提交 e6db484d 编写于 作者: L Luo Tao

make clear that current huber_cost is for two-classification

上级 493e1c04
......@@ -575,9 +575,9 @@ void MultiBinaryLabelCrossEntropy::backwardImp(Matrix& output,
//
// Huber loss for robust 2-classes classification
//
REGISTER_LAYER(huber, HuberTwoClass);
REGISTER_LAYER(huber, HuberTwoClassification);
bool HuberTwoClass::init(const LayerMap& layerMap,
bool HuberTwoClassification::init(const LayerMap& layerMap,
const ParameterMap& parameterMap) {
CostLayer::init(layerMap, parameterMap);
if (useGpu_) {
......@@ -589,7 +589,9 @@ bool HuberTwoClass::init(const LayerMap& layerMap,
return true;
}
void HuberTwoClass::forwardImp(Matrix& output, Argument& label, Matrix& cost) {
void HuberTwoClassification::forwardImp(Matrix& output,
Argument& label,
Matrix& cost) {
if (useGpu_) {
for (size_t i = 0; i < inputLayers_.size(); i++) {
tmpCpuInput_[i].resizeAndCopyFrom(
......@@ -600,10 +602,11 @@ void HuberTwoClass::forwardImp(Matrix& output, Argument& label, Matrix& cost) {
forwardImpIn(output, label, cost);
}
void HuberTwoClass::forwardImpIn(Matrix& output,
void HuberTwoClassification::forwardImpIn(Matrix& output,
Argument& label,
Matrix& target) {
size_t numSamples = target.getHeight();
CHECK(label.ids);
CHECK_EQ((*label.ids).getSize(), numSamples);
CHECK_EQ(output.getHeight(), numSamples);
CHECK_EQ(output.getWidth(), (size_t)1);
......@@ -624,7 +627,7 @@ void HuberTwoClass::forwardImpIn(Matrix& output,
target.copyFrom(cost.data(), numSamples);
}
void HuberTwoClass::backwardImp(Matrix& outputValue,
void HuberTwoClassification::backwardImp(Matrix& outputValue,
Argument& label,
Matrix& outputGrad) {
if (useGpu_) {
......@@ -636,7 +639,7 @@ void HuberTwoClass::backwardImp(Matrix& outputValue,
}
}
void HuberTwoClass::backwardImpIn(Matrix& output,
void HuberTwoClassification::backwardImpIn(Matrix& output,
Argument& label,
Matrix& outputG) {
size_t numSamples = output.getHeight();
......
......@@ -307,21 +307,17 @@ public:
/**
* Huber loss for robust 2-classes classification.
*
* For label={0, 1}, let y=2*label-1. Given output f, the loss is:
* \f[
* Loss =
* \left\{\begin{matrix}
* 4 * y * f & \textit{if} \ \ y* f < -1 \\
* (1 - y * f)^2 & \textit{if} \ \ -1 < y * f < 1 \\
* 0 & \textit{otherwise}
* \end{matrix}\right.
* \f]
* For label={0, 1}, let y=2*label-1. Given output f(x), the loss is:
* Loss = 4 * y * f, if y* f < -1 \\
* Loss = (1 - y * f)^2, if -1 < y * f < 1 \\
* Loss = 0, otherwise
*/
class HuberTwoClass : public CostLayer {
class HuberTwoClassification : public CostLayer {
std::vector<Argument> tmpCpuInput_;
public:
explicit HuberTwoClass(const LayerConfig& config) : CostLayer(config) {}
explicit HuberTwoClassification(const LayerConfig& config)
: CostLayer(config) {}
bool init(const LayerMap& layerMap,
const ParameterMap& parameterMap) override;
......
......@@ -830,7 +830,7 @@ TEST(Layer, square_error_weighted) {
TEST(Layer, huber_two_class) {
TestConfig config;
config.layerConfig.set_type("huber");
config.layerConfig.set_type("huber_classification");
config.biasSize = 0;
config.inputDefs.push_back({INPUT_DATA, "layer_0", 1, 0});
......
......@@ -2255,7 +2255,7 @@ define_cost('PnpairValidation', 'pnpair-validation')
define_cost('SumOfSquaresCostLayer', 'square_error')
define_cost('MultiBinaryLabelCrossEntropy', 'multi_binary_label_cross_entropy')
define_cost('SoftBinaryClassCrossEntropy', 'soft_binary_class_cross_entropy')
define_cost('HuberTwoClass', 'huber')
define_cost('HuberTwoClassification', 'huber_classification')
define_cost('SumCost', 'sum_cost')
define_cost('SmoothL1Cost', 'smooth_l1')
......
......@@ -108,7 +108,7 @@ __all__ = [
'sum_cost',
'rank_cost',
'lambda_cost',
'huber_cost',
'huber_classification_cost',
'block_expand_layer',
'maxout_layer',
'out_prod_layer',
......@@ -216,7 +216,7 @@ class LayerType(object):
RANK_COST = 'rank-cost'
LAMBDA_COST = 'lambda_cost'
HUBER = 'huber'
HUBER_CLASSIFICATION = 'huber_classification'
CROSS_ENTROPY = 'multi-class-cross-entropy'
CROSS_ENTROPY_WITH_SELFNORM = 'multi_class_cross_entropy_with_selfnorm'
SOFT_BIN_CLASS_CROSS_ENTROPY = 'soft_binary_class_cross_entropy'
......@@ -5605,16 +5605,26 @@ def sum_cost(input, name=None, layer_attr=None):
@wrap_name_default()
@layer_support()
def huber_cost(input, label, name=None, coeff=1.0, layer_attr=None):
def huber_classification_cost(input,
label,
name=None,
coeff=1.0,
layer_attr=None):
"""
A loss layer for huber loss.
For classification purposes, a variant of the Huber loss called modified Huber
is sometimes used. Given a prediction f(x) (a real-valued classifier score) and
a true binary class label :math:`y\in \left \{-1, 1 \right \}`, the modified Huber
loss is defined as:
.. math:
loss = \max \left ( 0, 1-yf(x) \right )^2, yf(x)\geq 1
loss = -4yf(x), \text{otherwise}
The example usage is:
.. code-block:: python
cost = huber_cost(input=input_layer,
label=label_layer)
cost = huber_classification_cost(input=input_layer, label=label_layer)
:param input: The first input layer.
:type input: LayerOutput.
......@@ -5634,11 +5644,12 @@ def huber_cost(input, label, name=None, coeff=1.0, layer_attr=None):
assert input.size == 1
Layer(
name=name,
type=LayerType.HUBER,
type=LayerType.HUBER_CLASSIFICATION,
inputs=[input.name, label.name],
coeff=coeff,
**ExtraLayerAttribute.to_kwargs(layer_attr))
return LayerOutput(name, LayerType.HUBER, parents=[input, label], size=1)
return LayerOutput(
name, LayerType.HUBER_CLASSIFICATION, parents=[input, label], size=1)
@wrap_name_default()
......
......@@ -180,8 +180,8 @@ layers {
active_type: ""
}
layers {
name: "__huber_cost_0__"
type: "huber"
name: "__huber_classification_cost_0__"
type: "huber_classification"
size: 1
active_type: ""
inputs {
......@@ -300,7 +300,7 @@ output_layer_names: "__rank_cost_0__"
output_layer_names: "__lambda_cost_0__"
output_layer_names: "__cross_entropy_0__"
output_layer_names: "__cross_entropy_with_selfnorm_0__"
output_layer_names: "__huber_cost_0__"
output_layer_names: "__huber_classification_cost_0__"
output_layer_names: "__multi_binary_label_cross_entropy_0__"
output_layer_names: "__sum_cost_0__"
output_layer_names: "__nce_layer_0__"
......@@ -326,7 +326,7 @@ sub_models {
layer_names: "__cross_entropy_with_selfnorm_0__"
layer_names: "huber_probs"
layer_names: "huber_label"
layer_names: "__huber_cost_0__"
layer_names: "__huber_classification_cost_0__"
layer_names: "__multi_binary_label_cross_entropy_0__"
layer_names: "__sum_cost_0__"
layer_names: "__nce_layer_0__"
......@@ -349,7 +349,7 @@ sub_models {
output_layer_names: "__lambda_cost_0__"
output_layer_names: "__cross_entropy_0__"
output_layer_names: "__cross_entropy_with_selfnorm_0__"
output_layer_names: "__huber_cost_0__"
output_layer_names: "__huber_classification_cost_0__"
output_layer_names: "__multi_binary_label_cross_entropy_0__"
output_layer_names: "__sum_cost_0__"
output_layer_names: "__nce_layer_0__"
......
......@@ -33,7 +33,7 @@ outputs(
input=probs, label=xe_label),
cross_entropy_with_selfnorm(
input=probs, label=xe_label),
huber_cost(
huber_classification_cost(
input=data_layer(
name='huber_probs', size=1),
label=data_layer(
......
Markdown is supported
0% .
You are about to add 0 people to the discussion. Proceed with caution.
先完成此消息的编辑!
想要评论请 注册