
vector
<
real
>
num
(
range
);
iota
(
begin
(
num
),
end
(
num
),
0.
);
if
(
range
==
n
)
return
num
;
random_shuffle
(
begin
(
num
),
end
(
num
));
num
.
resize
(
n
);
sort
(
begin
(
num
),
end
(
num
));
return
num
;
}
TEST
(
Layer
,
SubNestedSequenceLayer
)
{
// layer size is not crutial for this layer,
// so use a small layer size in unittest
const
int
layerSize
=
4
;
const
int
maxSeqNum
=
50
;
const
int
maxSeqLen
=
50
;
const
int
maxBeamSize
=
32
;
srand
((
size_t
)(
time
(
NULL
)));
int
beamSize
=
1
+
(
rand
()
%
maxBeamSize
);
TestConfig
config
;
config
.
layerConfig
.
set_type
(
"sub_nested_seq"
);
config
.
layerConfig
.
set_name
(
"sub_nested_seq_layer"
);
config
.
layerConfig
.
set_size
(
layerSize
);
int
seqNum
=
1
+
(
rand
()
%
maxSeqNum
);
// sequence information for the first input, it is a nested sequence
vector
<
int
>
seqStartPos
(
seqNum
+
1
,
0
);
vector
<
int
>
subSeqStartPos
(
1
,
0
);
// selected indices
MatrixPtr
selectedIndices
=
Matrix
::
create
(
seqNum
,
beamSize
,
false
,
false
);
selectedIndices
->
one
();
selectedIndices
->
mulScalar
(
-
1.
);
real
*
indicesData
=
selectedIndices
->
getData
();
for
(
int
i
=
0
;
i
<
seqNum
;
++
i
)
{
int
subSeqNum
=
1
+
(
rand
()
%
maxSeqNum
);
for
(
int
j
=
0
;
j
<
subSeqNum
;
++
j
)
{
subSeqStartPos
.
push_back
(
subSeqStartPos
.
back
()
+
(
1
+
(
rand
()
%
maxSeqLen
)));
}
vector
<
real
>
selSeqs
=
randSampling
(
static_cast
<
real
>
(
subSeqNum
),
min
(
beamSize
,
subSeqNum
));
memcpy
(
indicesData
+
(
i
*
beamSize
),
selSeqs
.
data
(),
selSeqs
.
size
()
*
sizeof
(
real
));
seqStartPos
[
i
+
1
]
=
subSeqStartPos
.
back
();
}
MatrixPtr
seqInputPtr
=
Matrix
::
create
(
seqStartPos
.
back
(),
layerSize
,
false
,
false
);
seqInputPtr
->
randomizeUniform
();
config
.
inputDefs
.
push_back
({
INPUT_SELF_DEFINE_DATA
,
"nested_seq_input"
,
seqInputPtr
,
seqStartPos
,
subSeqStartPos
});
config
.
layerConfig
.
add_inputs
();
config
.
inputDefs
.
push_back
(
{
INPUT_SELF_DEFINE_DATA
,
"selected_indices"
,
selectedIndices
});
config
.
layerConfig
.
add_inputs
();
for
(
auto
useGpu
:
{
false
,
true
})
{
testLayerGrad
(
config
,
"sub_nested_seq"
,
/* batchSize */
seqNum
,
/* trans */
false
,
/* useGpu*/
useGpu
,
/* useWeight */
false
);
}
}
TEST
(
Layer
,
ClipLayer
)
{
const
size_t
batchSize
=
128
;
const
size_t
size
=
512
;
...
...
paddle/operators/CMakeLists.txt
浏览文件 @
91e8aada
...
...
@@ -67,3 +67,5 @@ op_library(fc_op
op_library
(
recurrent_op SRCS recurrent_op.cc rnn/recurrent_op_utils.cc
DEPS op_desc tensor op_registry operator net_op
)
cc_test
(
recurrent_op_test SRCS recurrent_op_test.cc DEPS recurrent_op gtest mul_op add_op
)
op_library
(
uniform_random_op
SRCS uniform_random_op.cc uniform_random_op.cu
)
paddle/operators/cross_entropy_op.cc
浏览文件 @
91e8aada
...
...
@@ -70,9 +70,10 @@ OnehotCrossEntropy Operator.
namespace
ops
=
paddle
::
operators
;
REGISTER_OP
(
onehot_cross_entropy
,
ops
::
OnehotCrossEntropyOp
,
ops
::
OnehotCrossEntropyOpMaker
);
REGISTER_OP_CPU_KERNEL
(
onehot_cross_entropy
,
ops
::
OnehotCrossEntropyOpKernel
<
paddle
::
platform
::
CPUPlace
,
float
>
);
REGISTER_OP_CPU_KERNEL
(
onehot_cross_entropy
,
ops
::
OnehotCrossEntropyOpKernel
<
paddle
::
platform
::
CPUPlace
,
float
>
);
REGISTER_GRADIENT_OP
(
onehot_cross_entropy
,
onehot_cross_entropy_grad
,
ops
::
OnehotCrossEntropyGradientOp
);
REGISTER_OP_CPU_KERNEL
(
onehot_cross_entropy_grad
,
ops
::
OnehotCrossEntropyGradientOpKernel
<
paddle
::
platform
::
CPUPlace
,
float
>
);
paddle/operators/net_op.h
浏览文件 @
91e8aada
...
...
@@ -59,6 +59,15 @@ class NetOp : public framework::OperatorBase {
}
}
bool
SupportGPU
()
const
override
{
for
(
auto
&
op
:
ops_
)
{
if
(
!
op
->
SupportGPU
())
{
return
false
;
}
}
return
true
;
}
/**
* @brief Add an operator by ptr
*/
...
...
paddle/operators/uniform_random_op.cc
0 → 100644
浏览文件 @
91e8aada
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
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 <random>
#include <type_traits>
#include "paddle/framework/op_registry.h"
#include "paddle/framework/operator.h"
namespace
paddle
{
namespace
operators
{
// It seems that Eigen::Tensor::random in GPU will SEGFAULT.
// Use std::random and thrust::random(thrust is a std library in CUDA) to
// implement uniform random.
template
<
typename
T
>
class
CPUUniformRandomKernel
:
public
framework
::
OpKernel
{
public:
void
Compute
(
const
framework
::
ExecutionContext
&
context
)
const
override
{
auto
*
tensor
=
context
.
Output
<
framework
::
Tensor
>
(
0
);
T
*
data
=
tensor
->
mutable_data
<
T
>
(
context
.
GetPlace
());
unsigned
int
seed
=
static_cast
<
unsigned
int
>
(
context
.
op_
.
GetAttr
<
int
>
(
"seed"
));
std
::
minstd_rand
engine
;
if
(
seed
==
0
)
{
seed
=
std
::
random_device
()();
}
engine
.
seed
(
seed
);
std
::
uniform_real_distribution
<
T
>
dist
(
static_cast
<
T
>
(
context
.
op_
.
GetAttr
<
float
>
(
"min"
)),
static_cast
<
T
>
(
context
.
op_
.
GetAttr
<
float
>
(
"max"
)));
for
(
ssize_t
i
=
0
;
i
<
framework
::
product
(
tensor
->
dims
());
++
i
)
{
data
[
i
]
=
dist
(
engine
);
}
}
};
class
UniformRandomOp
:
public
framework
::
OperatorWithKernel
{
protected:
void
InferShape
(
const
framework
::
InferShapeContext
&
ctx
)
const
override
{
PADDLE_ENFORCE
(
GetAttr
<
float
>
(
"min"
)
<
GetAttr
<
float
>
(
"max"
),
"uniform_random's min must less then max"
);
auto
*
tensor
=
ctx
.
Output
<
framework
::
Tensor
>
(
0
);
auto
dims
=
GetAttr
<
std
::
vector
<
int
>>
(
"dims"
);
tensor
->
Resize
(
framework
::
make_ddim
(
dims
));
}
};
class
UniformRandomOpMaker
:
public
framework
::
OpProtoAndCheckerMaker
{
public:
UniformRandomOpMaker
(
framework
::
OpProto
*
proto
,
framework
::
OpAttrChecker
*
op_checker
)
:
framework
::
OpProtoAndCheckerMaker
(
proto
,
op_checker
)
{
AddOutput
(
"Out"
,
"The output tensor of uniform random op"
);
AddComment
(
R"DOC(Uniform random operator.
Used to initialize tensor with uniform random generator.
)DOC"
);
AddAttr
<
std
::
vector
<
int
>>
(
"dims"
,
"the dimension of random tensor"
);
AddAttr
<
float
>
(
"min"
,
"Minimum value of uniform random"
).
SetDefault
(
-
1.0
f
);
AddAttr
<
float
>
(
"max"
,
"Maximun value of uniform random"
).
SetDefault
(
1.0
f
);
AddAttr
<
int
>
(
"seed"
,
"Random seed of uniform random. "
"0 means generate a seed by system"
)
.
SetDefault
(
0
);
}
};
}
// namespace operators
}
// namespace paddle
REGISTER_OP
(
uniform_random
,
paddle
::
operators
::
UniformRandomOp
,
paddle
::
operators
::
UniformRandomOpMaker
);
REGISTER_OP_CPU_KERNEL
(
uniform_random
,
paddle
::
operators
::
CPUUniformRandomKernel
<
float
>
);
paddle/operators/uniform_random_op.cu
0 → 100644
浏览文件 @
91e8aada
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
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 <thrust/device_ptr.h>
#include <thrust/iterator/counting_iterator.h>
#include <thrust/random.h>
#include <thrust/transform.h>
#include "paddle/framework/op_registry.h"
#include "paddle/framework/operator.h"
namespace
paddle
{
namespace
operators
{
template
<
typename
T
>
struct
UniformGenerator
{
T
min_
,
max_
;
unsigned
int
seed_
;
__host__
__device__
UniformGenerator
(
T
min
,
T
max
,
int
seed
)
:
min_
(
min
),
max_
(
max
),
seed_
(
seed
)
{}
__host__
__device__
T
operator
()(
const
unsigned
int
n
)
const
{
thrust
::
minstd_rand
rng
;
rng
.
seed
(
seed_
);
thrust
::
uniform_real_distribution
<
T
>
dist
(
min_
,
max_
);
rng
.
discard
(
n
);
return
dist
(
rng
);
}
};
// It seems that Eigen::Tensor::random in GPU will SEGFAULT.
// Use std::random and thrust::random(thrust is a std library in CUDA) to
// implement uniform random.
template
<
typename
T
>
class
GPUUniformRandomKernel
:
public
framework
::
OpKernel
{
public:
void
Compute
(
const
framework
::
ExecutionContext
&
context
)
const
override
{
auto
*
tensor
=
context
.
Output
<
framework
::
Tensor
>
(
0
);
T
*
data
=
tensor
->
mutable_data
<
T
>
(
context
.
GetPlace
());
unsigned
int
seed
=
static_cast
<
unsigned
int
>
(
context
.
op_
.
GetAttr
<
int
>
(
"seed"
));
if
(
seed
==
0
)
{
seed
=
std
::
random_device
()();
}
T
min
=
static_cast
<
T
>
(
context
.
op_
.
GetAttr
<
float
>
(
"min"
));
T
max
=
static_cast
<
T
>
(
context
.
op_
.
GetAttr
<
float
>
(
"max"
));
thrust
::
counting_iterator
<
unsigned
int
>
index_sequence_begin
(
0
);
ssize_t
N
=
framework
::
product
(
tensor
->
dims
());
thrust
::
transform
(
index_sequence_begin
,
index_sequence_begin
+
N
,
thrust
::
device_ptr
<
T
>
(
data
),
UniformGenerator
<
T
>
(
min
,
max
,
seed
));
}
};
}
// namespace operators
}
// namespace paddle
REGISTER_OP_GPU_KERNEL
(
uniform_random
,
paddle
::
operators
::
GPUUniformRandomKernel
<
float
>
);
paddle/parameter/Argument.cpp
浏览文件 @
91e8aada
...
...
@@ -666,4 +666,24 @@ void Argument::subArgFrom(const Argument& input,
}
}
void
Argument
::
reorganizeSeqInfo
(
const
ICpuGpuVectorPtr
seqStartPos
,
const
ICpuGpuVectorPtr
subSeqStartPos
,
std
::
vector
<
std
::
vector
<
int
>>&
reorganizedSeqInfo
)
{
int
*
seqStarts
=
seqStartPos
->
getMutableData
(
false
);
int
*
subSeqStarts
=
subSeqStartPos
->
getMutableData
(
false
);
int
seqNum
=
seqStartPos
->
getSize
()
-
1
;
reorganizedSeqInfo
.
resize
(
seqNum
,
std
::
vector
<
int
>
());
int
seqIdx
=
0
;
for
(
size_t
i
=
0
;
i
<
subSeqStartPos
->
getSize
();
++
i
)
{
reorganizedSeqInfo
[
seqIdx
].
push_back
(
subSeqStarts
[
i
]);
if
(
subSeqStarts
[
i
]
==
seqStarts
[
seqIdx
+
1
])
{
seqIdx
++
;
if
(
seqIdx
==
seqNum
)
return
;
reorganizedSeqInfo
[
seqIdx
].
push_back
(
subSeqStarts
[
i
]);
}
}
}
}
// namespace paddle
paddle/parameter/Argument.h
浏览文件 @
91e8aada
...
...
@@ -317,6 +317,30 @@ struct Argument {
*/
void
printValueString
(
std
::
ostream
&
stream
,
const
std
::
string
&
prefix
=
""
)
const
;
/**
* @brief reorganizeSeqInfo will reorganize sequenceStartPositions and
* subSequenceStartPositions into a 2 dimensional arrary: reorganizedSeqInfo.
*
* @param seqStartPos: sequenceStartPositions of an Argument.
* @param subSeqStartPos: subSequenceStartPositions of an Argument.
* @param the reorganized sequence start position information.
*
* Examples:
* seqStartPos: [0, 4, 15, 20, 28]
* subSeqStartPos: [0, 3, 4, 5, 7, 10, 15, 20, 22, 23, 25, 28]
* reorganizedSeqInfo:
* [
* [0,3,4],
* [4,5,7,10,15],
* [15,20],
* [20,22,23,25,28]
* ]
*/
static
void
reorganizeSeqInfo
(
const
ICpuGpuVectorPtr
seqStartPos
,
const
ICpuGpuVectorPtr
subSeqStartPos
,
std
::
vector
<
std
::
vector
<
int
>>&
reorganizedSeqInfo
);
};
}
// namespace paddle
paddle/scripts/travis/build_doc.sh
浏览文件 @
91e8aada
...
...
@@ -5,15 +5,9 @@ set -e
mkdir
-p
$TRAVIS_BUILD_DIR
/build
cd
$TRAVIS_BUILD_DIR
/build
# Compile paddle binaries first
cmake ..
-DCMAKE_BUILD_TYPE
=
Debug
-DWITH_GPU
=
OFF
-DWITH_DOC
=
OFF
-DWITH_MKLDNN
=
OFF
-DWITH_MKLML
=
OFF
-DWITH_GOLANG
=
ON
-DWITH_STYLE_CHECK
=
OFF
mkdir
output
make
-j
`
nproc
`
find ..
-name
'*whl'
| xargs pip
install
# install all wheels.
rm
-rf
*
# Compile Documentation only.
cmake ..
-DCMAKE_BUILD_TYPE
=
Debug
-DWITH_GPU
=
OFF
-DWITH_MKLDNN
=
OFF
-DWITH_MKLML
=
OFF
-DWITH_DOC
=
ON
make
-j
`
nproc
`
gen_proto_py
make
-j
`
nproc
`
paddle_docs paddle_docs_cn
# check websites for broken links
...
...
@@ -35,6 +29,7 @@ TARGET_BRANCH="gh-pages"
SOURCE_BRANCH
=
"master"
# Clone the repo to output directory
mkdir
output
git clone
$REPO
output
cd
output
...
...
proto/CMakeLists.txt
浏览文件 @
91e8aada
...
...
@@ -17,7 +17,7 @@ foreach(filename ${proto_filenames})
COMMAND
${
PROTOBUF_PROTOC_EXECUTABLE
}
ARGS
"--python_out=
${
PROJ_ROOT
}
/python/paddle/proto"
"-I"
${
CMAKE_CURRENT_SOURCE_DIR
}
${
ABS_FIL
}
DEPENDS
${
ABS_FIL
}
${
external_project_dependencies
}
)
DEPENDS
${
ABS_FIL
}
protoc
)
endforeach
()
add_custom_target
(
gen_proto_py ALL DEPENDS
${
PROTO_GEN_PY
}
)
python/paddle/trainer/config_parser.py
浏览文件 @
91e8aada
...
...
@@ -2657,6 +2657,31 @@ class SubSequenceLayer(LayerBase):
self
.
create_bias_parameter
(
bias
,
size
)
@
config_layer
(
'sub_nested_seq'
)
class
SubNestedSequenceLayer
(
LayerBase
):
def
__init__
(
self
,
name
,
inputs
,
selected_indices
,
bias
=
False
,
**
xargs
):
if
isinstance
(
inputs
,
list
):
assert
len
(
inputs
)
==
1
,
(
'the first input of sub_nested_seq '
'layer is a single nested sequence.'
)
inputs
=
inputs
[
0
]
if
isinstance
(
selected_indices
,
list
):
assert
len
(
selected_indices
)
==
1
,
(
'the second input of '
'sub_nested_seq layer is a single layer which is a '
'set of selected indices.'
)
selected_indices
=
selected_indices
[
0
]
super
(
SubNestedSequenceLayer
,
self
).
__init__
(
name
,
'sub_nested_seq'
,
0
,
inputs
=
[
inputs
,
selected_indices
],
**
xargs
)
input_layer0
=
self
.
get_input_layer
(
0
)
size
=
input_layer0
.
size
self
.
set_layer_size
(
size
)
@
config_layer
(
'out_prod'
)
class
OuterProdLayer
(
LayerBase
):
def
__init__
(
self
,
name
,
inputs
,
device
=
None
):
...
...
@@ -3223,6 +3248,16 @@ class CTCLayer(LayerBase):
config_assert
(
len
(
self
.
inputs
)
==
2
,
'CTCLayer must have 2 inputs'
)
@
config_layer
(
'kmax_seq_score'
)
class
KmaxSeqScoreLayer
(
LayerBase
):
def
__init__
(
self
,
name
,
inputs
,
beam_size
,
**
xargs
):
super
(
KmaxSeqScoreLayer
,
self
).
__init__
(
name
,
'kmax_seq_score'
,
0
,
inputs
=
inputs
,
**
xargs
)
config_assert
(
len
(
self
.
inputs
)
==
1
,
'KmaxSeqScoreLayer has only one input.'
)
self
.
config
.
beam_size
=
beam_size
@
config_layer
(
'warp_ctc'
)
class
WarpCTCLayer
(
LayerBase
):
def
__init__
(
self
,
...
...
python/paddle/trainer_config_helpers/layers.py
浏览文件 @
91e8aada
...
...
@@ -129,8 +129,10 @@ __all__ = [
'prelu_layer'
,
'gated_unit_layer'
,
'crop_layer'
,
'sub_nested_seq_layer'
,
'clip_layer'
,
'slice_projection'
,
'kmax_sequence_score_layer'
,
]
...
...
@@ -224,8 +226,11 @@ class LayerType(object):
PRELU
=
'prelu'
CROP_LAYER
=
'crop'
SUB_NESTED_SEQ
=
'sub_nested_seq'
CLIP_LAYER
=
'clip'
KMAX_SEQ_SCORE
=
'kmax_seq_score'
@
staticmethod
def
is_layer_type
(
type_name
):
"""
...
...
@@ -6088,6 +6093,53 @@ def crop_layer(input, offset, axis=2, shape=None, name=None, layer_attr=None):
size
=
l
.
config
.
size
)
@
wrap_name_default
()
@
layer_support
()
def
sub_nested_seq_layer
(
input
,
selected_indices
,
name
=
None
):
"""
The sub_nested_seq_layer accepts two inputs: the first one is a nested
sequence; the second one is a set of selceted indices in the nested sequence.
Then sub_nest_seq_layer trims the first nested sequence input according
to the selected indices to form a new output. This layer is useful in
beam training.
The example usage is:
.. code-block:: python
sub_nest_seq = sub_nested_seq_layer(input=[data, selected_indices])
:param input: A nested sequence.
:type input: LayerOutput
:param selected_indices: a set of sequence indices in the nested sequence.
:type input: LayerOutput
:param name: name of this layer.
:type name: basestring
:return: LayerOutput object.
:rtype: LayerOutput
"""
assert
isinstance
(
input
,
LayerOutput
),
(
'The first input of '
'sub_nested_seq_layer must be a Paddle layer.'
)
assert
isinstance
(
selected_indices
,
LayerOutput
),
(
'The second input of '
'sub_nested_seq_layer must be a Paddle layer.'
)
l
=
Layer
(
inputs
=
input
.
name
,
selected_indices
=
selected_indices
.
name
,
name
=
name
,
type
=
LayerType
.
SUB_NESTED_SEQ
)
return
LayerOutput
(
name
=
name
,
layer_type
=
LayerType
.
SUB_NESTED_SEQ
,
parents
=
input
,
size
=
l
.
config
.
size
)
@
wrap_name_default
(
"clip"
)
def
clip_layer
(
input
,
min
,
max
,
name
=
None
):
"""
...
...
@@ -6109,7 +6161,8 @@ def clip_layer(input, min, max, name=None):
:type min: double
:param max: The upper threshold for clipping.
:type max: double
:return: LayerOutput
:return: LayerOutput object.
:rtype: LayerOutput
"""
Layer
(
name
=
name
,
...
...
@@ -6119,3 +6172,41 @@ def clip_layer(input, min, max, name=None):
max
=
max
)
return
LayerOutput
(
name
,
LayerType
.
CLIP_LAYER
,
parents
=
[
input
],
size
=
input
.
size
)
@
wrap_name_default
()
@
layer_support
()
def
kmax_sequence_score_layer
(
input
,
name
=
None
,
beam_size
=
1
):
"""
This layer accepts one input which are scores over a sequence or a nested
sequence, and returns indices of beam_size sequences with highest scores.
.. code-block:: python
kmax_indices = kmax_sequence_score_layer(input=input_layer, beam_size)
:param name: The Layer Name.
:type name: basestring
:param input: The input layer. It stores scores over a sequence or a nested
sequence and its size must be 1.
:type input: LayerOutput.
:param beam_size: squence indices with top beam_size scores are returned.
:type beam_size: double
:return: LayerOutput object.
:rtype: LayerOutput
"""
assert
isinstance
(
input
,
LayerOutput
),
(
"kmax_sequence_score_layer "
"accepts only one input."
)
assert
input
.
size
==
1
,
(
"input of kmax_sequence_score_layer is a score"
"over a sequence or a nested sequence, so its width must be 1."
)
Layer
(
name
=
name
,
type
=
LayerType
.
KMAX_SEQ_SCORE
,
inputs
=
[
input
.
name
],
beam_size
=
beam_size
)
return
LayerOutput
(
name
,
LayerType
.
KMAX_SEQ_SCORE
,
parents
=
[
input
],
size
=
input
.
size
)
python/paddle/trainer_config_helpers/tests/configs/file_list.sh
浏览文件 @
91e8aada
...
...
@@ -7,6 +7,7 @@ test_rnn_group shared_fc shared_lstm shared_gru test_cost_layers_with_weight
test_spp_layer test_bilinear_interp test_maxout test_bi_grumemory math_ops
test_seq_concat_reshape test_pad test_smooth_l1 test_multiplex_layer
test_prelu_layer test_row_conv test_detection_output_layer test_multibox_loss_layer
test_recursive_topology test_gated_unit_layer test_clip_layer test_row_l2_norm_layer
)
test_recursive_topology test_gated_unit_layer test_clip_layer test_row_l2_norm_layer
test_kmax_seq_socre_layer test_seq_select_layers
)
export
whole_configs
=(
test_split_datasource
)
python/paddle/trainer_config_helpers/tests/configs/protostr/test_kmax_seq_socre_layer.protostr
0 → 100644
浏览文件 @
91e8aada
type: "nn"
layers {
name: "input"
type: "data"
size: 300
active_type: ""
}
layers {
name: "data"
type: "data"
size: 128
active_type: ""
}
layers {
name: "__fc_layer_0__"
type: "fc"
size: 1
active_type: "exponential"
inputs {
input_layer_name: "data"
input_parameter_name: "___fc_layer_0__.w0"
}
bias_parameter_name: "___fc_layer_0__.wbias"
}
layers {
name: "__kmax_sequence_score_layer_0__"
type: "kmax_seq_score"
active_type: ""
inputs {
input_layer_name: "__fc_layer_0__"
}
beam_size: 5
}
parameters {
name: "___fc_layer_0__.w0"
size: 128
initial_mean: 0.0
initial_std: 0.0883883476483
dims: 128
dims: 1
initial_strategy: 0
initial_smart: true
}
parameters {
name: "___fc_layer_0__.wbias"
size: 1
initial_mean: 0.0
initial_std: 0.0
dims: 1
dims: 1
initial_strategy: 0
initial_smart: false
}
input_layer_names: "data"
output_layer_names: "__kmax_sequence_score_layer_0__"
sub_models {
name: "root"
layer_names: "input"
layer_names: "data"
layer_names: "__fc_layer_0__"
layer_names: "__kmax_sequence_score_layer_0__"
input_layer_names: "data"
output_layer_names: "__kmax_sequence_score_layer_0__"
is_recurrent_layer_group: false
}
python/paddle/trainer_config_helpers/tests/configs/protostr/test_seq_select_layers.protostr
0 → 100644
浏览文件 @
91e8aada
type: "nn"
layers {
name: "input_seq"
type: "data"
size: 300
active_type: ""
}
layers {
name: "input"
type: "data"
size: 5
active_type: ""
}
layers {
name: "__sub_nested_seq_layer_0__"
type: "sub_nested_seq"
size: 300
active_type: ""
inputs {
input_layer_name: "input_seq"
}
inputs {
input_layer_name: "input"
}
}
input_layer_names: "input_seq"
output_layer_names: "__sub_nested_seq_layer_0__"
sub_models {
name: "root"
layer_names: "input_seq"
layer_names: "input"
layer_names: "__sub_nested_seq_layer_0__"
input_layer_names: "input_seq"
output_layer_names: "__sub_nested_seq_layer_0__"
is_recurrent_layer_group: false
}
python/paddle/trainer_config_helpers/tests/configs/test_kmax_seq_socre_layer.py
0 → 100644
浏览文件 @
91e8aada
#!/usr/bin/env python
#coding=utf-8
from
paddle.trainer_config_helpers
import
*
data
=
data_layer
(
name
=
'input'
,
size
=
300
)
data
=
data_layer
(
name
=
"data"
,
size
=
128
)
scores
=
fc_layer
(
input
=
data
,
size
=
1
,
act
=
ExpActivation
())
kmax_seq_id
=
kmax_sequence_score_layer
(
input
=
scores
,
beam_size
=
5
)
outputs
(
kmax_seq_id
)
python/paddle/trainer_config_helpers/tests/configs/test_seq_select_layers.py
0 → 100644
浏览文件 @
91e8aada
#!/usr/bin/env python
#coding=utf-8
from
paddle.trainer_config_helpers
import
*
beam_size
=
5
data
=
data_layer
(
name
=
'input_seq'
,
size
=
300
)
selected_ids
=
data_layer
(
name
=
'input'
,
size
=
beam_size
)
sub_nest_seq
=
sub_nested_seq_layer
(
input
=
data
,
selected_indices
=
selected_ids
)
outputs
(
sub_nest_seq
)
python/paddle/v2/framework/tests/CMakeLists.txt
浏览文件 @
91e8aada
...
...
@@ -13,6 +13,7 @@ py_test(test_protobuf SRCS test_protobuf.py)
py_test
(
test_add_two_op SRCS test_add_two_op.py
)
py_test
(
test_sigmoid_op SRCS test_sigmoid_op.py
)
py_test
(
test_softmax_op SRCS test_softmax_op.py
)
py_test
(
test_cross_entropy_op SRCS test_cross_entropy_op.py
)
py_test
(
test_fill_zeros_like_op SRCS test_fill_zeros_like_op.py
)
py_test
(
gradient_checker SRCS gradient_checker.py
)
...
...
@@ -21,3 +22,4 @@ py_test(test_rowwise_add_op SRCS test_rowwise_add_op.py)
py_test
(
test_default_scope_funcs SRCS test_default_scope_funcs.py
)
py_test
(
test_operator SRCS test_operator.py
)
py_test
(
test_uniform_random_op SRCS test_uniform_random_op.py
)
python/paddle/v2/framework/tests/gradient_checker.py
浏览文件 @
91e8aada
import
unittest
import
numpy
import
paddle.v2.framework.core
as
core
from
paddle.v2.framework.op
import
Operator
import
numpy
import
unittest
__all__
=
[
'get_numeric_gradient'
]
def
create_op
(
op_type
):
kwargs
=
dict
()
for
in_name
in
Operator
.
get_op_input_names
(
op_type
):
kwargs
[
in_name
]
=
in_name
for
out_name
in
Operator
.
get_op_output_names
(
op_type
):
kwargs
[
out_name
]
=
out_name
return
Operator
(
op_type
,
**
kwargs
)
def
grad_var_name
(
var_name
):
return
var_name
+
"@GRAD"
def
get_numeric_gradient
(
op
,
input_values
,
output_name
,
input_to_check
,
delta
=
1e-2
,
delta
=
0.005
,
local_scope
=
None
):
"""
Get Numeric Gradient for an operator's input.
...
...
@@ -76,6 +91,113 @@ def get_numeric_gradient(op,
return
gradient_flat
.
reshape
(
tensor_to_check
.
get_dims
())
class
GradientChecker
(
unittest
.
TestCase
):
def
__is_close
(
self
,
numeric_grads
,
scope
,
max_relative_error
):
for
name
in
numeric_grads
:
op_grad
=
numpy
.
array
(
scope
.
find_var
(
grad_var_name
(
name
)).
get_tensor
())
is_close
=
numpy
.
allclose
(
numeric_grads
[
name
],
op_grad
,
rtol
=
max_relative_error
,
atol
=
100
)
if
not
is_close
:
return
False
return
True
def
check_grad
(
self
,
forward_op
,
input_vars
,
inputs_to_check
,
output_name
,
no_grad_set
=
None
,
only_cpu
=
False
,
max_relative_error
=
0.005
):
"""
:param forward_op: used to create backward_op
:param input_vars: numpy value of input variable. The following
computation will use these variables.
:param inputs_to_check: inputs var names that should check gradient.
:param output_name: output name that used to
:param max_relative_error: The relative tolerance parameter.
:param no_grad_set: used when create backward ops
:param only_cpu: only compute and check gradient on cpu kernel.
:return:
"""
if
no_grad_set
is
None
:
no_grad_set
=
set
()
tmp_outs
=
forward_op
.
temp_outputs
()
no_tmp_out
=
filter
(
lambda
name
:
name
not
in
tmp_outs
,
forward_op
.
outputs
())
if
len
(
no_tmp_out
)
!=
1
:
raise
ValueError
(
"non temp out_names should be 1"
)
in_names
=
forward_op
.
inputs
()
for
no_grad
in
no_grad_set
:
if
no_grad
not
in
in_names
:
raise
ValueError
(
"no_grad should be in in_names"
)
backward_op
=
core
.
Operator
.
backward
(
forward_op
,
no_grad_set
)
places
=
[
core
.
CPUPlace
()]
if
not
only_cpu
and
core
.
is_compile_gpu
()
and
backward_op
.
support_gpu
():
places
.
append
(
core
.
GPUPlace
(
0
))
numeric_grad
=
dict
()
# get numeric gradient
for
check_name
in
inputs_to_check
:
numeric_grad
[
check_name
]
=
\
get_numeric_gradient
(
forward_op
,
input_vars
,
output_name
,
check_name
)
# get operator gradient according to different device
for
place
in
places
:
scope
=
core
.
Scope
()
ctx
=
core
.
DeviceContext
.
create
(
place
)
# create input var and set value
for
name
,
value
in
input_vars
.
iteritems
():
if
name
not
in
in_names
:
raise
ValueError
(
name
+
" not in op.inputs_"
)
var
=
scope
.
new_var
(
name
).
get_tensor
()
var
.
set_dims
(
value
.
shape
)
var
.
set
(
value
,
place
)
# create output var
for
out_name
in
forward_op
.
outputs
():
scope
.
new_var
(
out_name
).
get_tensor
()
# infer the shape of output var and compute/set value of output var
forward_op
.
infer_shape
(
scope
)
forward_op
.
run
(
scope
,
ctx
)
# create output grad var
# set shape as the output var
# set value of this grad to ones
for
name
in
forward_op
.
outputs
():
out_tensor
=
scope
.
find_var
(
name
).
get_tensor
()
grad_tensor
=
scope
.
new_var
(
grad_var_name
(
name
)).
get_tensor
()
grad_tensor
.
set_dims
(
out_tensor
.
shape
())
data
=
1.0
*
numpy
.
ones
(
out_tensor
.
shape
())
grad_tensor
.
set
(
data
,
place
)
# create input grad var
for
name
in
backward_op
.
outputs
():
scope
.
new_var
(
name
).
get_tensor
()
# infer the shape of input gradient var and compute/set it's value
# with backward op
backward_op
.
infer_shape
(
scope
)
backward_op
.
run
(
scope
,
ctx
)
if
isinstance
(
place
,
core
.
CPUPlace
):
msg
=
"CPU kernel gradient is not close to numeric gradient"
else
:
if
isinstance
(
place
,
core
.
GPUPlace
):
msg
=
"GPU kernel gradient is not close to numeric gradient"
else
:
raise
ValueError
(
"unknown place "
+
type
(
place
))
self
.
assertTrue
(
self
.
__is_close
(
numeric_grad
,
scope
,
max_relative_error
),
msg
)
if
__name__
==
'__main__'
:
class
GetNumericGradientTest
(
unittest
.
TestCase
):
...
...
@@ -87,4 +209,28 @@ if __name__ == '__main__':
arr
=
get_numeric_gradient
(
add_op
,
{
'X'
:
x
,
"Y"
:
y
},
'Z'
,
'X'
)
self
.
assertAlmostEqual
(
arr
.
mean
(),
1.0
,
delta
=
1e-2
)
def
test_softmax_op
(
self
):
def
stable_softmax
(
x
):
"""Compute the softmax of vector x in a numerically stable way."""
shiftx
=
x
-
numpy
.
max
(
x
)
exps
=
numpy
.
exp
(
shiftx
)
return
exps
/
numpy
.
sum
(
exps
)
def
label_softmax_grad
(
Y
,
dY
):
dX
=
Y
*
0.0
for
i
in
range
(
Y
.
shape
[
0
]):
d
=
numpy
.
dot
(
Y
[
i
,
:],
dY
[
i
,
:])
dX
[
i
,
:]
=
Y
[
i
,
:]
*
(
dY
[
i
,
:]
-
d
)
return
dX
softmax_op
=
Operator
(
"softmax"
,
X
=
"X"
,
Y
=
"Y"
)
X
=
numpy
.
random
.
random
((
2
,
2
)).
astype
(
"float32"
)
Y
=
numpy
.
apply_along_axis
(
stable_softmax
,
1
,
X
)
dY
=
numpy
.
ones
(
Y
.
shape
)
dX
=
label_softmax_grad
(
Y
,
dY
)
arr
=
get_numeric_gradient
(
softmax_op
,
{
"X"
:
X
},
'Y'
,
'X'
)
numpy
.
testing
.
assert_almost_equal
(
arr
,
dX
,
decimal
=
1e-2
)
unittest
.
main
()
python/paddle/v2/framework/tests/op_test_util.py
浏览文件 @
91e8aada
import
paddle.v2.framework.core
as
core
import
unittest
import
numpy
import
paddle.v2.framework.core
as
core
from
paddle.v2.framework.op
import
Operator
...
...
@@ -24,7 +23,7 @@ class OpTestMeta(type):
scope
=
core
.
Scope
()
kwargs
=
dict
()
places
=
[
core
.
CPUPlace
()]
if
core
.
is_compile_gpu
()
and
core
.
Operator
.
support_gpu
(
self
.
type
)
:
if
core
.
is_compile_gpu
():
places
.
append
(
core
.
GPUPlace
(
0
))
for
place
in
places
:
...
...
@@ -53,6 +52,8 @@ class OpTestMeta(type):
kwargs
[
attr_name
]
=
self
.
attrs
[
attr_name
]
op
=
Operator
(
self
.
type
,
**
kwargs
)
if
isinstance
(
place
,
core
.
GPUPlace
)
and
not
op
.
support_gpu
():
return
op
.
infer_shape
(
scope
)
...
...
python/paddle/v2/framework/tests/test_cross_entropy_op.py
浏览文件 @
91e8aada
import
unittest
import
numpy
from
op_test_util
import
OpTestMeta
from
gradient_checker
import
GradientChecker
,
create_op
class
Test
SGD
(
unittest
.
TestCase
):
class
Test
CrossEntropy
(
unittest
.
TestCase
):
__metaclass__
=
OpTestMeta
def
setUp
(
self
):
...
...
@@ -20,7 +21,18 @@ class TestSGD(unittest.TestCase):
self
.
outputs
=
{
'Y'
:
numpy
.
array
(
Y
).
astype
(
"float32"
)}
# TODO(superjom) add gradient check
class
CrossEntropyGradOpTest
(
GradientChecker
):
def
test_softmax_grad
(
self
):
op
=
create_op
(
"onehot_cross_entropy"
)
batch_size
=
100
class_num
=
10
inputs
=
{
"X"
:
numpy
.
random
.
uniform
(
0.1
,
1.0
,
[
batch_size
,
class_num
]).
astype
(
"float32"
),
"label"
:
(
class_num
/
2
)
*
numpy
.
ones
(
batch_size
).
astype
(
"int32"
)
}
self
.
check_grad
(
op
,
inputs
,
set
(
"X"
),
"Y"
)
if
__name__
==
"__main__"
:
unittest
.
main
()
python/paddle/v2/framework/tests/test_softmax_op.py
浏览文件 @
91e8aada
import
unittest
import
numpy
as
np
import
paddle.v2.framework.core
as
core
from
paddle.v2.framework.op
import
Operator
from
gradient_checker
import
GradientChecker
,
create_op
from
op_test_util
import
OpTestMeta
...
...
@@ -25,62 +24,11 @@ class TestSoftmaxOp(unittest.TestCase):
}
class
TestSoftmaxGradOp
(
unittest
.
TestCase
):
def
test_softmax_grad
(
self
):
op
=
Operator
(
'softmax'
,
X
=
"X"
,
Y
=
"Y"
)
backward_op
=
core
.
Operator
.
backward
(
op
,
set
())
self
.
assertEqual
(
backward_op
.
type
(),
"softmax_grad"
)
expected
=
'''Op(softmax_grad), inputs:(X, Y, Y@GRAD), outputs:(X@GRAD).'''
self
.
assertEqual
(
expected
,
str
(
backward_op
))
batch_size
=
3
class_num
=
5
# Initialize X and add 1e-2 for numerical stability
Y
=
np
.
random
.
rand
(
batch_size
,
class_num
).
astype
(
np
.
float32
)
Y
=
Y
+
1e-2
dY
=
np
.
random
.
rand
(
batch_size
,
class_num
).
astype
(
np
.
float32
)
# Reference implementation of cross entropy with soft labels
def
label_softmax_grad
(
Y
,
dY
):
dX
=
Y
*
0.0
for
i
in
range
(
batch_size
):
d
=
np
.
dot
(
Y
[
i
,
:],
dY
[
i
,
:])
dX
[
i
,
:]
=
Y
[
i
,
:]
*
(
dY
[
i
,
:]
-
d
)
return
dX
expected
=
label_softmax_grad
(
Y
,
dY
)
scope
=
core
.
Scope
()
places
=
[]
places
.
append
(
core
.
CPUPlace
())
if
core
.
is_compile_gpu
():
places
.
append
(
core
.
GPUPlace
(
0
))
for
place
in
places
:
y
=
scope
.
new_var
(
"Y"
)
y_tensor
=
y
.
get_tensor
()
y_tensor
.
set_dims
([
batch_size
,
class_num
])
y_tensor
.
alloc_float
(
place
)
y_tensor
.
set
(
Y
,
place
)
dy
=
scope
.
new_var
(
"Y@GRAD"
)
dy_tensor
=
dy
.
get_tensor
()
dy_tensor
.
set_dims
([
batch_size
,
class_num
])
dy_tensor
.
alloc_float
(
place
)
dy_tensor
.
set
(
dY
,
place
)
x
=
scope
.
new_var
(
"X"
)
dx
=
scope
.
new_var
(
"X@GRAD"
)
tensor
=
scope
.
find_var
(
"X@GRAD"
).
get_tensor
()
backward_op
.
infer_shape
(
scope
)
self
.
assertEqual
([
batch_size
,
class_num
],
tensor
.
shape
())
ctx
=
core
.
DeviceContext
.
create
(
place
)
backward_op
.
run
(
scope
,
ctx
)
actual
=
np
.
array
(
tensor
)
np
.
testing
.
assert_almost_equal
(
actual
,
expected
,
decimal
=
3
)
class
SoftmaxGradOpTest
(
GradientChecker
):
def
test_softmax
(
self
):
op
=
create_op
(
"softmax"
)
inputs
=
{
"X"
:
np
.
random
.
uniform
(
0.1
,
1
,
[
10
,
10
]).
astype
(
"float32"
)}
self
.
check_grad
(
op
,
inputs
,
set
(
"X"
),
"Y"
)
if
__name__
==
'__main__'
:
...
...
python/paddle/v2/framework/tests/test_uniform_random_op.py
0 → 100644
浏览文件 @
91e8aada
import
unittest
from
paddle.v2.framework.op
import
Operator
import
paddle.v2.framework.core
as
core
import
numpy
class
UniformRandomTest
(
unittest
.
TestCase
):
def
test_uniform_random_cpu
(
self
):
self
.
uniform_random_test
(
place
=
core
.
CPUPlace
())
def
test_uniform_random_gpu
(
self
):
if
core
.
is_compile_gpu
():
self
.
uniform_random_test
(
place
=
core
.
GPUPlace
(
0
))
def
uniform_random_test
(
self
,
place
):
scope
=
core
.
Scope
()
scope
.
new_var
(
"X"
).
get_tensor
()
op
=
Operator
(
"uniform_random"
,
Out
=
"X"
,
dims
=
[
1000
,
784
],
min
=-
5.0
,
max
=
10.0
,
seed
=
10
)
op
.
infer_shape
(
scope
)
ctx
=
core
.
DeviceContext
.
create
(
place
)
op
.
run
(
scope
,
ctx
)
tensor
=
numpy
.
array
(
scope
.
find_var
(
"X"
).
get_tensor
())
self
.
assertAlmostEqual
(
tensor
.
mean
(),
2.5
,
delta
=
0.1
)
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