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
OPTHREE
Paddle
提交
01dc15ce
P
Paddle
项目概览
OPTHREE
/
Paddle
与 Fork 源项目一致
Fork自
PaddlePaddle / Paddle
通知
1
Star
1
Fork
0
代码
文件
提交
分支
Tags
贡献者
分支图
Diff
Issue
1
列表
看板
标记
里程碑
合并请求
0
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
Issue
1
列表
看板
标记
里程碑
合并请求
0
合并请求
0
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
01dc15ce
编写于
1月 16, 2019
作者:
X
Xin Pan
提交者:
GitHub
1月 16, 2019
浏览文件
操作
浏览文件
下载
差异文件
Merge pull request #15329 from panyx0718/imperative2
add imperative mode design
上级
16cb3ebd
6b762f65
变更
1
隐藏空白更改
内联
并排
Showing
1 changed file
with
212 addition
and
0 deletion
+212
-0
paddle/fluid/imperative/README.md
paddle/fluid/imperative/README.md
+212
-0
未找到文件。
paddle/fluid/imperative/README.md
0 → 100644
浏览文件 @
01dc15ce
# Overview
Imperative Programming is easier to learn, debug and try new ideas.
# Related Works
## Pytorch
https://pytorch.org/
## TensorFlow Eager
https://www.tensorflow.org/guide/eager
# Design
## API
```
python
class
Layer
(
object
):
def
__call__
(
inputs
):
# build some parameter once.
# ...
return
self
.
apply
(
inputs
):
def
forward
(
inputs
):
# forward logic with paddle operators. backward auto-generated.
class
PyLayer
(
core
.
PyLayer
):
def
__call__
(
cls
,
inputs
):
# trace the logic.
@
staticmethod
def
forward
(
inputs
):
# any forward logic implemented with numpy io.
@
staticmethod
def
backward
(
inputs
):
# any backward logic implemented with numpy io.
```
## Tracer
Current: Python Variable -> C++ VarBase -> C++ Variable -> C++ Tensor
Longer term.
```
python
# Parent class.
class
PyVarBase
(
object
):
pass
# Current python variable.
class
Variable
(
PyVarBase
):
pass
class
IVariable
(
PyVarBase
):
def
__init__
(
self
):
self
.
_ivar
=
core
.
VarBase
()
# Move var to a device.
def
to
(
device
):
pass
# Get var value.
def
value
():
pass
# Trigger backward.
def
backward
():
pass
# Get var's gradient value.
def
gradient_value
():
pass
# operators to override.
```
```
cpp
class
Tracer
{
public:
explicit
Tracer
(
framework
::
BlockDesc
*
root_block
)
:
root_block_
(
root_block
)
{}
virtual
~
Tracer
()
{}
void
Trace
(
OpBase
*
op
,
const
std
::
map
<
std
::
string
,
std
::
vector
<
VarBase
*>>&
inputs
,
const
std
::
map
<
std
::
string
,
std
::
vector
<
VarBase
*>>&
outputs
,
framework
::
BlockDesc
*
block
,
const
bool
stop_gradient
=
false
);
std
::
vector
<
VarBase
*>
PyTrace
(
OpBase
*
op
,
const
std
::
vector
<
VarBase
*>&
inputs
,
bool
stop_gradient
=
false
);
};
```
*
Trace forward operations
*
Perform quick shape/type infer, push kernel execution engine and return to user.
*
Perform autograd to generate gradients.
*
Clear trace.
*
Apply gradients with optimizers
## Autodiff
Lots of research already.
https://autodiff-workshop.github.io/
https://en.wikipedia.org/wiki/Automatic_differentiation
Basically, trace the forward execution, and perform autodiff
when needed.
*
Can be triggered by
`backward()`
.
*
Can select a block of code to trace and autodiff.
*
Use
`require_grad`
to drop some forward subgraph that doesn't need autodiff.
## Execution Engine
Lazy execution of pushed C++ operations.
## Device Placement
*
Operator executes on the inputs' device.
*
All inputs should live on the same device.
*
use
`Var.to()`
to explicitly move var to a device.
## Save/Load Models
TODO
## I/O
TODO
## Refactor
*
All function layers with parameters converted to class Layers.
*
Existing models converted to imperative mode.
*
All op tests run once in static graph, once in imperative mode.
# Examples
```
python
class
MyLayer
(
fluid
.
imperative
.
Layer
):
def
__init__
(
self
):
super
(
MyLayer
,
self
).
__init__
()
def
forward
(
self
,
inputs
):
x
=
fluid
.
layers
.
relu
(
inputs
)
x
=
fluid
.
layers
.
elementwise_mul
(
x
,
x
)
x
=
fluid
.
layers
.
reduce_sum
(
x
)
return
[
x
]
class
MyPyLayer
(
fluid
.
imperative
.
PyLayer
):
def
__init__
(
self
):
super
(
MyPyLayer
,
self
).
__init__
()
@
staticmethod
def
forward
(
inputs
):
return
np
.
tanh
(
inputs
[
0
])
@
staticmethod
def
backward
(
inputs
):
return
np
.
array
(
dout
)
*
(
1
-
np
.
square
(
np
.
array
(
out
)))
np_inp
=
np
.
ones
([
2
,
2
],
np
.
float32
)
with
fluid
.
imperative
.
guard
():
my_py_layer
=
MyPyLayer
()
outs
=
my_py_layer
(
np_inp
)
dy_out
=
np
.
sum
(
outs
[
0
].
_numpy
())
outs
[
0
].
_backward
()
dy_grad
=
var_inp
.
_gradient
()
class
MLP
(
fluid
.
imperative
.
Layer
):
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
)
x
=
self
.
_fc2
(
x
)
x
=
fluid
.
layers
.
reduce_sum
(
x
)
return
x
np_inp
=
np
.
array
([[
1.0
,
2.0
],
[
3.0
,
4.0
]],
dtype
=
np
.
float32
)
with
fluid
.
imperative
.
guard
():
var_inp
=
fluid
.
imperative
.
base
.
to_variable
(
np_inp
)
mlp
=
MLP
()
out
=
mlp
(
var_inp
)
dy_out
=
out
.
_numpy
()
out
.
_backward
()
```
# Plan
2.
1,3 fulltime, Can run a few simple models. (Currently, 2 20% engs)
4.
1, 4 fulltime, Can run 6 models, Performance 70% Pytorch. Release alpha.
6.
1, 5 fulltime, Performance close to Pytorch, can run multi-devices. Release Beta.
8.
1, 5 fulltime, Works in general. Update existing models. Can compile to static graph, support more optimizations.
12.
1 Done.
# Discussion
TODO.
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