utils/summary.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.

// lbann_summary - Write summary statistics to Tensorboard

#ifndef LBANN_SUMMARY_HPP_INCLUDED
#define LBANN_SUMMARY_HPP_INCLUDED

#include "lbann/base.hpp"
#include "lbann/comm.hpp"
#include <string>
#include <vector>

#ifdef LBANN_HAS_TBINF
#include "TBinf.hpp"
#endif

namespace lbann {

template <typename T, typename U>
using BiggerOf = typename std::conditional<(sizeof(T) > sizeof(U)), T, U>::type;

#ifdef LBANN_HAS_TBINF

class lbann_summary
{
public:
  lbann_summary(std::string logdir, lbann_comm* comm);
  ~lbann_summary();

  template <typename TensorDataType>
  void reduce_mean(const std::string tag,
                   const El::AbstractDistMatrix<TensorDataType>& mat,
                   int step);
  template <typename TensorDataType>
  void reduce_min(const std::string tag,
                  const El::AbstractDistMatrix<TensorDataType>& mat,
                  int step);
  template <typename TensorDataType>
  void reduce_max(const std::string tag,
                  const El::AbstractDistMatrix<TensorDataType>& mat,
                  int step);
  template <typename TensorDataType>
  void reduce_stdev(const std::string tag,
                    const El::AbstractDistMatrix<TensorDataType>& mat,
                    int step);
  template <typename TensorDataType>
  void reduce_scalar(const std::string tag, TensorDataType s, int step);
  template <typename TensorDataType>
  void sum_reduce_scalar(const std::string tag, TensorDataType s, int step);
  template <typename TensorDataType>
  void reduce_scalar_all(const std::string tag, TensorDataType s, int step);
  template <typename TensorDataType>
  void reduce_histogram(const std::string tag,
                        const El::AbstractDistMatrix<TensorDataType>& mat,
                        int step);
  template <typename TensorDataType>
  void reduce_2norm(const std::string tag,
                    const El::AbstractDistMatrix<TensorDataType>& mat,
                    int step);
  void report_image(std::string const& /*tag*/,
                    std::string const& /*img_format*/,
                    CPUMat const& /*image*/,
                    std::vector<int> const& /*dims*/,
                    int /*step*/);
  void flush();

private:
  lbann_comm* m_comm;
  TBinf::SummaryWriter* m_sw;

  struct pending_op
  {
    pending_op(const std::string tag_,
               int step_,
               float local_,
               float local2_ = 0.0f,
               int num_ = 0)
      : tag(tag_), step(step_), local(local_), local2(local2_), num(num_)
    {}
    const std::string tag;
    int step;
    float local;
    float local2;
    int num;
  };
  struct pending_histogram
  {
    pending_histogram(const std::string tag_,
                      int step_,
                      std::vector<double> buckets_,
                      double min_,
                      double max_,
                      double num_,
                      double sum_,
                      double sqsum_)
      : tag(tag_),
        step(step_),
        buckets(buckets_),
        min(min_),
        max(max_),
        num(num_),
        sum(sum_),
        sqsum(sqsum_)
    {}
    const std::string tag;
    int step;
    std::vector<double> buckets;
    double min;
    double max;
    double num;
    double sum;
    double sqsum;
  };

  std::vector<pending_op> m_pending_means;
  std::vector<pending_op> m_pending_mins;
  std::vector<pending_op> m_pending_maxes;
  std::vector<pending_op> m_pending_stdevs;
  std::vector<pending_op> m_pending_scalars;
  std::vector<pending_op> m_pending_sum_scalars;
  std::vector<pending_op> m_pending_scalar_alls;
  std::vector<double> m_histogram_buckets;
  std::vector<pending_histogram> m_pending_histograms;

  void flush_means();
  void flush_mins();
  void flush_maxes();
  void flush_stdevs();
  void flush_scalars();
  void flush_sum_scalars();
  void flush_scalar_alls();
  void flush_histograms();

  template <typename TensorDataType>
  auto local_sum(const El::AbstractMatrix<TensorDataType>& mat) const
    -> BiggerOf<TensorDataType, float>;
  template <typename TensorDataType, typename AccumT>
  void local_sum_sqsum(const El::AbstractMatrix<TensorDataType>& mat,
                       AccumT& sum,
                       AccumT& sqsum) const;
  template <typename TensorDataType>
  auto local_min(const El::AbstractMatrix<TensorDataType>& mat) const
    -> BiggerOf<TensorDataType, float>;
  template <typename TensorDataType>
  auto local_max(const El::AbstractMatrix<TensorDataType>& mat) const
    -> BiggerOf<TensorDataType, float>;
  template <typename TensorDataType>
  auto local_2norm(const El::AbstractMatrix<TensorDataType>& mat) const
    -> BiggerOf<TensorDataType, float>;
  std::string prepend_model(const std::string tag, int model) const;
  void gather_scalar_summary(const std::string tag, float s, int step);
  void gather_scalar_summary(const std::vector<pending_op>& ops,
                             std::vector<float>& scalars);
};

#else

class lbann_summary
{
public:
  lbann_summary(std::string logdir, lbann_comm* comm) {}

  void report_image(std::string const& tag,
                    std::string const& img_format,
                    CPUMat const& image,
                    std::vector<int> const& dims,
                    int step)
  {}

  template <typename TensorDataType>
  void reduce_mean(const std::string tag,
                   const El::AbstractDistMatrix<TensorDataType>& mat,
                   int step)
  {}
  template <typename TensorDataType>
  void reduce_min(const std::string tag,
                  const El::AbstractDistMatrix<TensorDataType>& mat,
                  int step)
  {}
  template <typename TensorDataType>
  void reduce_max(const std::string tag,
                  const El::AbstractDistMatrix<TensorDataType>& mat,
                  int step)
  {}
  template <typename TensorDataType>
  void reduce_stdev(const std::string tag,
                    const El::AbstractDistMatrix<TensorDataType>& mat,
                    int step)
  {}
  template <typename TensorDataType>
  void reduce_scalar(const std::string tag, TensorDataType s, int step)
  {}
  template <typename TensorDataType>
  void sum_reduce_scalar(const std::string tag, TensorDataType s, int step)
  {}
  template <typename TensorDataType>
  void reduce_scalar_all(const std::string tag, TensorDataType s, int step)
  {}
  template <typename TensorDataType>
  void reduce_histogram(const std::string tag,
                        const El::AbstractDistMatrix<TensorDataType>& mat,
                        int step)
  {}
  template <typename TensorDataType>
  void reduce_2norm(const std::string tag,
                    const El::AbstractDistMatrix<TensorDataType>& mat,
                    int step)
  {}
  void flush() {}
};

#endif // LBANN_HAS_TBINF

} // namespace lbann

#endif // LBANN_SUMMARY_HPP_INCLUDED