vuron/adept: Платформа для обучения и вывода нейронных сетей, с API приближенным к PyTorch.

src/nn/adam.cpp

74 строки2 KB

kolkir

Fix training issues

06 мар 2025, 23:54

06 мар 2025, 23:54efe26e8

#include <adept/nn/adam.hpp>
#include <adept/threading.hpp>

#include <cmath>

namespace adept {

Adam::Adam(std::vector<Variable> parameters,
           float32_t lr,
           float32_t beta1,
           float32_t beta2,
           float32_t eps)
    : parameters_(parameters), lr_(lr), beta1_(beta1), beta2_(beta2), eps_(eps) {}

void Adam::step() {
  // TODO: consider to use parallel_for? needs benchmarking.
  for (auto& param : parameters_) {
    if (param.defined() && param.requires_grad()) {
      auto grad = param.grad();
      auto data = param.data();
      auto param_state = state_.find(data.impl().get());

      // initialize state
      if (param_state == state_.end()) {
        auto state = std::make_unique<detail::AdamParamState>();
        state->step = 0;
        // exponential moving average of gradient values
        state->exp_avg = Tensor::zero(grad.properties());
        // exponential moving average of squared gradient values
        state->exp_avg_sq = Tensor::zero(grad.properties());
        state_[data.impl().get()] = std::move(state);
      }

      auto& state = static_cast<detail::AdamParamState&>(*state_[data.impl().get()]);
      auto& exp_avg = state.exp_avg;
      auto& exp_avg_sq = state.exp_avg_sq;

      state.step += 1;

      auto bias_correction1 = 1 - std::pow(beta1_, state.step);
      auto bias_correction2 = 1 - std::pow(beta2_, state.step);

      // decay the first and second moment running average coefficient
      exp_avg.mul_(beta1_).add_(grad * (1 - beta1_));
      exp_avg_sq.mul_(beta2_).add_((grad * grad) * (1 - beta2_));

      auto denom = (exp_avg_sq.sqrt() / std::sqrt(bias_correction2)).add_(eps_);

      auto step_size = lr_ / bias_correction1;
      data -= (exp_avg / denom) * step_size;
    }
  }
}

void Adam::zero_grad() {
  parallel_for_each(parameters_, [](auto& param) {
    if (param.defined())
      param.zero_grad();
  });
}

void Adam::set_lr(float32_t lr) {
  lr_ = lr;
}

float32_t Adam::lr() const {
  return lr_;
}

const detail::AdamParamState* Adam::state(void* key) const {
  return state_.at(key).get();
}

}  // namespace adept