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
PaddlePaddle
PaddleDetection
提交
f23691db
P
PaddleDetection
项目概览
PaddlePaddle
/
PaddleDetection
大约 1 年 前同步成功
通知
695
Star
11112
Fork
2696
代码
文件
提交
分支
Tags
贡献者
分支图
Diff
Issue
184
列表
看板
标记
里程碑
合并请求
40
Wiki
0
Wiki
分析
仓库
DevOps
项目成员
Pages
P
PaddleDetection
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
184
Issue
184
列表
看板
标记
里程碑
合并请求
40
合并请求
40
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
未验证
提交
f23691db
编写于
1月 15, 2018
作者:
Y
Yang yaming
提交者:
GitHub
1月 15, 2018
浏览文件
操作
浏览文件
下载
差异文件
Merge pull request #7434 from pkuyym/fix-7195
Add static_input for DynamicRNN
上级
535fefb7
25fee871
变更
2
隐藏空白更改
内联
并排
Showing
2 changed file
with
212 addition
and
0 deletion
+212
-0
python/paddle/v2/fluid/layers/control_flow.py
python/paddle/v2/fluid/layers/control_flow.py
+20
-0
python/paddle/v2/fluid/tests/test_dynrnn_static_input.py
python/paddle/v2/fluid/tests/test_dynrnn_static_input.py
+192
-0
未找到文件。
python/paddle/v2/fluid/layers/control_flow.py
浏览文件 @
f23691db
...
...
@@ -1291,6 +1291,26 @@ class DynamicRNN(object):
outputs
=
{
'Out'
:
input_array
})
return
array_read
(
array
=
input_array
,
i
=
self
.
step_idx
)
def
static_input
(
self
,
x
):
self
.
_assert_in_rnn_block_
(
"static_input"
)
if
not
isinstance
(
x
,
Variable
):
raise
TypeError
(
"static_input() can only take a Variable as its input"
)
if
self
.
lod_rank_table
is
None
:
raise
RuntimeError
(
"static_input() must be called after step_input()."
)
parent_block
=
self
.
_parent_block_
()
x_reordered
=
parent_block
.
create_var
(
name
=
unique_name
(
"dynamic_rnn_static_input_reordered"
),
type
=
core
.
VarDesc
.
VarType
.
LOD_TENSOR
,
dtype
=
x
.
dtype
)
parent_block
.
append_op
(
type
=
'reorder_lod_tensor_by_rank'
,
inputs
=
{
'X'
:
[
x
],
'RankTable'
:
[
self
.
lod_rank_table
]},
outputs
=
{
'Out'
:
[
x_reordered
]})
return
shrink_memory
(
x_reordered
,
self
.
step_idx
,
self
.
lod_rank_table
)
@
contextlib
.
contextmanager
def
block
(
self
):
if
self
.
status
!=
DynamicRNN
.
BEFORE_RNN
:
...
...
python/paddle/v2/fluid/tests/test_dynrnn_static_input.py
0 → 100644
浏览文件 @
f23691db
import
unittest
import
paddle.v2
as
paddle
import
paddle.v2.fluid.core
as
core
import
paddle.v2.fluid
as
fluid
from
paddle.v2.fluid.backward
import
append_backward
import
paddle.v2.fluid.framework
as
framework
from
paddle.v2.fluid.framework
import
Program
,
switch_main_program
import
bisect
import
numpy
as
np
fluid
.
default_startup_program
().
random_seed
=
1
class
TestDyRnnStaticInput
(
unittest
.
TestCase
):
def
setUp
(
self
):
self
.
_delta
=
0.005
self
.
_max_sequence_len
=
3
self
.
_program
=
Program
()
switch_main_program
(
self
.
_program
)
self
.
output_dim
=
10
self
.
place
=
core
.
CPUPlace
()
self
.
prepare_x_tensor
()
self
.
prepare_static_input_tensor
()
self
.
exe
=
fluid
.
Executor
(
self
.
place
)
def
prepare_x_tensor
(
self
):
self
.
x_tensor_dim
=
10
lod
=
[[
0
,
2
,
3
,
6
]]
shape
=
[
lod
[
0
][
-
1
],
self
.
x_tensor_dim
]
self
.
x_tensor_data
=
np
.
random
.
random
(
shape
).
astype
(
'float32'
)
self
.
x_tensor
=
core
.
LoDTensor
()
self
.
x_tensor
.
set_lod
(
lod
)
self
.
x_tensor
.
set
(
self
.
x_tensor_data
,
self
.
place
)
def
prepare_static_input_tensor
(
self
):
self
.
static_input_tensor_dim
=
4
lod
=
[[
0
,
1
,
3
,
6
]]
shape
=
[
lod
[
0
][
-
1
],
self
.
static_input_tensor_dim
]
self
.
static_input_data
=
np
.
random
.
random
(
shape
).
astype
(
'float32'
)
self
.
static_input_tensor
=
core
.
LoDTensor
()
self
.
static_input_tensor
.
set_lod
(
lod
)
self
.
static_input_tensor
.
set
(
self
.
static_input_data
,
self
.
place
)
def
fetch_value
(
self
,
var
):
fetch_outs
=
self
.
exe
.
run
(
feed
=
{
'x_tensor'
:
self
.
x_tensor
,
'static_input_tensor'
:
self
.
static_input_tensor
},
fetch_list
=
[
var
],
return_numpy
=
False
)
return
self
.
_lodtensor_to_ndarray
(
fetch_outs
[
0
])
def
_lodtensor_to_ndarray
(
self
,
lod_tensor
):
dims
=
lod_tensor
.
get_dims
()
ndarray
=
np
.
zeros
(
shape
=
dims
).
astype
(
'float32'
)
for
i
in
xrange
(
np
.
product
(
dims
)):
ndarray
.
ravel
()[
i
]
=
lod_tensor
.
get_float_element
(
i
)
return
ndarray
,
lod_tensor
.
lod
()
def
build_graph
(
self
,
only_forward
=
False
):
x_tensor
=
fluid
.
layers
.
data
(
name
=
'x_tensor'
,
shape
=
[
self
.
x_tensor_dim
],
dtype
=
'float32'
,
lod_level
=
1
)
x_tensor
.
stop_gradient
=
False
static_input_tensor
=
fluid
.
layers
.
data
(
name
=
'static_input_tensor'
,
shape
=
[
self
.
static_input_tensor_dim
],
dtype
=
'float32'
,
lod_level
=
1
)
static_input_tensor
.
stop_gradient
=
False
if
only_forward
:
static_input_out_array
=
self
.
_program
.
global_block
().
create_var
(
name
=
'static_input_out_array'
,
type
=
core
.
VarDesc
.
VarType
.
LOD_TENSOR_ARRAY
,
dtype
=
'float32'
)
static_input_out_array
.
stop_gradient
=
True
rnn
=
fluid
.
layers
.
DynamicRNN
()
with
rnn
.
block
():
step_x
=
rnn
.
step_input
(
x_tensor
)
step_static_input
=
rnn
.
static_input
(
static_input_tensor
)
if
only_forward
:
fluid
.
layers
.
array_write
(
x
=
step_static_input
,
i
=
rnn
.
step_idx
,
array
=
static_input_out_array
)
last
=
fluid
.
layers
.
sequence_pool
(
input
=
step_static_input
,
pool_type
=
'last'
)
projected
=
fluid
.
layers
.
fc
(
input
=
[
step_x
,
last
],
size
=
self
.
output_dim
)
rnn
.
output
(
projected
)
if
only_forward
:
static_input_step_outs
=
[]
step_idx
=
fluid
.
layers
.
fill_constant
(
shape
=
[
1
],
dtype
=
'int64'
,
value
=
0
)
step_idx
.
stop_gradient
=
True
for
i
in
xrange
(
self
.
_max_sequence_len
):
step_out
=
fluid
.
layers
.
array_read
(
static_input_out_array
,
step_idx
)
step_out
.
stop_gradient
=
True
static_input_step_outs
.
append
(
step_out
)
fluid
.
layers
.
increment
(
x
=
step_idx
,
value
=
1.0
,
in_place
=
True
)
if
only_forward
:
return
static_input_step_outs
last
=
fluid
.
layers
.
sequence_pool
(
input
=
rnn
(),
pool_type
=
'last'
)
loss
=
fluid
.
layers
.
mean
(
x
=
last
)
append_backward
(
loss
)
static_input_grad
=
self
.
_program
.
global_block
().
var
(
framework
.
grad_var_name
(
'static_input_tensor'
))
return
static_input_grad
,
loss
def
get_seq_len_from_lod
(
self
,
lod
):
return
[
lod
[
0
][
i
+
1
]
-
lod
[
0
][
i
]
for
i
in
xrange
(
len
(
lod
[
0
])
-
1
)]
def
get_expected_static_step_outs
(
self
):
x_lod
=
self
.
x_tensor
.
lod
()
x_seq_len
=
self
.
get_seq_len_from_lod
(
x_lod
)
x_seq_len_sorted
=
sorted
(
x_seq_len
)
x_sorted_indices
=
np
.
argsort
(
x_seq_len
)[::
-
1
]
static_lod
=
self
.
static_input_tensor
.
lod
()
static_sliced
=
[
self
.
static_input_data
[
static_lod
[
0
][
i
]:
static_lod
[
0
][
i
+
1
]]
for
i
in
xrange
(
len
(
static_lod
[
0
])
-
1
)
]
static_seq_len
=
self
.
get_seq_len_from_lod
(
static_lod
)
static_reordered
=
[]
for
i
in
xrange
(
len
(
x_sorted_indices
)):
static_reordered
.
extend
(
static_sliced
[
x_sorted_indices
[
i
]].
tolist
())
static_seq_len_reordered
=
[
static_seq_len
[
x_sorted_indices
[
i
]]
for
i
in
xrange
(
len
(
x_sorted_indices
))
]
static_step_outs
=
[]
static_step_lods
=
[]
for
i
in
xrange
(
self
.
_max_sequence_len
):
end
=
len
(
x_seq_len
)
-
bisect
.
bisect_left
(
x_seq_len_sorted
,
i
+
1
)
lod
=
[
0
]
for
i
in
xrange
(
end
):
lod
.
append
(
static_seq_len_reordered
[
i
]
+
lod
[
-
1
])
static_step_lods
.
append
([
lod
])
end
=
lod
[
-
1
]
static_step_outs
.
append
(
np
.
array
(
static_reordered
[:
end
]).
astype
(
'float32'
))
return
static_step_outs
,
static_step_lods
def
test_step_out
(
self
):
static_step_outs
=
self
.
build_graph
(
only_forward
=
True
)
self
.
exe
.
run
(
framework
.
default_startup_program
())
expected_outs
,
expected_lods
=
self
.
get_expected_static_step_outs
()
for
i
in
xrange
(
self
.
_max_sequence_len
):
step_out
,
lod
=
self
.
fetch_value
(
static_step_outs
[
i
])
self
.
assertTrue
(
np
.
allclose
(
step_out
,
expected_outs
[
i
]))
self
.
assertTrue
(
np
.
allclose
(
lod
,
expected_lods
[
i
]))
def
test_network_gradient
(
self
):
static_input_grad
,
loss
=
self
.
build_graph
()
self
.
exe
.
run
(
framework
.
default_startup_program
())
actual_gradients
,
actual_lod
=
self
.
fetch_value
(
static_input_grad
)
static_input_shape
=
self
.
static_input_tensor
.
get_dims
()
numeric_gradients
=
np
.
zeros
(
shape
=
static_input_shape
).
astype
(
'float32'
)
# calculate numeric gradients
tensor_size
=
np
.
product
(
static_input_shape
)
for
i
in
xrange
(
tensor_size
):
origin
=
self
.
static_input_tensor
.
get_float_element
(
i
)
x_pos
=
origin
+
self
.
_delta
self
.
static_input_tensor
.
set_float_element
(
i
,
x_pos
)
y_pos
=
self
.
fetch_value
(
loss
)[
0
][
0
]
x_neg
=
origin
-
self
.
_delta
self
.
static_input_tensor
.
set_float_element
(
i
,
x_neg
)
y_neg
=
self
.
fetch_value
(
loss
)[
0
][
0
]
self
.
static_input_tensor
.
set_float_element
(
i
,
origin
)
numeric_gradients
.
ravel
()[
i
]
=
(
y_pos
-
y_neg
)
/
self
.
_delta
/
2
self
.
assertTrue
(
np
.
allclose
(
actual_gradients
,
numeric_gradients
,
0.001
))
self
.
assertTrue
(
np
.
allclose
(
actual_lod
,
self
.
static_input_tensor
.
lod
()))
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