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
69eed34d
P
Paddle
项目概览
PaddlePaddle
/
Paddle
大约 1 年 前同步成功
通知
2298
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 >>
未验证
提交
69eed34d
编写于
9月 29, 2021
作者:
Z
zhaoyingli
提交者:
GitHub
9月 29, 2021
浏览文件
操作
浏览文件
下载
电子邮件补丁
差异文件
add optest for adamw (#36148)
* update func name * skip cpu * update unittest * update unittest
上级
3eb50715
变更
2
隐藏空白更改
内联
并排
Showing
2 changed file
with
165 addition
and
7 deletion
+165
-7
python/paddle/fluid/tests/unittests/test_adamw_op.py
python/paddle/fluid/tests/unittests/test_adamw_op.py
+162
-4
python/paddle/optimizer/adamw.py
python/paddle/optimizer/adamw.py
+3
-3
未找到文件。
python/paddle/fluid/tests/unittests/test_adamw_op.py
浏览文件 @
69eed34d
...
...
@@ -14,9 +14,153 @@
import
unittest
import
paddle
import
random
import
numpy
as
np
import
paddle.fluid
as
fluid
from
op_test
import
OpTest
from
functools
import
partial
from
paddle.framework
import
core
def
adamw_step
(
inputs
,
attributes
):
param
=
inputs
[
'Param'
]
grad
=
inputs
[
'Grad'
]
moment1
=
inputs
[
'Moment1'
]
moment2
=
inputs
[
'Moment2'
]
lr
=
inputs
[
'LearningRate'
]
beta1_pow
=
inputs
[
'Beta1Pow'
]
beta2_pow
=
inputs
[
'Beta2Pow'
]
epsilon
=
attributes
[
'epsilon'
]
if
'lr_ratio'
in
attributes
:
lr
=
lr
*
attributes
[
'lr_ratio'
]
if
attributes
[
"with_decay"
]:
coeff
=
attributes
[
"coeff"
]
decay
=
1.0
-
lr
*
coeff
param2
=
param
*
decay
param
=
param2
.
copy
()
if
'beta1'
in
attributes
:
beta1
=
attributes
[
'beta1'
]
else
:
beta1
=
inputs
[
'Beta1Tensor'
][
0
]
if
'beta2'
in
attributes
:
beta2
=
attributes
[
'beta2'
]
else
:
beta2
=
inputs
[
'Beta2Tensor'
][
0
]
moment1_out
=
beta1
*
moment1
+
(
1
-
beta1
)
*
grad
moment2_out
=
beta2
*
moment2
+
(
1
-
beta2
)
*
np
.
square
(
grad
)
lr_t
=
lr
*
np
.
sqrt
(
1
-
beta2_pow
)
/
(
1
-
beta1_pow
)
param_out
=
param
-
lr_t
*
(
moment1_out
/
(
np
.
sqrt
(
moment2_out
)
+
epsilon
))
return
param_out
,
moment1_out
,
moment2_out
class
TestAdamW
(
OpTest
):
def
setUp
(
self
):
'''Test AdamW Op with supplied attributes
'''
self
.
op_type
=
"adamw"
param
=
np
.
random
.
uniform
(
-
1
,
1
,
(
102
,
105
)).
astype
(
"float32"
)
grad
=
np
.
random
.
uniform
(
-
1
,
1
,
(
102
,
105
)).
astype
(
"float32"
)
moment1
=
np
.
random
.
uniform
(
-
1
,
1
,
(
102
,
105
)).
astype
(
"float32"
)
# The second moment is positive
moment2
=
np
.
random
.
random
((
102
,
105
)).
astype
(
"float32"
)
learning_rate
=
0.004
beta1
=
0.78
beta2
=
0.836
epsilon
=
1e-4
beta1_pow
=
beta1
**
10
beta2_pow
=
beta2
**
10
self
.
inputs
=
{
'Param'
:
param
,
'Grad'
:
grad
,
'Moment1'
:
moment1
,
'Moment2'
:
moment2
,
'LearningRate'
:
np
.
array
([
learning_rate
]).
astype
(
"float32"
),
'Beta1Pow'
:
np
.
array
([
beta1_pow
]).
astype
(
"float32"
),
'Beta2Pow'
:
np
.
array
([
beta2_pow
]).
astype
(
"float32"
)
}
self
.
attrs
=
{
'epsilon'
:
epsilon
,
'beta1'
:
beta1
,
'beta2'
:
beta2
,
"coeff"
:
0.5
,
"with_decay"
:
True
}
param_out
,
moment1_out
,
\
moment2_out
=
adamw_step
(
self
.
inputs
,
self
.
attrs
)
self
.
outputs
=
{
'Moment1Out'
:
moment1_out
,
'Moment2Out'
:
moment2_out
,
'ParamOut'
:
param_out
,
'Beta1PowOut'
:
np
.
array
([
beta1_pow
]).
astype
(
"float32"
)
*
beta1
,
'Beta2PowOut'
:
np
.
array
([
beta2_pow
]).
astype
(
"float32"
)
*
beta2
}
def
test_check_output
(
self
):
self
.
check_output
()
@
unittest
.
skipIf
(
not
core
.
is_compiled_with_cuda
(),
"core is not compiled with CUDA"
)
class
TestAdamW2
(
OpTest
):
def
setUp
(
self
):
'''Test AdamW Op with supplied attributes
'''
self
.
op_type
=
"adamw"
param
=
np
.
random
.
uniform
(
-
1
,
1
,
(
2
,
2
)).
astype
(
"float32"
)
grad
=
np
.
random
.
uniform
(
-
1
,
1
,
(
2
,
2
)).
astype
(
"float32"
)
moment1
=
np
.
random
.
uniform
(
-
1
,
1
,
(
2
,
2
)).
astype
(
"float32"
)
# The second moment is positive
moment2
=
np
.
random
.
random
((
2
,
2
)).
astype
(
"float32"
)
learning_rate
=
0.004
beta1
=
0.78
beta2
=
0.836
epsilon
=
1e-4
beta1_pow
=
beta1
**
10
beta2_pow
=
beta2
**
10
self
.
inputs
=
{
'Param'
:
param
,
'Grad'
:
grad
,
'Moment1'
:
moment1
,
'Moment2'
:
moment2
,
'LearningRate'
:
np
.
array
([
learning_rate
]).
astype
(
"float32"
),
'Beta1Pow'
:
np
.
array
([
beta1_pow
]).
astype
(
"float32"
),
'Beta2Pow'
:
np
.
array
([
beta2_pow
]).
astype
(
"float32"
)
}
self
.
attrs
=
{
'epsilon'
:
epsilon
,
'beta1'
:
beta1
,
'beta2'
:
beta2
,
"lr_ratio"
:
0.1
,
"coeff"
:
0.5
,
"with_decay"
:
True
}
param_out
,
moment1_out
,
moment2_out
=
adamw_step
(
self
.
inputs
,
self
.
attrs
)
self
.
outputs
=
{
'Moment1Out'
:
moment1_out
,
'Moment2Out'
:
moment2_out
,
'ParamOut'
:
param_out
,
'Beta1PowOut'
:
np
.
array
([
beta1_pow
]).
astype
(
"float32"
)
*
beta1
,
'Beta2PowOut'
:
np
.
array
([
beta2_pow
]).
astype
(
"float32"
)
*
beta2
}
def
test_check_output
(
self
):
self
.
check_output_with_place
(
core
.
CUDAPlace
(
0
))
class
TestAdamWOp
(
unittest
.
TestCase
):
...
...
@@ -160,7 +304,14 @@ def simple_lr_setting(param, decay_rate, n_layers):
return
decay_rate
**
(
n_layers
+
2
-
depth
)
@
unittest
.
skipIf
(
not
core
.
is_compiled_with_cuda
(),
"core is not compiled with CUDA"
)
class
TestAdamWOpLayerwiseLR
(
TestAdamWOp
):
def
setUp
(
self
):
random
.
seed
(
2021
)
np
.
random
.
seed
(
2021
)
paddle
.
seed
(
2021
)
def
test_adamw_op_dygraph
(
self
):
paddle
.
disable_static
()
value
=
np
.
arange
(
26
).
reshape
(
2
,
13
).
astype
(
"float32"
)
...
...
@@ -181,17 +332,20 @@ class TestAdamWOpLayerwiseLR(TestAdamWOp):
weight_decay
=
0.01
,
lr_ratio
=
simple_lr_fun
)
for
_
in
range
(
2
):
loss_ref
=
np
.
array
(
[
4.8383293
,
3.0854003
,
1.33299
,
-
0.418993
,
-
2.171043
])
for
i
in
range
(
5
):
a1
=
linear1
(
a
)
out
=
linear2
(
a1
)
out
=
paddle
.
mean
(
out
)
out
.
backward
()
adam
.
step
()
adam
.
clear_gradients
()
np
.
testing
.
assert_allclose
(
out
[
0
].
numpy
(),
loss_ref
[
i
],
rtol
=
1e-6
)
def
test_adamw_op
(
self
):
paddle
.
enable_static
()
place
=
fluid
.
CUDAPlace
(
0
)
if
fluid
.
is_compiled_with_cuda
()
\
else
fluid
.
CPUPlace
()
place
=
fluid
.
CUDAPlace
(
0
)
train_prog
=
fluid
.
Program
()
startup
=
fluid
.
Program
()
with
fluid
.
program_guard
(
train_prog
,
startup
):
...
...
@@ -223,7 +377,10 @@ class TestAdamWOpLayerwiseLR(TestAdamWOp):
exe
=
fluid
.
Executor
(
place
)
exe
.
run
(
startup
)
for
_
in
range
(
2
):
loss_ref
=
np
.
array
(
[
0.36120513
,
0.2720821
,
0.67208904
,
0.14607805
,
0.24098626
])
for
i
in
range
(
5
):
inputs
=
np
.
random
.
random
(
size
=
[
8
,
10
]).
astype
(
'float32'
)
outputs
=
np
.
random
.
random
(
size
=
[
8
,
1
]).
astype
(
'float32'
)
rets
=
exe
.
run
(
train_prog
,
...
...
@@ -231,6 +388,7 @@ class TestAdamWOpLayerwiseLR(TestAdamWOp):
"y"
:
outputs
},
fetch_list
=
[
avg_cost
])
assert
rets
[
0
]
is
not
None
np
.
testing
.
assert_allclose
(
rets
[
0
],
loss_ref
[
i
],
rtol
=
1e-6
)
paddle
.
disable_static
()
...
...
python/paddle/optimizer/adamw.py
浏览文件 @
69eed34d
...
...
@@ -171,9 +171,9 @@ class AdamW(Adam):
self
.
_lr_to_coeff
=
dict
()
if
lr_ratio
is
not
None
:
assert
isinstance
(
lr_ratio
,
Callable
)
if
core
.
is_compiled_with_xpu
()
or
core
.
is_compiled_with_npu
():
if
not
core
.
is_compiled_with_cuda
():
raise
NotImplementedError
(
"'lr_ratio' is unimplemented in XPU and NPU"
)
"'lr_ratio' is unimplemented in
CPU,
XPU and NPU"
)
self
.
_lr_ratio
=
lr_ratio
super
(
AdamW
,
self
).
__init__
(
...
...
@@ -305,7 +305,7 @@ class AdamW(Adam):
'epsilon'
,
self
.
_epsilon
,
'lazy_mode'
,
self
.
_lazy_mode
,
'min_row_size_to_use_multithread'
,
1000
,
'beta1'
,
_beta1
,
'beta2'
,
_beta2
,
'coeff'
,
self
.
_coeff
,
'multi_precision'
,
find_master
,
"lr_ratio"
,
lr_ratio_
)
find_master
,
'lr_ratio'
,
lr_ratio_
)
return
None
...
...
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