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
OPTHREE
Paddle
提交
e2192354
P
Paddle
项目概览
OPTHREE
/
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
e2192354
编写于
1月 05, 2018
作者:
G
guosheng
浏览文件
操作
浏览文件
下载
电子邮件补丁
差异文件
Enhance reorder_lod_tensor_by_rank_op to support Tensor
上级
e57a40b8
变更
4
隐藏空白更改
内联
并排
Showing
4 changed file
with
196 addition
and
42 deletion
+196
-42
paddle/operators/reduce_op.cc
paddle/operators/reduce_op.cc
+1
-0
paddle/operators/reorder_lod_tensor_by_rank_op.cc
paddle/operators/reorder_lod_tensor_by_rank_op.cc
+23
-10
python/paddle/v2/fluid/layers/control_flow.py
python/paddle/v2/fluid/layers/control_flow.py
+1
-1
python/paddle/v2/fluid/tests/test_reorder_lod_tensor.py
python/paddle/v2/fluid/tests/test_reorder_lod_tensor.py
+171
-31
未找到文件。
paddle/operators/reduce_op.cc
浏览文件 @
e2192354
...
...
@@ -77,6 +77,7 @@ class ReduceGradOp : public framework::OperatorWithKernel {
auto
x_grad_name
=
framework
::
GradVarName
(
"X"
);
if
(
ctx
->
HasOutput
(
x_grad_name
))
{
ctx
->
SetOutputDim
(
x_grad_name
,
x_dims
);
ctx
->
ShareLoD
(
"X"
,
/*->*/
x_grad_name
);
}
}
};
...
...
paddle/operators/reorder_lod_tensor_by_rank_op.cc
浏览文件 @
e2192354
...
...
@@ -88,20 +88,33 @@ class ReorderLoDTensorByRankTableBase : public framework::OperatorBase {
std
::
vector
<
AbsoluteRankTableItem
>
GetAbsoluteOffsetAndLengthByLoDRankTable
(
const
framework
::
LoDTensor
&
x
)
const
{
std
::
vector
<
AbsoluteRankTableItem
>
absolute_table
;
size_t
level
=
0
;
size_t
size
=
x
.
lod
()[
level
].
size
();
for
(
size_t
i
=
0
;
i
<
size
-
1
;
++
i
)
{
auto
lod_offset
=
framework
::
GetSubLoDAndAbsoluteOffset
(
x
.
lod
(),
i
,
i
+
1
,
level
);
if
(
x
.
lod
().
empty
())
{
// For Tensor without lod, such as the output of sequence_pool_op
size_t
size
=
x
.
dims
()[
0
];
absolute_table
.
reserve
(
size
);
for
(
size_t
i
=
0
;
i
<
size
;
++
i
)
{
absolute_table
.
emplace_back
();
absolute_table
.
back
().
length
=
1
;
absolute_table
.
back
().
offset
=
i
;
}
}
else
{
size_t
level
=
0
;
size_t
size
=
x
.
lod
()[
level
].
size
();
for
(
size_t
i
=
0
;
i
<
size
-
1
;
++
i
)
{
auto
lod_offset
=
framework
::
GetSubLoDAndAbsoluteOffset
(
x
.
lod
(),
i
,
i
+
1
,
level
);
auto
&
offset
=
lod_offset
.
second
;
auto
&
offset
=
lod_offset
.
second
;
absolute_table
.
emplace_back
();
absolute_table
.
back
().
length
=
offset
.
second
-
offset
.
first
;
absolute_table
.
back
().
offset
=
offset
.
first
;
absolute_table
.
back
().
lod
=
lod_offset
.
first
;
absolute_table
.
emplace_back
();
absolute_table
.
back
().
length
=
offset
.
second
-
offset
.
first
;
absolute_table
.
back
().
offset
=
offset
.
first
;
absolute_table
.
back
().
lod
=
lod_offset
.
first
;
}
}
return
absolute_table
;
}
...
...
python/paddle/v2/fluid/layers/control_flow.py
浏览文件 @
e2192354
...
...
@@ -464,7 +464,7 @@ def lod_rank_table(x, level=0):
"""LoD Rank Table Operator. Given an input variable **x** and a level number
of LoD, this layer creates a LodRankTable object. A LoDRankTable object
contains a list of bi-element tuples. Each tuple consists of an index and
a length, both of which are int type. Ref
f
ering to specified level of LoD,
a length, both of which are int type. Refering to specified level of LoD,
the index is the sequence index number and the length representes the
sequence length. Please note that the list is ranked in descending order by
the length. The following is an example:
...
...
python/paddle/v2/fluid/tests/test_reorder_lod_tensor.py
浏览文件 @
e2192354
import
unittest
import
paddle.v2.fluid
as
fluid
import
paddle.v2.fluid.core
as
core
import
numpy
class
TestReorderLoDTensor
(
unittest
.
TestCase
):
def
test_reorder
(
self
):
dat
=
fluid
.
layers
.
data
(
name
=
'input'
,
shape
=
[
1
],
lod_level
=
2
)
num_seq
=
5
# [name, dim, lod_level] pair indicating data info of source and target
data_desc
=
([
'input'
,
9
,
0
],
[
'ref'
,
5
,
1
])
@
classmethod
def
setUpClass
(
cls
):
cls
.
set_program
()
@
classmethod
def
set_program
(
cls
):
dat
=
fluid
.
layers
.
data
(
name
=
cls
.
data_desc
[
0
][
0
],
shape
=
[
cls
.
data_desc
[
0
][
1
]])
dat
.
stop_gradient
=
False
rank_dat
=
fluid
.
layers
.
data
(
name
=
'ref'
,
shape
=
[
1
],
lod_level
=
1
)
rank_dat
=
fluid
.
layers
.
data
(
name
=
cls
.
data_desc
[
1
][
0
],
shape
=
[
cls
.
data_desc
[
1
][
1
]])
table
=
fluid
.
layers
.
lod_rank_table
(
rank_dat
)
new_dat
=
fluid
.
layers
.
reorder_lod_tensor_by_rank
(
x
=
dat
,
rank_table
=
table
)
loss
=
fluid
.
layers
.
mean
(
x
=
new_dat
)
loss
=
fluid
.
layers
.
reduce_sum
(
new_dat
)
fluid
.
backward
.
append_backward
(
loss
=
loss
)
cls
.
fetch_list
=
[
new_dat
,
cls
.
data_desc
[
0
][
0
]
+
'@GRAD'
]
def
run_program
(
self
):
outputs
=
[]
input_grads
=
[]
places
=
[
core
.
CPUPlace
()]
if
core
.
is_compile_gpu
():
places
.
append
(
core
.
CUDAPlace
(
0
))
for
place
in
places
:
self
.
set_inputs
(
place
)
exe
=
fluid
.
Executor
(
place
)
output
,
input_grad
=
exe
.
run
(
fluid
.
default_main_program
(),
feed
=
self
.
inputs
,
fetch_list
=
self
.
fetch_list
,
return_numpy
=
False
)
outputs
.
append
(
output
)
input_grads
.
append
(
input_grad
)
self
.
actual_outputs
=
outputs
self
.
actual_grads
=
input_grads
def
set_data
(
self
):
self
.
data
=
{}
for
desc
in
self
.
data_desc
:
data_name
=
desc
[
0
]
data_dim
=
desc
[
1
]
data_lod_level
=
desc
[
2
]
data_lod
=
[]
for
i
in
range
(
data_lod_level
):
lod_level_i
=
numpy
.
random
.
randint
(
low
=
1
,
high
=
5
,
size
=
self
.
num_seq
if
i
==
0
else
lod_level_i
[
-
1
])
lod_level_i
=
[
0
]
+
numpy
.
cumsum
(
lod_level_i
).
tolist
()
data_lod
.
append
(
lod_level_i
)
data_value
=
numpy
.
random
.
random
(
size
=
[
data_lod
[
-
1
][
-
1
]
if
data_lod
else
self
.
num_seq
,
data_dim
]).
astype
(
'float32'
)
self
.
data
[
data_name
]
=
(
data_value
,
data_lod
)
def
set_inputs
(
self
,
place
):
self
.
inputs
=
{}
for
desc
in
self
.
data_desc
:
tensor
=
fluid
.
Tensor
()
tensor
.
set
(
self
.
data
[
desc
[
0
]][
0
],
place
)
if
self
.
data
[
desc
[
0
]][
1
]:
tensor
.
set_lod
(
self
.
data
[
desc
[
0
]][
1
])
self
.
inputs
[
desc
[
0
]]
=
tensor
def
reorder
(
self
):
level
=
0
# compute the rank_table according to ref_lod
ref_lod
=
self
.
data
[
self
.
data_desc
[
1
][
0
]][
1
][
level
]
rank_table
=
[]
# list of (index, length)
for
i
in
range
(
len
(
ref_lod
)
-
1
):
rank_table
.
append
((
i
,
ref_lod
[
i
+
1
]
-
ref_lod
[
i
]))
rank_table
=
sorted
(
rank_table
,
lambda
x
,
y
:
y
[
1
]
-
x
[
1
])
# compute the input sequence info according to input_lod
input_value
,
input_lod
=
self
.
data
[
self
.
data_desc
[
0
][
0
]]
input_table
=
[]
# list of (offset, length, sub_lod)
if
input_lod
:
for
i
in
range
(
len
(
input_lod
[
level
])
-
1
):
start_idx
=
i
end_idx
=
i
+
1
sub_lod
=
[]
for
lod_level_i
in
input_lod
[
level
:]:
sub_lod_i
=
[]
for
idx
in
range
(
start_idx
,
end_idx
):
sub_lod_i
.
append
(
lod_level_i
[
idx
+
1
]
-
lod_level_i
[
idx
])
sub_lod
.
append
(
sub_lod_i
)
start_idx
=
lod_level_i
[
start_idx
]
end_idx
=
lod_level_i
[
end_idx
]
input_table
.
append
((
start_idx
,
end_idx
-
start_idx
,
sub_lod
))
else
:
input_table
=
[(
i
,
1
,
[])
for
i
in
range
(
len
(
rank_table
))]
# reorder by rank_table
output_value
=
numpy
.
zeros_like
(
input_value
)
output_lod
=
[]
offset
=
0
for
index
,
length
in
rank_table
:
input_seq_start
=
input_table
[
index
][
0
]
input_seq_len
=
input_table
[
index
][
1
]
input_seq_end
=
input_seq_start
+
input_seq_len
output_value
[
offset
:
offset
+
input_seq_len
]
=
input_value
[
input_seq_start
:
input_seq_end
]
offset
+=
input_seq_len
input_seq_sub_lod
=
input_table
[
index
][
2
]
if
len
(
output_lod
)
==
0
:
output_lod
=
[[
0
]
for
i
in
input_seq_sub_lod
]
for
i
,
sub_lod_i
in
enumerate
(
input_seq_sub_lod
):
for
idx_sub
in
sub_lod_i
:
output_lod
[
i
].
append
(
output_lod
[
i
][
-
1
]
+
idx_sub
)
return
output_value
,
output_lod
def
test_reorder_lod_tensor
(
self
):
self
.
data_desc
[
0
][
-
1
]
=
2
# input is lod_tensor
self
.
set_data
()
self
.
run_program
()
# check output
expect_output
,
expect_output_lod
=
self
.
reorder
()
for
actual_output
in
self
.
actual_outputs
:
self
.
assertTrue
(
numpy
.
allclose
(
numpy
.
array
(
actual_output
),
expect_output
,
atol
=
0.001
))
self
.
assertEqual
(
expect_output_lod
,
actual_output
.
lod
())
# check gradient
expect_grad
=
numpy
.
ones_like
(
self
.
data
[
self
.
data_desc
[
0
][
0
]][
0
])
expect_grad_lod
=
self
.
data
[
self
.
data_desc
[
0
][
0
]][
1
]
for
actual_grad
in
self
.
actual_grads
:
self
.
assertTrue
(
numpy
.
allclose
(
numpy
.
array
(
actual_grad
),
expect_grad
,
atol
=
0.001
))
self
.
assertEqual
(
expect_grad_lod
,
actual_grad
.
lod
())
def
test_reorder_tensor
(
self
):
self
.
data_desc
[
0
][
-
1
]
=
0
# input is tensor
self
.
set_data
()
self
.
run_program
()
# check output
expect_output
,
expect_output_lod
=
self
.
reorder
()
for
actual_output
in
self
.
actual_outputs
:
self
.
assertTrue
(
numpy
.
allclose
(
numpy
.
array
(
actual_output
),
expect_output
,
atol
=
0.001
))
self
.
assertEqual
(
expect_output_lod
,
actual_output
.
lod
())
# check gradient
expect_grad
=
numpy
.
ones_like
(
self
.
data
[
self
.
data_desc
[
0
][
0
]][
0
])
expect_grad_lod
=
self
.
data
[
self
.
data_desc
[
0
][
0
]][
1
]
for
actual_grad
in
self
.
actual_grads
:
self
.
assertTrue
(
numpy
.
allclose
(
numpy
.
array
(
actual_grad
),
expect_grad
,
atol
=
0.001
))
self
.
assertEqual
(
expect_grad_lod
,
actual_grad
.
lod
())
global
outputs_from_tensor_implicit_lod
outputs_from_tensor_implicit_lod
=
self
.
actual_outputs
cpu
=
fluid
.
CPUPlace
()
exe
=
fluid
.
Executor
(
cpu
)
exe
.
run
(
fluid
.
default_startup_program
())
ref
=
fluid
.
Tensor
()
ref_lod
=
[
0
,
3
,
4
,
7
,
8
,
14
]
ref
.
set_lod
([
ref_lod
])
ref
.
set
(
numpy
.
random
.
random
(
size
=
[
14
,
1
]).
astype
(
'float32'
),
cpu
)
input
=
fluid
.
Tensor
()
lod_level_0
=
numpy
.
random
.
randint
(
low
=
1
,
high
=
5
,
size
=
5
)
lod_level_0
=
[
0
]
+
numpy
.
cumsum
(
lod_level_0
).
tolist
()
lod_level_1
=
numpy
.
random
.
randint
(
low
=
1
,
high
=
5
,
size
=
lod_level_0
[
-
1
])
lod_level_1
=
[
0
]
+
numpy
.
cumsum
(
lod_level_1
).
tolist
()
input
.
set_lod
([
lod_level_0
,
lod_level_1
])
input
.
set
(
numpy
.
random
.
random
(
size
=
[
lod_level_1
[
-
1
],
1
]).
astype
(
'float32'
),
cpu
)
ig
=
exe
.
run
(
fluid
.
default_main_program
(),
feed
=
{
'input'
:
input
,
'ref'
:
ref
},
fetch_list
=
[
'input@GRAD'
],
return_numpy
=
False
)[
0
]
self
.
assertAlmostEqual
(
numpy
.
array
(
ig
).
sum
(),
1.0
,
delta
=
0.001
)
self
.
assertEqual
(
input
.
lod
(),
ig
.
lod
())
# compare outputs between LodTensors with explicit and implicit lod
# use the same data but set the input lod explicitly
input_lod
=
[[
i
for
i
in
range
(
len
(
self
.
data
[
self
.
data_desc
[
0
][
0
]][
0
])
+
1
)
]]
self
.
inputs
[
self
.
data_desc
[
0
][
0
]].
set_lod
(
input_lod
)
# preserve the output of LodTensor with implicit lod to compare
expect_output
=
[
numpy
.
array
(
actual_output
)
for
actual_output
in
self
.
actual_outputs
]
self
.
run_program
()
for
actual_output
in
self
.
actual_outputs
:
self
.
assertTrue
(
numpy
.
allclose
(
numpy
.
array
(
actual_output
),
expect_output
,
atol
=
0.001
))
if
__name__
==
'__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