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
4825addd
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
4825addd
编写于
1月 12, 2022
作者:
S
Sylwester Fraczek
提交者:
GitHub
1月 12, 2022
浏览文件
操作
浏览文件
下载
电子邮件补丁
差异文件
Fix conv act int8 scale (#38331)
* fix conv act int8 scale * add unit test for conv+hard_swish
上级
d296456c
变更
4
隐藏空白更改
内联
并排
Showing
4 changed file
with
90 addition
and
31 deletion
+90
-31
paddle/fluid/framework/ir/mkldnn/fc_act_mkldnn_fuse_pass.h
paddle/fluid/framework/ir/mkldnn/fc_act_mkldnn_fuse_pass.h
+1
-1
paddle/fluid/operators/mkldnn/conv_mkldnn_op.cc
paddle/fluid/operators/mkldnn/conv_mkldnn_op.cc
+33
-21
python/paddle/fluid/contrib/slim/quantization/quant2_int8_mkldnn_pass.py
...luid/contrib/slim/quantization/quant2_int8_mkldnn_pass.py
+1
-0
python/paddle/fluid/tests/unittests/mkldnn/test_conv2d_int8_mkldnn_op.py
...luid/tests/unittests/mkldnn/test_conv2d_int8_mkldnn_op.py
+55
-9
未找到文件。
paddle/fluid/framework/ir/mkldnn/fc_act_mkldnn_fuse_pass.h
浏览文件 @
4825addd
...
...
@@ -42,4 +42,4 @@ class FuseFCActOneDNNPass : public FusePassBase {
}
// namespace ir
}
// namespace framework
}
// namespace paddle
a
}
// namespace paddle
paddle/fluid/operators/mkldnn/conv_mkldnn_op.cc
浏览文件 @
4825addd
...
...
@@ -218,13 +218,15 @@ class ConvMKLDNNHandlerT
:
dnnl
::
prop_kind
::
forward_training
;
float
sum_scale
=
1.0
f
;
float
activation_scale
=
1.0
f
;
std
::
vector
<
float
>
output_shift_scale
;
if
(
platform
::
is_int8
<
T
>
())
std
::
tie
(
sum_scale
,
output_shift_scale
)
=
get_int8_scales
(
ctx
);
std
::
tie
(
sum_scale
,
output_shift_scale
,
activation_scale
)
=
get_int8_scales
(
ctx
);
const
dnnl
::
primitive_attr
conv_attr
=
CreatePostOps
(
fuse_activation
,
fuse_alpha
,
fuse_beta
,
fuse_residual_conn
,
output_shift_scale
,
sum_scale
);
// for INT8 only!
output_shift_scale
,
sum_scale
,
activation_scale
);
// for INT8 only!
if
(
bias
)
{
auto
bias_tz
=
framework
::
vectorize
(
bias
->
dims
());
...
...
@@ -432,7 +434,7 @@ class ConvMKLDNNHandlerT
return
bias_scale_tuple
;
}
std
::
tuple
<
float
,
std
::
vector
<
float
>>
get_int8_scales
(
std
::
tuple
<
float
,
std
::
vector
<
float
>
,
float
>
get_int8_scales
(
const
framework
::
ExecutionContext
&
ctx
)
const
{
const
auto
*
filter
=
ctx
.
Input
<
Tensor
>
(
"Filter"
);
const
auto
&
weights_tz
=
framework
::
vectorize
(
filter
->
dims
());
...
...
@@ -445,8 +447,14 @@ class ConvMKLDNNHandlerT
const
auto
&
scale_in_eltwise_data
=
ctx
.
Attr
<
float
>
(
"Scale_in_eltwise"
);
auto
scale_weights_data
=
ctx
.
Attr
<
std
::
vector
<
float
>>
(
"Scale_weights"
);
bool
is_multi_channel
=
scale_weights_data
.
size
()
>
1
;
bool
has_activation
=
!
ctx
.
Attr
<
std
::
string
>
(
"fuse_activation"
).
empty
();
float
activation_scale
=
force_fp32_output
?
1.0
f
:
has_activation
?
ctx
.
Attr
<
float
>
(
"Scale_out"
)
:
1.0
f
;
auto
scale_out_data
=
force_fp32_output
?
1.0
f
:
ctx
.
Attr
<
float
>
(
"Scale_out"
);
force_fp32_output
?
1.0
f
:
has_activation
?
1.0
f
:
ctx
.
Attr
<
float
>
(
"Scale_out"
);
float
sum_scale
=
fuse_residual_conn
?
scale_out_data
/
scale_in_eltwise_data
:
1.0
f
;
int
count
=
...
...
@@ -468,13 +476,13 @@ class ConvMKLDNNHandlerT
static_cast
<
double
>
(
scale_weights_data
[
i
])));
}
return
std
::
make_tuple
(
sum_scale
,
output_shift_scale
);
return
std
::
make_tuple
(
sum_scale
,
output_shift_scale
,
activation_scale
);
}
dnnl
::
primitive_attr
CreatePostOps
(
std
::
string
fuse_activation
,
float
fuse_alpha
,
float
fuse_beta
,
bool
fuse_residual_conn
,
const
std
::
vector
<
float
>
output_shift_scale
=
{},
float
sum_scale
=
1.0
f
)
{
float
sum_scale
=
1.0
f
,
float
activation_scale
=
1.0
f
)
{
dnnl
::
primitive_attr
conv_attr
;
dnnl
::
post_ops
post_operations
;
if
(
output_shift_scale
.
size
()
>
0
)
{
...
...
@@ -492,30 +500,34 @@ class ConvMKLDNNHandlerT
}
// Fusion with ReLU layer is executed through the PostOps feature. Create a
// PostOps object and configure it to execute an eltwise relu operation.
constexpr
float
scale
=
1.0
f
;
if
(
fuse_activation
==
"relu"
||
fuse_activation
==
"leaky_relu"
)
{
post_operations
.
append_eltwise
(
scale
,
dnnl
::
algorithm
::
eltwise_relu
,
fuse_alpha
,
fuse_beta
);
post_operations
.
append_eltwise
(
activation_scale
,
dnnl
::
algorithm
::
eltwise_relu
,
fuse_alpha
,
fuse_beta
);
}
else
if
(
fuse_activation
==
"relu6"
)
{
post_operations
.
append_eltwise
(
scale
,
dnnl
::
algorithm
::
eltwise_bounded_relu
,
fuse_alpha
,
fuse_beta
);
}
else
if
(
fuse_activation
==
"swish"
)
{
post_operations
.
append_eltwise
(
scale
,
dnnl
::
algorithm
::
eltwise_swish
,
post_operations
.
append_eltwise
(
activation_scale
,
dnnl
::
algorithm
::
eltwise_bounded_relu
,
fuse_alpha
,
fuse_beta
);
}
else
if
(
fuse_activation
==
"swish"
)
{
post_operations
.
append_eltwise
(
activation_scale
,
dnnl
::
algorithm
::
eltwise_swish
,
fuse_alpha
,
fuse_beta
);
}
else
if
(
fuse_activation
==
"hard_swish"
)
{
post_operations
.
append_eltwise
(
scale
,
dnnl
::
algorithm
::
eltwise_hardswish
,
post_operations
.
append_eltwise
(
activation_scale
,
dnnl
::
algorithm
::
eltwise_hardswish
,
fuse_alpha
,
fuse_beta
);
}
else
if
(
fuse_activation
==
"hard_sigmoid"
)
{
post_operations
.
append_eltwise
(
scale
,
dnnl
::
algorithm
::
eltwise_linear
,
post_operations
.
append_eltwise
(
activation_scale
,
dnnl
::
algorithm
::
eltwise_linear
,
fuse_alpha
,
fuse_beta
);
post_operations
.
append_eltwise
(
scale
,
dnnl
::
algorithm
::
eltwise_clip
,
0.0
f
,
1.0
f
);
post_operations
.
append_eltwise
(
activation_scale
,
dnnl
::
algorithm
::
eltwise_clip
,
0.0
f
,
1.0
f
);
}
else
if
(
fuse_activation
==
"gelu_tanh"
)
{
post_operations
.
append_eltwise
(
scale
,
dnnl
::
algorithm
::
eltwise_gelu_tanh
,
0.0
f
,
0.0
f
);
post_operations
.
append_eltwise
(
activation_scale
,
dnnl
::
algorithm
::
eltwise_gelu_tanh
,
0.0
f
,
0.0
f
);
}
else
if
(
fuse_activation
==
"gelu_erf"
)
{
post_operations
.
append_eltwise
(
scale
,
dnnl
::
algorithm
::
eltwise_gelu_erf
,
0.0
f
,
0.0
f
);
post_operations
.
append_eltwise
(
activation_scale
,
dnnl
::
algorithm
::
eltwise_gelu_erf
,
0.0
f
,
0.0
f
);
}
conv_attr
.
set_post_ops
(
post_operations
);
return
conv_attr
;
...
...
python/paddle/fluid/contrib/slim/quantization/quant2_int8_mkldnn_pass.py
浏览文件 @
4825addd
...
...
@@ -426,6 +426,7 @@ class Quant2Int8MkldnnPass(object):
graph
=
self
.
_apply_pass
(
graph
,
'conv_elementwise_add_mkldnn_fuse_pass'
)
graph
=
self
.
_apply_pass
(
graph
,
'conv_relu_mkldnn_fuse_pass'
)
graph
=
self
.
_apply_pass
(
graph
,
'conv_relu6_mkldnn_fuse_pass'
)
graph
=
self
.
_apply_pass
(
graph
,
'conv_hard_swish_mkldnn_fuse_pass'
)
graph
=
self
.
_apply_pass
(
graph
,
'fc_fuse_pass'
,
[
'use_gpu'
,
'use_fc_padding'
],
[
False
,
False
])
graph
=
self
.
_apply_pass
(
graph
,
'repeated_fc_relu_fuse_pass'
)
...
...
python/paddle/fluid/tests/unittests/mkldnn/test_conv2d_int8_mkldnn_op.py
浏览文件 @
4825addd
...
...
@@ -43,7 +43,7 @@ class TestConv2DInt8Op(TestConv2DOp):
self
.
init_group
()
self
.
init_dilation
()
self
.
init_test_case
()
self
.
init_fuse_
relu
()
self
.
init_fuse_
activation
()
self
.
init_fuse_residual
()
self
.
init_data_type
()
...
...
@@ -54,7 +54,9 @@ class TestConv2DInt8Op(TestConv2DOp):
}
# This implementation of convolution quantization is based on OneDNN documentation
# https://oneapi-src.github.io/oneDNN/dev_guide_int8_computations.html#doxid-dev-guide-int8-computations-1dg-i8-comp-s11
scale_output_shift
=
(
self
.
scale_out
/
inner_scale
=
1.
if
self
.
fuse_activation
!=
""
else
self
.
scale_out
activation_scale
=
self
.
scale_out
if
self
.
fuse_activation
!=
""
else
1.
scale_output_shift
=
(
inner_scale
/
(
self
.
scale_in
*
self
.
scale_weights
[
0
]))
filter
=
np
.
random
.
random
(
self
.
filter_size
).
astype
(
self
.
weighttype
)
...
...
@@ -78,7 +80,7 @@ class TestConv2DInt8Op(TestConv2DOp):
init_low
,
init_high
,
self
.
input_residual_size
).
astype
(
self
.
srctype
)
return
(
output_
+
input_residual_
*
(
self
.
scale_out
/
self
.
scale_in_eltwise
)),
input_residual_
(
inner_scale
/
self
.
scale_in_eltwise
)),
input_residual_
if
self
.
srctype
==
np
.
int8
:
init_low
,
init_high
=
(
-
5
,
5
)
...
...
@@ -101,12 +103,24 @@ class TestConv2DInt8Op(TestConv2DOp):
output
,
input_residual
=
residual_helper
(
init_low
,
init_high
,
output
)
output
=
np
.
round
(
output
)
if
self
.
fuse_activation
==
"relu"
:
output
=
np
.
maximum
(
output
,
0
)
if
self
.
fuse_activation
==
""
:
pass
elif
self
.
fuse_activation
==
"relu"
:
output
=
activation_scale
*
np
.
maximum
(
output
,
0
)
elif
self
.
fuse_activation
==
"hard_swish"
:
output
=
activation_scale
*
output
/
6.
*
np
.
minimum
(
np
.
maximum
(
0
,
output
+
3.
),
6
)
elif
self
.
fuse_activation
==
"relu6"
:
output
=
activation_scale
*
np
.
maximum
(
0
,
np
.
minimum
(
6
,
output
))
elif
self
.
fuse_activation
==
"swish"
:
output
=
activation_scale
*
output
/
(
1.
+
np
.
exp
(
-
1.
*
output
))
elif
self
.
fuse_activation
==
"leaky_relu"
:
output
=
activation_scale
*
np
.
maximum
(
output
,
0.02
*
output
)
else
:
raise
NotImplementedError
(
"test for "
+
self
.
fuse_activation
+
" activation not implemented"
)
output
=
output
.
astype
(
self
.
dsttype
)
output
=
np
.
round
(
output
)
.
astype
(
self
.
dsttype
)
self
.
inputs
=
{
'Input'
:
...
...
@@ -131,6 +145,8 @@ class TestConv2DInt8Op(TestConv2DOp):
'Scale_weights'
:
self
.
scale_weights
,
'Scale_in_eltwise'
:
self
.
scale_in_eltwise
,
'fuse_activation'
:
self
.
fuse_activation
,
'fuse_alpha'
:
self
.
fuse_alpha
,
'fuse_beta'
:
self
.
fuse_beta
,
'fuse_residual_connection'
:
self
.
fuse_residual
,
'mkldnn_data_type'
:
self
.
mkldnn_data_type
}
...
...
@@ -165,8 +181,10 @@ class TestConv2DInt8Op(TestConv2DOp):
self
.
srctype
=
np
.
uint8
self
.
dsttype
=
np
.
int8
def
init_fuse_
relu
(
self
):
def
init_fuse_
activation
(
self
):
self
.
fuse_activation
=
"relu"
self
.
fuse_alpha
=
0
self
.
fuse_beta
=
0
def
init_fuse_residual
(
self
):
self
.
fuse_residual
=
True
...
...
@@ -190,6 +208,34 @@ class TestConv2D(TestConv2DInt8Op):
self
.
scale_in_eltwise
=
0.6
class
TestWithHardSwish
(
TestConv2D
):
def
init_fuse_activation
(
self
):
self
.
fuse_activation
=
"hard_swish"
self
.
fuse_alpha
=
0
self
.
fuse_beta
=
0
class
TestWithRelu6
(
TestConv2D
):
def
init_fuse_activation
(
self
):
self
.
fuse_activation
=
"relu6"
self
.
fuse_alpha
=
6
self
.
fuse_beta
=
0
class
TestWithSwish
(
TestConv2D
):
def
init_fuse_activation
(
self
):
self
.
fuse_activation
=
"swish"
self
.
fuse_alpha
=
1
self
.
fuse_beta
=
0
class
TestWithLeakyRelu
(
TestConv2D
):
def
init_fuse_activation
(
self
):
self
.
fuse_activation
=
"leaky_relu"
self
.
fuse_alpha
=
0.02
self
.
fuse_beta
=
0
class
TestWithPad
(
TestConv2D
):
def
init_test_case
(
self
):
TestConv2D
.
init_test_case
(
self
)
...
...
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