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
极客李华
Paddle
提交
c550e0ce
P
Paddle
项目概览
极客李华
/
Paddle
与 Fork 源项目一致
Fork自
PaddlePaddle / Paddle
通知
1
Star
0
Fork
0
代码
文件
提交
分支
Tags
贡献者
分支图
Diff
Issue
0
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
0
Issue
0
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 >>
提交
c550e0ce
编写于
12月 14, 2018
作者:
M
minqiyang
浏览文件
操作
浏览文件
下载
电子邮件补丁
差异文件
Add python interface for huber regression loss
test=develop
上级
6776e928
变更
2
隐藏空白更改
内联
并排
Showing
2 changed file
with
63 addition
and
13 deletion
+63
-13
paddle/fluid/operators/huber_loss_op.cc
paddle/fluid/operators/huber_loss_op.cc
+4
-3
python/paddle/fluid/layers/nn.py
python/paddle/fluid/layers/nn.py
+59
-10
未找到文件。
paddle/fluid/operators/huber_loss_op.cc
浏览文件 @
c550e0ce
...
...
@@ -124,8 +124,9 @@ REGISTER_OPERATOR(huber_loss, ops::HuberLossOp, ops::HuberLossOpMaker<float>,
paddle
::
framework
::
DefaultGradOpDescMaker
<
true
>
);
REGISTER_OPERATOR
(
huber_loss_grad
,
ops
::
HuberLossGradOp
);
REGISTER_OP_CPU_KERNEL
(
huber_loss
,
ops
::
HuberLossKernel
<
paddle
::
platform
::
CPUDeviceContext
,
float
>
);
huber_loss
,
ops
::
HuberLossKernel
<
paddle
::
platform
::
CPUDeviceContext
,
float
>
,
ops
::
HuberLossKernel
<
paddle
::
platform
::
CPUDeviceContext
,
double
>
);
REGISTER_OP_CPU_KERNEL
(
huber_loss_grad
,
ops
::
HuberLossGradKernel
<
paddle
::
platform
::
CPUDeviceContext
,
float
>
);
ops
::
HuberLossGradKernel
<
paddle
::
platform
::
CPUDeviceContext
,
float
>
,
ops
::
HuberLossGradKernel
<
paddle
::
platform
::
CPUDeviceContext
,
double
>
);
python/paddle/fluid/layers/nn.py
浏览文件 @
c550e0ce
...
...
@@ -169,6 +169,7 @@ __all__ = [
'log_loss'
,
'add_position_encoding'
,
'bilinear_tensor_product'
,
'huber_regression_loss'
,
]
...
...
@@ -4595,7 +4596,7 @@ def hsigmoid(input,
"""
The hierarchical sigmoid operator is used to accelerate the training
process of language model. This operator organizes the classes into a
complete binary tree, or you can use is_custom to pass your own tree to
complete binary tree, or you can use is_custom to pass your own tree to
implement hierarchical. Each leaf node represents a class(a word) and each
internal node acts as a binary classifier. For each word there's a unique
path from root to it's leaf node, hsigmoid calculate the cost for each
...
...
@@ -4611,7 +4612,7 @@ def hsigmoid(input,
2. build a dict to store word_id -> word's leaf to root path, we call it path_table.
3. build a dict to store word_id -> code of word's leaf to root path, we call it path_code. Code
means label of each binary classification, using 1 indicate true, 0 indicate false.
4. now, each word should has its path and code along the path, you can pass a batch of path and code
4. now, each word should has its path and code along the path, you can pass a batch of path and code
related to the same batch of inputs.
...
...
@@ -4621,8 +4622,8 @@ def hsigmoid(input,
and :math:`D` is the feature size.
label (Variable): The tensor variable contains labels of training data.
It's a tensor with shape is :math:`[N
\\
times 1]`.
num_classes: (int), The number of classes, must not be less than 2. with default tree this has to be set,
it should never be None under is_custom=False, but while is_custom is true, it should be non leaf num
num_classes: (int), The number of classes, must not be less than 2. with default tree this has to be set,
it should never be None under is_custom=False, but while is_custom is true, it should be non leaf num
which indicates the num of classes using by binary classify.
param_attr (ParamAttr|None): The parameter attribute for learnable parameters/weights
of hsigmoid. If it is set to None or one attribute of ParamAttr, hsigmoid
...
...
@@ -4635,15 +4636,15 @@ def hsigmoid(input,
is not set, the bias is initialized zero. Default: None.
name (str|None): A name for this layer(optional). If set None, the layer
will be named automatically. Default: None.
path_table: (Variable|None) this variable can store each batch of samples' path to root,
path_table: (Variable|None) this variable can store each batch of samples' path to root,
it should be in leaf -> root order
path_table should have the same shape with path_code, and for each sample i path_table[i] indicates a np.array like
structure and each element in this array is indexes in parent nodes' Weight Matrix.
path_code: (Variable|None) this variable can store each batch of samples' code,
path_table should have the same shape with path_code, and for each sample i path_table[i] indicates a np.array like
structure and each element in this array is indexes in parent nodes' Weight Matrix.
path_code: (Variable|None) this variable can store each batch of samples' code,
each code consist with every code of parent nodes. it should be in leaf -> root order
is_custom: (bool|False)using user defined binary tree instead of default complete binary tree, if costum is
is_custom: (bool|False)using user defined binary tree instead of default complete binary tree, if costum is
set you need to set path_table/path_code/num_classes, otherwise num_classes should be set
is_sparse: (bool|False)using sparse update instead of dense update, if set, the gradient
is_sparse: (bool|False)using sparse update instead of dense update, if set, the gradient
of W and input will be sparse.
Returns:
...
...
@@ -8770,3 +8771,51 @@ def bilinear_tensor_product(x,
# add activation
return
helper
.
append_activation
(
out
)
def
huber_regression_loss
(
input
,
label
,
delta
):
"""
Huber regression loss is a loss function used in robust regression.
Huber regression loss can evaluate the fitness of input to label.
Different from MSE loss, Huber regression loss is more robust for outliers.
When the difference between input and label is large than delta
.. math::
huber\_regression\_loss = delta * (label - input) - 0.5 * delta * delta
When the difference between input and label is less than delta
.. math::
huber\_regression\_loss = 0.5 * (label - input) * (label - input)
Args:
input (Variable): This input is a probability computed by the previous operator.
The first dimension is batch size, and the last dimension is 1.
label (Variable): The groud truth whose first dimension is batch size
and last dimension is 1.
delta (float): The parameter of huber regression loss, which controls
the range of outliers
Returns:
huber\_regression\_loss (Variable): The huber regression loss with shape [batch_size, 1].
Examples:
.. code-block:: python
predictions = fluid.layers.softmax(x)
loss = fluid.layers.huber_regression_loss(input=predictions, label=label, 1.0)
"""
helper
=
LayerHelper
(
'huber_regression_loss'
,
**
locals
())
residual
=
helper
.
create_variable_for_type_inference
(
dtype
=
helper
.
input_dtype
())
out
=
helper
.
create_variable_for_type_inference
(
dtype
=
helper
.
input_dtype
())
helper
.
append_op
(
type
=
'huber_loss'
,
inputs
=
{
'X'
:
input
,
'Y'
:
label
},
outputs
=
{
'Out'
:
out
,
'Residual'
:
residual
},
attrs
=
{
'delta'
:
delta
})
return
out
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