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
wmsofts
Paddle
提交
84d6434d
P
Paddle
项目概览
wmsofts
/
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
0
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
0
Issue
0
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
84d6434d
编写于
8月 14, 2017
作者:
D
dangqingqing
浏览文件
操作
浏览文件
下载
电子邮件补丁
差异文件
Compare the gradient consistency between GPU and CPU calculations.
上级
ffbb0be2
变更
4
隐藏空白更改
内联
并排
Showing
4 changed file
with
98 addition
and
101 deletion
+98
-101
paddle/operators/sigmoid_op.cc
paddle/operators/sigmoid_op.cc
+2
-1
python/paddle/v2/framework/tests/CMakeLists.txt
python/paddle/v2/framework/tests/CMakeLists.txt
+1
-0
python/paddle/v2/framework/tests/gradient_checker.py
python/paddle/v2/framework/tests/gradient_checker.py
+77
-96
python/paddle/v2/framework/tests/test_sigmoid_op.py
python/paddle/v2/framework/tests/test_sigmoid_op.py
+18
-4
未找到文件。
paddle/operators/sigmoid_op.cc
浏览文件 @
84d6434d
...
...
@@ -44,7 +44,8 @@ class SigmoidOpGrad : public framework::OperatorWithKernel {
protected:
void
InferShape
(
const
framework
::
InferShapeContext
&
ctx
)
const
override
{
ctx
.
Output
<
Tensor
>
(
0
)
->
Resize
(
ctx
.
Input
<
Tensor
>
(
0
)
->
dims
());
ctx
.
Output
<
Tensor
>
(
framework
::
GradVarName
(
"X"
))
->
Resize
(
ctx
.
Input
<
Tensor
>
(
"Y"
)
->
dims
());
}
};
...
...
python/paddle/v2/framework/tests/CMakeLists.txt
浏览文件 @
84d6434d
...
...
@@ -25,3 +25,4 @@ py_test(test_operator SRCS test_operator.py)
# py_test(test_gaussian_random_op SRCS test_gaussian_random_op.py)
py_test
(
test_uniform_random_op SRCS test_uniform_random_op.py
)
py_test
(
test_recurrent_op SRCS test_recurrent_op.py
)
py_test
(
test_gradient_checker SRCS test_gradient_checker.py
)
python/paddle/v2/framework/tests/gradient_checker.py
浏览文件 @
84d6434d
import
unittest
import
numpy
import
itertools
import
paddle.v2.framework.core
as
core
from
paddle.v2.framework.op
import
Operator
...
...
@@ -8,6 +9,7 @@ __all__ = ['get_numeric_gradient']
def
create_op
(
op_type
):
# TODO need to set attrs
kwargs
=
dict
()
for
in_name
in
Operator
.
get_op_input_names
(
op_type
):
kwargs
[
in_name
]
=
in_name
...
...
@@ -66,7 +68,6 @@ def get_numeric_gradient(op,
local_scope
.
find_var
(
output
).
get_tensor
().
alloc_float
(
core
.
CPUPlace
(
))
# TODO(yuyang18): Only CPU is support now.
cpu_ctx
=
core
.
DeviceContext
.
create
(
core
.
CPUPlace
())
def
get_output
():
...
...
@@ -109,12 +110,71 @@ def get_numeric_gradient(op,
class
GradientChecker
(
unittest
.
TestCase
):
def
assert_is_close
(
self
,
numeric_grads
,
scope
,
max_relative_error
,
msg_prefix
):
for
name
in
numeric_grads
:
b
=
numpy
.
array
(
scope
.
find_var
(
grad_var_name
(
name
)).
get_tensor
())
a
=
numeric_grads
[
name
]
def
get_grad
(
self
,
forward_op
,
backward_op
,
input_vars
,
grad_names
,
place
):
scope
=
core
.
Scope
()
ctx
=
core
.
DeviceContext
.
create
(
place
)
inputs
=
forward_op
.
inputs
()
in_names
=
[
item
for
k
in
inputs
for
item
in
inputs
[
k
]]
outputs
=
forward_op
.
outputs
()
out_names
=
[
item
for
k
in
outputs
for
item
in
outputs
[
k
]]
# create input var and set value
for
name
,
value
in
input_vars
.
iteritems
():
if
name
not
in
in_names
:
raise
ValueError
(
name
+
"does not exist in Op's inputs."
)
var
=
scope
.
new_var
(
name
).
get_tensor
()
var
.
set_dims
(
value
.
shape
)
var
.
set
(
value
,
place
)
# run forward op
for
out_name
in
out_names
:
scope
.
new_var
(
out_name
)
forward_op
.
infer_shape
(
scope
)
forward_op
.
run
(
scope
,
ctx
)
# set output var's shape
# set output grad to ones
for
name
in
out_names
:
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
=
numpy
.
ones
(
out_tensor
.
shape
(),
dtype
=
numpy
.
float32
)
grad_tensor
.
set
(
data
,
place
)
# run backward op
for
name
in
backward_op
.
outputs
():
scope
.
new_var
(
name
)
backward_op
.
infer_shape
(
scope
)
backward_op
.
run
(
scope
,
ctx
)
outs
=
[
numpy
.
array
(
scope
.
find_var
(
name
).
get_tensor
())
for
name
in
grad_names
]
return
outs
def
compare_grad
(
self
,
forward_op
,
inputs
):
backward_op
=
core
.
Operator
.
backward
(
forward_op
,
set
())
if
not
(
core
.
is_compile_gpu
()
and
backward_op
.
support_gpu
()):
return
outputs
=
backward_op
.
outputs
()
out_names
=
[
item
for
k
in
outputs
for
item
in
outputs
[
k
]]
cpu_grads
=
self
.
get_grad
(
forward_op
,
backward_op
,
inputs
,
out_names
,
core
.
CPUPlace
())
gpu_grads
=
self
.
get_grad
(
forward_op
,
backward_op
,
inputs
,
out_names
,
core
.
GPUPlace
(
0
))
for
c_grad
,
g_grad
,
name
in
itertools
.
izip
(
cpu_grads
,
gpu_grads
,
out_names
):
self
.
assertTrue
(
numpy
.
allclose
(
c_grad
,
g_grad
),
"output name: "
+
name
+
" has diff"
)
def
assert_is_close
(
self
,
numeric_grads
,
analytic_grads
,
names
,
max_relative_error
,
msg_prefix
):
for
a
,
b
,
name
in
itertools
.
izip
(
numeric_grads
,
analytic_grads
,
names
):
abs_a
=
numpy
.
abs
(
a
)
# if abs_a is nearly zero, then use abs error for a, not relative
# error.
...
...
@@ -159,106 +219,27 @@ class GradientChecker(unittest.TestCase):
inputs
=
forward_op
.
inputs
()
in_names
=
[
item
for
k
in
inputs
for
item
in
inputs
[
k
]]
outputs
=
forward_op
.
outputs
()
out_names
=
[
item
for
k
in
outputs
for
item
in
outputs
[
k
]]
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
)
bwd_outputs
=
backward_op
.
outputs
()
bwd_out_names
=
[
item
for
k
in
bwd_outputs
for
item
in
bwd_outputs
[
k
]]
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 numerical gradients
numeric_grads
=
[
get_numeric_gradient
(
forward_op
,
input_vars
,
output_name
,
name
)
for
name
in
inputs_to_check
]
# get operator gradient according to different device
check_names
=
[
grad_var_name
(
name
)
for
name
in
inputs_to_check
]
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
out_names
:
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
out_names
:
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
bwd_out_names
:
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
)
self
.
assert_is_close
(
numeric_grad
,
scope
,
max_relative_error
,
# get analytical gradients according to different device
analytic_grads
=
self
.
get_grad
(
forward_op
,
backward_op
,
input_vars
,
check_grad_names
,
place
)
self
.
assert_is_close
(
numeric_grads
,
analytic_grads
,
check_names
,
max_relative_error
,
"Gradient Check On %s"
%
str
(
place
))
if
__name__
==
'__main__'
:
class
GetNumericGradientTest
(
unittest
.
TestCase
):
def
test_add_op
(
self
):
add_op
=
Operator
(
'add_two'
,
X
=
"X"
,
Y
=
"Y"
,
Out
=
"Z"
)
x
=
numpy
.
random
.
random
((
10
,
1
)).
astype
(
"float32"
)
y
=
numpy
.
random
.
random
((
10
,
1
)).
astype
(
"float32"
)
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/test_sigmoid_op.py
浏览文件 @
84d6434d
import
unittest
from
op_test_util
import
OpTestMeta
import
numpy
as
np
from
op_test_util
import
OpTestMeta
from
gradient_checker
import
GradientChecker
,
create_op
class
TestSigmoidOp
(
unittest
.
TestCase
):
...
...
@@ -8,12 +9,25 @@ class TestSigmoidOp(unittest.TestCase):
def
setUp
(
self
):
self
.
type
=
"sigmoid"
self
.
inputs
=
{
'X'
:
np
.
random
.
random
((
32
,
100
)).
astype
(
"float32"
)}
self
.
inputs
=
{
'X'
:
np
.
random
.
random
((
15
,
31
)).
astype
(
"float32"
)}
self
.
outputs
=
{
'Y'
:
1
/
(
1
+
np
.
exp
(
-
self
.
inputs
[
'X'
]))}
#class TestSigmoidGradOp(unittest.TestCase):
#TODO(qingqing) add unit test
class
TestSigmoidGradOp
(
GradientChecker
):
def
test_compare_grad
(
self
):
op
=
create_op
(
"sigmoid"
)
inputs
=
{
"X"
:
np
.
random
.
random
((
11
,
17
)).
astype
(
"float32"
)}
# compare gpu and cpu results for backward op
self
.
compare_grad
(
op
,
inputs
)
def
test_check_grad
(
self
):
op
=
create_op
(
"sigmoid"
)
inputs
=
{
"X"
:
np
.
random
.
random
((
11
,
17
)).
astype
(
"float32"
)}
# check gradients
self
.
check_grad
(
op
,
inputs
,
set
(
"X"
),
"Y"
)
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