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
afeixing77
Paddle
提交
a4b9daf9
P
Paddle
项目概览
afeixing77
/
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 >>
未验证
提交
a4b9daf9
编写于
12月 28, 2020
作者:
L
Leo Chen
提交者:
GitHub
12月 28, 2020
浏览文件
操作
浏览文件
下载
电子邮件补丁
差异文件
fix optimizer dtype (#29917)
上级
9602a182
变更
5
显示空白变更内容
内联
并排
Showing
5 changed file
with
53 addition
and
8 deletion
+53
-8
python/paddle/fluid/optimizer.py
python/paddle/fluid/optimizer.py
+9
-2
python/paddle/fluid/tests/unittests/test_optimizer.py
python/paddle/fluid/tests/unittests/test_optimizer.py
+34
-1
python/paddle/optimizer/adam.py
python/paddle/optimizer/adam.py
+0
-1
python/paddle/optimizer/adamw.py
python/paddle/optimizer/adamw.py
+0
-1
python/paddle/optimizer/optimizer.py
python/paddle/optimizer/optimizer.py
+10
-3
未找到文件。
python/paddle/fluid/optimizer.py
浏览文件 @
a4b9daf9
...
...
@@ -108,8 +108,12 @@ class Optimizer(object):
self
.
regularization
=
regularization
self
.
_grad_clip
=
grad_clip
self
.
_learning_rate
=
learning_rate
# the learning rate type should be inferenced from loss
self
.
_dtype
=
None
# Infer the dtype form parameter
if
self
.
_parameter_list
:
self
.
_dtype
=
self
.
_parameter_list
[
0
].
dtype
# each program should have a independent learning rate
# program -> Variable(learning_rate)
self
.
_learning_rate_map
=
dict
()
...
...
@@ -768,7 +772,10 @@ class Optimizer(object):
else
:
act_no_grad_set
=
self
.
_get_no_grad_set
(
loss
,
no_grad_set
)
# Infer dtype by loss if None
if
self
.
_dtype
is
None
:
self
.
_dtype
=
loss
.
dtype
if
framework
.
in_dygraph_mode
():
parameter_list
=
parameter_list
if
parameter_list
\
else
self
.
_parameter_list
...
...
python/paddle/fluid/tests/unittests/test_optimizer.py
浏览文件 @
a4b9daf9
...
...
@@ -23,7 +23,8 @@ import paddle.fluid.core as core
import
paddle.compat
as
cpt
import
numpy
as
np
from
paddle.fluid.backward
import
append_backward
from
paddle.fluid.framework
import
Program
,
program_guard
from
paddle.fluid.framework
import
Program
,
program_guard
,
convert_np_dtype_to_dtype_
import
paddle
class
TestOptimizer
(
unittest
.
TestCase
):
...
...
@@ -1042,5 +1043,37 @@ class TestGradientMergeOptimizer(unittest.TestCase):
[
'sgd'
,
'sgd'
])
class
TestOptimizerDtype
(
unittest
.
TestCase
):
'''
The dtype of optimizer should be inferred by parameters, and the learning rate
is cteated with the same dtype.
'''
def
check_with_dtype
(
self
,
dtype
):
class
MyLayer
(
paddle
.
nn
.
Layer
):
def
__init__
(
self
,
dtype
):
super
(
MyLayer
,
self
).
__init__
()
self
.
_w
=
self
.
create_parameter
([
2
,
3
],
dtype
=
dtype
)
self
.
_b
=
self
.
create_parameter
([
2
,
3
],
dtype
=
dtype
)
def
forward
(
self
,
x
):
return
x
*
self
.
_w
+
self
.
_b
with
paddle
.
fluid
.
dygraph
.
guard
():
model
=
MyLayer
(
dtype
)
x
=
paddle
.
rand
([
10
,
2
,
3
],
dtype
=
dtype
)
loss
=
model
(
x
)
adam
=
paddle
.
optimizer
.
Adam
(
parameters
=
model
.
parameters
())
loss
.
backward
()
adam
.
step
()
self
.
assertEqual
(
adam
.
_dtype
,
convert_np_dtype_to_dtype_
(
dtype
))
def
test_float64
(
self
):
self
.
check_with_dtype
(
'float64'
)
def
test_float32
(
self
):
self
.
check_with_dtype
(
'float32'
)
if
__name__
==
'__main__'
:
unittest
.
main
()
python/paddle/optimizer/adam.py
浏览文件 @
a4b9daf9
...
...
@@ -270,7 +270,6 @@ class Adam(Optimizer):
adam.step()
adam.clear_grad()
"""
self
.
_dtype
=
None
params_grads
=
[]
for
param
in
self
.
_parameter_list
:
if
not
param
.
trainable
:
...
...
python/paddle/optimizer/adamw.py
浏览文件 @
a4b9daf9
...
...
@@ -210,7 +210,6 @@ class AdamW(Adam):
@
framework
.
dygraph_only
@
imperative_base
.
no_grad
def
step
(
self
):
self
.
_dtype
=
None
params_grads
=
[]
for
param
in
self
.
_parameter_list
:
if
not
param
.
trainable
:
...
...
python/paddle/optimizer/optimizer.py
浏览文件 @
a4b9daf9
...
...
@@ -132,8 +132,12 @@ class Optimizer(object):
self
.
regularization
=
weight_decay
self
.
_grad_clip
=
grad_clip
self
.
_learning_rate
=
learning_rate
# the learning rate type should be inferenced from loss
self
.
_dtype
=
None
# Infer the dtype form parameter
if
self
.
_parameter_list
:
self
.
_dtype
=
self
.
_parameter_list
[
0
].
dtype
# each program should have a independent learning rate
# program -> tensor(learning_rate)
self
.
_learning_rate_map
=
dict
()
...
...
@@ -675,7 +679,10 @@ class Optimizer(object):
else
:
act_no_grad_set
=
self
.
_get_no_grad_set
(
loss
,
no_grad_set
)
# Infer dtype by loss if None
if
self
.
_dtype
is
None
:
self
.
_dtype
=
loss
.
dtype
if
framework
.
in_dygraph_mode
():
parameter_list
=
parameters
if
parameters
\
else
self
.
_parameter_list
...
...
@@ -885,6 +892,7 @@ class Optimizer(object):
return
optimize_ops
,
params_grads
@
imperative_base
.
no_grad
@
framework
.
dygraph_only
def
step
(
self
):
"""
...
...
@@ -910,7 +918,6 @@ class Optimizer(object):
adam.step()
adam.clear_grad()
"""
self
.
_dtype
=
None
params_grads
=
[]
for
param
in
self
.
_parameter_list
:
if
not
param
.
trainable
:
...
...
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