nnet-am-average.cc File Reference

#include <algorithm>
#include "base/kaldi-common.h"
#include "util/common-utils.h"
#include "hmm/transition-model.h"
#include "nnet2/combine-nnet-a.h"
#include "nnet2/am-nnet.h"

Include dependency graph for nnet-am-average.cc:

Go to the source code of this file.

Namespaces
	kaldi
	This code computes Goodness of Pronunciation (GOP) and extracts phone-level pronunciation feature for mispronunciations detection tasks, the reference:

Functions
void	GetWeights (const std::string &weights_str, int32 num_inputs, std::vector< BaseFloat > *weights)
std::vector< bool >	GetSkipLayers (const std::string &skip_layers_str, const int32 first_layer_idx, const int32 last_layer_idx)
int	main (int argc, char *argv[])

Function Documentation

main()

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

Definition at line 112 of file nnet-am-average.cc.

References UpdatableComponent::Add(), NonlinearComponent::Add(), ParseOptions::GetArg(), Nnet::GetComponent(), AmNnet::GetNnet(), kaldi::GetSkipLayers(), kaldi::GetWeights(), rnnlm::i, KALDI_ASSERT, KALDI_LOG, KALDI_VLOG, Nnet::LastUpdatableComponent(), ParseOptions::NumArgs(), Nnet::NumComponents(), ParseOptions::PrintUsage(), AmNnet::Read(), ParseOptions::Read(), TransitionModel::Read(), ParseOptions::Register(), UpdatableComponent::Scale(), NonlinearComponent::Scale(), Output::Stream(), Input::Stream(), AmNnet::Write(), and TransitionModel::Write().

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

    const char *usage =
        "This program averages (or sums, if --sum=true) the parameters over a\n"
        "number of neural nets.  If you supply the option --skip-last-layer=true,\n"
        "the parameters of the last updatable layer are copied from <model1> instead\n"
        "of being averaged (useful in multi-language scenarios).\n"
        "The --weights option can be used to weight each model differently.\n"
        "\n"
        "Usage:  nnet-am-average [options] <model1> <model2> ... <modelN> <model-out>\n"
        "\n"
        "e.g.:\n"
        " nnet-am-average 1.1.nnet 1.2.nnet 1.3.nnet 2.nnet\n";

    bool binary_write = true;
    bool sum = false;

    ParseOptions po(usage);
    po.Register("sum", &sum, "If true, sums instead of averages.");
    po.Register("binary", &binary_write, "Write output in binary mode");
    string weights_str;
    bool skip_last_layer = false;
    string skip_layers_str;
    po.Register("weights", &weights_str, "Colon-separated list of weights, one "
                "for each input model.  These will be normalized to sum to one.");
    po.Register("skip-last-layer", &skip_last_layer, "If true, averaging of "
                "the last updatable layer is skipped (result comes from model1)");
    po.Register("skip-layers", &skip_layers_str, "Colon-separated list of "
                "indices of the layers that should be skipped during averaging."
                "Be careful: this parameter uses an absolute indexing of "
                "layers, i.e. iterates over all components, not over updatable "
                "ones only.");

    po.Read(argc, argv);

    if (po.NumArgs() < 2) {
      po.PrintUsage();
      exit(1);
    }

    std::string
        nnet1_rxfilename = po.GetArg(1),
        nnet_wxfilename = po.GetArg(po.NumArgs());

    TransitionModel trans_model1;
    AmNnet am_nnet1;
    {
      bool binary_read;
      Input ki(nnet1_rxfilename, &binary_read);
      trans_model1.Read(ki.Stream(), binary_read);
      am_nnet1.Read(ki.Stream(), binary_read);
    }

    int32 num_inputs = po.NumArgs() - 1;

    std::vector<BaseFloat> model_weights;
    GetWeights(weights_str, num_inputs, &model_weights);

    int32 c_begin = 0,
        c_end = (skip_last_layer ?
                 am_nnet1.GetNnet().LastUpdatableComponent() :
                 am_nnet1.GetNnet().NumComponents());
    KALDI_ASSERT(c_end != -1 && "Network has no updatable components.");

    int32 last_layer_idx = am_nnet1.GetNnet().NumComponents();
    std::vector<bool> skip_layers = GetSkipLayers(skip_layers_str,
                                             0,
                                             last_layer_idx);

    // scale the components - except the last layer, if skip_last_layer == true.
    for (int32 c = c_begin; c < c_end; c++) {
      if (skip_layers[c]) {
        KALDI_VLOG(2) << "Not averaging layer " << c << " (as requested)";
        continue;
      }
      bool updated = false;
      UpdatableComponent *uc =
        dynamic_cast<UpdatableComponent*>(&(am_nnet1.GetNnet().GetComponent(c)));
      if (uc != NULL)  {
        KALDI_VLOG(2) << "Averaging layer " << c << " (UpdatableComponent)";
        uc->Scale(model_weights[0]);
        updated = true;
      }
      NonlinearComponent *nc =
        dynamic_cast<NonlinearComponent*>(&(am_nnet1.GetNnet().GetComponent(c)));
      if (nc != NULL) {
        KALDI_VLOG(2) << "Averaging layer " << c << " (NonlinearComponent)";
        nc->Scale(model_weights[0]);
        updated = true;
      }
      if (! updated) {
        KALDI_VLOG(2) << "Not averaging layer " << c
          << " (unscalable component)";
      }
    }

    for (int32 i = 2; i <= num_inputs; i++) {
      bool binary_read;
      Input ki(po.GetArg(i), &binary_read);
      TransitionModel trans_model;
      trans_model.Read(ki.Stream(), binary_read);
      AmNnet am_nnet;
      am_nnet.Read(ki.Stream(), binary_read);

      for (int32 c = c_begin; c < c_end; c++) {
        if (skip_layers[c]) continue;

        UpdatableComponent *uc_average =
          dynamic_cast<UpdatableComponent*>(&(am_nnet1.GetNnet().GetComponent(c)));
        const UpdatableComponent *uc_this =
          dynamic_cast<const UpdatableComponent*>(&(am_nnet.GetNnet().GetComponent(c)));
        if (uc_average != NULL) {
          KALDI_ASSERT(uc_this != NULL &&
                       "Networks must have the same structure.");
          uc_average->Add(model_weights[i-1], *uc_this);
        }

        NonlinearComponent *nc_average =
          dynamic_cast<NonlinearComponent*>(&(am_nnet1.GetNnet().GetComponent(c)));
        const NonlinearComponent *nc_this =
          dynamic_cast<const NonlinearComponent*>(&(am_nnet.GetNnet().GetComponent(c)));
        if (nc_average != NULL) {
          KALDI_ASSERT(nc_this != NULL &&
                       "Networks must have the same structure.");
          nc_average->Add(model_weights[i-1], *nc_this);
        }
      }
    }

    {
      Output ko(nnet_wxfilename, binary_write);
      trans_model1.Write(ko.Stream(), binary_write);
      am_nnet1.Write(ko.Stream(), binary_write);
    }

    KALDI_LOG << "Averaged parameters of " << num_inputs
              << " neural nets, and wrote to " << nnet_wxfilename;
    return 0; // it will throw an exception if there are any problems.
  } catch(const std::exception &e) {
    std::cerr << e.what() << '\n';
    return -1;
  }
}