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
4ec9ecae
P
PaddleDetection
项目概览
PaddlePaddle
/
PaddleDetection
大约 1 年 前同步成功
通知
692
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
体验新版 GitCode,发现更多精彩内容 >>
未验证
提交
4ec9ecae
编写于
6月 19, 2018
作者:
Q
Qiao Longfei
提交者:
GitHub
6月 19, 2018
浏览文件
操作
浏览文件
下载
差异文件
Merge pull request #11547 from jacquesqiao/support-ftrl-optimizer
add ftrl optimizer
上级
c22ebb3b
6caea459
变更
2
隐藏空白更改
内联
并排
Showing
2 changed file
with
178 addition
and
4 deletion
+178
-4
python/paddle/fluid/optimizer.py
python/paddle/fluid/optimizer.py
+112
-4
python/paddle/fluid/tests/unittests/test_optimizer.py
python/paddle/fluid/tests/unittests/test_optimizer.py
+66
-0
未找到文件。
python/paddle/fluid/optimizer.py
浏览文件 @
4ec9ecae
...
...
@@ -26,10 +26,10 @@ from clip import append_gradient_clip_ops, error_clip_callback
from
contextlib
import
contextmanager
__all__
=
[
'SGD'
,
'Momentum'
,
'Adagrad'
,
'Adam'
,
'Adamax'
,
'DecayedAdagrad'
,
'SGD'
,
'Momentum'
,
'Adagrad'
,
'Adam'
,
'Adamax'
,
'DecayedAdagrad'
,
'Ftrl'
,
'SGDOptimizer'
,
'MomentumOptimizer'
,
'AdagradOptimizer'
,
'AdamOptimizer'
,
'AdamaxOptimizer'
,
'DecayedAdagradOptimizer'
,
'RMSPropOptimizer'
,
'Adadelta'
,
'ModelAverage'
,
'Optimizer'
'
FtrlOptimizer'
,
'
Adadelta'
,
'ModelAverage'
,
'Optimizer'
]
...
...
@@ -628,7 +628,7 @@ class AdadeltaOptimizer(Optimizer):
E(dx_t^2) &=
\\
rho * E(dx_{t-1}^2) + (1-
\\
rho) * (-g*learning
\\
_rate)^2
Args:
learning_rate(float): global le
ra
ning rate
learning_rate(float): global le
ar
ning rate
rho(float): rho in equation
epsilon(float): epsilon in equation
...
...
@@ -729,7 +729,7 @@ class RMSPropOptimizer(Optimizer):
Args:
learning_rate(float): global le
ra
ning rate.
learning_rate(float): global le
ar
ning rate.
rho(float): rho is :math: `
\\
rho` in equation, set 0.95 by default.
epsilon(float): :math: `
\\
epsilon` in equation is smoothing term to
avoid division by zero, set 1e-6 by default.
...
...
@@ -810,6 +810,113 @@ class RMSPropOptimizer(Optimizer):
return
rmsprop_op
class
FtrlOptimizer
(
Optimizer
):
"""
FTRL (Follow The Regularized Leader) Optimizer.
The paper that proposed Follow The Regularized Leader (FTRL):
(https://www.eecs.tufts.edu/~dsculley/papers/ad-click-prediction.pdf)
.. math::
&new\_accum = squared\_accum + grad^2
&if (lr\_power == -0.5):
&\quad linear\_accum += grad -
\\
frac{
\\
sqrt{new\_accum} -
\\
sqrt{squared\_accum}}{learning\_rate * param}
&else:
&\quad linear\_accum += grad -
\\
frac{new\_accum^{-lr\_power} - accum^{-lr\_power}}{learning\_rate * param}
&x = l1 * sign(linear\_accum) - linear\_accum
&if (lr\_power == -0.5):
&\quad y =
\\
frac{
\\
sqrt{new\_accum}}{learning\_rate} + (2 * l2)
&\quad pre\_shrink =
\\
frac{x}{y}
&\quad param = (abs(linear\_accum) > l1).select(pre\_shrink, 0.0)
&else:
&\quad y =
\\
frac{new\_accum^{-lr\_power}}{learning\_rate} + (2 * l2)
&\quad pre\_shrink =
\\
frac{x}{y}
&\quad param = (abs(linear\_accum) > l1).select(pre\_shrink, 0.0)
&squared\_accum += grad^2
Args:
learning_rate (float|Variable): global learning rate.
l1 (float):
l2 (float):
lr_power (float):
Raises:
ValueError: If learning_rate, rho, epsilon, momentum are None.
Examples:
.. code-block:: python
optimizer = fluid.optimizer.Ftrl(0.0001)
_, params_grads = optimizer.minimize(cost)
"""
_squared_acc_str
=
"squared"
_linear_acc_str
=
"linear"
def
__init__
(
self
,
learning_rate
,
l1
=
0.0
,
l2
=
0.0
,
lr_power
=-
0.5
,
**
kwargs
):
super
(
FtrlOptimizer
,
self
).
__init__
(
learning_rate
=
learning_rate
,
**
kwargs
)
if
learning_rate
is
None
:
raise
ValueError
(
"learning_rate is not set."
)
self
.
type
=
"ftrl"
self
.
_l1
=
l1
self
.
_l2
=
l2
self
.
_lr_power
=
lr_power
def
_create_accumulators
(
self
,
block
,
parameters
):
if
not
isinstance
(
block
,
framework
.
Block
):
raise
TypeError
(
"block is not instance of framework.Block."
)
for
p
in
parameters
:
self
.
_add_accumulator
(
self
.
_squared_acc_str
,
p
)
self
.
_add_accumulator
(
self
.
_linear_acc_str
,
p
)
def
_append_optimize_op
(
self
,
block
,
param_and_grad
):
if
not
isinstance
(
block
,
framework
.
Block
):
raise
TypeError
(
"block is not instance of framework.Block."
)
squared_acc
=
self
.
_get_accumulator
(
self
.
_squared_acc_str
,
param_and_grad
[
0
])
linear_acc
=
self
.
_get_accumulator
(
self
.
_linear_acc_str
,
param_and_grad
[
0
])
ftrl_op
=
block
.
append_op
(
type
=
self
.
type
,
inputs
=
{
"Param"
:
param_and_grad
[
0
],
"Grad"
:
param_and_grad
[
1
],
"SquaredAccumulator"
:
squared_acc
,
"LinearAccumulator"
:
linear_acc
,
"LearningRate"
:
self
.
_create_param_lr
(
param_and_grad
),
},
outputs
=
{
"ParamOut"
:
param_and_grad
[
0
],
"SquaredAccumOut"
:
squared_acc
,
"LinearAccumOut"
:
linear_acc
},
attrs
=
{
"l1"
:
self
.
_l1
,
"l2"
:
self
.
_l1
,
"lr_power"
:
self
.
_lr_power
})
return
ftrl_op
# We short the class name, since users will use the optimizer with the package
# name. The sample code:
#
...
...
@@ -826,6 +933,7 @@ Adamax = AdamaxOptimizer
DecayedAdagrad
=
DecayedAdagradOptimizer
Adadelta
=
AdadeltaOptimizer
RMSProp
=
RMSPropOptimizer
Ftrl
=
FtrlOptimizer
class
ModelAverage
(
Optimizer
):
...
...
python/paddle/fluid/tests/unittests/test_optimizer.py
浏览文件 @
4ec9ecae
...
...
@@ -434,5 +434,71 @@ class TestDecayedAdagradOptimizer(unittest.TestCase):
self
.
assertAlmostEqual
(
init_ops
[
1
].
attr
(
'value'
),
0.0
)
class
TestFtrlOptimizer
(
unittest
.
TestCase
):
class
MockFtrl
(
optimizer
.
FtrlOptimizer
):
def
get_accumulators
(
self
):
return
self
.
_accumulators
def
get_squared_str
(
self
):
return
self
.
_squared_acc_str
def
get_linear_str
(
self
):
return
self
.
_linear_acc_str
def
test_ftrl_optimizer
(
self
):
init_program
=
framework
.
Program
()
program
=
framework
.
Program
()
block
=
program
.
global_block
()
mul_x
=
block
.
create_parameter
(
dtype
=
"float32"
,
shape
=
[
5
,
10
],
lod_level
=
0
,
name
=
"mul.x"
,
optimize_attr
=
{
'learning_rate'
:
1.1
})
mul_y
=
block
.
create_var
(
dtype
=
"float32"
,
shape
=
[
10
,
8
],
lod_level
=
0
,
name
=
"mul.y"
)
mul_out
=
block
.
create_var
(
dtype
=
"float32"
,
shape
=
[
5
,
8
],
lod_level
=
0
,
name
=
"mul.out"
)
block
.
append_op
(
type
=
"mul"
,
inputs
=
{
"X"
:
mul_x
,
"Y"
:
mul_y
},
outputs
=
{
"Out"
:
mul_out
},
attrs
=
{
"x_num_col_dims"
:
1
})
mean_out
=
block
.
create_var
(
dtype
=
"float32"
,
shape
=
[
1
],
lod_level
=
0
,
name
=
"mean.out"
)
block
.
append_op
(
type
=
"mean"
,
inputs
=
{
"X"
:
mul_out
},
outputs
=
{
"Out"
:
mean_out
})
learning_rate
=
0.01
ftrl_optimizer
=
self
.
MockFtrl
(
learning_rate
=
learning_rate
,
l1
=
0.0
,
l2
=
0.0
,
lr_power
=-
0.5
)
params_grads
=
append_backward
(
mean_out
)
self
.
assertEqual
(
len
(
params_grads
),
1
)
self
.
assertEqual
(
len
(
ftrl_optimizer
.
get_accumulators
()),
0
)
opts
=
ftrl_optimizer
.
create_optimization_pass
(
params_grads
,
mul_out
,
init_program
)
self
.
assertEqual
(
len
(
opts
),
3
)
self
.
assertEqual
([
op
.
type
for
op
in
opts
],
[
"fill_constant"
,
"elementwise_mul"
,
"ftrl"
])
# Check accumulators
accumulators
=
ftrl_optimizer
.
get_accumulators
()
self
.
assertEqual
(
len
(
accumulators
),
2
)
self
.
assertTrue
(
ftrl_optimizer
.
get_squared_str
()
in
accumulators
)
self
.
assertTrue
(
ftrl_optimizer
.
get_linear_str
()
in
accumulators
)
squared_acc
=
accumulators
[
ftrl_optimizer
.
get_squared_str
()]
linear_acc
=
accumulators
[
ftrl_optimizer
.
get_linear_str
()]
self
.
assertEqual
(
len
(
squared_acc
),
1
)
self
.
assertEqual
(
len
(
linear_acc
),
1
)
self
.
assertTrue
(
mul_x
.
name
in
squared_acc
)
self
.
assertTrue
(
mul_x
.
name
in
linear_acc
)
# Check init_program
init_ops
=
init_program
.
global_block
().
ops
self
.
assertEqual
(
len
(
init_ops
),
3
)
self
.
assertEqual
(
init_ops
[
0
].
type
,
"fill_constant"
)
self
.
assertAlmostEqual
(
init_ops
[
0
].
attr
(
'value'
),
learning_rate
)
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