nnet-forward.cc File Reference

#include <limits>
#include "nnet/nnet-nnet.h"
#include "nnet/nnet-loss.h"
#include "nnet/nnet-pdf-prior.h"
#include "base/kaldi-common.h"
#include "util/common-utils.h"
#include "base/timer.h"

Include dependency graph for nnet-forward.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 30 of file nnet-forward.cc.

References CuMatrixBase< Real >::Add(), CuMatrixBase< Real >::ApplyLog(), PdfPriorOptions::class_frame_counts, SequentialTableReader< Holder >::Done(), Timer::Elapsed(), Nnet::Feedforward(), ParseOptions::GetArg(), Nnet::GetLastComponent(), Component::GetType(), kaldi::GetVerboseLevel(), KALDI_ERR, KALDI_ISFINITE, KALDI_LOG, KALDI_VLOG, KALDI_WARN, Component::kBlockSoftmax, SequentialTableReader< Holder >::Key(), Component::kSoftmax, CuMatrixBase< Real >::Max(), CuMatrixBase< Real >::Min(), SequentialTableReader< Holder >::Next(), ParseOptions::NumArgs(), MatrixBase< Real >::NumRows(), ParseOptions::PrintUsage(), ParseOptions::Read(), Nnet::Read(), PdfPriorOptions::Register(), ParseOptions::Register(), Nnet::RemoveLastComponent(), Nnet::SetDropoutRate(), PdfPrior::SubtractOnLogpost(), MatrixBase< Real >::Sum(), CuMatrixBase< Real >::Sum(), Component::TypeToMarker(), SequentialTableReader< Holder >::Value(), and TableWriter< Holder >::Write().

                                 {
  using namespace kaldi;
  using namespace kaldi::nnet1;
  try {
    const char *usage =
      "Perform forward pass through Neural Network.\n"
      "Usage: nnet-forward [options] <nnet1-in> <feature-rspecifier> <feature-wspecifier>\n"
      "e.g.: nnet-forward final.nnet ark:input.ark ark:output.ark\n";

    ParseOptions po(usage);

    PdfPriorOptions prior_opts;
    prior_opts.Register(&po);

    std::string feature_transform;
    po.Register("feature-transform", &feature_transform,
        "Feature transform in front of main network (in nnet format)");

    bool no_softmax = false;
    po.Register("no-softmax", &no_softmax,
        "Removes the last component with Softmax, if found. The pre-softmax "
        "activations are the output of the network. Decoding them leads to "
        "the same lattices as if we had used 'log-posteriors'.");

    bool apply_log = false;
    po.Register("apply-log", &apply_log, "Transform NN output by log()");

    std::string use_gpu="no";
    po.Register("use-gpu", &use_gpu,
        "yes|no|optional, only has effect if compiled with CUDA");

    using namespace kaldi;
    using namespace kaldi::nnet1;
    typedef kaldi::int32 int32;

    po.Read(argc, argv);

    if (po.NumArgs() != 3) {
      po.PrintUsage();
      exit(1);
    }

    std::string model_filename = po.GetArg(1),
        feature_rspecifier = po.GetArg(2),
        feature_wspecifier = po.GetArg(3);

    // Select the GPU
#if HAVE_CUDA == 1
    CuDevice::Instantiate().SelectGpuId(use_gpu);
#endif

    Nnet nnet_transf;
    if (feature_transform != "") {
      nnet_transf.Read(feature_transform);
    }

    Nnet nnet;
    nnet.Read(model_filename);
    // optionally remove softmax,
    Component::ComponentType last_comp_type = nnet.GetLastComponent().GetType();
    if (no_softmax) {
      if (last_comp_type == Component::kSoftmax ||
          last_comp_type == Component::kBlockSoftmax) {
        KALDI_LOG << "Removing " << Component::TypeToMarker(last_comp_type)
                  << " from the nnet " << model_filename;
        nnet.RemoveLastComponent();
      } else {
        KALDI_WARN << "Last component 'NOT-REMOVED' by --no-softmax=true, "
          << "the component was " << Component::TypeToMarker(last_comp_type);
      }
    }

    // avoid some bad option combinations,
    if (apply_log && no_softmax) {
      KALDI_ERR << "Cannot use both --apply-log=true --no-softmax=true, "
                << "use only one of the two!";
    }

    // we will subtract log-priors later,
    PdfPrior pdf_prior(prior_opts);

    // disable dropout,
    nnet_transf.SetDropoutRate(0.0);
    nnet.SetDropoutRate(0.0);

    kaldi::int64 tot_t = 0;

    SequentialBaseFloatMatrixReader feature_reader(feature_rspecifier);
    BaseFloatMatrixWriter feature_writer(feature_wspecifier);

    CuMatrix<BaseFloat> feats, feats_transf, nnet_out;
    Matrix<BaseFloat> nnet_out_host;

    Timer time;
    double time_now = 0;
    int32 num_done = 0;

    // main loop,
    for (; !feature_reader.Done(); feature_reader.Next()) {
      // read
      Matrix<BaseFloat> mat = feature_reader.Value();
      std::string utt = feature_reader.Key();
      KALDI_VLOG(2) << "Processing utterance " << num_done+1
                    << ", " << utt
                    << ", " << mat.NumRows() << "frm";


      if (!KALDI_ISFINITE(mat.Sum())) {  // check there's no nan/inf,
        KALDI_ERR << "NaN or inf found in features for " << utt;
      }

      // push it to gpu,
      feats = mat;

      // fwd-pass, feature transform,
      nnet_transf.Feedforward(feats, &feats_transf);
      if (!KALDI_ISFINITE(feats_transf.Sum())) {  // check there's no nan/inf,
        KALDI_ERR << "NaN or inf found in transformed-features for " << utt;
      }

      // fwd-pass, nnet,
      nnet.Feedforward(feats_transf, &nnet_out);
      if (!KALDI_ISFINITE(nnet_out.Sum())) {  // check there's no nan/inf,
        KALDI_ERR << "NaN or inf found in nn-output for " << utt;
      }

      // convert posteriors to log-posteriors,
      if (apply_log) {
        if (!(nnet_out.Min() >= 0.0 && nnet_out.Max() <= 1.0)) {
          KALDI_WARN << "Applying 'log()' to data which don't seem to be "
                     << "probabilities," << utt;
        }
        nnet_out.Add(1e-20);  // avoid log(0),
        nnet_out.ApplyLog();
      }

      // subtract log-priors from log-posteriors or pre-softmax,
      if (prior_opts.class_frame_counts != "") {
        pdf_prior.SubtractOnLogpost(&nnet_out);
      }

      // download from GPU,
      nnet_out_host = Matrix<BaseFloat>(nnet_out);

      // write,
      if (!KALDI_ISFINITE(nnet_out_host.Sum())) {  // check there's no nan/inf,
        KALDI_ERR << "NaN or inf found in final output nn-output for " << utt;
      }
      feature_writer.Write(feature_reader.Key(), nnet_out_host);

      // progress log,
      if (num_done % 100 == 0) {
        time_now = time.Elapsed();
        KALDI_VLOG(1) << "After " << num_done << " utterances: time elapsed = "
                      << time_now/60 << " min; processed " << tot_t/time_now
                      << " frames per second.";
      }
      num_done++;
      tot_t += mat.NumRows();
    }

    // final message,
    KALDI_LOG << "Done " << num_done << " files"
              << " in " << time.Elapsed()/60 << "min,"
              << " (fps " << tot_t/time.Elapsed() << ")";

#if HAVE_CUDA == 1
    if (GetVerboseLevel() >= 1) {
      CuDevice::Instantiate().PrintProfile();
    }
#endif

    if (num_done == 0) return -1;
    return 0;
  } catch(const std::exception &e) {
    std::cerr << e.what();
    return -1;
  }
}