Add activation gelu (#14569)

paddle/fluid/operators/activation_op.cc

@@ -149,6 +149,13 @@ $out = \max(x, 0)$
 
 )DOC";
 
+UNUSED constexpr char GeluDoc[] = R"DOC(
+Gelu Activation Operator.
+
+$out = \\frac{1 + erf(\\frac{x}{\\sqrt{2}})}{2} x$
+
+)DOC";
+
 UNUSED constexpr char TanhDoc[] = R"DOC(
 Tanh Activation Operator.
 
@@ -472,6 +479,7 @@ REGISTER_ACTIVATION_OP_MAKER(Sigmoid, SigmoidDoc);
 REGISTER_ACTIVATION_OP_MAKER(LogSigmoid, LogSigmoidDoc);
 REGISTER_ACTIVATION_OP_MAKER(Exp, ExpDoc);
 REGISTER_ACTIVATION_OP_MAKER(Relu, ReluDoc);
+REGISTER_ACTIVATION_OP_MAKER(Gelu, GeluDoc);
 REGISTER_ACTIVATION_OP_MAKER(Tanh, TanhDoc);
 REGISTER_ACTIVATION_OP_MAKER(TanhShrink, TanhShrinkDoc);
 REGISTER_ACTIVATION_OP_MAKER(Sqrt, SqrtDoc);
@@ -489,6 +497,7 @@ REGISTER_ACTIVATION_OP_MAKER(Softsign, SoftsignDoc);
 
 REGISTER_ACTIVATION_OP_GRAD_MAKER(Sigmoid, sigmoid);
 REGISTER_ACTIVATION_OP_GRAD_MAKER(Relu, relu);
+REGISTER_ACTIVATION_OP_GRAD_MAKER(Gelu, gelu);
 REGISTER_ACTIVATION_OP_GRAD_MAKER(Exp, exp);
 REGISTER_ACTIVATION_OP_GRAD_MAKER(Tanh, tanh);
 REGISTER_ACTIVATION_OP_GRAD_MAKER(Ceil, ceil);
@@ -525,6 +534,7 @@ namespace ops = paddle::operators;
   __macro(Round, round);             \
   __macro(Log, log);                 \
   __macro(Square, square);           \
+  __macro(Gelu, gelu);               \
   __macro(BRelu, brelu);             \
   __macro(Pow, pow);                 \
   __macro(STanh, stanh);             \

paddle/fluid/operators/activation_op.h

@@ -16,6 +16,11 @@ limitations under the License. */
 #include <utility>
 #include <vector>
 
+#include <cmath>
+#ifndef _USE_MATH_DEFINES
+#define _USE_MATH_DEFINES
+#endif
+
 #include "paddle/fluid/framework/eigen.h"
 #include "paddle/fluid/framework/op_registry.h"
 #include "paddle/fluid/operators/detail/safe_ref.h"
@@ -212,6 +217,31 @@ struct ReluGradFunctor : public BaseActivationFunctor<T> {
   }
 };
 
+// gelu(x) = 0.5 * x *  (1 + erf(x / sqrt(2)))
+template <typename T>
+struct GeluFunctor : public BaseActivationFunctor<T> {
+  template <typename Device, typename X, typename Out>
+  void operator()(Device d, X x, Out out) const {
+    auto temp =
+        ((x * static_cast<T>(M_SQRT1_2)).erf()).template cast<T>().eval();
+    out.device(d) = x * static_cast<T>(0.5) * (static_cast<T>(1) + temp);
+  }
+};
+
+template <typename T>
+struct GeluGradFunctor : BaseActivationFunctor<T> {
+  bool Inplace() const { return IsInplace("gelu"); }
+  template <typename Device, typename X, typename Out, typename dOut,
+            typename dX>
+  void operator()(Device d, X x, Out out, dOut dout, dX dx) const {
+    auto temp = (static_cast<T>(0.5 * M_2_SQRTPI * M_SQRT1_2) * x *
+                 ((-static_cast<T>(0.5) * x.square()).exp()))
+                    .template cast<T>()
+                    .eval();
+    dx.device(d) = dout * (out / x + temp);
+  }
+};
+
 // tanh(x) = (exp(x) - exp(-x)) / (exp(x) + exp(-x))
 template <typename T>
 struct TanhFunctor : public BaseActivationFunctor<T> {
@@ -877,6 +907,7 @@ struct SwishGradFunctor : public BaseActivationFunctor<T> {
   __macro(logsigmoid, LogSigmoidFunctor, LogSigmoidGradFunctor);     \
   __macro(exp, ExpFunctor, ExpGradFunctor);                          \
   __macro(relu, ReluFunctor, ReluGradFunctor);                       \
+  __macro(gelu, GeluFunctor, GeluGradFunctor);                       \
   __macro(tanh, TanhFunctor, TanhGradFunctor);                       \
   __macro(softshrink, SoftShrinkFunctor, SoftShrinkGradFunctor);     \
   __macro(sqrt, SqrtFunctor, SqrtGradFunctor);                       \

paddle/fluid/platform/float16.h

@@ -1039,6 +1039,11 @@ HOSTDEVICE inline float16 exp(const float16& a) {
   return float16(::expf(static_cast<float>(a)));
 }
 
+template <>
+HOSTDEVICE inline float16 erf(const float16& a) {
+  return float16(::erff(static_cast<float>(a)));
+}
+
 template <>
 HOSTDEVICE inline float16 log(const float16& a) {
   return float16(::logf(static_cast<float>(a)));

python/paddle/fluid/tests/unittests/test_activation_op.py

@@ -18,7 +18,7 @@ import unittest
 import numpy as np
 import paddle.fluid.core as core
 from op_test import OpTest
-from scipy.special import expit
+from scipy.special import expit, erf
 
 
 class TestActivation(OpTest):
@@ -295,6 +295,23 @@ class TestRelu(TestActivation):
         self.check_grad(['X'], 'Out', max_relative_error=0.007)
 
 
+class TestGelu(TestActivation):
+    def setUp(self):
+        self.op_type = "gelu"
+        self.init_dtype()
+
+        x = np.random.uniform(-1, 1, [11, 17]).astype(self.dtype)
+        out = 0.5 * x * (1.0 + erf(x / np.sqrt(2.0)))
+
+        self.inputs = {'X': OpTest.np_dtype_to_fluid_dtype(x)}
+        self.outputs = {'Out': out}
+
+    def test_check_grad(self):
+        if self.dtype == np.float16:
+            return
+        self.check_grad(['X'], 'Out', max_relative_error=0.007)
+
+
 class TestBRelu(TestActivation):
     def setUp(self):
         self.op_type = "brelu"
@@ -628,6 +645,7 @@ create_test_act_fp16_class(TestCos, grad_atol=0.85)
 create_test_act_fp16_class(TestSin)
 create_test_act_fp16_class(TestRound, grad_check=False)
 create_test_act_fp16_class(TestRelu)
+create_test_act_fp16_class(TestGelu)
 create_test_act_fp16_class(TestBRelu)
 create_test_act_fp16_class(TestRelu6)
 create_test_act_fp16_class(TestSoftRelu)