trainer.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_TRAINER_HPP
#define LBANN_TRAINER_HPP

#include "lbann/base.hpp"
#include "lbann/detect_El_mpi.hpp"
#include "lbann/io/persist.hpp"
#include "lbann/proto/lbann.pb.h"
#include "lbann/utils/hash.hpp"
#include "lbann/utils/threads/thread_pool.hpp"
#include <memory>
#include <string>
#include <unordered_map>
#include <vector>

namespace lbann {

// Forward-declarations
class data_coordinator;
class description;
class lbann_comm;
class callback_base;
class ExecutionContext;
class generic_data_reader;
class TrainingAlgorithm;
class TerminationCriteria;
class model;

class trainer
{
public:

  trainer(lbann_comm* comm,
          std::unique_ptr<data_coordinator> dc,
          size_t mini_batch_size,
          std::unique_ptr<TrainingAlgorithm> alg = nullptr);

  ~trainer();



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



  void set_name(std::string const& name);

  void set_random_seeds(int root_random_seed,
                        int random_seed,
                        int data_seq_random_seed)
  {
    m_root_random_seed = root_random_seed;
    m_random_seed = random_seed;
    m_data_seq_random_seed = data_seq_random_seed;
  }

  void add_callback(std::shared_ptr<callback_base> cb)
  {
    if (cb == nullptr) {
      throw lbann_exception(
        "model: Attempted to add null pointer as a callback.");
    }
    m_callbacks.push_back(std::move(cb));
  }

  void setup(std::unique_ptr<thread_pool> io_thread_pool,
             std::map<execution_mode, generic_data_reader*> data_readers);

  void allow_background_io_activity(bool enable)
  {
    m_background_io_allowed = enable;
  }



  std::string get_name() const { return m_name; }

  description get_description() const;

  int get_random_seed() const noexcept { return m_random_seed; }
  int get_data_seq_random_seed() const noexcept
  {
    return m_data_seq_random_seed;
  }

  std::vector<observer_ptr<callback_base>> get_callbacks() const
  {
    std::vector<observer_ptr<callback_base>> callback_list;
    callback_list.reserve(m_callbacks.size());
    for (const auto& ptr : m_callbacks) {
      callback_list.push_back(ptr.get());
    }
    return callback_list;
  }

  std::vector<std::shared_ptr<callback_base>>& get_callbacks_with_ownership()
  {
    return m_callbacks;
  }

  const data_coordinator& get_data_coordinator() const
  {
    if (m_data_coordinator == nullptr) {
      LBANN_ERROR("data_coordinator is nullptr");
    }
    return *m_data_coordinator;
  }

  data_coordinator& get_data_coordinator()
  {
    return const_cast<data_coordinator&>(
      static_cast<const trainer&>(*this).get_data_coordinator());
  }

  thread_pool& get_io_thread_pool() const
  {
    if (!m_io_thread_pool) {
      LBANN_ERROR("m_io_thread_pool is null");
    }
    return *(m_io_thread_pool.get());
  }

  lbann_comm* get_comm() const noexcept { return m_comm; }

  persist& get_persist_obj() noexcept { return m_persist; }

  size_t get_max_mini_batch_size() const noexcept
  {
    return m_max_mini_batch_size;
  }

  bool background_io_activity_allowed() const noexcept
  {
    return m_background_io_allowed;
  }


  using execution_context_key_pair_t =
    typename std::pair<observer_ptr<model>, execution_mode>;

  execution_context_key_pair_t
  check_and_build_execution_context(TrainingAlgorithm& alg,
                                    observer_ptr<model> model,
                                    execution_mode mode);

  execution_context_key_pair_t
  check_and_build_execution_context(ExecutionContext& c,
                                    model& model,
                                    execution_mode mode);

  ExecutionContext& get_execution_context(observer_ptr<model> model,
                                          execution_mode mode);

  ExecutionContext& get_execution_context(execution_context_key_pair_t key);

  bool execution_context_valid(model& m, execution_mode mode) const noexcept;

  bool execution_context_valid(execution_context_key_pair_t key) const noexcept;


  void
  train(observer_ptr<model> model, El::Int num_epochs, El::Int num_batches = 0);

  void evaluate(observer_ptr<model> model,
                execution_mode mode,
                El::Int num_batches = 0);


  std::vector<El::Grid*> get_grids() const;
  void add_grid(std::unique_ptr<El::Grid> g);


  bool save_to_checkpoint_shared();

  bool load_from_checkpoint_shared(persist& p);

  bool load_from_checkpoint_shared(model& m, ExecutionContext& c);

  bool save_to_checkpoint_distributed();

  bool load_from_checkpoint_distributed(persist& p);

  bool load_from_checkpoint_distributed(model& m, ExecutionContext& c);

  void write_proto(lbann_data::Trainer& proto);


private:
  void delete_execution_context(execution_context_key_pair_t key);

  void for_each_execution_context(
    std::function<void(observer_ptr<ExecutionContext>)> fn);

private:
  persist m_persist;

  using model_execution_context_hash_t =
    pair_hash<observer_ptr<model>,
              execution_mode,
              std::hash<observer_ptr<model>>,
              enum_hash<execution_mode>>;

  using ModelContextMapType =
    std::unordered_map<std::pair<observer_ptr<model>, execution_mode>,
                       std::unique_ptr<ExecutionContext>,
                       model_execution_context_hash_t>;

  ModelContextMapType m_model_execution_context;

  std::string m_name;

  std::vector<std::shared_ptr<callback_base>> m_callbacks;

  std::unique_ptr<thread_pool> m_io_thread_pool;

  std::unique_ptr<data_coordinator> m_data_coordinator;

  std::unique_ptr<TrainingAlgorithm> m_training_alg;

  lbann_comm* m_comm;

  std::vector<std::unique_ptr<El::Grid>> m_grids;

  size_t m_max_mini_batch_size;

  int m_root_random_seed;

  int m_random_seed;

  int m_data_seq_random_seed;

  bool m_background_io_allowed;
};

trainer& get_trainer();

trainer const& get_const_trainer();

} // namespace lbann

#endif // LBANN_TRAINER_HPP