Merge pull request #1592 from Noplz/smooth_l1_layer

Smooth l1 layer for SSD

paddle/gserver/layers/CostLayer.cpp

@@ -192,6 +192,59 @@ void SumOfSquaresCostLayer::backwardImp(Matrix& output,
   outputG.sumOfSquaresBp(output, *label.value);
 }
 
+//
+// class SmoothL1CostLayer
+//
+
+REGISTER_LAYER(smooth_l1, SmoothL1CostLayer);
+
+bool SmoothL1CostLayer::init(const LayerMap& layerMap,
+                             const ParameterMap& parameterMap) {
+  return CostLayer::init(layerMap, parameterMap);
+}
+
+void SmoothL1CostLayer::forwardImp(Matrix& output,
+                                   Argument& label,
+                                   Matrix& target) {
+  MatrixPtr targetCpu, outputCpu, labelCpu;
+  if (useGpu_) {
+    targetCpu =
+        Matrix::create(target.getHeight(), target.getWidth(), false, false);
+    outputCpu =
+        Matrix::create(output.getHeight(), output.getWidth(), false, false);
+    labelCpu = Matrix::create(
+        label.value->getHeight(), label.value->getWidth(), false, false);
+    targetCpu->copyFrom(target);
+    outputCpu->copyFrom(output);
+    labelCpu->copyFrom(*label.value);
+    targetCpu->smoothL1(*outputCpu, *(labelCpu));
+    target.copyFrom(*targetCpu);
+  } else {
+    target.smoothL1(output, *label.value);
+  }
+}
+
+void SmoothL1CostLayer::backwardImp(Matrix& output,
+                                    Argument& label,
+                                    Matrix& outputG) {
+  MatrixPtr outputGCpu, outputCpu, labelCpu;
+  if (useGpu_) {
+    outputGCpu =
+        Matrix::create(outputG.getHeight(), outputG.getWidth(), false, false);
+    outputCpu =
+        Matrix::create(output.getHeight(), output.getWidth(), false, false);
+    labelCpu = Matrix::create(
+        label.value->getHeight(), label.value->getWidth(), false, false);
+    outputGCpu->copyFrom(outputG);
+    outputCpu->copyFrom(output);
+    labelCpu->copyFrom(*label.value);
+    outputGCpu->smoothL1Bp(*outputCpu, *labelCpu);
+    outputG.copyFrom(*outputGCpu);
+  } else {
+    outputG.smoothL1Bp(output, *label.value);
+  }
+}
+
 //
 // class RankingCost
 //

paddle/gserver/layers/CostLayer.h

@@ -159,6 +159,29 @@ public:
                    Matrix& outputGrad) override;
 };
 
+/**
+ * This cost layer compute smooth L1 loss for real-valued regression
+ * tasks.
+ * \f[
+ * L =
+ *   (output - label)^2 * 0.5  / -1 < (output - label) < 1 /
+ *   (output - label) - 0.5    / otherwise  /
+ * \f]
+ */
+class SmoothL1CostLayer : public CostLayer {
+public:
+  explicit SmoothL1CostLayer(const LayerConfig& config) : CostLayer(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;
+};
+
 /**
  * A cost layer for learning to rank (LTR) task. This layer contains at leat
  * three inputs.

paddle/gserver/tests/test_LayerGrad.cpp

@@ -1602,6 +1602,20 @@ TEST(Layer, PadLayer) {
   }
 }
 
+TEST(Layer, smooth_l1) {
+  TestConfig config;
+  config.layerConfig.set_type("smooth_l1");
+
+  config.inputDefs.push_back({INPUT_DATA, "layer_0", 1, 0});
+  config.inputDefs.push_back({INPUT_DATA_TARGET, "layer_1", 1, 0});
+  config.layerConfig.add_inputs();
+  config.layerConfig.add_inputs();
+
+  for (auto useGpu : {false, true}) {
+    testLayerGrad(config, "smooth_l1", 100, false, useGpu, false, 2.0);
+  }
+}
+
 int main(int argc, char** argv) {
   testing::InitGoogleTest(&argc, argv);
   initMain(argc, argv);

paddle/math/Matrix.cpp

@@ -3590,6 +3590,55 @@ void CpuMatrix::sumOfSquaresBp(Matrix& output, Matrix& label) {
   }
 }
 
+void CpuMatrix::smoothL1(Matrix& output, Matrix& label) {
+  CHECK(output.useGpu_ == false && label.useGpu_ == false)
+      << "Matrix type are not equal";
+
+  size_t numSamples = getHeight();
+  size_t dim = output.getWidth();
+  CHECK_EQ(label.getHeight(), numSamples);
+  CHECK_EQ(output.getHeight(), numSamples);
+  CHECK_EQ(label.getWidth(), dim);
+  CHECK_EQ(getWidth(), (size_t)1);
+  real* out = output.getData();
+  real* cost = getData();
+  real* lbl = label.getData();
+
+  for (size_t i = 0; i < numSamples; ++i, out += dim, cost += dim, lbl += dim) {
+    for (size_t j = 0; j < dim; ++j) {
+      cost[j] = std::fabs(out[j] - lbl[j]);
+      if (cost[j] < 1.0)
+        cost[j] = 0.5 * cost[j] * cost[j];
+      else
+        cost[j] = cost[j] - 0.5;
+    }
+  }
+}
+
+void CpuMatrix::smoothL1Bp(Matrix& output, Matrix& label) {
+  CHECK(output.useGpu_ == false && label.useGpu_ == false)
+      << "Matrix type are not equal";
+
+  size_t numSamples = getHeight();
+  size_t dim = output.getWidth();
+  CHECK_EQ(label.getHeight(), numSamples);
+  CHECK_EQ(output.getHeight(), numSamples);
+  CHECK_EQ(label.getWidth(), dim);
+  CHECK_EQ(getWidth(), (size_t)1);
+  real* out = output.getData();
+  real* cost = getData();
+  real* lbl = label.getData();
+
+  // f'(x) = x         if |x| < 1
+  //       = sign(x)   otherwise
+  for (size_t i = 0; i < numSamples; ++i, out += dim, cost += dim, lbl += dim) {
+    for (size_t j = 0; j < dim; ++j) {
+      cost[j] = out[j] - lbl[j];
+      if (std::fabs(cost[j]) >= 1) cost[j] = (0 < cost[j]) - (cost[j] < 0);
+    }
+  }
+}
+
 void CpuMatrix::tanh(Matrix& output) {
   CHECK(isContiguous());
   CHECK(output.isContiguous());

paddle/math/Matrix.h

@@ -783,6 +783,14 @@ public:
     LOG(FATAL) << "Not implemented";
   }
 
+  virtual void smoothL1(Matrix& output, Matrix& label) {
+    LOG(FATAL) << "Not implemented";
+  }
+
+  virtual void smoothL1Bp(Matrix& outputV, Matrix& label) {
+    LOG(FATAL) << "Not implemented";
+  }
+
   virtual void tanh(Matrix& output) { LOG(FATAL) << "Not implemented"; }
 
   virtual void tanhDerivative(Matrix& output) {
@@ -1720,6 +1728,9 @@ public:
   /// gradient of sumOfSquares.
   void sumOfSquaresBp(Matrix& outputV, Matrix& label);
 
+  void smoothL1(Matrix& output, Matrix& label);
+  void smoothL1Bp(Matrix& output, Matrix& label);
+
   void tanh(Matrix& output);
   void tanhDerivative(Matrix& output);