blob: d852b6de526827e8cdc1c6ba0e860decf0b0013b [file] [log] [blame]
#include "caffe2/operators/mish_op.h"
#include <string>
#include <vector>
#include "caffe2/core/types.h"
#include "caffe2/utils/eigen_utils.h"
#include "caffe2/utils/math.h"
namespace caffe2 {
template <>
template <typename T>
bool MishFunctor<CPUContext>::
operator()(const int N, const T* X, T* Y, CPUContext* /* context */) const {
ConstEigenVectorArrayMap<T> X_arr(X, N);
EigenVectorArrayMap<T>(Y, N) = X_arr *
((T(1) + X_arr.exp()) - T(1) / (T(1) + X_arr.exp())) /
((T(1) + X_arr.exp()) + T(1) / (T(1) + X_arr.exp()));
return true;
}
template <>
template <typename T>
bool MishGradientOp<CPUContext>::DoRunWithType() {
auto& Xin = Input(X);
auto& Yin = Input(Y);
auto& DYin = Input(DY);
CAFFE_ENFORCE_EQ(Xin.numel(), Yin.numel());
CAFFE_ENFORCE_EQ(DYin.numel(), Yin.numel());
auto* DXout = Output(DX, Yin.sizes(), at::dtype<float>());
const float* Xdata = Xin.template data<float>();
const float* Ydata = Yin.template data<float>();
const float* dYdata = DYin.template data<float>();
float* dXdata = DXout->template mutable_data<float>();
EigenVectorArrayMap<float> dXvec(dXdata, DXout->numel());
ConstEigenVectorArrayMap<float> Xvec(Xdata, Xin.numel());
ConstEigenVectorArrayMap<float> Yvec(Ydata, Yin.numel());
ConstEigenVectorArrayMap<float> dYvec(dYdata, DYin.numel());
// w = e^(3x) + 4*e^2x + e^x * (6 + 4x) + 4(1 + x)
// q = (e^x + 1)^2 + 1
// dX = dY * e^x * w / q^2
dXvec = dYvec * Xvec.exp() *
((T(3) * Xvec).exp() + T(4) * (T(2) * Xvec).exp() +
Xvec.exp() * (T(6) + T(4) * Xvec) + T(4) * (T(1) + Xvec)) /
(((Xvec.exp() + T(1)) * (Xvec.exp() + T(1)) + T(1)) *
((Xvec.exp() + T(1)) * (Xvec.exp() + T(1)) + T(1)));
return true;
}
REGISTER_CPU_OPERATOR(
Mish,
UnaryElementwiseOp<
TensorTypes<float>,
CPUContext,
MishFunctor<CPUContext>>);
REGISTER_CPU_OPERATOR(MishGradient, MishGradientOp<CPUContext>);
// Input: X, output: Y
OPERATOR_SCHEMA(Mish)
.NumInputs(1)
.NumOutputs(1)
.IdenticalTypeAndShape()
.SetDoc(R"DOC(
Mish takes one input data (Tensor) and produces one output data
(Tensor) where the Mish function, y = x / (1 + exp(-x)), is applied to the
tensor elementwise.
)DOC")
.Input(0, "X", "1D input tensor")
.Output(0, "Y", "1D output tensor");
// Input: X, Y, dY, output: dX
OPERATOR_SCHEMA(MishGradient)
.NumInputs(3)
.NumOutputs(1)
.AllowInplace({{2, 0}})
.SetDoc(R"DOC(
MishGradient takes X, Y and dY and uses this to update dX according to the
chain rule and derivatives of the Mish function.
)DOC");
namespace {
class GetMishGradient : public GradientMakerBase {
using GradientMakerBase::GradientMakerBase;
std::vector<OperatorDef> GetGradientDefs() override {
return SingleGradientDef(
"MishGradient",
"",
std::vector<std::string>{I(0), O(0), GO(0)},
std::vector<std::string>{GI(0)});
}
};
} // namespace
REGISTER_GRADIENT(Mish, GetMishGradient);
} // namespace caffe2