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
1bc47c84
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
1bc47c84
编写于
7月 14, 2022
作者:
Y
Yao Zihang
提交者:
GitHub
7月 14, 2022
浏览文件
操作
浏览文件
下载
电子邮件补丁
差异文件
Optimize batchnorm1d using 2D kernel (#43530)
上级
a2c4c86b
变更
3
展开全部
隐藏空白更改
内联
并排
Showing
3 changed file
with
549 addition
and
38 deletion
+549
-38
paddle/phi/kernels/gpu/batch_norm_grad_kernel.cu
paddle/phi/kernels/gpu/batch_norm_grad_kernel.cu
+4
-2
paddle/phi/kernels/gpu/batch_norm_kernel.cu
paddle/phi/kernels/gpu/batch_norm_kernel.cu
+511
-11
python/paddle/fluid/tests/unittests/test_batch_norm_op_v2.py
python/paddle/fluid/tests/unittests/test_batch_norm_op_v2.py
+34
-25
未找到文件。
paddle/phi/kernels/gpu/batch_norm_grad_kernel.cu
浏览文件 @
1bc47c84
...
...
@@ -591,10 +591,12 @@ void BatchNormGradRawKernel(const Context &ctx,
// ctx.GetPlace()),
// epsilon, saved_mean_data, saved_var_data));
#else
// CUDNN
PER_ACTIVATION mode
only support small batch size
// CUDNN only support small batch size
const
size_t
CUDNN_PER_ACTIVATION_THRESHOLD
=
131070
;
const
size_t
CUDNN_SPATIAL_THRESHOLD
=
880801
;
const
bool
use_native_kernel
=
(
x_dims
.
size
()
==
2
&&
N
>=
CUDNN_PER_ACTIVATION_THRESHOLD
);
((
x_dims
.
size
()
==
2
&&
N
>=
CUDNN_PER_ACTIVATION_THRESHOLD
)
||
(
x_dims
.
size
()
==
3
&&
N
>=
CUDNN_SPATIAL_THRESHOLD
));
if
(
use_native_kernel
)
{
if
(
compute_format
==
DataLayout
::
kNCHW
)
{
BNBackward
<
T
,
block
,
DataLayout
::
kNCHW
>
...
...
paddle/phi/kernels/gpu/batch_norm_kernel.cu
浏览文件 @
1bc47c84
此差异已折叠。
点击以展开。
python/paddle/fluid/tests/unittests/test_batch_norm_op_v2.py
浏览文件 @
1bc47c84
...
...
@@ -82,50 +82,58 @@ class TestBatchNorm(unittest.TestCase):
self
.
assertRaises
(
ValueError
,
error2d_dataformat
)
self
.
assertRaises
(
ValueError
,
error3d_dataformat
)
def
test_eager_api
(
self
):
places
=
[
fluid
.
CPUPlace
()]
if
core
.
is_compiled_with_cuda
():
places
.
append
(
fluid
.
CUDAPlace
(
0
))
for
p
in
places
:
shape
=
[
4
,
10
,
4
,
4
]
def
test_large_batch
(
self
):
def
compute_v1
(
x
):
with
fluid
.
dygraph
.
guard
(
p
):
bn
=
fluid
.
dygraph
.
BatchNorm
(
shape
[
1
])
#bn = paddle.nn.BatchNorm2D(shape[1])
def
compute_baseline
(
x
):
with
fluid
.
dygraph
.
guard
(
p
):
bn
=
fluid
.
dygraph
.
BatchNorm
(
shape
[
1
])
x1
=
paddle
.
to_tensor
(
x
)
x1
.
stop_gradient
=
False
y
=
bn
(
x1
)
y
.
backward
()
return
y
.
numpy
(),
x1
.
gradient
()
def
compute_1d
(
x
):
with
fluid
.
dygraph
.
guard
(
p
):
with
_test_eager_guard
():
bn
=
paddle
.
nn
.
BatchNorm1D
(
shape
[
1
])
x1
=
paddle
.
to_tensor
(
x
)
x1
.
stop_gradient
=
False
y
=
bn
(
x1
)
y
.
backward
()
return
y
.
numpy
(),
x1
.
gradient
()
def
compute_v2
(
x
):
with
fluid
.
dygraph
.
guard
(
p
):
with
_test_eager_guard
():
print
(
"v2"
)
bn
=
paddle
.
nn
.
BatchNorm2D
(
shape
[
1
])
x1
=
paddle
.
to_tensor
(
x
)
x1
.
stop_gradient
=
False
y
=
bn
(
x1
)
y
.
backward
()
return
y
.
numpy
(),
x1
.
gradient
()
places
=
[
fluid
.
CPUPlace
()]
if
core
.
is_compiled_with_cuda
():
places
.
append
(
fluid
.
CUDAPlace
(
0
))
for
p
in
places
:
# [N, C]
shape
=
[
200000
,
4
]
x
=
np
.
random
.
randn
(
*
shape
).
astype
(
"float32"
)
y1
,
g1
=
compute_baseline
(
x
)
y2
,
g2
=
compute_1d
(
x
)
self
.
assertTrue
(
np
.
allclose
(
g1
,
g2
))
self
.
assertTrue
(
np
.
allclose
(
y1
,
y2
))
# [N, C, L]
shape
=
[
1000000
,
4
,
4
]
x
=
np
.
random
.
randn
(
*
shape
).
astype
(
"float32"
)
y1
,
g1
=
compute_
v1
(
x
)
y2
,
g2
=
compute_
v2
(
x
)
y1
,
g1
=
compute_
baseline
(
x
)
y2
,
g2
=
compute_
1d
(
x
)
self
.
assertTrue
(
np
.
allclose
(
g1
,
g2
))
self
.
assertTrue
(
np
.
allclose
(
y1
,
y2
))
def
test_eager_api
_1d
(
self
):
def
test_eager_api
(
self
):
places
=
[
fluid
.
CPUPlace
()]
if
core
.
is_compiled_with_cuda
():
places
.
append
(
fluid
.
CUDAPlace
(
0
))
for
p
in
places
:
shape
=
[
200000
,
4
]
shape
=
[
4
,
10
,
4
,
4
]
def
compute_v1
(
x
):
with
fluid
.
dygraph
.
guard
(
p
):
bn
=
fluid
.
dygraph
.
BatchNorm
(
shape
[
1
])
#bn = paddle.nn.BatchNorm2D(shape[1])
x1
=
paddle
.
to_tensor
(
x
)
x1
.
stop_gradient
=
False
y
=
bn
(
x1
)
...
...
@@ -135,7 +143,8 @@ class TestBatchNorm(unittest.TestCase):
def
compute_v2
(
x
):
with
fluid
.
dygraph
.
guard
(
p
):
with
_test_eager_guard
():
bn
=
paddle
.
nn
.
BatchNorm1D
(
shape
[
1
])
print
(
"v2"
)
bn
=
paddle
.
nn
.
BatchNorm2D
(
shape
[
1
])
x1
=
paddle
.
to_tensor
(
x
)
x1
.
stop_gradient
=
False
y
=
bn
(
x1
)
...
...
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