nnet-am-compute.cc File Reference

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

Include dependency graph for nnet-am-compute.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-am-compute.cc.

References MatrixBase< Real >::ApplyFloor(), MatrixBase< Real >::ApplyLog(), SequentialTableReader< Holder >::Done(), ParseOptions::GetArg(), AmNnet::GetNnet(), rnnlm::i, KALDI_ASSERT, KALDI_LOG, KALDI_WARN, SequentialTableReader< Holder >::Key(), Nnet::LeftContext(), MatrixBase< Real >::MulColsVec(), SequentialTableReader< Holder >::Next(), kaldi::nnet2::NnetComputation(), kaldi::nnet2::NnetComputationChunked(), ParseOptions::NumArgs(), AmNnet::NumPdfs(), MatrixBase< Real >::NumRows(), Nnet::OutputDim(), ParseOptions::PrintUsage(), AmNnet::Priors(), AmNnet::Read(), ParseOptions::Read(), TransitionModel::Read(), ParseOptions::Register(), Nnet::RightContext(), VectorBase< Real >::Scale(), Input::Stream(), VectorBase< Real >::Sum(), CuMatrix< Real >::Swap(), SequentialTableReader< Holder >::Value(), and TableWriter< Holder >::Write().

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

    const char *usage =
        "Does the neural net computation for each file of input features, and\n"
        "outputs as a matrix the result.  Used mostly for debugging.\n"
        "Note: if you want it to apply a log (e.g. for log-likelihoods), use\n"
        "--apply-log=true\n"
        "\n"
        "Usage:  nnet-am-compute [options] <model-in> <feature-rspecifier> "
        "<feature-or-loglikes-wspecifier>\n"
        "See also: nnet-compute, nnet-logprob\n";

    bool divide_by_priors = false;
    bool apply_log = false;
    bool pad_input = true;
    std::string use_gpu = "no";
    int32 chunk_size = 0;
    ParseOptions po(usage);
    po.Register("divide-by-priors", &divide_by_priors, "If true, "
                "divide by the priors stored in the model and re-normalize, apply-log may follow");
    po.Register("apply-log", &apply_log, "Apply a log to the result of the computation "
                "before outputting.");
    po.Register("pad-input", &pad_input, "If true, duplicate the first and last frames "
                "of input features as required for temporal context, to prevent #frames "
                "of output being less than those of input.");
    po.Register("use-gpu", &use_gpu,
                "yes|no|optional|wait, only has effect if compiled with CUDA");
    po.Register("chunk-size", &chunk_size, "Process the feature matrix in chunks.  "
                "This is useful when processing large feature files in the GPU.  "
                "If chunk-size > 0, pad-input must be true.");

    po.Read(argc, argv);

    if (po.NumArgs() != 3) {
      po.PrintUsage();
      exit(1);
    }
    // If chunk_size is greater than 0, pad_input needs to be true.
    KALDI_ASSERT(chunk_size < 0 || pad_input);

#if HAVE_CUDA==1
    CuDevice::Instantiate().SelectGpuId(use_gpu);
#endif

    std::string nnet_rxfilename = po.GetArg(1),
        features_rspecifier = po.GetArg(2),
        features_or_loglikes_wspecifier = 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);
    }

    Nnet &nnet = am_nnet.GetNnet();

    int64 num_done = 0, num_frames = 0;

    Vector<BaseFloat> inv_priors(am_nnet.Priors());
    KALDI_ASSERT((!divide_by_priors || inv_priors.Dim() == am_nnet.NumPdfs()) &&
                 "Priors in neural network not set up.");
    inv_priors.ApplyPow(-1.0);

    SequentialBaseFloatMatrixReader feature_reader(features_rspecifier);
    BaseFloatMatrixWriter writer(features_or_loglikes_wspecifier);

    for (; !feature_reader.Done();  feature_reader.Next()) {
      std::string utt = feature_reader.Key();
      const Matrix<BaseFloat> &feats  = feature_reader.Value();

      int32 output_frames = feats.NumRows(), output_dim = nnet.OutputDim();
      if (!pad_input)
        output_frames -= nnet.LeftContext() + nnet.RightContext();
      if (output_frames <= 0) {
        KALDI_WARN << "Skipping utterance " << utt << " because output "
                   << "would be empty.";
        continue;
      }

      Matrix<BaseFloat> output(output_frames, output_dim);
      CuMatrix<BaseFloat> cu_feats(feats);
      CuMatrix<BaseFloat> cu_output(output);
      if (chunk_size > 0 && chunk_size < feats.NumRows()) {      
        NnetComputationChunked(nnet, cu_feats, chunk_size, &cu_output);
      } else {
        NnetComputation(nnet, cu_feats, pad_input, &cu_output);
      }
      cu_output.Swap(&output);
      
      if (divide_by_priors) {
        output.MulColsVec(inv_priors); // scales each column by the corresponding element
        // of inv_priors.
        for (int32 i = 0; i < output.NumRows(); i++) {
          SubVector<BaseFloat> frame(output, i);
          BaseFloat p = frame.Sum();
          if (!(p > 0.0)) {
            KALDI_WARN << "Bad sum of probabilities " << p;
          } else {
            frame.Scale(1.0 / p); // re-normalize to sum to one.
          }
        }
      }

      if (apply_log) {
        output.ApplyFloor(1.0e-20);
        output.ApplyLog();
      }
      writer.Write(utt, output);
      num_frames += feats.NumRows();
      num_done++;
    }
#if HAVE_CUDA==1
    CuDevice::Instantiate().PrintProfile();
#endif

    KALDI_LOG << "Processed " << num_done << " feature files, "
              << num_frames << " frames of input were processed.";

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