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
f45818e7
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 >>
提交
f45818e7
编写于
4月 13, 2018
作者:
L
Luo Tao
浏览文件
操作
浏览文件
下载
电子邮件补丁
差异文件
create new varible in scope
上级
6e735e1e
变更
3
隐藏空白更改
内联
并排
Showing
3 changed file
with
41 addition
and
17 deletion
+41
-17
python/paddle/fluid/__init__.py
python/paddle/fluid/__init__.py
+1
-0
python/paddle/fluid/inference_transpiler.py
python/paddle/fluid/inference_transpiler.py
+35
-13
python/paddle/fluid/tests/book/test_image_classification.py
python/paddle/fluid/tests/book/test_image_classification.py
+5
-4
未找到文件。
python/paddle/fluid/__init__.py
浏览文件 @
f45818e7
...
...
@@ -67,6 +67,7 @@ __all__ = framework.__all__ + executor.__all__ + concurrency.__all__ + [
'clip'
,
'SimpleDistributeTranspiler'
,
'DistributeTranspiler'
,
'InferenceTranspiler'
,
'memory_optimize'
,
'release_memory'
,
'profiler'
,
...
...
python/paddle/fluid/inference_transpiler.py
浏览文件 @
f45818e7
...
...
@@ -21,7 +21,20 @@ from . import core
class
InferenceTranspiler
:
def
transpile
(
self
,
program
,
scope
,
place
):
'''
Transpile the program to a inference program by fused batch normalization.
Transpile the program. Support only fuse batch normalization now.
:param program: program to transpile
:type program: Program
:param scope: inference scope
:type scope: Scope
:param place: inference place
:type place: Place
'''
self
.
fuse_batch_norm
(
program
,
scope
,
place
)
def
fuse_batch_norm
(
self
,
program
,
scope
,
place
):
'''
Transpile the program by fused batch normalization.
The batch normalization followed the convolution or fully connected layer
can be integrated with them. Doing so will give us a forward acceleration,
...
...
@@ -57,8 +70,6 @@ class InferenceTranspiler:
:type scope: Scope
:param place: inference place
:type place: Place
:return: program by fused batch normalization
:rtype: Program
'''
self
.
scope
=
scope
self
.
place
=
place
...
...
@@ -96,7 +107,7 @@ class InferenceTranspiler:
# TODO(luotao): use clone() method to flush the program.desc in force,
# since some large program.desc will not be flushed immediately.
# And a better solution will be considered later.
return
program
.
clone
()
program
=
program
.
clone
()
# ====================== private transpiler functions =====================
def
_insert_bias_op
(
self
,
index
,
current_op
,
bn_op
):
...
...
@@ -142,11 +153,25 @@ class InferenceTranspiler:
:type with_bias: Int
'''
def
_load_tensor
(
param_name
):
return
self
.
scope
.
find_var
(
param_name
[
0
]).
get_tensor
()
def
_update_param
(
op
,
old_param_name
,
new_param
):
# For the sake of remaining the original variables the same as before,
# create new variables in scope to store the new parameters.
old_param_name
=
old_param_name
[
0
]
old_var
=
self
.
block
.
vars
[
old_param_name
]
new_param_name
=
old_param_name
+
'_fuse_bn'
new_var
=
self
.
block
.
create_parameter
(
name
=
new_param_name
.
encode
(
'ascii'
),
type
=
old_var
.
type
,
dtype
=
old_var
.
dtype
,
shape
=
old_var
.
shape
)
op
.
rename_input
(
old_param_name
,
new_param_name
)
self
.
scope
.
var
(
new_param_name
)
tensor
=
self
.
scope
.
find_var
(
new_param_name
).
get_tensor
()
tensor
.
set
(
np
.
array
(
new_param
),
self
.
place
)
def
_load_param
(
param_name
):
return
np
.
array
(
_load_tensor
(
param_name
))
return
np
.
array
(
self
.
scope
.
find_var
(
param_name
[
0
]).
get_tensor
(
))
bias_bn
=
_load_param
(
bn_op
.
input
(
"Bias"
))
#Bias
scale_bn
=
_load_param
(
bn_op
.
input
(
"Scale"
))
#Scale
...
...
@@ -155,8 +180,6 @@ class InferenceTranspiler:
# TODO(luotao1): consider only conv2d now. fc would be delt later.
current_param
=
_load_param
(
current_op
.
input
(
"Filter"
))
current_tensor
=
_load_tensor
(
current_op
.
input
(
"Filter"
))
std_bn
=
np
.
float32
(
np
.
sqrt
(
np
.
add
(
var_bn
,
1e-5
)))
tmp
=
np
.
float32
(
np
.
divide
(
scale_bn
,
std_bn
))
...
...
@@ -167,8 +190,6 @@ class InferenceTranspiler:
bias
=
np
.
zeros
(
bias_bn
.
shape
)
bias
=
np
.
float32
(
np
.
add
(
np
.
multiply
(
np
.
subtract
(
bias
,
mean_bn
),
tmp
),
bias_bn
))
bias_tensor
=
_load_tensor
(
bias_op
.
input
(
"Y"
))
bias_tensor
.
set
(
bias
,
self
.
place
)
# re-compute weight of conv2d
tmp
=
tmp
.
reshape
(
tmp
.
shape
[
0
],
-
1
)
...
...
@@ -176,8 +197,9 @@ class InferenceTranspiler:
dst_param
=
np
.
float32
(
np
.
multiply
(
dst_param
,
tmp
))
dst_param
=
dst_param
.
reshape
(
current_param
.
shape
)
# set the updated parameters
current_tensor
.
set
(
np
.
array
(
dst_param
),
self
.
place
)
# update parameters
_update_param
(
current_op
,
current_op
.
input
(
"Filter"
),
dst_param
)
_update_param
(
bias_op
,
bias_op
.
input
(
"Y"
),
bias
)
# collect the renamed input
self
.
input_map
[
bn_op
.
output
(
"Y"
)[
0
]]
=
bias_op
.
output
(
"Out"
)[
0
]
...
...
python/paddle/fluid/tests/book/test_image_classification.py
浏览文件 @
f45818e7
...
...
@@ -226,16 +226,17 @@ def infer(use_cuda, save_dirname=None):
batch_size
=
1
tensor_img
=
numpy
.
random
.
rand
(
batch_size
,
3
,
32
,
32
).
astype
(
"float32"
)
# Use inference_transpiler to speedup
inference_transpiler_program
=
inference_program
.
clone
()
t
=
fluid
.
InferenceTranspiler
()
t
.
transpile
(
inference_transpiler_program
,
inference_scope
,
place
)
# Construct feed as a dictionary of {feed_target_name: feed_target_data}
# and results will contain a list of data corresponding to fetch_targets.
results
=
exe
.
run
(
inference_program
,
feed
=
{
feed_target_names
[
0
]:
tensor_img
},
fetch_list
=
fetch_targets
)
# Use inference_transpiler to speedup
t
=
fluid
.
InferenceTranspiler
()
inference_transpiler_program
=
t
.
transpile
(
inference_program
,
inference_scope
,
place
)
transpiler_results
=
exe
.
run
(
inference_transpiler_program
,
feed
=
{
feed_target_names
[
0
]:
tensor_img
},
fetch_list
=
fetch_targets
)
...
...
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