updating the PR

Author: KarelVesely84

egs/wsj/s5/steps/nnet3/chain/train.py

@@ -239,9 +239,9 @@ def process_args(args):
  If you have GPUs and have nvcc installed, go to src/ and do
  ./configure; make""")
 
- run_opts.train_queue_opt = "--gpu 1 --mem 10G"
+ run_opts.train_queue_opt = "--gpu 1"
  run_opts.parallel_train_opts = "--use-gpu={}".format(args.use_gpu)
- run_opts.combine_queue_opt = "--gpu 1 --mem 10G"
+ run_opts.combine_queue_opt = "--gpu 1"
  run_opts.combine_gpu_opt = "--use-gpu={}".format(args.use_gpu)
 
  else:

src/nnet3/decodable-online-looped.cc

@@ -162,11 +162,6 @@ void DecodableNnetLoopedOnlineBase::AdvanceChunk() {
  input_frame = num_feature_frames_ready - 1;
  input_features_->GetFrame(input_frame, &this_row);
  }
-
- // dump 1st chunk of features (debug),
- if (num_chunks_computed_ == 0)
- KALDI_VLOG(100) << "feature-dump " << this_feats;
-
  feats_chunk.Swap(&this_feats);
  }
  computer_.AcceptInput("input", &feats_chunk);
@@ -205,11 +200,6 @@ void DecodableNnetLoopedOnlineBase::AdvanceChunk() {
  Matrix<BaseFloat> ivectors(num_ivectors,
  ivector.Dim());
  ivectors.CopyRowsFromVec(ivector);
-
- // dump 1st chunk of ivectors (debug),
- if (num_chunks_computed_ == 0)
- KALDI_VLOG(100) << "ivector-dump " << ivectors;
-
  CuMatrix<BaseFloat> cu_ivectors;
  cu_ivectors.Swap(&ivectors);
  computer_.AcceptInput("ivector", &cu_ivectors);

src/online2bin/online2-tcp-nnet3-decode-faster.cc

@@ -38,7 +38,6 @@
 #include <arpa/inet.h>
 #include <unistd.h>
 #include <string>
-#include <iomanip>
 
 namespace kaldi {
 
@@ -86,25 +85,20 @@ std::string LatticeToString(const Lattice &lat, const fst::SymbolTable &word_sym
  return msg.str();
 }
 
-std::string GetTimeString(int32 t_beg, int32 t_end, int32 subsamp) {
+std::string GetTimeString(int32 t_beg, int32 t_end, BaseFloat time_unit) {
  char buffer[100];
- double t_beg2 = t_beg * subsamp / 100.;
- double t_end2 = t_end * subsamp / 100.;
- snprintf(buffer, 100, "{ %5.2f --> %5.2f }", t_beg2, t_end2);
+ double t_beg2 = t_beg * time_unit;
+ double t_end2 = t_end * time_unit;
+ snprintf(buffer, 100, "%.2f %.2f", t_beg2, t_end2);
  return std::string(buffer);
 }
 
 int32 GetLatticeTimeSpan(const Lattice& lat) {
- // convert the lattice,
- CompactLattice clat;
- ConvertLattice(lat, &clat);
- // get the word-alignment,
+ LatticeWeight weight;
+ std::vector<int32> alignment;
  std::vector<int32> words;
- std::vector<int32> begin_times;
- std::vector<int32> lengths;
- CompactLatticeToWordAlignment(clat, &words, &begin_times, &lengths);
- // get ending time of last word,
- return begin_times.back() + lengths.back();
+ GetLinearSymbolSequence(lat, &alignment, &words, &weight);
+ return alignment.size();
 }
 
 std::string LatticeToString(const CompactLattice &clat, const fst::SymbolTable &word_syms) {
@@ -169,7 +163,7 @@ int main(int argc, char *argv[]) {
  po.Register("port-num", &port_num,
  "Port number the server will listen on.");
  po.Register("produce-time", &produce_time,
- "Send 'sentence' begin/end times based on end-points");
+ "Prepend begin/end times between endpoints (e.g. '5.46 6.81 <text_output>0', in seconds)");
 
  feature_opts.Register(&po);
  decodable_opts.Register(&po);
@@ -189,6 +183,9 @@ int main(int argc, char *argv[]) {
 
  OnlineNnet2FeaturePipelineInfo feature_info(feature_opts);
 
+ BaseFloat frame_shift = feature_info.FrameShiftInSeconds();
+ int32 frame_subsampling = decodable_opts.frame_subsampling_factor;
+
  KALDI_VLOG(1) << "Loading AM...";
 
  TransitionModel trans_model;
@@ -256,7 +253,6 @@ int main(int argc, char *argv[]) {
  eos = !server.ReadChunk(chunk_len);
 
  if (eos) {
- KALDI_VLOG(2) << "eos detected";
  feature_pipeline.InputFinished();
  decoder.AdvanceDecoding();
  decoder.FinalizeDecoding();
@@ -270,8 +266,7 @@ int main(int argc, char *argv[]) {
  if (produce_time) {
  int32 t_beg = frame_offset - decoder.NumFramesDecoded();
  int32 t_end = frame_offset;
- int32 s = decodable_opts.frame_subsampling_factor;
- msg = GetTimeString(t_beg, t_end, s) + " " + msg;
+ msg = GetTimeString(t_beg, t_end, frame_shift * frame_subsampling) + " " + msg;
  }
 
  KALDI_VLOG(1) << "EndOfAudio, sending message: " << msg;
@@ -284,13 +279,7 @@ int main(int argc, char *argv[]) {
 
  Vector<BaseFloat> wave_part = server.GetChunk();
  feature_pipeline.AcceptWaveform(samp_freq, wave_part);
-
- // dump 1st chunk of 'raw' audio (debug),
- if (samp_count == 0)
- KALDI_VLOG(100) << "raw audio " << wave_part;
-
  samp_count += chunk_len;
- KALDI_VLOG(2) << "samp_count " << samp_count;
 
  if (silence_weighting.Active() &&
  feature_pipeline.IvectorFeature() != NULL) {
@@ -299,10 +288,8 @@ int main(int argc, char *argv[]) {
  &delta_weights);
  feature_pipeline.UpdateFrameWeights(delta_weights,
  frame_offset * decodable_opts.frame_subsampling_factor);
- KALDI_VLOG(2) << "silence weighting";
  }
 
- KALDI_VLOG(2) << "Advance decoding";
  decoder.AdvanceDecoding();
 
  if (samp_count > check_count) {
@@ -315,8 +302,7 @@ int main(int argc, char *argv[]) {
  if (produce_time) {
  int32 t_beg = frame_offset;
  int32 t_end = frame_offset + GetLatticeTimeSpan(lat);
- int32 s = decodable_opts.frame_subsampling_factor;
- msg = GetTimeString(t_beg, t_end, s) + " " + msg;
+ msg = GetTimeString(t_beg, t_end, frame_shift * frame_subsampling) + " " + msg;
  }
 
  KALDI_VLOG(1) << "Temporary transcript: " << msg;
@@ -336,8 +322,7 @@ int main(int argc, char *argv[]) {
  if (produce_time) {
  int32 t_beg = frame_offset - decoder.NumFramesDecoded();
  int32 t_end = frame_offset;
- int32 s = decodable_opts.frame_subsampling_factor;
- msg = GetTimeString(t_beg, t_end, s) + " " + msg;
+ msg = GetTimeString(t_beg, t_end, frame_shift * frame_subsampling) + " " + msg;
  }
 
  KALDI_VLOG(1) << "Endpoint, sending message: " << msg;