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
66d1c6ce
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
66d1c6ce
编写于
11月 02, 2017
作者:
A
Abhinav Arora
提交者:
GitHub
11月 02, 2017
浏览文件
操作
浏览文件
下载
电子邮件补丁
差异文件
Adding the Xavier Initializer (#5270)
* Adding the Xavier Initializer * Addressing code review feedback
上级
27ce7291
变更
2
隐藏空白更改
内联
并排
Showing
2 changed file
with
237 addition
and
1 deletion
+237
-1
python/paddle/v2/framework/initializer.py
python/paddle/v2/framework/initializer.py
+130
-1
python/paddle/v2/framework/tests/test_initializer.py
python/paddle/v2/framework/tests/test_initializer.py
+107
-0
未找到文件。
python/paddle/v2/framework/initializer.py
浏览文件 @
66d1c6ce
import
paddle.v2.framework.framework
as
framework
import
numpy
as
np
__all__
=
[
'ConstantInitializer'
,
'UniformInitializer'
]
__all__
=
[
'ConstantInitializer'
,
'UniformInitializer'
,
'NormalInitializer'
,
'XavierInitializer'
]
class
Initializer
(
object
):
...
...
@@ -20,6 +24,41 @@ class Initializer(object):
"""
raise
NotImplementedError
()
def
_compute_fans
(
self
,
var
):
"""Compute the fan_in and the fan_out for layers
This method computes the fan_in and the fan_out
for neural network layers, if not specified. It is
not possible to perfectly estimate fan_in and fan_out.
This method will estimate it correctly for matrix multiply and
convolutions.
Args:
var: variable for which fan_in and fan_out have to be computed
Returns:
tuple of two integers (fan_in, fan_out)
"""
shape
=
var
.
shape
if
not
shape
or
len
(
shape
)
==
0
:
fan_in
=
fan_out
=
1
elif
len
(
shape
)
==
1
:
fan_in
=
fan_out
=
shape
[
0
]
elif
len
(
shape
)
==
2
:
# This is the case for simple matrix multiply
fan_in
=
shape
[
0
]
fan_out
=
shape
[
1
]
else
:
# Assume this to be a convolutional kernel
# In PaddlePaddle, the shape of the kernel is like:
# [num_filters, num_filter_channels, ...] where the remaining
# dimensions are the filter_size
receptive_field_size
=
np
.
prod
(
shape
[
2
:])
fan_in
=
shape
[
1
]
*
receptive_field_size
fan_out
=
shape
[
0
]
*
receptive_field_size
return
(
fan_in
,
fan_out
)
class
ConstantInitializer
(
Initializer
):
"""Implements the constant initializer
...
...
@@ -156,3 +195,93 @@ class NormalInitializer(Initializer):
})
var
.
op
=
op
return
op
class
XavierInitializer
(
Initializer
):
"""Implements the Xavier initializer
This class implements the Xavier weight initializer from the paper
Understanding the difficulty of training deep feedforward neural
networks[1] by Xavier Glorot and Yoshua Bengio.
This initializer is designed to keep the scale of the gradients
approximately same in all the layers. In case of Uniform distribution,
the range is [-x, x], where x = sqrt(6 / (fan_in + fan_out)).
In case of Normal distribution, the mean is 0 and the standard deviation
is sqrt(2/ (fan_in + fan_out)).
References:
[1] Understanding the difficulty of training deep feedforward neural
networks. International conference on artificial intelligence and
statistics.
(http://proceedings.mlr.press/v9/glorot10a.html)
"""
def
__init__
(
self
,
uniform
=
True
,
fan_in
=
None
,
fan_out
=
None
,
seed
=
0
):
"""Constructor for XavierInitializer
Args:
uniform: whether to use uniform or normal distribution
fan_in: fan_in for Xavier initialization. If None, it is
inferred from the variable.
fan_out: fan_out for Xavier initialization. If None, it is
inferred from the variable.
seed: random seed
Note: It is recommended to set fan_in and fan_out to None for
most cases.
"""
assert
uniform
is
not
None
assert
seed
is
not
None
super
(
XavierInitializer
,
self
).
__init__
()
self
.
_uniform
=
uniform
self
.
_fan_in
=
fan_in
self
.
_fan_out
=
fan_out
self
.
_seed
=
seed
def
__call__
(
self
,
var
,
block
):
"""Add xavier initialization ops for a variable
Args:
var: Variable that needs to be initialized
block: The block in which initialization ops
should be added
Returns:
the initialization op
"""
assert
isinstance
(
var
,
framework
.
Variable
)
assert
isinstance
(
block
,
framework
.
Block
)
f_in
,
f_out
=
self
.
_compute_fans
(
var
)
# If fan_in and fan_out are passed, use them
fan_in
=
f_in
if
self
.
_fan_in
is
None
else
self
.
_fan_in
fan_out
=
f_out
if
self
.
_fan_out
is
None
else
self
.
_fan_out
if
self
.
_uniform
:
limit
=
np
.
sqrt
(
6.0
/
float
(
fan_in
+
fan_out
))
op
=
block
.
prepend_op
(
type
=
"uniform_random"
,
outputs
=
{
"Out"
:
var
},
attrs
=
{
"shape"
:
var
.
shape
,
"data_type"
:
int
(
var
.
data_type
),
"min"
:
-
limit
,
"max"
:
limit
,
"seed"
:
self
.
_seed
})
else
:
std
=
np
.
sqrt
(
2.0
/
float
(
fan_in
+
fan_out
))
op
=
block
.
prepend_op
(
type
=
"gaussian_random"
,
outputs
=
{
"Out"
:
var
},
attrs
=
{
"shape"
:
var
.
shape
,
"data_type"
:
int
(
var
.
data_type
),
"mean"
:
0.0
,
"std"
:
std
,
"seed"
:
self
.
_seed
})
var
.
op
=
op
return
op
python/paddle/v2/framework/tests/test_initializer.py
浏览文件 @
66d1c6ce
import
numpy
as
np
import
unittest
import
paddle.v2.framework.framework
as
framework
...
...
@@ -116,5 +117,111 @@ class TestNormalInitializer(unittest.TestCase):
self
.
assertEqual
(
init_op
.
attr
(
'seed'
),
123
)
class
TestXavierInitializer
(
unittest
.
TestCase
):
def
test_uniform_xavier_initializer
(
self
):
"""Test Xavier initializer with uniform distribution on
for matrix multiply.
"""
program
=
framework
.
Program
()
block
=
program
.
global_block
()
param
=
block
.
create_parameter
(
dtype
=
"float32"
,
shape
=
[
5
,
10
],
lod_level
=
0
,
name
=
"param"
,
initializer
=
initializer
.
XavierInitializer
())
self
.
assertEqual
(
len
(
block
.
ops
),
1
)
init_op
=
block
.
ops
[
0
]
self
.
assertEqual
(
init_op
.
type
,
'uniform_random'
)
limit
=
np
.
sqrt
(
6.0
/
(
param
.
shape
[
0
]
+
param
.
shape
[
1
]))
self
.
assertAlmostEqual
(
init_op
.
attr
(
'min'
),
-
limit
,
delta
=
DELTA
)
self
.
assertAlmostEqual
(
init_op
.
attr
(
'max'
),
limit
,
delta
=
DELTA
)
self
.
assertEqual
(
init_op
.
attr
(
'seed'
),
0
)
def
test_uniform_xavier_initializer_conv
(
self
):
"""Test Xavier initializer with uniform distribution on
for convolutions.
"""
program
=
framework
.
Program
()
block
=
program
.
global_block
()
param
=
block
.
create_parameter
(
dtype
=
"float32"
,
shape
=
[
5
,
10
,
15
,
20
],
lod_level
=
0
,
name
=
"param"
,
initializer
=
initializer
.
XavierInitializer
())
self
.
assertEqual
(
len
(
block
.
ops
),
1
)
init_op
=
block
.
ops
[
0
]
self
.
assertEqual
(
init_op
.
type
,
'uniform_random'
)
receptive_field_size
=
float
(
15
*
20
)
limit
=
np
.
sqrt
(
6.0
/
(
(
param
.
shape
[
0
]
+
param
.
shape
[
1
])
*
receptive_field_size
))
self
.
assertAlmostEqual
(
init_op
.
attr
(
'min'
),
-
limit
,
delta
=
DELTA
)
self
.
assertAlmostEqual
(
init_op
.
attr
(
'max'
),
limit
,
delta
=
DELTA
)
self
.
assertEqual
(
init_op
.
attr
(
'seed'
),
0
)
def
test_normal_xavier_initializer
(
self
):
"""Test Xavier initializer with normal distribution on
for matrix multiply.
"""
program
=
framework
.
Program
()
block
=
program
.
global_block
()
param
=
block
.
create_parameter
(
dtype
=
"float32"
,
shape
=
[
5
,
10
],
lod_level
=
0
,
name
=
"param"
,
initializer
=
initializer
.
XavierInitializer
(
uniform
=
False
))
self
.
assertEqual
(
len
(
block
.
ops
),
1
)
init_op
=
block
.
ops
[
0
]
self
.
assertEqual
(
init_op
.
type
,
'gaussian_random'
)
std
=
np
.
sqrt
(
2.0
/
(
param
.
shape
[
0
]
+
param
.
shape
[
1
]))
self
.
assertAlmostEqual
(
init_op
.
attr
(
'mean'
),
0.0
,
delta
=
DELTA
)
self
.
assertAlmostEqual
(
init_op
.
attr
(
'std'
),
std
,
delta
=
DELTA
)
self
.
assertEqual
(
init_op
.
attr
(
'seed'
),
0
)
def
test_normal_xavier_initializer_conv
(
self
):
"""Test Xavier initializer with normal distribution on
for convolutions.
"""
program
=
framework
.
Program
()
block
=
program
.
global_block
()
param
=
block
.
create_parameter
(
dtype
=
"float32"
,
shape
=
[
5
,
10
,
15
,
20
],
lod_level
=
0
,
name
=
"param"
,
initializer
=
initializer
.
XavierInitializer
(
uniform
=
False
))
self
.
assertEqual
(
len
(
block
.
ops
),
1
)
init_op
=
block
.
ops
[
0
]
self
.
assertEqual
(
init_op
.
type
,
'gaussian_random'
)
receptive_field_size
=
float
(
15
*
20
)
std
=
np
.
sqrt
(
2.0
/
(
(
param
.
shape
[
0
]
+
param
.
shape
[
1
])
*
receptive_field_size
))
self
.
assertAlmostEqual
(
init_op
.
attr
(
'mean'
),
0.0
,
delta
=
DELTA
)
self
.
assertAlmostEqual
(
init_op
.
attr
(
'std'
),
std
,
delta
=
DELTA
)
self
.
assertEqual
(
init_op
.
attr
(
'seed'
),
0
)
def
test_xavier_initializer_supplied_arguments
(
self
):
"""Test the Xavier initializer with supplied arguments
"""
program
=
framework
.
Program
()
block
=
program
.
global_block
()
block
.
create_parameter
(
dtype
=
"float32"
,
shape
=
[
5
,
10
],
lod_level
=
0
,
name
=
"param"
,
initializer
=
initializer
.
XavierInitializer
(
fan_in
=
12
,
fan_out
=
23
,
seed
=
134
))
self
.
assertEqual
(
len
(
block
.
ops
),
1
)
init_op
=
block
.
ops
[
0
]
self
.
assertEqual
(
init_op
.
type
,
'uniform_random'
)
limit
=
np
.
sqrt
(
6.0
/
(
12
+
23
))
self
.
assertAlmostEqual
(
init_op
.
attr
(
'min'
),
-
limit
,
delta
=
DELTA
)
self
.
assertAlmostEqual
(
init_op
.
attr
(
'max'
),
limit
,
delta
=
DELTA
)
self
.
assertEqual
(
init_op
.
attr
(
'seed'
),
134
)
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