nnet-train-parallel.cc File Reference

#include "base/kaldi-common.h"
#include "util/common-utils.h"
#include "hmm/transition-model.h"
#include "nnet2/nnet-update-parallel.h"
#include "nnet2/am-nnet.h"

Include dependency graph for nnet-train-parallel.cc:

Go to the source code of this file.

Functions
int	main (int argc, char *argv[])

Function Documentation

main()

int main	(	int	argc,
		char *	argv[]
	)

Definition at line 27 of file nnet-train-parallel.cc.

References kaldi::nnet2::DoBackpropParallel(), kaldi::g_num_threads, ParseOptions::GetArg(), AmNnet::GetNnet(), KALDI_ASSERT, KALDI_LOG, ParseOptions::NumArgs(), ParseOptions::PrintUsage(), AmNnet::Read(), ParseOptions::Read(), TransitionModel::Read(), ParseOptions::Register(), Output::Stream(), Input::Stream(), AmNnet::Write(), TransitionModel::Write(), and Nnet::ZeroStats().

                                 {
  try {
    using namespace kaldi;
    using namespace kaldi::nnet2;
    typedef kaldi::int32 int32;
    typedef kaldi::int64 int64;

    const char *usage =
        "Train the neural network parameters with backprop and stochastic\n"
        "gradient descent using minibatches.  As nnet-train-simple, but\n"
        "uses multiple threads in a Hogwild type of update (for CPU, not GPU).\n"
        "\n"
        "Usage:  nnet-train-parallel [options] <model-in> <training-examples-in> <model-out>\n"
        "\n"
        "e.g.:\n"
        "nnet-train-parallel --num-threads=8 1.nnet ark:1.1.egs 2.nnet\n";
    
    bool binary_write = true;
    bool zero_stats = true;
    int32 minibatch_size = 1024;
    int32 srand_seed = 0;
    
    ParseOptions po(usage);
    po.Register("binary", &binary_write, "Write output in binary mode");
    po.Register("zero-stats", &zero_stats, "If true, zero stats "
                "stored with the neural net (only affects mixing up).");
    po.Register("srand", &srand_seed,
                "Seed for random number generator (e.g., for dropout)");
    po.Register("num-threads", &g_num_threads, "Number of training threads to use "
                "in the parallel update. [Note: if you use a parallel "
                "implementation of BLAS, the actual number of threads may be larger.]");
    po.Register("minibatch-size", &minibatch_size, "Number of examples to use for "
                "each minibatch during training.");
    
    po.Read(argc, argv);
    srand(srand_seed);

    if (po.NumArgs() != 3) {
      po.PrintUsage();
      exit(1);
    }
    
    std::string nnet_rxfilename = po.GetArg(1),
        examples_rspecifier = po.GetArg(2),
        nnet_wxfilename = po.GetArg(3);

    TransitionModel trans_model;
    AmNnet am_nnet;
    {
      bool binary_read;
      Input ki(nnet_rxfilename, &binary_read);
      trans_model.Read(ki.Stream(), binary_read);
      am_nnet.Read(ki.Stream(), binary_read);
    }

    KALDI_ASSERT(minibatch_size > 0);

    if (zero_stats) am_nnet.GetNnet().ZeroStats();

    double num_examples = 0;
    SequentialNnetExampleReader example_reader(examples_rspecifier);
    

    DoBackpropParallel(am_nnet.GetNnet(),
                       minibatch_size,
                       &example_reader,
                       &num_examples,
                       &(am_nnet.GetNnet()));
    
    {
      Output ko(nnet_wxfilename, binary_write);
      trans_model.Write(ko.Stream(), binary_write);
      am_nnet.Write(ko.Stream(), binary_write);
    }
    
    KALDI_LOG << "Finished training, processed " << num_examples
              << " training examples (weighted).  Wrote model to "
              << nnet_wxfilename;
    return (num_examples == 0 ? 1 : 0);
  } catch(const std::exception &e) {
    std::cerr << e.what() << '\n';
    return -1;
  }
}