Fine Tuning

[Armin234]

Hi,
is there an example for train / retrain / fine tune a network created with the "Neural Network Console" ?
I tryed with the result.nnp net.nntxt and model.nnp, but i didn't find the right way.

It would be great if I could use the networks developed with the "Neural Network Console", simply with the Nabla-Lib in a C++ program on my Win10 computer and retrain or fine tune the networks with other data sets and then perform the inference with the same C++ program.

Regards
Armin

[TomonobuTsujikawa]

Here is the sample code how to train the model and monitor its performance:

// Copyright 2018,2019,2020,2021 Sony Group Corporation.
//
// 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
//
//     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.
#include <nbla_utils/nnp.hpp>

#ifdef WITH_CUDA
#include <nbla/cuda/cudnn/init.hpp>
#include <nbla/cuda/init.hpp>
#endif

#ifdef TIMING
#include <chrono>
#endif

#include <cassert>
#include <fstream>
#include <iomanip>
#include <iostream>
#include <sstream>
#include <string>

using namespace nbla;

/******************************************/
// Example of mnist training
/******************************************/
int main(int argc, char *argv[]) {

  if (argc != 5) {
    std::cerr << "Usage: " << argv[0] << " nnp_file input_pgm" << std::endl;
    std::cerr << std::endl;
    std::cerr << "Positional arguments: " << std::endl;
    std::cerr << "  nnp_file  : .nnp file created by "
                 "examples/vision/mnist/save_nnp_classification.py."
              << std::endl;
    std::cerr << "  optimizer : Executor name in nnp file."
              << std::endl;
    std::cerr << "  monitor : Executor name in nnp file."
              << std::endl;
    std::cerr << "  iter_num  : iteration number"
              << std::endl;
    return -1;
  }
  const std::string nnp_file(argv[1]);

  std::string optimizer_name;
  optimizer_name = argv[2];

  std::string monitor_name;
  monitor_name = argv[3];

  int iter_num = atoi(argv[4]);

  // Create a context (the following setting is recommended.)
  nbla::Context cpu_ctx{{"cpu:float"}, "CpuCachedArray", "0"};
#ifdef WITH_CUDA
  nbla::init_cudnn();
  nbla::Context ctx{
      {"cudnn:float", "cuda:float", "cpu:float"}, "CudaCachedArray", "0"};
#else
  nbla::Context ctx = cpu_ctx;
#endif

  // Create a Nnp object
  nbla::utils::nnp::Nnp nnp(ctx);

  // Set nnp file to Nnp object.
  nnp.add(nnp_file);

  auto optimizer = nnp.get_optimizer(optimizer_name);
  auto monitor = nnp.get_monitor(monitor_name);

#ifdef TIMING
#ifdef WITH_CUDA
  nbla::cuda_device_synchronize("0");
#endif
  // Timing starts
  auto start = std::chrono::steady_clock::now();
#endif

  for (int i = 0; i < iter_num; ++i) {
    float loss = optimizer->update(i);
    printf("loss=%f\n", loss);
  }

  // called for each epoch
  float avg = monitor->monitor_epoch();
  printf("epoch_avg=%f\n", avg);


#ifdef TIMING
#ifdef WITH_CUDA
  nbla::cuda_device_synchronize("0");
#endif
  // Timing ends
  auto end = std::chrono::steady_clock::now();
  std::cout << "Elapsed time: "
            << std::chrono::duration_cast<std::chrono::microseconds>(end - start).count() * 0.001
            << " [ms]." << std::endl;
#endif

  return 0;
}