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
Paddle
提交
7ce0d45e
P
Paddle
项目概览
PaddlePaddle
/
Paddle
大约 1 年 前同步成功
通知
2297
Star
20931
Fork
5422
代码
文件
提交
分支
Tags
贡献者
分支图
Diff
Issue
1423
列表
看板
标记
里程碑
合并请求
543
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
1,423
Issue
1,423
列表
看板
标记
里程碑
合并请求
543
合并请求
543
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
7ce0d45e
编写于
7月 11, 2018
作者:
Q
qiaolongfei
浏览文件
操作
浏览文件
下载
电子邮件补丁
差异文件
fix adam and adamax optimizer
上级
e91ecd5d
变更
1
显示空白变更内容
内联
并排
Showing
1 changed file
with
61 addition
and
61 deletion
+61
-61
python/paddle/fluid/optimizer.py
python/paddle/fluid/optimizer.py
+61
-61
未找到文件。
python/paddle/fluid/optimizer.py
浏览文件 @
7ce0d45e
...
@@ -123,7 +123,7 @@ class Optimizer(object):
...
@@ -123,7 +123,7 @@ class Optimizer(object):
"""
"""
pass
pass
def
_finish_update
(
self
,
block
):
def
_finish_update
(
self
,
block
,
parameters
):
"""Finish any custom updates needed
"""Finish any custom updates needed
before completing an optimization step
before completing an optimization step
...
@@ -132,7 +132,7 @@ class Optimizer(object):
...
@@ -132,7 +132,7 @@ class Optimizer(object):
parameters: list of parameter variables for the optimizer
parameters: list of parameter variables for the optimizer
Returns:
Returns:
list of finish ops or
None
None
"""
"""
pass
pass
...
@@ -236,7 +236,8 @@ class Optimizer(object):
...
@@ -236,7 +236,8 @@ class Optimizer(object):
# Get custom finish ops for subclasses
# Get custom finish ops for subclasses
# FIXME: Need to fix this once we figure out how to handle dependencies
# FIXME: Need to fix this once we figure out how to handle dependencies
self
.
_finish_update
(
loss
.
block
)
self
.
_finish_update
(
loss
.
block
,
[
p
[
0
]
for
p
in
parameters_and_grads
])
end
=
len
(
global_block
.
ops
)
end
=
len
(
global_block
.
ops
)
return
global_block
.
slice_ops
(
start
,
end
)
return
global_block
.
slice_ops
(
start
,
end
)
...
@@ -486,6 +487,8 @@ class AdamOptimizer(Optimizer):
...
@@ -486,6 +487,8 @@ class AdamOptimizer(Optimizer):
"""
"""
_moment1_acc_str
=
"moment1"
_moment1_acc_str
=
"moment1"
_moment2_acc_str
=
"moment2"
_moment2_acc_str
=
"moment2"
_beta1_pow_acc_str
=
"beta1_pow_acc"
_beta2_pow_acc_str
=
"beta2_pow_acc"
def
__init__
(
self
,
def
__init__
(
self
,
learning_rate
=
0.001
,
learning_rate
=
0.001
,
...
@@ -507,32 +510,22 @@ class AdamOptimizer(Optimizer):
...
@@ -507,32 +510,22 @@ class AdamOptimizer(Optimizer):
def
_create_accumulators
(
self
,
block
,
parameters
):
def
_create_accumulators
(
self
,
block
,
parameters
):
assert
isinstance
(
block
,
framework
.
Block
)
assert
isinstance
(
block
,
framework
.
Block
)
main_block
=
block
.
program
.
global_block
()
# Create beta1 and beta2 power tensors
beta_shape
=
[
1
]
self
.
_beta1_pow_acc
=
self
.
helper
.
create_global_variable
(
name
=
unique_name
.
generate
(
'beta1_pow_acc'
),
dtype
=
'float32'
if
self
.
_dtype
==
None
else
self
.
_dtype
,
shape
=
beta_shape
,
lod_level
=
0
,
persistable
=
True
)
self
.
helper
.
set_variable_initializer
(
self
.
_beta1_pow_acc
,
initializer
=
Constant
(
self
.
_beta1
))
self
.
_beta2_pow_acc
=
self
.
helper
.
create_global_variable
(
name
=
unique_name
.
generate
(
'beta2_pow_acc'
),
dtype
=
'float32'
if
self
.
_dtype
==
None
else
self
.
_dtype
,
shape
=
beta_shape
,
lod_level
=
0
,
persistable
=
True
)
self
.
helper
.
set_variable_initializer
(
self
.
_beta2_pow_acc
,
initializer
=
Constant
(
self
.
_beta2
))
# Create accumulator tensors for first and second moments
# Create accumulator tensors for first and second moments
for
p
in
parameters
:
for
p
in
parameters
:
self
.
_add_accumulator
(
self
.
_moment1_acc_str
,
p
)
self
.
_add_accumulator
(
self
.
_moment1_acc_str
,
p
)
self
.
_add_accumulator
(
self
.
_moment2_acc_str
,
p
)
self
.
_add_accumulator
(
self
.
_moment2_acc_str
,
p
)
self
.
_add_accumulator
(
name
=
self
.
_beta1_pow_acc_str
,
param
=
p
,
dtype
=
'float32'
,
fill_value
=
self
.
_beta1
,
shape
=
[
1
])
self
.
_add_accumulator
(
name
=
self
.
_beta2_pow_acc_str
,
param
=
p
,
dtype
=
'float32'
,
fill_value
=
self
.
_beta2
,
shape
=
[
1
])
def
_append_optimize_op
(
self
,
block
,
param_and_grad
):
def
_append_optimize_op
(
self
,
block
,
param_and_grad
):
assert
isinstance
(
block
,
framework
.
Block
)
assert
isinstance
(
block
,
framework
.
Block
)
...
@@ -541,6 +534,11 @@ class AdamOptimizer(Optimizer):
...
@@ -541,6 +534,11 @@ class AdamOptimizer(Optimizer):
param_and_grad
[
0
])
param_and_grad
[
0
])
moment2
=
self
.
_get_accumulator
(
self
.
_moment2_acc_str
,
moment2
=
self
.
_get_accumulator
(
self
.
_moment2_acc_str
,
param_and_grad
[
0
])
param_and_grad
[
0
])
beta1_pow_acc
=
self
.
_get_accumulator
(
self
.
_beta1_pow_acc_str
,
param_and_grad
[
0
])
beta2_pow_acc
=
self
.
_get_accumulator
(
self
.
_beta2_pow_acc_str
,
param_and_grad
[
0
])
# create the adam optimize op
# create the adam optimize op
adam_op
=
block
.
append_op
(
adam_op
=
block
.
append_op
(
type
=
self
.
type
,
type
=
self
.
type
,
...
@@ -550,8 +548,8 @@ class AdamOptimizer(Optimizer):
...
@@ -550,8 +548,8 @@ class AdamOptimizer(Optimizer):
"LearningRate"
:
self
.
_create_param_lr
(
param_and_grad
),
"LearningRate"
:
self
.
_create_param_lr
(
param_and_grad
),
"Moment1"
:
moment1
,
"Moment1"
:
moment1
,
"Moment2"
:
moment2
,
"Moment2"
:
moment2
,
"Beta1Pow"
:
self
.
_
beta1_pow_acc
,
"Beta1Pow"
:
beta1_pow_acc
,
"Beta2Pow"
:
self
.
_
beta2_pow_acc
"Beta2Pow"
:
beta2_pow_acc
},
},
outputs
=
{
outputs
=
{
"ParamOut"
:
param_and_grad
[
0
],
"ParamOut"
:
param_and_grad
[
0
],
...
@@ -566,25 +564,28 @@ class AdamOptimizer(Optimizer):
...
@@ -566,25 +564,28 @@ class AdamOptimizer(Optimizer):
return
adam_op
return
adam_op
def
_finish_update
(
self
,
block
):
def
_finish_update
(
self
,
block
,
parameters
):
"""Update Beta1 and Beta2 Power accumulators
"""Update Beta1 and Beta2 Power accumulators
"""
"""
assert
isinstance
(
block
,
framework
.
Block
)
assert
isinstance
(
block
,
framework
.
Block
)
main_block
=
block
.
program
.
global_block
()
main_block
=
block
.
program
.
global_block
()
scale_beta1
=
main_block
.
append_op
(
for
param
in
parameters
:
beta1_pow_acc
=
self
.
_get_accumulator
(
self
.
_beta1_pow_acc_str
,
param
)
beta2_pow_acc
=
self
.
_get_accumulator
(
self
.
_beta2_pow_acc_str
,
param
)
main_block
.
append_op
(
type
=
"scale"
,
type
=
"scale"
,
inputs
=
{
"X"
:
self
.
_
beta1_pow_acc
},
inputs
=
{
"X"
:
beta1_pow_acc
},
outputs
=
{
"Out"
:
self
.
_
beta1_pow_acc
},
outputs
=
{
"Out"
:
beta1_pow_acc
},
attrs
=
{
"scale"
:
self
.
_beta1
})
attrs
=
{
"scale"
:
self
.
_beta1
})
scale_beta2
=
main_block
.
append_op
(
main_block
.
append_op
(
type
=
"scale"
,
type
=
"scale"
,
inputs
=
{
"X"
:
self
.
_
beta2_pow_acc
},
inputs
=
{
"X"
:
beta2_pow_acc
},
outputs
=
{
"Out"
:
self
.
_
beta2_pow_acc
},
outputs
=
{
"Out"
:
beta2_pow_acc
},
attrs
=
{
"scale"
:
self
.
_beta2
})
attrs
=
{
"scale"
:
self
.
_beta2
})
return
[
scale_beta1
,
scale_beta2
]
class
AdamaxOptimizer
(
Optimizer
):
class
AdamaxOptimizer
(
Optimizer
):
"""
"""
...
@@ -626,6 +627,7 @@ class AdamaxOptimizer(Optimizer):
...
@@ -626,6 +627,7 @@ class AdamaxOptimizer(Optimizer):
"""
"""
_moment_acc_str
=
"moment"
_moment_acc_str
=
"moment"
_inf_norm_acc_str
=
"inf_norm"
_inf_norm_acc_str
=
"inf_norm"
_beta1_pow_acc_str
=
"beta1_pow_acc"
def
__init__
(
self
,
def
__init__
(
self
,
learning_rate
=
0.001
,
learning_rate
=
0.001
,
...
@@ -645,21 +647,16 @@ class AdamaxOptimizer(Optimizer):
...
@@ -645,21 +647,16 @@ class AdamaxOptimizer(Optimizer):
self
.
_epsilon
=
epsilon
self
.
_epsilon
=
epsilon
def
_create_accumulators
(
self
,
block
,
parameters
):
def
_create_accumulators
(
self
,
block
,
parameters
):
# Create beta1 power accumulator tensor
beta_shape
=
[
1
]
self
.
_beta1_pow_acc
=
self
.
helper
.
create_global_variable
(
name
=
unique_name
.
generate
(
'beta1_pow_acc'
),
dtype
=
'float32'
if
self
.
_dtype
==
None
else
self
.
_dtype
,
shape
=
beta_shape
,
lod_level
=
0
,
persistable
=
True
)
self
.
helper
.
set_variable_initializer
(
self
.
_beta1_pow_acc
,
initializer
=
Constant
(
self
.
_beta1
))
# Create accumulator tensors for first moment and infinity norm
# Create accumulator tensors for first moment and infinity norm
for
p
in
parameters
:
for
p
in
parameters
:
self
.
_add_accumulator
(
self
.
_moment_acc_str
,
p
)
self
.
_add_accumulator
(
self
.
_moment_acc_str
,
p
)
self
.
_add_accumulator
(
self
.
_inf_norm_acc_str
,
p
)
self
.
_add_accumulator
(
self
.
_inf_norm_acc_str
,
p
)
self
.
_add_accumulator
(
name
=
self
.
_beta1_pow_acc_str
,
param
=
p
,
dtype
=
'float32'
,
fill_value
=
self
.
_beta1
,
shape
=
[
1
])
def
_append_optimize_op
(
self
,
block
,
param_and_grad
):
def
_append_optimize_op
(
self
,
block
,
param_and_grad
):
assert
isinstance
(
block
,
framework
.
Block
)
assert
isinstance
(
block
,
framework
.
Block
)
...
@@ -667,6 +664,8 @@ class AdamaxOptimizer(Optimizer):
...
@@ -667,6 +664,8 @@ class AdamaxOptimizer(Optimizer):
moment
=
self
.
_get_accumulator
(
self
.
_moment_acc_str
,
param_and_grad
[
0
])
moment
=
self
.
_get_accumulator
(
self
.
_moment_acc_str
,
param_and_grad
[
0
])
inf_norm
=
self
.
_get_accumulator
(
self
.
_inf_norm_acc_str
,
inf_norm
=
self
.
_get_accumulator
(
self
.
_inf_norm_acc_str
,
param_and_grad
[
0
])
param_and_grad
[
0
])
beta1_pow_acc
=
self
.
_get_accumulator
(
self
.
_beta1_pow_acc_str
,
param_and_grad
[
0
])
# create the adamax optimize op
# create the adamax optimize op
adamax_op
=
block
.
append_op
(
adamax_op
=
block
.
append_op
(
type
=
self
.
type
,
type
=
self
.
type
,
...
@@ -676,7 +675,7 @@ class AdamaxOptimizer(Optimizer):
...
@@ -676,7 +675,7 @@ class AdamaxOptimizer(Optimizer):
"LearningRate"
:
self
.
_create_param_lr
(
param_and_grad
),
"LearningRate"
:
self
.
_create_param_lr
(
param_and_grad
),
"Moment"
:
moment
,
"Moment"
:
moment
,
"InfNorm"
:
inf_norm
,
"InfNorm"
:
inf_norm
,
"Beta1Pow"
:
self
.
_
beta1_pow_acc
"Beta1Pow"
:
beta1_pow_acc
},
},
outputs
=
{
outputs
=
{
"ParamOut"
:
param_and_grad
[
0
],
"ParamOut"
:
param_and_grad
[
0
],
...
@@ -691,19 +690,20 @@ class AdamaxOptimizer(Optimizer):
...
@@ -691,19 +690,20 @@ class AdamaxOptimizer(Optimizer):
return
adamax_op
return
adamax_op
def
_finish_update
(
self
,
block
):
def
_finish_update
(
self
,
block
,
parameters
):
"""Update Beta1 Power accumulator
"""Update Beta1 Power accumulator
"""
"""
assert
isinstance
(
block
,
framework
.
Block
)
assert
isinstance
(
block
,
framework
.
Block
)
main_block
=
block
.
program
.
global_block
()
main_block
=
block
.
program
.
global_block
()
scale_beta1
=
main_block
.
append_op
(
for
param
in
parameters
:
beta1_pow_acc
=
self
.
_get_accumulator
(
self
.
_beta1_pow_acc_str
,
param
)
main_block
.
append_op
(
type
=
"scale"
,
type
=
"scale"
,
inputs
=
{
"X"
:
self
.
_
beta1_pow_acc
},
inputs
=
{
"X"
:
beta1_pow_acc
},
outputs
=
{
"Out"
:
self
.
_
beta1_pow_acc
},
outputs
=
{
"Out"
:
beta1_pow_acc
},
attrs
=
{
"scale"
:
self
.
_beta1
})
attrs
=
{
"scale"
:
self
.
_beta1
})
return
[
scale_beta1
]
class
DecayedAdagradOptimizer
(
Optimizer
):
class
DecayedAdagradOptimizer
(
Optimizer
):
"""
"""
...
...
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