sgd.hpp

Go to the documentation of this file.

// Copyright (c) 2014-2023, Lawrence Livermore National Security, LLC.
// Produced at the Lawrence Livermore National Laboratory.
// Written by the LBANN Research Team (B. Van Essen, et al.) listed in
// the CONTRIBUTORS file. <lbann-dev@llnl.gov>
//
// LLNL-CODE-697807.
// All rights reserved.
//
// This file is part of LBANN: Livermore Big Artificial Neural Network
// Toolkit. For details, see http://software.llnl.gov/LBANN or
// https://github.com/LLNL/LBANN.
//
// Licensed under the Apache License, Version 2.0 (the "Licensee"); you
// may not use this file except in compliance with the License.  You may
// obtain a copy of the License at:
//
// http://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.

#ifndef LBANN_OPTIMIZERS_SGD_HPP_INCLUDED
#define LBANN_OPTIMIZERS_SGD_HPP_INCLUDED

#include "lbann/io/persist.hpp"
#include "lbann/optimizers/data_type_optimizer.hpp"
#include "lbann/proto/optimizers.pb.h"

namespace lbann {

template <typename TensorDataType>
class sgd
  : public Cloneable<sgd<TensorDataType>, data_type_optimizer<TensorDataType>>
{
  using BaseType =
    Cloneable<sgd<TensorDataType>, data_type_optimizer<TensorDataType>>;

public:

  using AbsDistMatrixType = El::AbstractDistMatrix<TensorDataType>;

  using OptimizerType = data_type_optimizer<TensorDataType>;

  using WeightsType = data_type_weights<TensorDataType>;


public:

  sgd(TensorDataType learning_rate,
      TensorDataType momentum = 0,
      bool nesterov = false);
  sgd(const sgd& other);
  sgd& operator=(const sgd& other);
  ~sgd() override = default;


  template <class ArchiveT>
  void serialize(ArchiveT& ar);



  std::string get_type() const override { return "SGD"; }
  description get_description() const override;



  TensorDataType get_momentum() const noexcept { return m_momentum; }
  void set_momentum(TensorDataType momentum) { m_momentum = momentum; }

  bool using_nesterov() const noexcept { return m_nesterov; }
  void set_nesterov(bool nesterov) { m_nesterov = nesterov; }

  const AbsDistMatrixType& get_velocity() const;
  AbsDistMatrixType& get_velocity();



  using OptimizerType::setup;
  void setup(WeightsType* w = nullptr) override;


  void write_proto(lbann_data::Optimizer& opt) const final;

protected:
  sgd() : sgd(El::To<TensorDataType>(0.f), El::To<TensorDataType>(0.f), false)
  {}

  void step_compute(AbsDistMatrixType& values,
                    const AbsDistMatrixType& gradient) override;

private:
  TensorDataType m_momentum;
  bool m_nesterov;
  std::unique_ptr<AbsDistMatrixType> m_velocity;

  void momentum_step_cpu(AbsDistMatrixType& values,
                         const AbsDistMatrixType& gradient);
#ifdef LBANN_HAS_GPU

  void momentum_step_gpu(AbsDistMatrixType& values,
                         const AbsDistMatrixType& gradient);
#endif // LBANN_HAS_GPU

  friend class cereal::access;
};

template <typename TensorDataType>
std::unique_ptr<optimizer>
build_sgd_optimizer_from_pbuf(google::protobuf::Message const&);

} // namespace lbann

#endif // LBANN_OPTIMIZERS_SGD_HPP_INCLUDED