hypergradient_adam.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_HYPERGRADIENT_ADAM_HPP_INCLUDED
#define LBANN_OPTIMIZERS_HYPERGRADIENT_ADAM_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 hypergradient_adam : public Cloneable<hypergradient_adam<TensorDataType>,
                                            data_type_optimizer<TensorDataType>>
{
  using BaseType = Cloneable<hypergradient_adam<TensorDataType>,
                             data_type_optimizer<TensorDataType>>;

public:

  using AbsDistMatrixType = El::AbstractDistMatrix<TensorDataType>;

  using WeightsType = data_type_weights<TensorDataType>;

  using OptimizerType = data_type_optimizer<TensorDataType>;


public:
  hypergradient_adam(
    TensorDataType init_learning_rate = El::To<TensorDataType>(1e-3),
    TensorDataType hyper_learning_rate = El::To<TensorDataType>(1e-7),
    TensorDataType beta1 = El::To<TensorDataType>(0.9),
    TensorDataType beta2 = El::To<TensorDataType>(0.99),
    TensorDataType eps = El::To<TensorDataType>(1e-8));
  hypergradient_adam(const hypergradient_adam& other);
  hypergradient_adam& operator=(const hypergradient_adam& other);
  ~hypergradient_adam() override = default;

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

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

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

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

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

private:
  TensorDataType m_hyper_learning_rate;
  TensorDataType m_beta1;
  TensorDataType m_beta2;
  TensorDataType m_eps;
  TensorDataType m_current_beta1;
  TensorDataType m_current_beta2;
  std::unique_ptr<AbsDistMatrixType> m_moment1;
  std::unique_ptr<AbsDistMatrixType> m_moment2;
  std::unique_ptr<AbsDistMatrixType> m_old_gradient;
};

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

} // namespace lbann

#endif // LBANN_OPTIMIZER_HYPERGRADIENT_ADAM_HPP_INCLUDED