Files
2026-07-13 13:33:03 +08:00

60 lines
2.1 KiB
C++

//
// ParameterOptimizer.hpp
// MNN
//
// Created by MNN on 2019/11/22.
// Copyright © 2018, Alibaba Group Holding Limited
//
#ifndef ParameterOptimizer_hpp
#define ParameterOptimizer_hpp
#include <MNN/expr/Expr.hpp>
#include <MNN/expr/Module.hpp>
#include <set>
namespace MNN {
namespace Train {
class MNN_PUBLIC ParameterOptimizer {
public:
enum RegularizationMethod {
L1,
L2,
L1L2,
};
ParameterOptimizer(std::shared_ptr<Express::Module> module);
virtual ~ParameterOptimizer() = default;
bool step(Express::VARP loss);
int currentStep();
void setCurrentStep(int step);
static void makeLoopModel(const char* mnnFileName, std::vector<Express::VARP> outputs, const std::pair<std::vector<Express::VARP>, std::vector<Express::VARP>>& parameters);
struct ParameterOptGrad {
Express::VARP parameter;
Express::VARP parameterGrad;
Express::VARP learningRate;
};
virtual std::map<Express::VARP, Express::VARP> onGetNextParameter(Express::VARP loss) = 0;
virtual std::pair<std::vector<Express::VARP>, std::vector<Express::VARP>> onMakeParameterUpdateGraphByGrad(const std::vector<ParameterOptGrad>& parameterGrads);
std::pair<std::vector<Express::VARP>, std::vector<Express::VARP>> makeParameterUpdateGraphByGrad(const std::vector<Express::VARP>& p, const std::vector<Express::VARP>& pd, const std::vector<Express::VARP>& lr);
static ParameterOptimizer* createSGD(std::shared_ptr<Express::Module> module, float lr, float momentum, float weightDecay, RegularizationMethod method);
static ParameterOptimizer* createADAM(std::shared_ptr<Express::Module> module, float lr, float momentum, float momentum2, float weightDecay, float eps, RegularizationMethod method);
protected:
const std::set<Express::VARP>& trainable() const {
return mTrainable;
}
std::shared_ptr<Express::Module> module() const {
return mModule;
}
private:
int mStep = 0;
std::shared_ptr<Express::Module> mModule;
std::set<Express::VARP> mTrainable;
};
} // namespace Train
} // namespace MNN
#endif