src/modules/optim/ema.cc¶

// Copyright 2026 The Tenzr Project Authors
//
// SPDX-License-Identifier: Apache-2.0
//
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.
// You may obtain a copy of the License at
//
//      https://www.apache.org/licenses/LICENSE-2.0
//
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS,
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
// See the License for the specific language governing permissions and
// limitations under the License.

#include "tenzr/modules/optim/ema.h"

#include <vector>

#include "tenzr/common/error.h"

namespace tenzr::optim {

  ExponentialMovingAverage::ExponentialMovingAverage(const std::unordered_map<std::string, Tensor>& params, double decay) : decay_(decay) {
    TENZR_CHECK(decay >= 0.0 && decay < 1.0, "Decay must be in [0, 1)");
    for (const auto& [name, p] : params) {
      shadow_params_[name] = p.Clone();
    }
  }

  void ExponentialMovingAverage::Update(const std::unordered_map<std::string, Tensor>& params) {
    double one_minus_d = 1.0 - decay_;

    for (const auto& [name, p] : params) {
      auto s_it = shadow_params_.find(name);
      if (s_it == shadow_params_.end() || !s_it->second.HasData() || !p.HasData()) {
        continue;
      }

      auto s_raw = s_it->second.ToVectorAs<double>();
      auto p_raw = p.ToVectorAs<double>();
      size_t n = s_raw.size();

      for (size_t i = 0; i < n; ++i) {
        s_raw[i] = decay_ * s_raw[i] + one_minus_d * p_raw[i];
      }
      s_it->second.BindData(Tensor::FromVector<double>(s_raw, s_it->second.shape()).Astype(s_it->second.Dtype()).Data());
    }
  }

  void ExponentialMovingAverage::ApplyShadow(std::unordered_map<std::string, Tensor>& params) {
    backup_params_.clear();
    for (auto& [name, p] : params) {
      auto s_it = shadow_params_.find(name);
      if (s_it != shadow_params_.end() && s_it->second.HasData() && p.HasData()) {
        backup_params_[name] = p.Clone();
        p.BindData(Tensor::FromVector<double>(s_it->second.ToVectorAs<double>(), p.shape()).Astype(p.Dtype()).Data());
      }
    }
  }

  void ExponentialMovingAverage::Restore(std::unordered_map<std::string, Tensor>& params) {
    for (auto& [name, p] : params) {
      auto b_it = backup_params_.find(name);
      if (b_it != backup_params_.end() && b_it->second.HasData() && p.HasData()) {
        p.BindData(Tensor::FromVector<double>(b_it->second.ToVectorAs<double>(), p.shape()).Astype(p.Dtype()).Data());
      }
    }
    backup_params_.clear();
  }

}  // namespace tenzr::optim