代码¶

/**
 * @file tiny_optimizer.hpp
 * @brief Optimizers for tiny_ai — SGD (with momentum) and Adam-lite.
 */

#pragma once

#include "tiny_tensor.hpp"

#ifdef __cplusplus

#include <vector>

namespace tiny
{

struct ParamGroup
{
    Tensor *param;
    Tensor *grad;
};

class Optimizer
{
public:
    virtual ~Optimizer() {}
    virtual void init(const std::vector<ParamGroup> &groups) = 0;
    virtual void step(std::vector<ParamGroup> &groups) = 0;
    virtual void zero_grad(std::vector<ParamGroup> &groups);
};

class SGD : public Optimizer
{
public:
    SGD(float lr = 0.01f, float momentum = 0.0f, float weight_decay = 0.0f);
    void init(const std::vector<ParamGroup> &groups) override;
    void step(std::vector<ParamGroup> &groups) override;
private:
    float lr_;
    float momentum_;
    float weight_decay_;
    std::vector<Tensor> velocity_;
};

class Adam : public Optimizer
{
public:
    Adam(float lr       = 1e-3f,
         float beta1    = 0.9f,
         float beta2    = 0.999f,
         float epsilon  = 1e-8f,
         float weight_decay = 0.0f);
    void init(const std::vector<ParamGroup> &groups) override;
    void step(std::vector<ParamGroup> &groups) override;
private:
    float lr_, beta1_, beta2_, epsilon_, weight_decay_;
    int   t_;
    std::vector<Tensor> m_, v_;
};

} // namespace tiny

#endif // __cplusplus

/**
 * @file tiny_optimizer.cpp
 * @brief SGD and Adam optimizer implementations.
 */

#include "tiny_optimizer.hpp"
#include <cmath>
#include <cstring>

#ifdef __cplusplus

namespace tiny
{

void Optimizer::zero_grad(std::vector<ParamGroup> &groups)
{
    for (auto &pg : groups)
        if (pg.grad) pg.grad->zero();
}

SGD::SGD(float lr, float momentum, float weight_decay)
    : lr_(lr), momentum_(momentum), weight_decay_(weight_decay) {}

void SGD::init(const std::vector<ParamGroup> &groups)
{
    velocity_.clear();
    velocity_.reserve(groups.size());
    for (const auto &pg : groups)
        velocity_.push_back(Tensor::zeros_like(*pg.param));
}

void SGD::step(std::vector<ParamGroup> &groups)
{
    for (size_t p = 0; p < groups.size(); p++)
    {
        Tensor &param = *groups[p].param;
        Tensor &grad  = *groups[p].grad;
        Tensor &vel   = velocity_[p];

        for (int i = 0; i < param.size; i++)
        {
            float g = grad.data[i];
            if (weight_decay_ > 0.0f) g += weight_decay_ * param.data[i];
            if (momentum_ > 0.0f)
            {
                vel.data[i] = momentum_ * vel.data[i] + g;
                g = vel.data[i];
            }
            param.data[i] -= lr_ * g;
        }
    }
}

Adam::Adam(float lr, float beta1, float beta2, float epsilon, float weight_decay)
    : lr_(lr), beta1_(beta1), beta2_(beta2), epsilon_(epsilon),
      weight_decay_(weight_decay), t_(0) {}

void Adam::init(const std::vector<ParamGroup> &groups)
{
    t_ = 0;
    m_.clear(); v_.clear();
    m_.reserve(groups.size()); v_.reserve(groups.size());
    for (const auto &pg : groups)
    {
        m_.push_back(Tensor::zeros_like(*pg.param));
        v_.push_back(Tensor::zeros_like(*pg.param));
    }
}

void Adam::step(std::vector<ParamGroup> &groups)
{
    t_++;
    float bc1  = 1.0f - powf(beta1_, (float)t_);
    float bc2  = 1.0f - powf(beta2_, (float)t_);
    float lr_t = lr_ * sqrtf(bc2) / (bc1 + TINY_MATH_MIN_DENOMINATOR);

    for (size_t p = 0; p < groups.size(); p++)
    {
        Tensor &param = *groups[p].param;
        Tensor &grad  = *groups[p].grad;
        Tensor &m     = m_[p];
        Tensor &v     = v_[p];

        for (int i = 0; i < param.size; i++)
        {
            float g = grad.data[i];
            if (weight_decay_ > 0.0f) g += weight_decay_ * param.data[i];

            m.data[i] = beta1_ * m.data[i] + (1.0f - beta1_) * g;
            v.data[i] = beta2_ * v.data[i] + (1.0f - beta2_) * g * g;

            param.data[i] -= lr_t * m.data[i] / (sqrtf(v.data[i]) + epsilon_);
        }
    }
}

} // namespace tiny

#endif // __cplusplus