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
cf3a07e8
P
Paddle
项目概览
PaddlePaddle
/
Paddle
12 个月 前同步成功
通知
2292
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
cf3a07e8
编写于
12月 19, 2018
作者:
X
Xin Pan
提交者:
GitHub
12月 19, 2018
浏览文件
操作
浏览文件
下载
差异文件
Merge pull request #14878 from panyx0718/imperative
MLP forward backward
上级
b37fb7a6
3f9c429e
变更
10
隐藏空白更改
内联
并排
Showing
10 changed file
with
166 addition
and
21 deletion
+166
-21
paddle/fluid/imperative/layer.cc
paddle/fluid/imperative/layer.cc
+4
-2
paddle/fluid/imperative/tracer.h
paddle/fluid/imperative/tracer.h
+19
-6
paddle/fluid/operators/mul_op.cc
paddle/fluid/operators/mul_op.cc
+2
-1
paddle/fluid/pybind/imperative.cc
paddle/fluid/pybind/imperative.cc
+3
-2
python/paddle/fluid/backward.py
python/paddle/fluid/backward.py
+5
-2
python/paddle/fluid/framework.py
python/paddle/fluid/framework.py
+3
-0
python/paddle/fluid/imperative/base.py
python/paddle/fluid/imperative/base.py
+2
-1
python/paddle/fluid/imperative/layers.py
python/paddle/fluid/imperative/layers.py
+8
-3
python/paddle/fluid/layers/nn.py
python/paddle/fluid/layers/nn.py
+45
-0
python/paddle/fluid/tests/unittests/test_imperative.py
python/paddle/fluid/tests/unittests/test_imperative.py
+75
-4
未找到文件。
paddle/fluid/imperative/layer.cc
浏览文件 @
cf3a07e8
...
...
@@ -188,11 +188,13 @@ std::vector<Variable*> OpBase::ApplyGrad(framework::Scope* scope) {
std
::
vector
<
Variable
*>
ret
;
for
(
size_t
i
=
0
;
i
<
input_vars_
->
size
();
++
i
)
{
bool
found
=
false
;
VarBase
*
origin_var
=
(
*
input_vars_
)[
i
];
for
(
const
std
::
string
&
outvar
:
grad_op_desc_
->
OutputArgumentNames
())
{
Variable
*
var
=
scope
->
FindVar
(
outvar
);
VarBase
*
origin_var
=
(
*
input_vars_
)[
i
];
std
::
string
orig_var
=
grad_to_var_
->
at
(
outvar
);
PADDLE_ENFORCE
(
origin_var
->
var_desc_
->
Name
()
==
orig_var
);
if
(
origin_var
->
var_desc_
->
Name
()
!=
orig_var
)
{
continue
;
}
VLOG
(
3
)
<<
"apply grad "
<<
outvar
<<
" with origin "
<<
orig_var
;
origin_var
->
ApplyGrad
(
scope
,
var
);
found
=
true
;
...
...
paddle/fluid/imperative/tracer.h
浏览文件 @
cf3a07e8
...
...
@@ -43,9 +43,12 @@ void CreateGradOp(const framework::OpDesc& op_desc,
class
Tracer
{
public:
explicit
Tracer
(
framework
::
BlockDesc
*
root_block
)
:
root_block_
(
root_block
)
{
explicit
Tracer
(
framework
::
BlockDesc
*
root_block
,
framework
::
BlockDesc
*
startup_block
)
:
root_block_
(
root_block
),
startup_block_
(
startup_block
)
{
root_scope_
=
new
framework
::
Scope
();
scopes_
[
root_block_
]
=
root_scope_
;
scopes_
[
startup_block_
]
=
root_scope_
;
}
virtual
~
Tracer
()
{
delete
root_scope_
;
}
...
...
@@ -80,6 +83,8 @@ class Tracer {
}
else
{
op
->
pre_ops_
->
push_back
(
nullptr
);
}
VLOG
(
3
)
<<
"input vname "
<<
vname
<<
" "
<<
var
->
Get
<
framework
::
LoDTensor
>
().
dims
().
size
();
}
*
op
->
output_vars_
=
outputs
;
...
...
@@ -98,12 +103,19 @@ class Tracer {
outputs
[
i
]
->
pre_op_
=
op
;
outputs
[
i
]
->
pre_op_out_idx_
=
i
;
}
VLOG
(
3
)
<<
"tracer running "
<<
op_desc
->
Type
();
op_base
->
Run
(
*
scope
,
platform
::
CPUPlace
());
framework
::
OpDesc
*
grad_op_desc
;
auto
grad_to_var
=
new
std
::
unordered_map
<
std
::
string
,
std
::
string
>
();
CreateGradOp
(
*
op_desc
,
{},
{
block
},
&
grad_op_desc
,
grad_to_var
);
op
->
grad_op_desc_
=
grad_op_desc
;
op
->
grad_to_var_
=
grad_to_var
;
if
(
block
==
startup_block_
)
{
op
->
grad_op_desc_
=
nullptr
;
op
->
grad_to_var_
=
nullptr
;
}
else
{
framework
::
OpDesc
*
grad_op_desc
;
auto
grad_to_var
=
new
std
::
unordered_map
<
std
::
string
,
std
::
string
>
();
CreateGradOp
(
*
op_desc
,
{},
{
block
},
&
grad_op_desc
,
grad_to_var
);
op
->
grad_op_desc_
=
grad_op_desc
;
op
->
grad_to_var_
=
grad_to_var
;
}
op
->
block_
=
block
;
}
...
...
@@ -121,6 +133,7 @@ class Tracer {
private:
std
::
map
<
framework
::
BlockDesc
*
,
framework
::
Scope
*>
scopes_
;
framework
::
BlockDesc
*
root_block_
;
framework
::
BlockDesc
*
startup_block_
;
framework
::
Scope
*
root_scope_
;
};
...
...
paddle/fluid/operators/mul_op.cc
浏览文件 @
cf3a07e8
...
...
@@ -49,7 +49,8 @@ class MulOp : public framework::OperatorWithKernel {
PADDLE_ENFORCE_GT
(
y_dims
.
size
(),
y_num_col_dims
,
"The input tensor Y's rank of MulOp should be larger than "
"y_num_col_dims."
);
"y_num_col_dims: %ld vs %ld"
,
y_dims
.
size
(),
y_num_col_dims
);
auto
x_mat_dims
=
framework
::
flatten_to_2d
(
x_dims
,
x_num_col_dims
);
auto
y_mat_dims
=
framework
::
flatten_to_2d
(
y_dims
,
y_num_col_dims
);
...
...
paddle/fluid/pybind/imperative.cc
浏览文件 @
cf3a07e8
...
...
@@ -24,8 +24,9 @@ namespace pybind {
void
BindTracer
(
pybind11
::
module
*
m
)
{
pybind11
::
class_
<
imperative
::
Tracer
>
(
*
m
,
"Tracer"
,
""
)
.
def
(
"__init__"
,
[](
imperative
::
Tracer
&
self
,
framework
::
BlockDesc
*
root_block
)
{
new
(
&
self
)
imperative
::
Tracer
(
root_block
);
[](
imperative
::
Tracer
&
self
,
framework
::
BlockDesc
*
root_block
,
framework
::
BlockDesc
*
startup_block
)
{
new
(
&
self
)
imperative
::
Tracer
(
root_block
,
startup_block
);
})
.
def
(
"trace"
,
&
imperative
::
Tracer
::
Trace
)
.
def
(
"get_scope"
,
&
imperative
::
Tracer
::
GetScope
,
...
...
python/paddle/fluid/backward.py
浏览文件 @
cf3a07e8
...
...
@@ -489,8 +489,11 @@ def append_backward(loss, parameter_list=None, no_grad_set=None,
grad_to_var
=
dict
()
op_desc
=
_create_op_desc_
(
"fill_constant"
,
{},
{
"Out"
:
[
_append_grad_suffix_
(
loss
.
name
)]},
{
"shape"
:
[
1
],
"fill_constant"
,
{},
{
"Out"
:
[
_append_grad_suffix_
(
loss
.
name
)]},
{
"shape"
:
[
1
],
# TODO(panyx0718): This can be loss.shape.
"value"
:
1.0
,
"dtype"
:
loss
.
dtype
,
"force_cpu"
:
False
,
...
...
python/paddle/fluid/framework.py
浏览文件 @
cf3a07e8
...
...
@@ -1324,6 +1324,9 @@ class Block(object):
def
_prepend_op
(
self
,
*
args
,
**
kwargs
):
op_desc
=
self
.
desc
.
_prepend_op
()
op
=
Operator
(
self
,
op_desc
,
*
args
,
**
kwargs
)
if
_in_imperative_mode
():
_imperative_tracer
().
trace
(
op
.
iop
,
[
v
.
_ivar
for
v
in
op
.
inputs
],
[
v
.
_ivar
for
v
in
op
.
outputs
],
self
.
desc
)
self
.
ops
.
insert
(
0
,
op
)
return
op
...
...
python/paddle/fluid/imperative/base.py
浏览文件 @
cf3a07e8
...
...
@@ -28,7 +28,8 @@ def enabled():
def
guard
():
train
=
framework
.
Program
()
startup
=
framework
.
Program
()
tracer
=
core
.
Tracer
(
train
.
current_block
().
desc
)
tracer
=
core
.
Tracer
(
train
.
current_block
().
desc
,
startup
.
current_block
().
desc
)
with
framework
.
program_guard
(
train
,
startup
):
with
framework
.
unique_name
.
guard
():
with
framework
.
_imperative_guard
(
tracer
):
...
...
python/paddle/fluid/imperative/layers.py
浏览文件 @
cf3a07e8
...
...
@@ -25,11 +25,9 @@ __all__ = ['PyLayer']
class
PyLayer
(
core
.
Layer
):
def
__init__
(
self
):
pass
self
.
_built
=
False
def
__call__
(
self
,
inputs
):
# TODO(panyx0718): Support declarative mode as well.
assert
base
.
enabled
()
if
not
isinstance
(
inputs
,
list
)
and
not
isinstance
(
inputs
,
tuple
):
inputs
=
[
inputs
]
...
...
@@ -37,8 +35,15 @@ class PyLayer(core.Layer):
for
x
in
inputs
:
py_var
=
base
.
to_variable
(
x
)
var_inputs
.
append
(
py_var
)
if
not
self
.
_built
:
self
.
_build_once
(
inputs
)
self
.
_built
=
True
outputs
=
self
.
forward
(
var_inputs
)
return
outputs
def
_build_once
(
self
,
inputs
):
pass
def
forward
(
self
,
inputs
):
return
[]
python/paddle/fluid/layers/nn.py
浏览文件 @
cf3a07e8
...
...
@@ -29,6 +29,7 @@ from . import utils
from
..
import
unique_name
from
functools
import
reduce
from
..
import
core
from
..imperative
import
layers
__all__
=
[
'fc'
,
...
...
@@ -9426,3 +9427,47 @@ def huber_loss(input, label, delta):
'Residual'
:
residual
},
attrs
=
{
'delta'
:
delta
})
return
out
class
FC
(
layers
.
PyLayer
):
def
__init__
(
self
,
size
,
param_attr
=
None
,
num_flatten_dims
=
1
,
dtype
=
core
.
VarDesc
.
VarType
.
FP32
):
super
(
FC
,
self
).
__init__
()
self
.
_size
=
size
self
.
_num_flatten_dims
=
num_flatten_dims
self
.
_dtype
=
dtype
self
.
_helper
=
LayerHelper
(
'FC'
,
param_attr
=
param_attr
)
def
_build_once
(
self
,
inputs
):
input_shape
=
inputs
[
0
].
shape
param_shape
=
[
reduce
(
lambda
a
,
b
:
a
*
b
,
input_shape
[
self
.
_num_flatten_dims
:],
1
)
]
+
[
self
.
_size
]
self
.
_w
=
self
.
_helper
.
create_parameter
(
attr
=
self
.
_helper
.
param_attr
,
shape
=
param_shape
,
dtype
=
self
.
_dtype
,
is_bias
=
False
)
def
forward
(
self
,
inputs
):
tmp
=
self
.
_helper
.
create_variable_for_type_inference
(
self
.
_dtype
)
self
.
_helper
.
append_op
(
type
=
"mul"
,
inputs
=
{
"X"
:
inputs
[
0
],
"Y"
:
self
.
_w
},
outputs
=
{
"Out"
:
tmp
},
attrs
=
{
"x_num_col_dims"
:
self
.
_num_flatten_dims
,
"y_num_col_dims"
:
1
})
out
=
self
.
_helper
.
create_variable_for_type_inference
(
self
.
_dtype
)
self
.
_helper
.
append_op
(
type
=
"sum"
,
inputs
=
{
"X"
:
[
tmp
]},
outputs
=
{
"Out"
:
out
},
attrs
=
{
"use_mkldnn"
:
False
})
return
out
python/paddle/fluid/tests/unittests/test_imperative.py
浏览文件 @
cf3a07e8
...
...
@@ -12,12 +12,23 @@
# See the License for the specific language governing permissions and
# limitations under the License.
import
contextlib
import
unittest
import
sys
import
numpy
as
np
import
paddle.fluid
as
fluid
from
paddle.fluid
import
core
from
paddle.fluid.layers.nn
import
FC
@
contextlib
.
contextmanager
def
new_program_scope
():
prog
=
fluid
.
Program
()
startup_prog
=
fluid
.
Program
()
scope
=
fluid
.
core
.
Scope
()
with
fluid
.
scope_guard
(
scope
):
with
fluid
.
program_guard
(
prog
,
startup_prog
):
yield
class
MyLayer
(
fluid
.
imperative
.
PyLayer
):
...
...
@@ -30,6 +41,23 @@ class MyLayer(fluid.imperative.PyLayer):
return
[
fluid
.
layers
.
elementwise_mul
(
x
,
x
)]
class
MLP
(
fluid
.
imperative
.
PyLayer
):
def
__init__
(
self
):
super
(
MLP
,
self
).
__init__
()
self
.
_fc1
=
FC
(
3
,
fluid
.
ParamAttr
(
initializer
=
fluid
.
initializer
.
Constant
(
value
=
0.1
)))
self
.
_fc2
=
FC
(
4
,
fluid
.
ParamAttr
(
initializer
=
fluid
.
initializer
.
Constant
(
value
=
0.1
)))
def
forward
(
self
,
inputs
):
x
=
self
.
_fc1
(
inputs
[
0
])
x
=
self
.
_fc2
(
x
)
x
=
fluid
.
layers
.
reduce_sum
(
x
)
return
x
class
TestImperative
(
unittest
.
TestCase
):
def
test_layer
(
self
):
with
fluid
.
imperative
.
guard
():
...
...
@@ -39,13 +67,56 @@ class TestImperative(unittest.TestCase):
l
.
forward
([])
def
test_layer_in_out
(
self
):
np_inp
=
np
.
array
([
1.0
,
2.0
,
-
1.0
],
dtype
=
np
.
float32
)
with
fluid
.
imperative
.
guard
():
l
=
MyLayer
()
x
=
l
(
np
.
array
([
1.0
,
2.0
,
-
1.0
],
dtype
=
np
.
float32
)
)[
0
]
x
=
l
(
np
_inp
)[
0
]
self
.
assertIsNotNone
(
x
)
sys
.
stderr
.
write
(
"%s output: %s
\n
"
%
(
x
,
x
.
_numpy
())
)
dy_out
=
x
.
_numpy
(
)
x
.
_backward
()
sys
.
stderr
.
write
(
"grad %s
\n
"
%
l
.
_x_for_debug
.
_gradient
())
dy_grad
=
l
.
_x_for_debug
.
_gradient
()
with
new_program_scope
():
inp
=
fluid
.
layers
.
data
(
name
=
"inp"
,
shape
=
[
3
],
append_batch_size
=
False
)
l
=
MyLayer
()
x
=
l
(
inp
)[
0
]
param_grads
=
fluid
.
backward
.
append_backward
(
x
,
parameter_list
=
[
l
.
_x_for_debug
.
name
])[
0
]
exe
=
fluid
.
Executor
(
fluid
.
CPUPlace
())
static_out
,
static_grad
=
exe
.
run
(
feed
=
{
inp
.
name
:
np_inp
},
fetch_list
=
[
x
.
name
,
param_grads
[
1
].
name
])
self
.
assertTrue
(
np
.
allclose
(
dy_out
,
static_out
))
self
.
assertTrue
(
np
.
allclose
(
dy_grad
,
static_grad
))
def
test_mlp
(
self
):
np_inp
=
np
.
array
([[
1.0
,
2.0
],
[
3.0
,
4.0
]],
dtype
=
np
.
float32
)
with
fluid
.
imperative
.
guard
():
mlp
=
MLP
()
out
=
mlp
(
np_inp
)
dy_out
=
out
.
_numpy
()
out
.
_backward
()
dy_grad
=
mlp
.
_fc1
.
_w
.
_gradient
()
with
new_program_scope
():
inp
=
fluid
.
layers
.
data
(
name
=
"inp"
,
shape
=
[
2
,
2
],
append_batch_size
=
False
)
mlp
=
MLP
()
out
=
mlp
(
inp
)
param_grads
=
fluid
.
backward
.
append_backward
(
out
,
parameter_list
=
[
mlp
.
_fc1
.
_w
.
name
])[
0
]
exe
=
fluid
.
Executor
(
fluid
.
CPUPlace
())
exe
.
run
(
fluid
.
default_startup_program
())
static_out
,
static_grad
=
exe
.
run
(
feed
=
{
inp
.
name
:
np_inp
},
fetch_list
=
[
out
.
name
,
param_grads
[
1
].
name
])
self
.
assertTrue
(
np
.
allclose
(
dy_out
,
static_out
))
self
.
assertTrue
(
np
.
allclose
(
dy_grad
,
static_grad
))
if
__name__
==
'__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