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
bfbc25bd
P
PaddleDetection
项目概览
PaddlePaddle
/
PaddleDetection
12 个月 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 >>
提交
bfbc25bd
编写于
3月 16, 2018
作者:
K
Kexin Zhao
浏览文件
操作
浏览文件
下载
电子邮件补丁
差异文件
add fp16 pool2d support
上级
02b3cfb1
变更
5
隐藏空白更改
内联
并排
Showing
5 changed file
with
131 addition
and
33 deletion
+131
-33
paddle/fluid/operators/pool_cudnn_op.cu.cc
paddle/fluid/operators/pool_cudnn_op.cu.cc
+11
-7
paddle/fluid/operators/pool_op.cc
paddle/fluid/operators/pool_op.cc
+7
-3
python/paddle/fluid/tests/unittests/op_test.py
python/paddle/fluid/tests/unittests/op_test.py
+3
-3
python/paddle/fluid/tests/unittests/test_conv2d_op.py
python/paddle/fluid/tests/unittests/test_conv2d_op.py
+25
-16
python/paddle/fluid/tests/unittests/test_pool2d_op.py
python/paddle/fluid/tests/unittests/test_pool2d_op.py
+85
-4
未找到文件。
paddle/fluid/operators/pool_cudnn_op.cu.cc
浏览文件 @
bfbc25bd
...
...
@@ -78,7 +78,8 @@ class PoolCUDNNOpKernel : public framework::OpKernel<T> {
// ------------------- cudnn pool algorithm ---------------------
auto
handle
=
ctx
.
cuda_device_context
().
cudnn_handle
();
T
alpha
=
1.0
f
,
beta
=
0.0
f
;
typename
platform
::
CudnnDataType
<
T
>::
ScalingParamType
alpha
=
1.0
f
,
beta
=
0.0
f
;
PADDLE_ENFORCE
(
platform
::
dynload
::
cudnnPoolingForward
(
handle
,
cudnn_pool_desc
,
&
alpha
,
cudnn_input_desc
,
input_data
,
&
beta
,
...
...
@@ -144,7 +145,8 @@ class PoolCUDNNGradOpKernel : public framework::OpKernel<T> {
// ------------------- cudnn pool algorithm ---------------------
auto
handle
=
ctx
.
cuda_device_context
().
cudnn_handle
();
T
alpha
=
1.0
f
,
beta
=
0.0
f
;
typename
platform
::
CudnnDataType
<
T
>::
ScalingParamType
alpha
=
1.0
f
,
beta
=
0.0
f
;
if
(
input_grad
)
{
T
*
input_grad_data
=
input_grad
->
mutable_data
<
T
>
(
ctx
.
GetPlace
());
...
...
@@ -162,17 +164,19 @@ class PoolCUDNNGradOpKernel : public framework::OpKernel<T> {
}
// namespace paddle
namespace
ops
=
paddle
::
operators
;
namespace
plat
=
paddle
::
platform
;
REGISTER_OP_KERNEL
(
pool2d
,
CUDNN
,
::
paddle
::
platform
::
CUDAPlace
,
REGISTER_OP_KERNEL
(
pool2d
,
CUDNN
,
plat
::
CUDAPlace
,
ops
::
PoolCUDNNOpKernel
<
float
>
,
ops
::
PoolCUDNNOpKernel
<
double
>
);
REGISTER_OP_KERNEL
(
pool2d_grad
,
CUDNN
,
::
paddle
::
platform
::
CUDAPlace
,
ops
::
PoolCUDNNOpKernel
<
double
>
,
ops
::
PoolCUDNNOpKernel
<
plat
::
float16
>
);
REGISTER_OP_KERNEL
(
pool2d_grad
,
CUDNN
,
plat
::
CUDAPlace
,
ops
::
PoolCUDNNGradOpKernel
<
float
>
,
ops
::
PoolCUDNNGradOpKernel
<
double
>
);
REGISTER_OP_KERNEL
(
pool3d
,
CUDNN
,
::
paddle
::
platform
::
CUDAPlace
,
REGISTER_OP_KERNEL
(
pool3d
,
CUDNN
,
plat
::
CUDAPlace
,
ops
::
PoolCUDNNOpKernel
<
float
>
,
ops
::
PoolCUDNNOpKernel
<
double
>
);
REGISTER_OP_KERNEL
(
pool3d_grad
,
CUDNN
,
::
paddle
::
platform
::
CUDAPlace
,
REGISTER_OP_KERNEL
(
pool3d_grad
,
CUDNN
,
plat
::
CUDAPlace
,
ops
::
PoolCUDNNGradOpKernel
<
float
>
,
ops
::
PoolCUDNNGradOpKernel
<
double
>
);
paddle/fluid/operators/pool_op.cc
浏览文件 @
bfbc25bd
...
...
@@ -124,11 +124,15 @@ framework::OpKernelType PoolOpGrad::GetExpectedKernelType(
}
#endif
auto
input_data_type
=
framework
::
ToDataType
(
ctx
.
Input
<
Tensor
>
(
"X"
)
->
type
());
if
(
input_data_type
==
framework
::
proto
::
VarType
::
FP16
)
{
PADDLE_ENFORCE_EQ
(
library_
,
framework
::
LibraryType
::
kCUDNN
,
"float16 can only be used when CUDNN is used"
);
}
std
::
string
data_format
=
ctx
.
Attr
<
std
::
string
>
(
"data_format"
);
framework
::
DataLayout
layout_
=
framework
::
StringToDataLayout
(
data_format
);
return
framework
::
OpKernelType
(
framework
::
ToDataType
(
ctx
.
Input
<
Tensor
>
(
"X"
)
->
type
()),
ctx
.
GetPlace
(),
layout_
,
library_
);
return
framework
::
OpKernelType
(
input_data_type
,
ctx
.
GetPlace
(),
layout_
,
library_
);
}
Pool2dOpMaker
::
Pool2dOpMaker
(
OpProto
*
proto
,
OpAttrChecker
*
op_checker
)
...
...
python/paddle/fluid/tests/unittests/op_test.py
浏览文件 @
bfbc25bd
...
...
@@ -483,9 +483,9 @@ class OpTest(unittest.TestCase):
input: input numpy array
Returns:
input:
if the dtype of input is np.float16, its dtype will be
changed to np.uint16 so that the internal memory will be
reinterpreted input
as of dtype np.uint16.
input:
The dtype of input will be changed to np.uint16 if
it is originally np.float16, such that the internal memory
of input will be reinterpreted
as of dtype np.uint16.
"""
if
input
.
dtype
==
np
.
float16
:
input
.
dtype
=
np
.
uint16
...
...
python/paddle/fluid/tests/unittests/test_conv2d_op.py
浏览文件 @
bfbc25bd
...
...
@@ -65,10 +65,10 @@ class TestConv2dOp(OpTest):
def
setUp
(
self
):
self
.
use_cudnn
=
False
self
.
use_mkldnn
=
False
self
.
dtype
=
np
.
float32
self
.
init_op_type
()
self
.
init_group
()
self
.
init_dilation
()
self
.
init_data_type
()
self
.
init_test_case
()
conv2d_param
=
{
...
...
@@ -159,9 +159,6 @@ class TestConv2dOp(OpTest):
f_c
=
self
.
input_size
[
1
]
/
self
.
groups
self
.
filter_size
=
[
6
,
f_c
,
3
,
3
]
def
init_data_type
(
self
):
self
.
dtype
=
np
.
float32
def
init_dilation
(
self
):
self
.
dilations
=
[
1
,
1
]
...
...
@@ -246,8 +243,10 @@ class TestCUDNN(TestConv2dOp):
self
.
op_type
=
"conv2d"
class
TestFP16CUDNN
(
TestCUDNN
):
def
init_data_type
(
self
):
class
TestFP16CUDNN
(
TestConv2dOp
):
def
init_op_type
(
self
):
self
.
use_cudnn
=
True
self
.
op_type
=
"conv2d"
self
.
dtype
=
np
.
float16
def
test_check_output
(
self
):
...
...
@@ -263,8 +262,10 @@ class TestCUDNNWithPad(TestWithPad):
self
.
op_type
=
"conv2d"
class
TestFP16CUDNNWithPad
(
TestCUDNNWithPad
):
def
init_data_type
(
self
):
class
TestFP16CUDNNWithPad
(
TestWithPad
):
def
init_op_type
(
self
):
self
.
use_cudnn
=
True
self
.
op_type
=
"conv2d"
self
.
dtype
=
np
.
float16
def
test_check_output
(
self
):
...
...
@@ -280,8 +281,10 @@ class TestCUDNNWithStride(TestWithStride):
self
.
op_type
=
"conv2d"
class
TestFP16CUDNNWithStride
(
TestCUDNNWithStride
):
def
init_data_type
(
self
):
class
TestFP16CUDNNWithStride
(
TestWithStride
):
def
init_op_type
(
self
):
self
.
use_cudnn
=
True
self
.
op_type
=
"conv2d"
self
.
dtype
=
np
.
float16
def
test_check_output
(
self
):
...
...
@@ -297,8 +300,10 @@ class TestCUDNNWithGroup(TestWithGroup):
self
.
op_type
=
"conv2d"
class
TestFP16CUDNNWithGroup
(
TestCUDNNWithGroup
):
def
init_data_type
(
self
):
class
TestFP16CUDNNWithGroup
(
TestWithGroup
):
def
init_op_type
(
self
):
self
.
use_cudnn
=
True
self
.
op_type
=
"conv2d"
self
.
dtype
=
np
.
float16
def
test_check_output
(
self
):
...
...
@@ -314,8 +319,10 @@ class TestCUDNNWith1x1(TestWith1x1):
self
.
op_type
=
"conv2d"
class
TestFP16CUDNNWith1x1
(
TestCUDNNWith1x1
):
def
init_data_type
(
self
):
class
TestFP16CUDNNWith1x1
(
TestWith1x1
):
def
init_op_type
(
self
):
self
.
use_cudnn
=
True
self
.
op_type
=
"conv2d"
self
.
dtype
=
np
.
float16
def
test_check_output
(
self
):
...
...
@@ -331,8 +338,10 @@ class TestCUDNNWithInput1x1Filter1x1(TestWithInput1x1Filter1x1):
self
.
op_type
=
"conv2d"
class
TestFP16CUDNNWithInput1x1Filter1x1
(
TestCUDNNWithInput1x1Filter1x1
):
def
init_data_type
(
self
):
class
TestFP16CUDNNWithInput1x1Filter1x1
(
TestWithInput1x1Filter1x1
):
def
init_op_type
(
self
):
self
.
use_cudnn
=
True
self
.
op_type
=
"conv2d"
self
.
dtype
=
np
.
float16
def
test_check_output
(
self
):
...
...
python/paddle/fluid/tests/unittests/test_pool2d_op.py
浏览文件 @
bfbc25bd
...
...
@@ -80,6 +80,7 @@ class TestPool2d_Op(OpTest):
def
setUp
(
self
):
self
.
use_cudnn
=
False
self
.
use_mkldnn
=
False
self
.
dtype
=
np
.
float32
self
.
init_test_case
()
self
.
init_global_pool
()
self
.
init_op_type
()
...
...
@@ -87,11 +88,11 @@ class TestPool2d_Op(OpTest):
self
.
init_ceil_mode
()
if
self
.
global_pool
:
self
.
paddings
=
[
0
for
_
in
range
(
len
(
self
.
paddings
))]
input
=
np
.
random
.
random
(
self
.
shape
).
astype
(
"float32"
)
input
=
np
.
random
.
random
(
self
.
shape
).
astype
(
self
.
dtype
)
output
=
self
.
pool2D_forward_naive
(
input
,
self
.
ksize
,
self
.
strides
,
self
.
paddings
,
self
.
global_pool
,
self
.
ceil_mode
).
astype
(
"float32"
)
self
.
inputs
=
{
'X'
:
input
}
self
.
ceil_mode
).
astype
(
self
.
dtype
)
self
.
inputs
=
{
'X'
:
OpTest
.
np_dtype_to_fluid_dtype
(
input
)
}
self
.
attrs
=
{
'strides'
:
self
.
strides
,
...
...
@@ -105,7 +106,7 @@ class TestPool2d_Op(OpTest):
'data_format'
:
'AnyLayout'
# TODO(dzhwinter) : should be fix latter
}
self
.
outputs
=
{
'Out'
:
output
.
astype
(
'float32'
)
}
self
.
outputs
=
{
'Out'
:
output
}
def
test_check_output
(
self
):
if
self
.
use_cudnn
:
...
...
@@ -115,6 +116,8 @@ class TestPool2d_Op(OpTest):
self
.
check_output
()
def
test_check_grad
(
self
):
if
self
.
dtype
==
np
.
float16
:
return
if
self
.
use_cudnn
and
self
.
pool_type
!=
"max"
:
place
=
core
.
CUDAPlace
(
0
)
self
.
check_grad_with_place
(
...
...
@@ -212,36 +215,114 @@ class TestCUDNNCase1(TestPool2d_Op):
self
.
op_type
=
"pool2d"
class
TestFP16CUDNNCase1
(
TestPool2d_Op
):
def
init_op_type
(
self
):
self
.
use_cudnn
=
True
self
.
op_type
=
"pool2d"
self
.
dtype
=
np
.
float16
def
test_check_output
(
self
):
if
core
.
is_compiled_with_cuda
():
place
=
core
.
CUDAPlace
(
0
)
if
core
.
is_float16_supported
(
place
):
self
.
check_output_with_place
(
place
,
atol
=
1e-3
)
class
TestCUDNNCase2
(
TestCase1
):
def
init_op_type
(
self
):
self
.
use_cudnn
=
True
self
.
op_type
=
"pool2d"
class
TestFP16CUDNNCase2
(
TestCase1
):
def
init_op_type
(
self
):
self
.
use_cudnn
=
True
self
.
op_type
=
"pool2d"
self
.
dtype
=
np
.
float16
def
test_check_output
(
self
):
if
core
.
is_compiled_with_cuda
():
place
=
core
.
CUDAPlace
(
0
)
if
core
.
is_float16_supported
(
place
):
self
.
check_output_with_place
(
place
,
atol
=
1e-3
)
class
TestCUDNNCase3
(
TestCase2
):
def
init_op_type
(
self
):
self
.
use_cudnn
=
True
self
.
op_type
=
"pool2d"
class
TestFP16CUDNNCase3
(
TestCase2
):
def
init_op_type
(
self
):
self
.
use_cudnn
=
True
self
.
op_type
=
"pool2d"
self
.
dtype
=
np
.
float16
def
test_check_output
(
self
):
if
core
.
is_compiled_with_cuda
():
place
=
core
.
CUDAPlace
(
0
)
if
core
.
is_float16_supported
(
place
):
self
.
check_output_with_place
(
place
,
atol
=
1e-3
)
class
TestCUDNNCase4
(
TestCase3
):
def
init_op_type
(
self
):
self
.
use_cudnn
=
True
self
.
op_type
=
"pool2d"
class
TestFP16CUDNNCase4
(
TestCase3
):
def
init_op_type
(
self
):
self
.
use_cudnn
=
True
self
.
op_type
=
"pool2d"
self
.
dtype
=
np
.
float16
def
test_check_output
(
self
):
if
core
.
is_compiled_with_cuda
():
place
=
core
.
CUDAPlace
(
0
)
if
core
.
is_float16_supported
(
place
):
self
.
check_output_with_place
(
place
,
atol
=
1e-3
)
class
TestCUDNNCase5
(
TestCase4
):
def
init_op_type
(
self
):
self
.
use_cudnn
=
True
self
.
op_type
=
"pool2d"
class
TestFP16CUDNNCase5
(
TestCase4
):
def
init_op_type
(
self
):
self
.
use_cudnn
=
True
self
.
op_type
=
"pool2d"
self
.
dtype
=
np
.
float16
def
test_check_output
(
self
):
if
core
.
is_compiled_with_cuda
():
place
=
core
.
CUDAPlace
(
0
)
if
core
.
is_float16_supported
(
place
):
self
.
check_output_with_place
(
place
,
atol
=
1e-3
)
class
TestCUDNNCase6
(
TestCase5
):
def
init_op_type
(
self
):
self
.
use_cudnn
=
True
self
.
op_type
=
"pool2d"
class
TestFP16CUDNNCase6
(
TestCase5
):
def
init_op_type
(
self
):
self
.
use_cudnn
=
True
self
.
op_type
=
"pool2d"
self
.
dtype
=
np
.
float16
def
test_check_output
(
self
):
if
core
.
is_compiled_with_cuda
():
place
=
core
.
CUDAPlace
(
0
)
if
core
.
is_float16_supported
(
place
):
self
.
check_output_with_place
(
place
,
atol
=
1e-3
)
class
TestCeilModeCase1
(
TestCUDNNCase1
):
def
init_ceil_mode
(
self
):
self
.
ceil_mode
=
True
...
...
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