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
102fc859
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
未验证
提交
102fc859
编写于
4月 16, 2019
作者:
Q
Qiyang Min
提交者:
GitHub
4月 16, 2019
浏览文件
操作
浏览文件
下载
差异文件
Merge pull request #16777 from velconia/dygraph_untrack_op
Imperative tracer does not hold op any more
上级
f4626ee4
9a3848a2
变更
4
隐藏空白更改
内联
并排
Showing
4 changed file
with
289 addition
and
34 deletion
+289
-34
python/paddle/fluid/dygraph/layers.py
python/paddle/fluid/dygraph/layers.py
+12
-0
python/paddle/fluid/dygraph/tracer.py
python/paddle/fluid/dygraph/tracer.py
+59
-8
python/paddle/fluid/framework.py
python/paddle/fluid/framework.py
+17
-26
python/paddle/fluid/tests/unittests/test_dygraph_multi_forward.py
...addle/fluid/tests/unittests/test_dygraph_multi_forward.py
+201
-0
未找到文件。
python/paddle/fluid/dygraph/layers.py
浏览文件 @
102fc859
...
...
@@ -48,6 +48,12 @@ class Layer(core.Layer):
self
.
_helper
=
LayerObjectHelper
(
self
.
_full_name
)
def
train
(
self
):
framework
.
_dygraph_tracer
().
_train_mode
()
def
eval
(
self
):
framework
.
_dygraph_tracer
().
_eval_mode
()
def
full_name
(
self
):
"""Full name for this layers.
...
...
@@ -254,6 +260,12 @@ class PyLayer(core.PyLayer):
def
__init__
(
self
):
super
(
PyLayer
,
self
).
__init__
()
def
train
(
self
):
framework
.
_dygraph_tracer
().
_train_mode
()
def
eval
(
self
):
framework
.
_dygraph_tracer
().
_eval_mode
()
@
classmethod
def
_do_forward
(
cls
,
inputs
):
return
cls
.
_to_tuple
(
cls
.
forward
(
inputs
))
...
...
python/paddle/fluid/dygraph/tracer.py
浏览文件 @
102fc859
...
...
@@ -24,7 +24,9 @@ __all__ = ['Tracer']
def
release_op
(
op
):
del
framework
.
_dygraph_tracer
().
_ops
[
op
.
_trace_id
]
del
framework
.
_dygraph_tracer
().
_ops
[
op
.
_trace_id
].
inputs
del
framework
.
_dygraph_tracer
().
_ops
[
op
.
_trace_id
].
outputs
del
framework
.
_dygraph_tracer
().
_ops
[
op
.
_trace_id
].
backward_refs
class
Tracer
(
core
.
Tracer
):
...
...
@@ -38,6 +40,7 @@ class Tracer(core.Tracer):
self
.
_ops
=
defaultdict
()
self
.
_vars
=
defaultdict
()
self
.
_trace_id
=
0
self
.
_train_mode
=
True
def
trace_var
(
self
,
name
,
var
):
self
.
_vars
[
name
]
=
var
...
...
@@ -46,15 +49,57 @@ class Tracer(core.Tracer):
return
list
((
item
for
name
,
item
in
six
.
iteritems
(
self
.
_vars
)
if
isinstance
(
item
,
framework
.
Parameter
)))
def
trace_op
(
self
,
op
,
stop_gradient
=
False
):
def
trace_op
(
self
,
op
,
inputs
,
outputs
,
stop_gradient
=
False
):
# TODO(minqiyang): remove this line after we take apart all
# backward grads and forward variables
if
self
.
_train_mode
:
op
.
inputs
=
inputs
inps
=
defaultdict
(
list
)
for
k
,
vars
in
six
.
iteritems
(
inputs
):
if
isinstance
(
vars
,
framework
.
Variable
):
inps
[
k
].
append
(
vars
.
_ivar
)
elif
isinstance
(
vars
,
list
)
or
isinstance
(
vars
,
tuple
):
for
var
in
vars
:
inps
[
k
].
append
(
var
.
_ivar
)
op
.
outputs
=
outputs
outs
=
defaultdict
(
list
)
for
k
,
vars
in
six
.
iteritems
(
outputs
):
if
isinstance
(
vars
,
framework
.
Variable
):
outs
[
k
].
append
(
vars
.
_ivar
)
elif
isinstance
(
vars
,
list
)
or
isinstance
(
vars
,
tuple
):
for
var
in
vars
:
outs
[
k
].
append
(
var
.
_ivar
)
else
:
inps
=
defaultdict
(
list
)
for
k
,
vars
in
six
.
iteritems
(
inputs
):
if
isinstance
(
vars
,
framework
.
Variable
):
op
.
previous_ops
.
append
(
vars
.
op
)
inps
[
k
].
append
(
vars
.
_ivar
)
elif
isinstance
(
vars
,
list
)
or
isinstance
(
vars
,
tuple
):
for
var
in
vars
:
op
.
previous_ops
.
append
(
var
.
op
)
inps
[
k
].
append
(
var
.
_ivar
)
op
.
outputs
=
outputs
outs
=
defaultdict
(
list
)
for
k
,
vars
in
six
.
iteritems
(
outputs
):
if
isinstance
(
vars
,
framework
.
Variable
):
vars
.
op
=
op
outs
[
k
].
append
(
vars
.
_ivar
)
elif
isinstance
(
vars
,
list
)
or
isinstance
(
vars
,
tuple
):
for
var
in
vars
:
var
.
op
=
op
outs
[
k
].
append
(
var
.
_ivar
)
# record op's trace id
op
.
iop
.
_trace_id
=
self
.
_trace_id
backward_refs
=
self
.
trace
(
op
.
iop
,
op
.
inputs
,
op
.
outp
uts
,
op
.
attrs
,
backward_refs
=
self
.
trace
(
op
.
iop
,
inps
,
o
uts
,
op
.
attrs
,
framework
.
_current_expected_place
(),
stop_gradient
)
if
not
stop_gradient
:
if
not
stop_gradient
and
self
.
_train_mode
:
self
.
_trace_id
+=
1
self
.
_ops
[
op
.
iop
.
_trace_id
]
=
op
...
...
@@ -65,10 +110,16 @@ class Tracer(core.Tracer):
# TODO(minqiyang): remove all inputs and outputs after separate
# var and grad
op
.
backward_refs
=
defaultdict
(
list
)
for
k
,
v
in
six
.
iteritems
(
op
.
inputs
):
for
k
,
v
in
six
.
iteritems
(
inputs
):
if
k
in
backward_refs
:
op
.
backward_refs
[
k
]
=
op
.
inputs
[
k
]
op
.
backward_refs
[
k
]
=
inputs
[
k
]
for
k
,
v
in
six
.
iteritems
(
o
p
.
o
utputs
):
for
k
,
v
in
six
.
iteritems
(
outputs
):
if
k
in
backward_refs
:
op
.
backward_refs
[
k
]
=
op
.
outputs
[
k
]
op
.
backward_refs
[
k
]
=
outputs
[
k
]
def
_train_mode
(
self
):
self
.
_train_mode
=
True
def
_eval_mode
(
self
):
self
.
_train_mode
=
False
python/paddle/fluid/framework.py
浏览文件 @
102fc859
...
...
@@ -411,6 +411,7 @@ class Variable(object):
if
persistable
else
False
)
if
persistable
:
_dygraph_tracer
().
trace_var
(
name
,
self
)
self
.
op
=
None
else
:
self
.
error_clip
=
error_clip
...
...
@@ -939,24 +940,7 @@ class Operator(object):
raise
ValueError
(
"`type` to initialized an Operator can not be None."
)
self
.
iop
=
core
.
OpBase
(
type
)
# TODO(minqiyang): remove these lines after we take apart all
# backward grads and forward variables
self
.
inputs
=
defaultdict
(
list
)
if
inputs
is
not
None
:
for
k
,
v
in
six
.
iteritems
(
inputs
):
if
isinstance
(
v
,
Variable
):
self
.
inputs
[
k
].
append
(
v
.
_ivar
)
elif
isinstance
(
v
,
list
)
or
isinstance
(
v
,
tuple
):
self
.
inputs
[
k
].
extend
([
var
.
_ivar
for
var
in
v
])
self
.
outputs
=
defaultdict
(
list
)
if
outputs
is
not
None
:
for
k
,
v
in
six
.
iteritems
(
outputs
):
if
isinstance
(
v
,
Variable
):
self
.
outputs
[
k
].
append
(
v
.
_ivar
)
elif
isinstance
(
v
,
list
)
or
isinstance
(
v
,
tuple
):
self
.
outputs
[
k
].
extend
([
var
.
_ivar
for
var
in
v
])
self
.
previous_ops
=
[]
self
.
attrs
=
attrs
if
attrs
else
{}
else
:
...
...
@@ -1647,15 +1631,18 @@ class Block(object):
block
=
self
,
desc
=
None
,
type
=
kwargs
.
get
(
"type"
,
None
),
inputs
=
kwargs
.
get
(
"inputs"
,
None
)
,
outputs
=
kwargs
.
get
(
"outputs"
,
None
)
,
attrs
=
kwargs
.
get
(
"attrs"
,
None
))
inputs
=
None
,
outputs
=
None
,
attrs
=
kwargs
.
get
(
"attrs"
,
{}
))
# record ops in tracer rather than blocks
#
# TODO(minqiyang): add op stop_gradient support in static mode too.
# currently, we only support stop_gradient in dygraph mode.
_dygraph_tracer
().
trace_op
(
op
,
kwargs
.
get
(
"stop_gradient"
,
False
))
_dygraph_tracer
().
trace_op
(
op
,
kwargs
.
get
(
"inputs"
,
{}),
kwargs
.
get
(
"outputs"
,
{}),
kwargs
.
get
(
"stop_gradient"
,
False
))
else
:
op_desc
=
self
.
desc
.
append_op
()
op
=
Operator
(
...
...
@@ -1719,10 +1706,14 @@ class Block(object):
self
,
None
,
type
=
kwargs
.
get
(
"type"
,
None
),
inputs
=
kwargs
.
get
(
"inputs"
,
None
),
outputs
=
kwargs
.
get
(
"outputs"
,
None
),
attrs
=
kwargs
.
get
(
"attrs"
,
None
))
_dygraph_tracer
().
trace_op
(
op
,
kwargs
.
get
(
"stop_gradient"
,
False
))
inputs
=
None
,
outputs
=
None
,
attrs
=
kwargs
.
get
(
"attrs"
,
{}))
_dygraph_tracer
().
trace_op
(
op
,
kwargs
.
get
(
"inputs"
,
{}),
kwargs
.
get
(
"outputs"
,
{}),
kwargs
.
get
(
"stop_gradient"
,
False
))
else
:
op_desc
=
self
.
desc
.
_prepend_op
()
op
=
Operator
(
...
...
python/paddle/fluid/tests/unittests/test_dygraph_multi_forward.py
0 → 100644
浏览文件 @
102fc859
# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
from
__future__
import
print_function
import
contextlib
import
unittest
import
numpy
as
np
import
six
import
paddle
import
paddle.fluid
as
fluid
from
paddle.fluid
import
core
from
paddle.fluid.optimizer
import
SGDOptimizer
from
paddle.fluid.dygraph.nn
import
Conv2D
,
Pool2D
,
FC
from
paddle.fluid.dygraph.base
import
to_variable
from
test_imperative_base
import
new_program_scope
class
SimpleImgConvPool
(
fluid
.
dygraph
.
Layer
):
def
__init__
(
self
,
name_scope
,
num_channels
,
num_filters
,
filter_size
,
pool_size
,
pool_stride
,
pool_padding
=
0
,
pool_type
=
'max'
,
global_pooling
=
False
,
conv_stride
=
1
,
conv_padding
=
0
,
conv_dilation
=
1
,
conv_groups
=
1
,
act
=
None
,
use_cudnn
=
False
,
param_attr
=
None
,
bias_attr
=
None
):
super
(
SimpleImgConvPool
,
self
).
__init__
(
name_scope
)
self
.
_conv2d
=
Conv2D
(
self
.
full_name
(),
num_channels
=
num_channels
,
num_filters
=
num_filters
,
filter_size
=
filter_size
,
stride
=
conv_stride
,
padding
=
conv_padding
,
dilation
=
conv_dilation
,
groups
=
conv_groups
,
param_attr
=
None
,
bias_attr
=
None
,
use_cudnn
=
use_cudnn
)
self
.
_pool2d
=
Pool2D
(
self
.
full_name
(),
pool_size
=
pool_size
,
pool_type
=
pool_type
,
pool_stride
=
pool_stride
,
pool_padding
=
pool_padding
,
global_pooling
=
global_pooling
,
use_cudnn
=
use_cudnn
)
def
forward
(
self
,
inputs
):
x
=
self
.
_conv2d
(
inputs
)
x
=
self
.
_pool2d
(
x
)
return
x
class
MNIST
(
fluid
.
dygraph
.
Layer
):
def
__init__
(
self
,
name_scope
):
super
(
MNIST
,
self
).
__init__
(
name_scope
)
self
.
_simple_img_conv_pool_1
=
SimpleImgConvPool
(
self
.
full_name
(),
1
,
20
,
5
,
2
,
2
,
act
=
"relu"
)
self
.
_simple_img_conv_pool_2
=
SimpleImgConvPool
(
self
.
full_name
(),
20
,
50
,
5
,
2
,
2
,
act
=
"relu"
)
pool_2_shape
=
50
*
4
*
4
SIZE
=
10
scale
=
(
2.0
/
(
pool_2_shape
**
2
*
SIZE
))
**
0.5
self
.
_fc
=
FC
(
self
.
full_name
(),
10
,
param_attr
=
fluid
.
param_attr
.
ParamAttr
(
initializer
=
fluid
.
initializer
.
NormalInitializer
(
loc
=
0.0
,
scale
=
scale
)),
act
=
"softmax"
)
def
forward
(
self
,
inputs
):
x
=
self
.
_simple_img_conv_pool_1
(
inputs
)
x
=
self
.
_simple_img_conv_pool_2
(
x
)
x
=
self
.
_fc
(
x
)
return
x
class
TestDygraphMultiForward
(
unittest
.
TestCase
):
def
test_mnist_forward_float32
(
self
):
seed
=
90
epoch_num
=
1
with
fluid
.
dygraph
.
guard
():
fluid
.
default_startup_program
().
random_seed
=
seed
fluid
.
default_main_program
().
random_seed
=
seed
mnist
=
MNIST
(
"mnist"
)
sgd
=
SGDOptimizer
(
learning_rate
=
1e-3
)
train_reader
=
paddle
.
batch
(
paddle
.
dataset
.
mnist
.
train
(),
batch_size
=
128
,
drop_last
=
True
)
dy_param_init_value
=
{}
mnist
.
eval
()
for
epoch
in
range
(
epoch_num
):
for
batch_id
,
data
in
enumerate
(
train_reader
()):
dy_x_data
=
np
.
array
(
[
x
[
0
].
reshape
(
1
,
28
,
28
)
for
x
in
data
]).
astype
(
'float32'
)
y_data
=
np
.
array
(
[
x
[
1
]
for
x
in
data
]).
astype
(
'int64'
).
reshape
(
128
,
1
)
img
=
to_variable
(
dy_x_data
)
label
=
to_variable
(
y_data
)
label
.
stop_gradient
=
True
cost
=
mnist
(
img
)
loss
=
fluid
.
layers
.
cross_entropy
(
cost
,
label
)
avg_loss
=
fluid
.
layers
.
mean
(
loss
)
dy_out
=
avg_loss
.
numpy
()
if
epoch
==
0
and
batch_id
==
0
:
for
param
in
mnist
.
parameters
():
dy_param_init_value
[
param
.
name
]
=
param
.
numpy
()
with
new_program_scope
():
fluid
.
default_startup_program
().
random_seed
=
seed
fluid
.
default_main_program
().
random_seed
=
seed
exe
=
fluid
.
Executor
(
fluid
.
CPUPlace
(
)
if
not
core
.
is_compiled_with_cuda
()
else
fluid
.
CUDAPlace
(
0
))
mnist
=
MNIST
(
"mnist"
)
sgd
=
SGDOptimizer
(
learning_rate
=
1e-3
)
train_reader
=
paddle
.
batch
(
paddle
.
dataset
.
mnist
.
train
(),
batch_size
=
128
,
drop_last
=
True
)
img
=
fluid
.
layers
.
data
(
name
=
'pixel'
,
shape
=
[
1
,
28
,
28
],
dtype
=
'float32'
)
label
=
fluid
.
layers
.
data
(
name
=
'label'
,
shape
=
[
1
],
dtype
=
'int64'
)
cost
=
mnist
(
img
)
loss
=
fluid
.
layers
.
cross_entropy
(
cost
,
label
)
avg_loss
=
fluid
.
layers
.
mean
(
loss
)
# initialize params and fetch them
static_param_init_value
=
{}
static_param_name_list
=
[]
for
param
in
mnist
.
parameters
():
static_param_name_list
.
append
(
param
.
name
)
out
=
exe
.
run
(
fluid
.
default_startup_program
(),
fetch_list
=
static_param_name_list
)
for
i
in
range
(
len
(
static_param_name_list
)):
static_param_init_value
[
static_param_name_list
[
i
]]
=
out
[
i
]
for
epoch
in
range
(
epoch_num
):
for
batch_id
,
data
in
enumerate
(
train_reader
()):
static_x_data
=
np
.
array
(
[
x
[
0
].
reshape
(
1
,
28
,
28
)
for
x
in
data
]).
astype
(
'float32'
)
y_data
=
np
.
array
(
[
x
[
1
]
for
x
in
data
]).
astype
(
'int64'
).
reshape
([
128
,
1
])
fetch_list
=
[
avg_loss
.
name
]
out
=
exe
.
run
(
fluid
.
default_main_program
(),
feed
=
{
"pixel"
:
static_x_data
,
"label"
:
y_data
},
fetch_list
=
fetch_list
)
static_out
=
out
[
0
]
self
.
assertTrue
(
np
.
allclose
(
dy_x_data
.
all
(),
static_x_data
.
all
()))
for
key
,
value
in
six
.
iteritems
(
static_param_init_value
):
self
.
assertTrue
(
np
.
allclose
(
value
,
dy_param_init_value
[
key
]))
self
.
assertTrue
(
np
.
allclose
(
static_out
,
dy_out
))
if
__name__
==
'__main__'
:
unittest
.
main
()
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