// nnetbin/nnet-train-mpe-sequential.cc // Copyright 2011-2016 Brno University of Technology (author: Karel Vesely); // Arnab Ghoshal // See ../../COPYING for clarification regarding multiple authors // // 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 // // THIS CODE IS PROVIDED *AS IS* BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY // KIND, EITHER EXPRESS OR IMPLIED, INCLUDING WITHOUT LIMITATION ANY IMPLIED // WARRANTIES OR CONDITIONS OF TITLE, FITNESS FOR A PARTICULAR PURPOSE, // MERCHANTABLITY OR NON-INFRINGEMENT. // See the Apache 2 License for the specific language governing permissions and // limitations under the License. #include "base/kaldi-common.h" #include "util/common-utils.h" #include "tree/context-dep.h" #include "hmm/transition-model.h" #include "fstext/fstext-lib.h" #include "decoder/faster-decoder.h" #include "decoder/decodable-matrix.h" #include "lat/kaldi-lattice.h" #include "lat/lattice-functions.h" #include "nnet/nnet-trnopts.h" #include "nnet/nnet-component.h" #include "nnet/nnet-activation.h" #include "nnet/nnet-nnet.h" #include "nnet/nnet-pdf-prior.h" #include "nnet/nnet-utils.h" #include "base/timer.h" #include "cudamatrix/cu-device.h" namespace kaldi { namespace nnet1 { void LatticeAcousticRescore(const Matrix &log_like, const TransitionModel &trans_model, const std::vector &state_times, Lattice *lat) { kaldi::uint64 props = lat->Properties(fst::kFstProperties, false); if (!(props & fst::kTopSorted)) KALDI_ERR << "Input lattice must be topologically sorted."; KALDI_ASSERT(!state_times.empty()); std::vector > time_to_state(log_like.NumRows()); for (size_t i = 0; i < state_times.size(); i++) { KALDI_ASSERT(state_times[i] >= 0); if (state_times[i] < log_like.NumRows()) // end state may be past this.. time_to_state[state_times[i]].push_back(i); else KALDI_ASSERT(state_times[i] == log_like.NumRows() && "There appears to be lattice/feature mismatch."); } for (int32 t = 0; t < log_like.NumRows(); t++) { for (size_t i = 0; i < time_to_state[t].size(); i++) { int32 state = time_to_state[t][i]; for (fst::MutableArcIterator aiter(lat, state); !aiter.Done(); aiter.Next()) { LatticeArc arc = aiter.Value(); int32 trans_id = arc.ilabel; if (trans_id != 0) { // Non-epsilon input label on arc int32 pdf_id = trans_model.TransitionIdToPdf(trans_id); arc.weight.SetValue2(-log_like(t, pdf_id) + arc.weight.Value2()); aiter.SetValue(arc); } } } } } } // namespace nnet1 } // namespace kaldi int main(int argc, char *argv[]) { using namespace kaldi; using namespace kaldi::nnet1; typedef kaldi::int32 int32; try { const char *usage = "Perform one iteration of MPE/sMBR training using SGD with per-utterance" "updates.\n" "Usage: nnet-train-mpe-sequential [options] " " " " []\n" "e.g.: nnet-train-mpe-sequential nnet.init trans.mdl scp:feats.scp " "scp:denlats.scp ark:ali.ark nnet.iter1\n"; ParseOptions po(usage); NnetTrainOptions trn_opts; trn_opts.learn_rate = 0.00001; // changing default, trn_opts.Register(&po); bool binary = true; po.Register("binary", &binary, "Write output in binary mode"); std::string feature_transform; po.Register("feature-transform", &feature_transform, "Feature transform in 'nnet1' format"); std::string silence_phones_str; po.Register("silence-phones", &silence_phones_str, "Colon-separated list of integer id's of silence phones, e.g. 46:47"); PdfPriorOptions prior_opts; prior_opts.Register(&po); BaseFloat acoustic_scale = 1.0, lm_scale = 1.0, old_acoustic_scale = 0.0; po.Register("acoustic-scale", &acoustic_scale, "Scaling factor for acoustic likelihoods"); po.Register("lm-scale", &lm_scale, "Scaling factor for \"graph costs\" (including LM costs)"); po.Register("old-acoustic-scale", &old_acoustic_scale, "Add in the scores in the input lattices with this scale, rather " "than discarding them."); bool one_silence_class = false; po.Register("one-silence-class", &one_silence_class, "If true, the newer behavior reduces insertions."); kaldi::int32 max_frames = 6000; po.Register("max-frames", &max_frames, "Maximum number of frames an utterance can have (skipped if longer)"); bool do_smbr = false; po.Register("do-smbr", &do_smbr, "Use state-level accuracies instead of phone accuracies."); std::string use_gpu="yes"; po.Register("use-gpu", &use_gpu, "yes|no|optional, only has effect if compiled with CUDA"); po.Read(argc, argv); if (po.NumArgs() != 6) { po.PrintUsage(); exit(1); } std::string model_filename = po.GetArg(1), transition_model_filename = po.GetArg(2), feature_rspecifier = po.GetArg(3), den_lat_rspecifier = po.GetArg(4), ref_ali_rspecifier = po.GetArg(5), target_model_filename = po.GetArg(6); std::vector silence_phones; if (!kaldi::SplitStringToIntegers(silence_phones_str, ":", false, &silence_phones)) { KALDI_ERR << "Invalid silence-phones string " << silence_phones_str; } kaldi::SortAndUniq(&silence_phones); if (silence_phones.empty()) { KALDI_LOG << "No silence phones specified."; } #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); // we will use pre-softmax activations, removing softmax, // - pre-softmax activations are equivalent to 'log-posterior + C_frame', // - all paths crossing a frame share same 'C_frame', // - with GMM, we also have the unnormalized acoustic likelihoods, if (nnet.GetLastComponent().GetType() == kaldi::nnet1::Component::kSoftmax) { KALDI_LOG << "Removing softmax from the nnet " << model_filename; nnet.RemoveLastComponent(); } else { KALDI_LOG << "The nnet was without softmax. " << "The last component in " << model_filename << " was " << Component::TypeToMarker(nnet.GetLastComponent().GetType()); } nnet.SetTrainOptions(trn_opts); // Read the class-frame-counts, compute priors, PdfPrior log_prior(prior_opts); // Read transition model, TransitionModel trans_model; ReadKaldiObject(transition_model_filename, &trans_model); SequentialBaseFloatMatrixReader feature_reader(feature_rspecifier); RandomAccessLatticeReader den_lat_reader(den_lat_rspecifier); RandomAccessInt32VectorReader ref_ali_reader(ref_ali_rspecifier); CuMatrix feats_transf, nnet_out, nnet_diff; Matrix nnet_out_h; Timer time; double time_now = 0; KALDI_LOG << "TRAINING STARTED"; int32 num_done = 0, num_no_ref_ali = 0, num_no_den_lat = 0, num_other_error = 0; kaldi::int64 total_frames = 0; double total_frame_acc = 0.0, utt_frame_acc; // main loop over utterances, for (; !feature_reader.Done(); feature_reader.Next()) { std::string utt = feature_reader.Key(); if (!den_lat_reader.HasKey(utt)) { KALDI_WARN << "Missing lattice for " << utt; num_no_den_lat++; continue; } if (!ref_ali_reader.HasKey(utt)) { KALDI_WARN << "Missing alignment for " << utt; num_no_ref_ali++; continue; } // 1) get the features, numerator alignment, const Matrix &mat = feature_reader.Value(); const std::vector &ref_ali = ref_ali_reader.Value(utt); // check duration of numerator alignments, if (static_cast(ref_ali.size()) != mat.NumRows()) { KALDI_WARN << "Duration mismatch!" << " alignment " << ref_ali.size() << " features " << mat.NumRows(); num_other_error++; continue; } if (mat.NumRows() > max_frames) { KALDI_WARN << "Skipping " << utt << " that has " << mat.NumRows() << " frames," << " it is longer than '--max-frames'" << max_frames; num_other_error++; continue; } // 2) get the denominator lattice, preprocess Lattice den_lat = den_lat_reader.Value(utt); if (den_lat.Start() == -1) { KALDI_WARN << "Empty lattice of " << utt << ", skipping."; num_other_error++; continue; } if (old_acoustic_scale != 1.0) { fst::ScaleLattice(fst::AcousticLatticeScale(old_acoustic_scale), &den_lat); } // optional sort it topologically kaldi::uint64 props = den_lat.Properties(fst::kFstProperties, false); if (!(props & fst::kTopSorted)) { if (fst::TopSort(&den_lat) == false) { KALDI_ERR << "Cycles detected in lattice."; } } // get the lattice length and times of states std::vector state_times; int32 max_time = kaldi::LatticeStateTimes(den_lat, &state_times); // check for temporal length of denominator lattices if (max_time != mat.NumRows()) { KALDI_WARN << "Duration mismatch!" << " denominator lattice " << max_time << " features " << mat.NumRows() << "," << " skipping " << utt; num_other_error++; continue; } // get dims, int32 num_frames = mat.NumRows(); // 3) get the pre-softmax outputs from NN, // apply transform, nnet_transf.Feedforward(CuMatrix(mat), &feats_transf); // propagate through the nnet (we know it's w/o softmax), nnet.Propagate(feats_transf, &nnet_out); // subtract the log_prior, if (prior_opts.class_frame_counts != "") { log_prior.SubtractOnLogpost(&nnet_out); } // transfer it back to the host, nnet_out_h = Matrix(nnet_out); // release the buffers we don't need anymore feats_transf.Resize(0, 0); nnet_out.Resize(0, 0); // 4) rescore the latice LatticeAcousticRescore(nnet_out_h, trans_model, state_times, &den_lat); if (acoustic_scale != 1.0 || lm_scale != 1.0) fst::ScaleLattice(fst::LatticeScale(lm_scale, acoustic_scale), &den_lat); kaldi::Posterior post; if (do_smbr) { // use state-level accuracies, i.e. sMBR estimation, utt_frame_acc = LatticeForwardBackwardMpeVariants( trans_model, silence_phones, den_lat, ref_ali, "smbr", one_silence_class, &post); } else { // use phone-level accuracies, i.e. MPFE (minimum phone frame error), utt_frame_acc = LatticeForwardBackwardMpeVariants( trans_model, silence_phones, den_lat, ref_ali, "mpfe", one_silence_class, &post); } // 6) convert the Posterior to a matrix, PosteriorToPdfMatrix(post, trans_model, &nnet_diff); nnet_diff.Scale(-1.0); // need to flip the sign of derivative, KALDI_VLOG(1) << "Lattice #" << num_done + 1 << " processed" << " (" << utt << "): found " << den_lat.NumStates() << " states and " << fst::NumArcs(den_lat) << " arcs."; KALDI_VLOG(1) << "Utterance " << utt << ": Average frame accuracy = " << (utt_frame_acc/num_frames) << " over " << num_frames << " frames," << " diff-range(" << nnet_diff.Min() << "," << nnet_diff.Max() << ")"; // 7) backpropagate through the nnet, update, nnet.Backpropagate(nnet_diff, NULL); nnet_diff.Resize(0, 0); // release GPU memory, // increase time counter total_frame_acc += utt_frame_acc; total_frames += num_frames; num_done++; if (num_done % 100 == 0) { time_now = time.Elapsed(); KALDI_VLOG(1) << "After " << num_done << " utterances: " << "time elapsed = " << time_now / 60 << " min; " << "processed " << total_frames / time_now << " frames per sec."; #if HAVE_CUDA == 1 // check that GPU computes accurately, CuDevice::Instantiate().CheckGpuHealth(); #endif } // GRADIENT LOGGING // First utterance, if (num_done == 1) { KALDI_VLOG(1) << nnet.InfoPropagate(); KALDI_VLOG(1) << nnet.InfoBackPropagate(); KALDI_VLOG(1) << nnet.InfoGradient(); } // Every 1000 utterances (--verbose=2), if (GetVerboseLevel() >= 2) { if (num_done % 1000 == 0) { KALDI_VLOG(2) << nnet.InfoPropagate(); KALDI_VLOG(2) << nnet.InfoBackPropagate(); KALDI_VLOG(2) << nnet.InfoGradient(); } } } // main loop over utterances, // After last utterance, KALDI_VLOG(1) << nnet.InfoPropagate(); KALDI_VLOG(1) << nnet.InfoBackPropagate(); KALDI_VLOG(1) << nnet.InfoGradient(); // Add the softmax layer back before writing, KALDI_LOG << "Appending the softmax " << target_model_filename; nnet.AppendComponentPointer(new Softmax(nnet.OutputDim(), nnet.OutputDim())); // Store the nnet, nnet.Write(target_model_filename, binary); time_now = time.Elapsed(); KALDI_LOG << "TRAINING FINISHED; " << "Time taken = " << time_now / 60 << " min; processed " << total_frames / time_now << " frames per second."; KALDI_LOG << "Done " << num_done << " files, " << num_no_ref_ali << " with no reference alignments, " << num_no_den_lat << " with no lattices, " << num_other_error << " with other errors."; KALDI_LOG << "Overall average frame-accuracy is " << total_frame_acc / total_frames << " over " << total_frames << " frames."; #if HAVE_CUDA == 1 CuDevice::Instantiate().PrintProfile(); #endif return 0; } catch(const std::exception &e) { std::cerr << e.what(); return -1; } }