Added Leaky Relu activation

paddle/operators/activation_op.cc

@@ -69,6 +69,22 @@ class ReluOpMaker : public framework::OpProtoAndCheckerMaker {
   }
 };
 
+template <typename AttrType>
+class LeakyReluOpMaker : public framework::OpProtoAndCheckerMaker {
+ public:
+  LeakyReluOpMaker(framework::OpProto *proto,
+                   framework::OpAttrChecker *op_checker)
+      : OpProtoAndCheckerMaker(proto, op_checker) {
+    AddInput("X", "Input of LeakyRelu operator");
+    AddOutput("Y", "Output of LeakyRelu operator");
+    AddComment(
+        "LeakyRelu activation operator, "
+        "leaky_relu = max(x, alpha * x)");
+    AddAttr<AttrType>("alpha", "The small negative slope")
+        .SetDefault(static_cast<AttrType>(0.02f));
+  }
+};
+
 class TanhOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
   TanhOpMaker(framework::OpProto *proto, framework::OpAttrChecker *op_checker)
@@ -240,6 +256,9 @@ REGISTER_OP(softsign, ops::ActivationOp, ops::SoftsignOpMaker, softsign_grad,
 REGISTER_OP(brelu, ops::ActivationOp, ops::BReluOpMaker<float>, brelu_grad,
             ops::ActivationOpGrad);
 
+REGISTER_OP(leaky_relu, ops::ActivationOp, ops::LeakyReluOpMaker<float>,
+            leaky_relu_grad, ops::ActivationOpGrad);
+
 REGISTER_OP(soft_relu, ops::ActivationOp, ops::SoftReluOpMaker<float>,
             soft_relu_grad, ops::ActivationOpGrad);
 

paddle/operators/activation_op.h

@@ -309,6 +309,33 @@ struct SoftReluGradFunctor : public BaseActivationFunctor<T> {
   }
 };
 
+template <typename T>
+struct LeakyReluFunctor : public BaseActivationFunctor<T> {
+  float alpha;
+  typename BaseActivationFunctor<T>::AttrPair GetAttrs() {
+    return {{"alpha", &alpha}};
+  }
+
+  template <typename Device, typename X, typename Y>
+  void operator()(Device d, X x, Y y) const {
+    y.device(d) = x.cwiseMax(alpha * x);
+  }
+};
+
+template <typename T>
+struct LeakyReluGradFunctor : public BaseActivationFunctor<T> {
+  float alpha;
+  typename BaseActivationFunctor<T>::AttrPair GetAttrs() {
+    return {{"alpha", &alpha}};
+  }
+  template <typename Device, typename X, typename Y, typename dY, typename dX>
+  void operator()(Device d, X x, Y y, dY dy, dX dx) const {
+    auto temp1 = alpha * (x < static_cast<T>(0)).template cast<T>().eval();
+    auto temp2 = (x >= static_cast<T>(0)).template cast<T>().eval();
+    dx.device(d) = dy * (temp1 + temp2).template cast<T>();
+  }
+};
+
 template <typename T>
 struct PowFunctor : public BaseActivationFunctor<T> {
   float factor;
@@ -379,4 +406,5 @@ struct STanhGradFunctor : public BaseActivationFunctor<T> {
   __macro(soft_relu, SoftReluFunctor, SoftReluGradFunctor);      \
   __macro(pow, PowFunctor, PowGradFunctor);                      \
   __macro(stanh, STanhFunctor, STanhGradFunctor);                \
-  __macro(softsign, SoftsignFunctor, SoftsignGradFunctor)
+  __macro(softsign, SoftsignFunctor, SoftsignGradFunctor);       \
+  __macro(leaky_relu, LeakyReluFunctor, LeakyReluGradFunctor)

python/paddle/v2/framework/tests/test_activation_op.py

@@ -122,6 +122,23 @@ class TestBRelu(OpTest):
         self.check_grad(['X'], 'Y', max_relative_error=0.02)
 
 
+class TestLeakyRelu(OpTest):
+    def setUp(self):
+        self.op_type = "leaky_relu"
+        alpha = 0.02
+        self.attrs = {'alpha': alpha}
+        self.inputs = {'X': np.random.uniform(-3, 3, [4, 4]).astype("float32")}
+        self.outputs = {
+            'Y': np.maximum(self.inputs['X'], alpha * self.inputs['X'])
+        }
+
+    def test_check_output(self):
+        self.check_output()
+
+    def test_check_grad(self):
+        self.check_grad(['X'], 'Y', max_relative_error=0.008)
+
+
 class TestSoftRelu(OpTest):
     def setUp(self):
         self.op_type = "soft_relu"