refine Huber loss, add huber_regression_cost

doc/api/v2/config/layer.rst

@@ -419,9 +419,14 @@ multi_binary_label_cross_entropy_cost
 ..  autoclass:: paddle.v2.layer.multi_binary_label_cross_entropy_cost
     :noindex:
 
-huber_cost
-----------
-..  autoclass:: paddle.v2.layer.huber_cost
+huber_regression_cost
+-------------------------
+..  autoclass:: paddle.v2.layer.huber_regression_cost
+    :noindex:
+
+huber_classification_cost
+-------------------------
+..  autoclass:: paddle.v2.layer.huber_classification_cost
     :noindex:
 
 lambda_cost

paddle/gserver/layers/CostLayer.cpp

@@ -572,13 +572,8 @@ void MultiBinaryLabelCrossEntropy::backwardImp(Matrix& output,
   }
 }
 
-//
-// Huber loss for robust 2-classes classification
-//
-REGISTER_LAYER(huber, HuberTwoClass);
-
-bool HuberTwoClass::init(const LayerMap& layerMap,
-                         const ParameterMap& parameterMap) {
+bool HuberCost::init(const LayerMap& layerMap,
+                     const ParameterMap& parameterMap) {
   CostLayer::init(layerMap, parameterMap);
   if (useGpu_) {
     tmpCpuInput_.reserve(inputLayers_.size());
@@ -589,69 +584,138 @@ bool HuberTwoClass::init(const LayerMap& layerMap,
   return true;
 }
 
-void HuberTwoClass::forwardImp(Matrix& output, Argument& label, Matrix& cost) {
+void HuberCost::forwardImp(Matrix& output, Argument& label, Matrix& cost) {
   if (useGpu_) {
     for (size_t i = 0; i < inputLayers_.size(); i++) {
       tmpCpuInput_[i].resizeAndCopyFrom(
           getInput(i), false, HPPL_STREAM_DEFAULT);
     }
     hl_stream_synchronize(HPPL_STREAM_DEFAULT);
   }
-  forwardImpIn(output, label, cost);
 }
 
-void HuberTwoClass::forwardImpIn(Matrix& output,
-                                 Argument& label,
-                                 Matrix& target) {
+//
+// Huber loss for robust regression.
+//
+REGISTER_LAYER(huber_regression, HuberRegressionLoss);
+
+bool HuberRegressionLoss::init(const LayerMap& layerMap,
+                               const ParameterMap& parameterMap) {
+  HuberCost::init(layerMap, parameterMap);
+  delta_ = config_.delta();
+  return true;
+}
+
+void HuberRegressionLoss::forwardImp(Matrix& output,
+                                     Argument& label,
+                                     Matrix& target) {
+  HuberCost::forwardImp(output, label, target);
   size_t numSamples = target.getHeight();
-  CHECK_EQ((*label.ids).getSize(), numSamples);
+  size_t dim = output.getWidth();
+  CHECK(label.value);
+  CHECK_EQ((*label.value).getHeight(), numSamples);
   CHECK_EQ(output.getHeight(), numSamples);
-  CHECK_EQ(output.getWidth(), (size_t)1);
+  CHECK_EQ(dim, (*label.value).getWidth());
   CHECK_EQ(target.getWidth(), (size_t)1);
 
   real* out = useGpu_ ? tmpCpuInput_[0].value->getData() : output.getData();
-  int* lbl = useGpu_ ? tmpCpuInput_[1].ids->getData() : (*label.ids).getData();
-  std::vector<real> cost(numSamples);
+  real* lbl =
+      useGpu_ ? tmpCpuInput_[1].value->getData() : (*label.value).getData();
+  std::vector<real> cost(numSamples, 0);
   for (size_t i = 0; i < numSamples; ++i) {
-    int y = 2 * lbl[i] - 1;
-    if (out[i] * y < -1)
-      cost[i] = -4 * out[i] * y;
-    else if (out[i] * y < 1)
-      cost[i] = (1 - out[i] * y) * (1 - out[i] * y);
-    else
-      cost[i] = 0;
+    for (size_t j = 0; j < dim; ++j) {
+      int index = i * dim + j;
+      real a = std::abs(lbl[index] - out[index]);
+      if (a <= delta_)
+        cost[i] += a * a / 2;
+      else
+        cost[i] += delta_ * (a - delta_ / 2);
+    }
   }
   target.copyFrom(cost.data(), numSamples);
 }
 
-void HuberTwoClass::backwardImp(Matrix& outputValue,
-                                Argument& label,
-                                Matrix& outputGrad) {
-  if (useGpu_) {
-    backwardImpIn(
-        *tmpCpuInput_[0].value, tmpCpuInput_[1], *tmpCpuInput_[0].grad);
-    outputGrad.copyFrom(*tmpCpuInput_[0].grad);
-  } else {
-    backwardImpIn(outputValue, label, outputGrad);
+void HuberRegressionLoss::backwardImp(Matrix& output,
+                                      Argument& label,
+                                      Matrix& outputG) {
+  size_t numSamples = output.getHeight();
+  size_t dim = output.getWidth();
+  real* out = useGpu_ ? tmpCpuInput_[0].value->getData() : output.getData();
+  real* lbl =
+      useGpu_ ? tmpCpuInput_[1].value->getData() : (*label.value).getData();
+  real* grad = useGpu_ ? tmpCpuInput_[0].grad->getData() : outputG.getData();
+  for (size_t i = 0; i < numSamples; ++i) {
+    for (size_t j = 0; j < dim; ++j) {
+      int index = i * dim + j;
+      real a = lbl[index] - out[index];
+      if (std::abs(a) <= delta_)
+        grad[index] += -a;
+      else
+        grad[index] += a > 0 ? -delta_ : delta_;
+    }
   }
+  if (useGpu_) outputG.copyFrom(grad, numSamples * dim);
 }
 
-void HuberTwoClass::backwardImpIn(Matrix& output,
-                                  Argument& label,
-                                  Matrix& outputG) {
-  size_t numSamples = output.getHeight();
-  real* out = output.getData();
-  real* grad = outputG.getData();
-  int* lbl = (*label.ids).getData();
+//
+// Huber loss for robust 2-classes classification
+//
+REGISTER_LAYER(huber_classification, HuberTwoClassification);
+
+bool HuberTwoClassification::init(const LayerMap& layerMap,
+                                  const ParameterMap& parameterMap) {
+  return HuberCost::init(layerMap, parameterMap);
+}
+
+void HuberTwoClassification::forwardImp(Matrix& output,
+                                        Argument& label,
+                                        Matrix& target) {
+  HuberCost::forwardImp(output, label, target);
+  size_t numSamples = target.getHeight();
+  size_t dim = output.getWidth();
+  CHECK(label.ids);
+  CHECK_EQ((*label.ids).getSize(), numSamples);
+  CHECK_EQ(output.getHeight(), numSamples);
+  CHECK_EQ(target.getWidth(), (size_t)1);
+
+  real* out = useGpu_ ? tmpCpuInput_[0].value->getData() : output.getData();
+  int* lbl = useGpu_ ? tmpCpuInput_[1].ids->getData() : (*label.ids).getData();
+  std::vector<real> cost(numSamples, 0);
   for (size_t i = 0; i < numSamples; ++i) {
     int y = 2 * lbl[i] - 1;
-    if (y * out[i] < -1)
-      grad[i] += -4 * y;
-    else if (y * out[i] < 1)
-      grad[i] += -2 * (1 - y * out[i]) * y;
+    for (size_t j = 0; j < dim; ++j) {
+      int index = i * dim + j;
+      real a = out[index] * y;
+      if (a < -1)
+        cost[i] += -4 * a;
+      else if (a < 1)
+        cost[i] += (1 - a) * (1 - a);
+    }
   }
+  target.copyFrom(cost.data(), numSamples);
 }
 
+void HuberTwoClassification::backwardImp(Matrix& output,
+                                         Argument& label,
+                                         Matrix& outputG) {
+  size_t numSamples = output.getHeight();
+  size_t dim = output.getWidth();
+  real* out = useGpu_ ? tmpCpuInput_[0].value->getData() : output.getData();
+  int* lbl = useGpu_ ? tmpCpuInput_[1].ids->getData() : (*label.ids).getData();
+  real* grad = useGpu_ ? tmpCpuInput_[0].grad->getData() : outputG.getData();
+  for (size_t i = 0; i < numSamples; ++i) {
+    int y = 2 * lbl[i] - 1;
+    for (size_t j = 0; j < dim; ++j) {
+      int index = i * dim + j;
+      real a = out[index] * y;
+      if (a < -1)
+        grad[index] += -4 * y;
+      else if (a < 1)
+        grad[index] += -2 * (1 - a) * y;
+    }
+  }
+  if (useGpu_) outputG.copyFrom(grad, numSamples * dim);
+}
 /**
  * This cost layer compute the sum of its input as loss.
  * \f[

paddle/gserver/layers/CostLayer.h

@@ -304,37 +304,68 @@ class MultiBinaryLabelCrossEntropy : public CostLayer {
                    Matrix& outputGrad) override;
 };
 
-/**
- * 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]
+/*
+ * A base layer for HuberRegressionLoss and HuberTwoClassification.
  */
-class HuberTwoClass : public CostLayer {
+class HuberCost : public CostLayer {
+public:
   std::vector<Argument> tmpCpuInput_;
 
-public:
-  explicit HuberTwoClass(const LayerConfig& config) : CostLayer(config) {}
+  explicit HuberCost(const LayerConfig& config) : CostLayer(config) {}
 
   bool init(const LayerMap& layerMap,
             const ParameterMap& parameterMap) override;
 
   void forwardImp(Matrix& output, Argument& label, Matrix& cost) override;
 
-  void forwardImpIn(Matrix& output, Argument& label, Matrix& cost);
+  void backwardImp(Matrix& outputValue, Argument& label, Matrix& outputGrad) {}
+};
+
+/**
+ * Huber loss for robust regression.
+ *
+ * Given output f(x), label y and delta, the loss is:
+ * Loss = 0.5 * (1 - y * f)^2, if abs(y - f) <= delta \\
+ * Loss = delta * abs(y - f) - 0.5 * delta^2, otherwise
+ */
+class HuberRegressionLoss : public HuberCost {
+public:
+  explicit HuberRegressionLoss(const LayerConfig& config) : HuberCost(config) {}
+
+  bool init(const LayerMap& layerMap,
+            const ParameterMap& parameterMap) override;
+
+  void forwardImp(Matrix& output, Argument& label, Matrix& cost) override;
 
   void backwardImp(Matrix& outputValue,
                    Argument& label,
                    Matrix& outputGrad) override;
 
-  void backwardImpIn(Matrix& outputValue, Argument& label, Matrix& outputGrad);
+protected:
+  real delta_;
+};
+
+/**
+ * Huber loss for robust 2-classes classification.
+ *
+ * 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 HuberTwoClassification : public HuberCost {
+public:
+  explicit HuberTwoClassification(const LayerConfig& config)
+      : HuberCost(config) {}
+
+  bool init(const LayerMap& layerMap,
+            const ParameterMap& parameterMap) override;
+
+  void forwardImp(Matrix& output, Argument& label, Matrix& cost) override;
+
+  void backwardImp(Matrix& outputValue,
+                   Argument& label,
+                   Matrix& outputGrad) override;
 };
 
 typedef std::shared_ptr<CostLayer> CostLayerPtr;

paddle/gserver/tests/test_LayerGrad.cpp

@@ -850,9 +850,27 @@ TEST(Layer, square_error_weighted) {
   }
 }
 
+TEST(Layer, huber_regression_loss) {
+  TestConfig config;
+  config.layerConfig.set_type("huber_regression");
+  config.biasSize = 0;
+
+  config.inputDefs.push_back({INPUT_DATA, "layer_0", 10, 0});
+  config.inputDefs.push_back({INPUT_DATA_TARGET, "layer_1", 10, 0});
+  config.layerConfig.add_inputs();
+  config.layerConfig.add_inputs();
+
+  for (auto useGpu : {false, true}) {
+    for (auto delta : {1, 3, 5}) {
+      config.layerConfig.set_delta(delta);
+      testLayerGrad(config, "huber_regression", 100, /* trans */ false, useGpu);
+    }
+  }
+}
+
 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});
@@ -861,7 +879,7 @@ TEST(Layer, huber_two_class) {
   config.layerConfig.add_inputs();
 
   for (auto useGpu : {false, true}) {
-    testLayerGrad(config, "huber", 100, /* trans */ false, useGpu);
+    testLayerGrad(config, "huber_two_class", 100, /* trans */ false, useGpu);
   }
 }
 

proto/ModelConfig.proto

@@ -499,6 +499,9 @@ message LayerConfig {
   optional int32 axis = 54 [ default = 2 ];
   repeated uint32 offset = 55;
   repeated uint32 shape = 56;
+
+  // for HuberRegressionLoss
+  optional double delta = 57 [ default = 1.0 ];
 }
 
 message EvaluatorConfig {

python/paddle/trainer/config_parser.py

@@ -2274,7 +2274,7 @@ def init(cls, name, inputs, device=None, coeff=1.):
 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')
 
@@ -2336,6 +2336,17 @@ def __init__(self, name, inputs, NDCG_num=5, max_sort_size=-1, device=None):
         self.config.max_sort_size = max_sort_size
 
 
+@config_layer('huber_regression')
+class HuberRegressionLoss(LayerBase):
+    def __init__(self, name, inputs, delta=1., coeff=1., device=None):
+        super(HuberRegressionLoss, self).__init__(
+            name, 'huber_regression', 1, inputs=inputs, device=device)
+        config_assert(
+            len(self.inputs) == 2, 'HuberRegression must have 2 inputs')
+        self.config.delta = delta
+        self.config.coeff = coeff
+
+
 @config_layer('nce')
 class NCELayer(LayerBase):
     def __init__(self,