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
DS小龙哥
Paddle
提交
20667e1e
P
Paddle
项目概览
DS小龙哥
/
Paddle
与 Fork 源项目一致
Fork自
PaddlePaddle / Paddle
通知
1
Star
0
Fork
0
代码
文件
提交
分支
Tags
贡献者
分支图
Diff
Issue
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
20667e1e
编写于
11月 06, 2017
作者:
Y
Yang Yang(Tony)
提交者:
GitHub
11月 06, 2017
浏览文件
操作
浏览文件
下载
电子邮件补丁
差异文件
add fill_constant_batch_size_like_op to Static RNN's h_boot (#5332)
上级
70154597
变更
4
显示空白变更内容
内联
并排
Showing
4 changed file
with
159 addition
and
65 deletion
+159
-65
paddle/operators/fill_constant_batch_size_like_op.cc
paddle/operators/fill_constant_batch_size_like_op.cc
+19
-12
python/paddle/v2/framework/layers.py
python/paddle/v2/framework/layers.py
+28
-8
python/paddle/v2/framework/tests/test_fill_constant_batch_size_like_op.py
.../framework/tests/test_fill_constant_batch_size_like_op.py
+8
-3
python/paddle/v2/framework/tests/test_recurrent_op.py
python/paddle/v2/framework/tests/test_recurrent_op.py
+104
-42
未找到文件。
paddle/operators/fill_constant_batch_size_like_op.cc
浏览文件 @
20667e1e
...
...
@@ -34,15 +34,18 @@ class FillConstantBatchSizeLikeOp : public framework::OperatorWithKernel {
std
::
vector
<
int64_t
>
shape_int64
(
shape
.
size
(),
0
);
std
::
transform
(
shape
.
begin
(),
shape
.
end
(),
shape_int64
.
begin
(),
[](
int
a
)
{
return
static_cast
<
int64_t
>
(
a
);
});
auto
dims
=
framework
::
make_ddim
(
shape_int64
);
auto
output_dim
=
framework
::
make_ddim
(
shape_int64
);
int
dim_idx
=
ctx
->
Attrs
().
Get
<
int
>
(
"dim_idx"
);
PADDLE_ENFORCE_GE
(
dim_idx
,
0
);
PADDLE_ENFORCE_GT
(
static_cast
<
int
>
(
shape
.
size
()),
dim_idx
);
PADDLE_ENFORCE_GT
(
ctx
->
GetInputDim
(
"Input"
).
size
(),
dim_idx
);
int
input_dim_idx
=
ctx
->
Attrs
().
Get
<
int
>
(
"input_dim_idx"
);
PADDLE_ENFORCE_GE
(
input_dim_idx
,
0
);
PADDLE_ENFORCE_GT
(
ctx
->
GetInputDim
(
"Input"
).
size
(),
input_dim_idx
);
dims
[
dim_idx
]
=
ctx
->
GetInputDim
(
"Input"
)[
dim_idx
];
ctx
->
SetOutputDim
(
"Out"
,
dims
);
int
output_dim_idx
=
ctx
->
Attrs
().
Get
<
int
>
(
"output_dim_idx"
);
PADDLE_ENFORCE_GE
(
output_dim_idx
,
0
);
PADDLE_ENFORCE_GT
(
static_cast
<
int
>
(
shape
.
size
()),
output_dim_idx
);
output_dim
[
output_dim_idx
]
=
ctx
->
GetInputDim
(
"Input"
)[
input_dim_idx
];
ctx
->
SetOutputDim
(
"Out"
,
output_dim
);
}
protected:
...
...
@@ -69,8 +72,11 @@ class FillConstantBatchSizeLikeOpMaker
"(Tensor) Tensor of specified shape will be filled "
"with the specified value"
);
AddAttr
<
std
::
vector
<
int
>>
(
"shape"
,
"(vector<int>) The shape of the output"
);
AddAttr
<
int
>
(
"dim_idx"
,
"(int, default 0) The index of batch size dimension"
)
AddAttr
<
int
>
(
"input_dim_idx"
,
"(int, default 0) the index of input's batch size dimension"
)
.
SetDefault
(
0
);
AddAttr
<
int
>
(
"output_dim_idx"
,
"(int, default 0) the index of output's batch size dimension"
)
.
SetDefault
(
0
);
AddAttr
<
float
>
(
"value"
,
"(float, default 0) The value to be filled"
)
.
SetDefault
(
0.0
f
);
...
...
@@ -86,8 +92,9 @@ Fill up a variable with specified constant value.
}
// namespace paddle
namespace
ops
=
paddle
::
operators
;
REGISTER_OP
_WITHOUT_GRADIENT
(
fill_constant_batch_size_like
,
REGISTER_OP
ERATOR
(
fill_constant_batch_size_like
,
ops
::
FillConstantBatchSizeLikeOp
,
paddle
::
framework
::
EmptyGradOpMaker
,
ops
::
FillConstantBatchSizeLikeOpMaker
);
REGISTER_OP_CPU_KERNEL
(
fill_constant_batch_size_like
,
...
...
python/paddle/v2/framework/layers.py
浏览文件 @
20667e1e
...
...
@@ -581,25 +581,45 @@ class StaticRNN(object):
if
self
.
status
!=
StaticRNN
.
IN_RNN_BLOCK
:
raise
ValueError
(
"You must invoke {0} in rnn block"
.
format
(
method
))
def
memory
(
self
,
init
=
None
,
shape
=
None
,
dtype
=
None
,
init_value
=
0
):
def
memory
(
self
,
init
=
None
,
shape
=
None
,
batch_ref
=
None
,
init_value
=
0.0
,
init_batch_dim_idx
=
0
,
ref_batch_dim_idx
=
1
):
'''
:param init: boot memory, if not set, a shape, batch_ref must be provided
:param shape: shape of the boot memory
:param batch_ref: batch size reference variable
:param init_value: the init value of boot memory
:param init_batch_dim_idx: the index of batch size in init's dimension
:param ref_batch_dim_idx: the index of batch size in batch_ref's dimension
:return: boot memory
'''
self
.
_assert_in_rnn_block_
(
'memory'
)
if
init
is
None
:
if
shape
is
None
or
dtype
is
None
:
if
shape
is
None
or
batch_ref
is
None
:
raise
ValueError
(
"if init is None, memory at least need shape and
dtype
"
)
"if init is None, memory at least need shape and
batch_ref
"
)
parent_block
=
self
.
parent_block
()
var_name
=
unique_name
(
"@"
.
join
([
self
.
helper
.
name
,
"memory_boot"
]))
boot_var
=
parent_block
.
create_var
(
name
=
var_name
,
shape
=
shape
,
dtype
=
dtype
,
persistable
=
False
)
name
=
var_name
,
shape
=
shape
,
dtype
=
batch_ref
.
data_type
,
persistable
=
False
)
parent_block
.
append_op
(
type
=
"fill_constant"
,
inputs
=
{},
type
=
"fill_constant
_batch_size_like
"
,
inputs
=
{
'Input'
:
[
batch_ref
]
},
outputs
=
{
'Out'
:
[
boot_var
]},
attrs
=
{
'value'
:
init_value
,
'shape'
:
[
40
]
+
list
(
boot_var
.
shape
[
1
:]),
'data_type'
:
boot_var
.
data_type
'shape'
:
boot_var
.
shape
,
'data_type'
:
boot_var
.
data_type
,
'input_dim_idx'
:
ref_batch_dim_idx
,
'output_dim_idx'
:
init_batch_dim_idx
})
return
self
.
memory
(
init
=
boot_var
)
...
...
python/paddle/v2/framework/tests/test_fill_constant_batch_size_like_op.py
浏览文件 @
20667e1e
...
...
@@ -21,9 +21,14 @@ class TestFillConstantBatchSizeLikeWhenSecondDimIsBatchSize(OpTest):
def
setUp
(
self
):
self
.
op_type
=
"fill_constant_batch_size_like"
self
.
inputs
=
{
'Input'
:
np
.
random
.
random
((
219
,
232
)).
astype
(
"float32"
)}
self
.
attrs
=
{
'value'
:
3.5
,
'shape'
:
[
132
,
-
1
,
7
],
'dim_idx'
:
1
}
out
=
np
.
random
.
random
((
132
,
232
,
7
)).
astype
(
"float32"
)
self
.
attrs
=
{
'value'
:
3.5
,
'shape'
:
[
132
,
-
1
,
7
],
'input_dim_idx'
:
0
,
'output_dim_idx'
:
1
}
out
=
np
.
random
.
random
((
132
,
219
,
7
)).
astype
(
"float32"
)
out
.
fill
(
3.5
)
self
.
outputs
=
{
'Out'
:
out
}
...
...
python/paddle/v2/framework/tests/test_recurrent_op.py
浏览文件 @
20667e1e
import
unittest
import
logging
from
op_test
import
get_numeric_gradient
from
paddle.v2.framework.layers
import
*
import
paddle.v2.framework.layers
as
layers
from
paddle.v2.framework.framework
import
Program
from
paddle.v2.framework.executor
import
Executor
from
paddle.v2.framework.backward
import
append_backward_ops
...
...
@@ -16,8 +13,8 @@ class PyRNNBase(object):
self
.
x
=
np
.
ones
(
shape
=
input_shape
).
astype
(
"float32"
)
self
.
y
=
np
.
zeros
(
shape
=
output_shape
).
astype
(
"float32"
)
def
step
(
self
):
pass
def
step
(
self
,
step_id
,
x
):
raise
NotImplementedError
def
forward
(
self
):
for
step_id
in
range
(
self
.
x
.
shape
[
0
]):
...
...
@@ -116,30 +113,30 @@ class RecurrentOpTest1(unittest.TestCase):
self
.
output_shape
=
(
self
.
sent_len
,
self
.
batch_size
,
self
.
input_dim
)
self
.
py_rnn
=
PySimpleRNN1
(
self
.
input_shape
,
self
.
output_shape
)
self
.
output
=
mean
(
x
=
self
.
create_rnn_op
(),
**
self
.
p_info
)
self
.
output
=
layers
.
mean
(
x
=
self
.
create_rnn_op
(),
**
self
.
p_info
)
def
create_rnn_op
(
self
):
x
=
data
(
x
=
layers
.
data
(
shape
=
[
self
.
sent_len
,
self
.
batch_size
,
self
.
input_dim
],
data_type
=
'float32'
,
name
=
'x'
,
append_batch_size
=
False
,
**
self
.
p_info
)
x
.
stop_gradient
=
False
h_boot
=
data
(
h_boot
=
layers
.
data
(
shape
=
[
self
.
input_dim
],
data_type
=
'float32'
,
name
=
'h_boot'
,
**
self
.
p_info
)
h_boot
.
stop_gradient
=
False
rnn
=
StaticRNN
(
main_program
=
self
.
main_program
)
rnn
=
layers
.
StaticRNN
(
main_program
=
self
.
main_program
)
with
rnn
.
step
():
h_pre
=
rnn
.
memory
(
init
=
h_boot
)
x_t
=
rnn
.
step_input
(
x
)
h
=
scale
(
x
=
elementwise_add
(
h
=
layers
.
scale
(
x
=
layers
.
elementwise_add
(
x
=
h_pre
,
y
=
x_t
,
**
self
.
p_info
),
scale
=
self
.
py_rnn
.
scale
,
**
self
.
p_info
)
...
...
@@ -249,41 +246,41 @@ class RecurrentOpTest2(RecurrentOpTest1):
self
.
output_shape
=
(
self
.
sent_len
,
self
.
batch_size
,
self
.
input_dim
)
self
.
py_rnn
=
PySimpleRNN2
(
self
.
input_shape
,
self
.
output_shape
)
self
.
output
=
mean
(
x
=
self
.
create_rnn_op
(),
**
self
.
p_info
)
self
.
output
=
layers
.
mean
(
x
=
self
.
create_rnn_op
(),
**
self
.
p_info
)
def
create_rnn_op
(
self
):
x
=
data
(
x
=
layers
.
data
(
shape
=
[
self
.
sent_len
,
self
.
batch_size
,
self
.
input_dim
],
data_type
=
'float32'
,
name
=
'x'
,
append_batch_size
=
False
,
**
self
.
p_info
)
x
.
stop_gradient
=
False
h_boot
=
data
(
h_boot
=
layers
.
data
(
shape
=
[
self
.
input_dim
],
data_type
=
'float32'
,
name
=
'h_boot'
,
**
self
.
p_info
)
h_boot
.
stop_gradient
=
False
rnn
=
StaticRNN
(
main_program
=
self
.
main_program
)
rnn
=
layers
.
StaticRNN
(
main_program
=
self
.
main_program
)
with
rnn
.
step
():
h_pre
=
rnn
.
memory
(
init
=
h_boot
)
x_t
=
rnn
.
step_input
(
x
)
temp_l
=
fc
(
input
=
x_t
,
temp_l
=
layers
.
fc
(
input
=
x_t
,
size
=
self
.
input_dim
,
param_attr
=
{
'name'
:
'W'
},
bias_attr
=
False
,
**
self
.
p_info
)
temp_r
=
fc
(
input
=
h_pre
,
temp_r
=
layers
.
fc
(
input
=
h_pre
,
size
=
self
.
input_dim
,
param_attr
=
{
'name'
:
'U'
},
bias_attr
=
False
,
**
self
.
p_info
)
h
=
sigmoid
(
x
=
elementwise_add
(
h
=
layers
.
sigmoid
(
x
=
layers
.
elementwise_add
(
x
=
temp_l
,
y
=
temp_r
,
**
self
.
p_info
),
**
self
.
p_info
)
...
...
@@ -293,7 +290,7 @@ class RecurrentOpTest2(RecurrentOpTest1):
return
rnn
()
class
RecurrentOp
Test3
(
RecurrentOpTest1
):
class
RecurrentOp
MultipleMemoryTest
(
RecurrentOpTest1
):
'''
Test RNNOp with two memories
equation:
...
...
@@ -310,8 +307,8 @@ class RecurrentOpTest3(RecurrentOpTest1):
class
PySimpleRNN3
(
PyRNNBase
):
def
__init__
(
self
,
input_shape
,
output_shape
):
super
(
RecurrentOp
Test3
.
PySimpleRNN3
,
self
).
__init__
(
input_shape
,
output_shape
)
super
(
RecurrentOp
MultipleMemoryTest
.
PySimpleRNN3
,
self
).
__init__
(
input_shape
,
output_shape
)
seq_len
,
batch_size
,
input_dim
=
input_shape
self
.
h_boot1
=
np
.
random
.
normal
(
size
=
(
batch_size
,
...
...
@@ -345,27 +342,27 @@ class RecurrentOpTest3(RecurrentOpTest1):
self
.
input_shape
=
(
self
.
sent_len
,
self
.
batch_size
,
self
.
input_dim
)
self
.
output_shape
=
(
self
.
sent_len
,
self
.
batch_size
,
self
.
input_dim
)
self
.
py_rnn
=
RecurrentOp
Test3
.
PySimpleRNN3
(
self
.
input_shape
,
self
.
output_shape
)
self
.
py_rnn
=
RecurrentOp
MultipleMemoryTest
.
PySimpleRNN3
(
self
.
input_shape
,
self
.
output_shape
)
self
.
output
=
mean
(
x
=
self
.
create_rnn_op
(),
**
self
.
p_info
)
self
.
output
=
layers
.
mean
(
x
=
self
.
create_rnn_op
(),
**
self
.
p_info
)
def
create_rnn_op
(
self
):
x
=
data
(
x
=
layers
.
data
(
shape
=
[
self
.
sent_len
,
self
.
batch_size
,
self
.
input_dim
],
data_type
=
'float32'
,
name
=
'x'
,
append_batch_size
=
False
,
**
self
.
p_info
)
x
.
stop_gradient
=
False
h_boot1
=
data
(
h_boot1
=
layers
.
data
(
shape
=
[
self
.
batch_size
,
self
.
input_dim
],
data_type
=
'float32'
,
name
=
'h_boot1'
,
append_batch_size
=
False
,
**
self
.
p_info
)
h_boot1
.
stop_gradient
=
False
h_boot2
=
data
(
h_boot2
=
layers
.
data
(
shape
=
[
self
.
batch_size
,
self
.
input_dim
],
data_type
=
'float32'
,
name
=
'h_boot2'
,
...
...
@@ -373,15 +370,15 @@ class RecurrentOpTest3(RecurrentOpTest1):
**
self
.
p_info
)
h_boot2
.
stop_gradient
=
False
rnn
=
StaticRNN
(
main_program
=
self
.
main_program
)
rnn
=
layers
.
StaticRNN
(
main_program
=
self
.
main_program
)
with
rnn
.
step
():
h_pre1
=
rnn
.
memory
(
init
=
h_boot1
)
h_pre2
=
rnn
.
memory
(
init
=
h_boot2
)
x_t
=
rnn
.
step_input
(
x
)
mem1
=
scale
(
x
=
h_pre1
,
scale
=
1.0
,
**
self
.
p_info
)
mem2
=
scale
(
x
=
h_pre2
,
scale
=
1.0
,
**
self
.
p_info
)
out
=
sums
(
input
=
[
mem1
,
x_t
,
mem2
],
**
self
.
p_info
)
mem1
=
layers
.
scale
(
x
=
h_pre1
,
scale
=
1.0
,
**
self
.
p_info
)
mem2
=
layers
.
scale
(
x
=
h_pre2
,
scale
=
1.0
,
**
self
.
p_info
)
out
=
layers
.
sums
(
input
=
[
mem1
,
x_t
,
mem2
],
**
self
.
p_info
)
rnn
.
update_memory
(
h_pre1
,
mem1
)
rnn
.
update_memory
(
h_pre2
,
mem2
)
...
...
@@ -390,5 +387,70 @@ class RecurrentOpTest3(RecurrentOpTest1):
return
rnn
()
class
RecurrentOpNoMemBootTest
(
RecurrentOpTest1
):
'''
Test RNNOp with two memories
equation:
mem = x + mem_pre
y = mem
vars:
- x
memories:
- mem
outputs:
- y
'''
class
PySimpleRNN4
(
PyRNNBase
):
def
__init__
(
self
,
input_shape
,
output_shape
):
super
(
RecurrentOpNoMemBootTest
.
PySimpleRNN4
,
self
).
__init__
(
input_shape
,
output_shape
)
men_dim
=
input_shape
self
.
mems
=
np
.
zeros
(
shape
=
men_dim
).
astype
(
"float32"
)
def
step
(
self
,
step_id
,
x
):
if
step_id
==
0
:
pre_mem
=
np
.
zeros_like
(
x
)
else
:
pre_mem
=
self
.
mems
[
step_id
-
1
]
self
.
mems
[
step_id
]
=
pre_mem
+
x
self
.
y
[
step_id
]
=
self
.
mems
[
step_id
]
input_dim
=
1
batch_size
=
1
sent_len
=
2
def
setUp
(
self
):
self
.
setup_program
()
self
.
data_field
=
{
"x"
}
self
.
input_shape
=
(
self
.
sent_len
,
self
.
batch_size
,
self
.
input_dim
)
self
.
output_shape
=
(
self
.
sent_len
,
self
.
batch_size
,
self
.
input_dim
)
self
.
py_rnn
=
RecurrentOpNoMemBootTest
.
PySimpleRNN4
(
self
.
input_shape
,
self
.
output_shape
)
self
.
output
=
layers
.
mean
(
x
=
self
.
create_rnn_op
(),
**
self
.
p_info
)
print
self
.
main_program
def
create_rnn_op
(
self
):
x
=
layers
.
data
(
shape
=
[
self
.
sent_len
,
self
.
batch_size
,
self
.
input_dim
],
data_type
=
'float32'
,
name
=
'x'
,
append_batch_size
=
False
,
**
self
.
p_info
)
x
.
stop_gradient
=
False
rnn
=
layers
.
StaticRNN
(
main_program
=
self
.
main_program
)
with
rnn
.
step
():
mem_pre
=
rnn
.
memory
(
shape
=
[
-
1
,
self
.
input_dim
],
batch_ref
=
x
)
x_t
=
rnn
.
step_input
(
x
)
mem
=
layers
.
elementwise_add
(
x
=
mem_pre
,
y
=
x_t
,
**
self
.
p_info
)
rnn
.
update_memory
(
mem_pre
,
mem
)
rnn
.
output
(
mem
)
return
rnn
()
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