20190325

.gitignore

@@ -73,15 +73,17 @@ GSYMS
 /src/kaldi.mk.bak
 
 # /egs/
-/egs/*/s*/mfcc
-/egs/*/s*/plp
-/egs/*/s*/exp
-/egs/*/s*/data
+/egs/*/*/mfcc
+/egs/*/*/plp
+/egs/*/*/exp
+/egs/*/*/data
 
 # /tools/
+/tools/pocolm/
 /tools/ATLAS/
 /tools/atlas3.8.3.tar.gz
 /tools/irstlm/
+/tools/mitlm/
 /tools/openfst
 /tools/openfst-1.3.2.tar.gz
 /tools/openfst-1.3.2/
@@ -143,3 +145,6 @@ GSYMS
 /tools/mmseg-1.3.0.tar.gz
 /tools/mmseg-1.3.0/
 /kaldiwin_vs*
+/tools/cub-1.8.0.zip
+/tools/cub-1.8.0/
+/tools/cub

.travis.yml

@@ -49,7 +49,7 @@ script:
   # for the explanation why extra switches needed for clang with ccache.
   - CXX="ccache clang++-3.8 -Qunused-arguments -fcolor-diagnostics -Wno-tautological-compare"
     CFLAGS=""
-    LDFLAGS="-llapack"
+    LDFLAGS="-llapack -Wl,-fuse-ld=gold"
     INCDIRS="$XROOT/usr/include"
     LIBDIRS="$XROOT/usr/lib"
       tools/extras/travis_script.sh

egs/aishell/s5/RESULTS

@@ -1,8 +1,18 @@
-%WER 33.82 [ 35432 / 104765, 743 ins, 3991 del, 30698 sub ] exp/mono/decode_test/cer_12_0.0
-%WER 19.39 [ 20310 / 104765, 903 ins, 1452 del, 17955 sub ] exp/tri1/decode_test/cer_13_0.5
-%WER 19.23 [ 20147 / 104765, 910 ins, 1287 del, 17950 sub ] exp/tri2/decode_test/cer_14_0.5
-%WER 17.14 [ 17961 / 104765, 812 ins, 1024 del, 16125 sub ] exp/tri3a/decode_test/cer_14_0.0
-%WER 13.64 [ 14294 / 104765, 669 ins, 736 del, 12889 sub ] exp/tri4a/decode_test/cer_14_0.5
-%WER 12.23 [ 12809 / 104765, 656 ins, 580 del, 11573 sub ] exp/tri5a/decode_test/cer_13_1.0
-%WER 8.45 [ 8849 / 104765, 312 ins, 538 del, 7999 sub ] exp/nnet3/tdnn_sp/decode_test/cer_13_1.0
-%WER 7.46 [ 7813 / 104765, 287 ins, 472 del, 7054 sub ] exp/chain/tdnn_1a_sp/decode_test/cer_10_1.0
+%WER 36.41 [ 38146 / 104765, 837 ins, 3114 del, 34195 sub ] exp/mono/decode_test/cer_10_0.0
+%WER 18.76 [ 19654 / 104765, 949 ins, 1152 del, 17553 sub ] exp/tri1/decode_test/cer_13_0.5
+%WER 18.64 [ 19531 / 104765, 941 ins, 1159 del, 17431 sub ] exp/tri2/decode_test/cer_14_0.5
+%WER 17.04 [ 17849 / 104765, 810 ins, 1021 del, 16018 sub ] exp/tri3a/decode_test/cer_14_0.5
+%WER 13.82 [ 14482 / 104765, 764 ins, 670 del, 13048 sub ] exp/tri4a/decode_test/cer_13_0.5
+%WER 12.12 [ 12694 / 104765, 751 ins, 523 del, 11420 sub ] exp/tri5a/decode_test/cer_13_0.5
+%WER 8.65 [ 9064 / 104765, 367 ins, 455 del, 8242 sub ] exp/nnet3/tdnn_sp/decode_test/cer_14_0.5
+%WER 7.48 [ 7839 / 104765, 285 ins, 454 del, 7100 sub ] exp/chain/tdnn_1a_sp/decode_test/cer_10_1.0
+
+# nnet3 tdnn with online pitch, local/nnet3/tuning/tun_tdnn_2a.sh
+%WER 8.64 [ 9050 / 104765, 349 ins, 521 del, 8180 sub ] exp/nnet3/tdnn_sp/decode_test/cer_15_0.5
+%WER 8.72 [ 9135 / 104765, 367 ins, 422 del, 8346 sub ] exp/nnet3/tdnn_sp_online/decode_test/cer_12_1.0
+%WER 9.36 [ 9807 / 104765, 386 ins, 441 del, 8980 sub ] exp/nnet3/tdnn_sp_online/decode_test_per_utt/cer_13_1.0
+
+# chain with online pitch, local/chain/tuning/run_tdnn_2a.sh
+%WER 7.45 [ 7807 / 104765, 340 ins, 497 del, 6970 sub ] exp/chain/tdnn_2a_sp/decode_test/cer_11_0.5
+%WER 7.43 [ 7780 / 104765, 341 ins, 469 del, 6970 sub ] exp/chain/tdnn_2a_sp_online/decode_test/cer_11_0.5
+%WER 7.92 [ 8296 / 104765, 384 ins, 472 del, 7440 sub ] exp/chain/tdnn_2a_sp_online/decode_test_per_utt/cer_11_0.5

egs/aishell/s5/conf/online_pitch.conf

@@ -0,0 +1,4 @@
+--sample-frequency=16000
+--simulate-first-pass-online=true
+--normalization-right-context=25
+--frames-per-chunk=10

egs/aishell/s5/local/aishell_prepare_dict.sh

@@ -15,21 +15,9 @@ mkdir -p $dict_dir
 cp $res_dir/lexicon.txt $dict_dir
 
 cat $dict_dir/lexicon.txt | awk '{ for(n=2;n<=NF;n++){ phones[$n] = 1; }} END{for (p in phones) print p;}'| \
-  sort -u |\
-  perl -e '
-  my %ph_cl;
-  while (<STDIN>) {
-    $phone = $_;
-    chomp($phone);
-    chomp($_);
-    $phone = $_;
-    next if ($phone eq "sil");
-    if (exists $ph_cl{$phone}) { push(@{$ph_cl{$phone}}, $_)  }
-    else { $ph_cl{$phone} = [$_]; }
-  }
-  foreach $key ( keys %ph_cl ) {
-     print "@{ $ph_cl{$key} }\n"
-  }
+  perl -e 'while(<>){ chomp($_); $phone = $_; next if ($phone eq "sil");
+    m:^([^\d]+)(\d*)$: || die "Bad phone $_"; $q{$1} .= "$phone "; }
+    foreach $l (values %q) {print "$l\n";}
   ' | sort -k1 > $dict_dir/nonsilence_phones.txt  || exit 1;
 
 echo sil > $dict_dir/silence_phones.txt

egs/aishell/s5/local/chain/tuning/run_tdnn_1a.sh

@@ -90,7 +90,7 @@ if [ $stage -le 10 ]; then
   echo "$0: creating neural net configs using the xconfig parser";
 
   num_targets=$(tree-info $treedir/tree |grep num-pdfs|awk '{print $2}')
-  learning_rate_factor=$(echo "print 0.5/$xent_regularize" | python)
+  learning_rate_factor=$(echo "print (0.5/$xent_regularize)" | python)
 
   mkdir -p $dir/configs
   cat <<EOF > $dir/configs/network.xconfig

egs/aishell/s5/local/chain/tuning/run_tdnn_2a.sh

@@ -0,0 +1,211 @@
+#!/bin/bash
+
+# This script is based on run_tdnn_1a.sh.
+# This setup used online pitch to train the neural network.
+# It requires a online_pitch.conf in the conf dir.
+
+set -e
+
+# configs for 'chain'
+affix=
+stage=0
+train_stage=-10
+get_egs_stage=-10
+dir=exp/chain/tdnn_2a  # Note: _sp will get added to this
+decode_iter=
+
+# training options
+num_epochs=4
+initial_effective_lrate=0.001
+final_effective_lrate=0.0001
+max_param_change=2.0
+final_layer_normalize_target=0.5
+num_jobs_initial=2
+num_jobs_final=12
+minibatch_size=128
+frames_per_eg=150,110,90
+remove_egs=true
+common_egs_dir=
+xent_regularize=0.1
+
+# End configuration section.
+echo "$0 $@"  # Print the command line for logging
+
+. ./cmd.sh
+. ./path.sh
+. ./utils/parse_options.sh
+
+if ! cuda-compiled; then
+  cat <<EOF && exit 1
+This script is intended to be used with GPUs but you have not compiled Kaldi with CUDA
+If you want to use GPUs (and have them), go to src/, and configure and make on a machine
+where "nvcc" is installed.
+EOF
+fi
+
+# The iVector-extraction and feature-dumping parts are the same as the standard
+# nnet3 setup, and you can skip them by setting "--stage 8" if you have already
+# run those things.
+
+dir=${dir}${affix:+_$affix}_sp
+train_set=train_sp
+ali_dir=exp/tri5a_sp_ali
+treedir=exp/chain/tri6_7d_tree_sp
+lang=data/lang_chain
+
+
+# if we are using the speed-perturbed data we need to generate
+# alignments for it.
+local/nnet3/run_ivector_common.sh --stage $stage --online true || exit 1;
+
+if [ $stage -le 7 ]; then
+  # Get the alignments as lattices (gives the LF-MMI training more freedom).
+  # use the same num-jobs as the alignments
+  nj=$(cat $ali_dir/num_jobs) || exit 1;
+  steps/align_fmllr_lats.sh --nj $nj --cmd "$train_cmd" data/$train_set \
+    data/lang exp/tri5a exp/tri5a_sp_lats
+  rm exp/tri5a_sp_lats/fsts.*.gz # save space
+fi
+
+if [ $stage -le 8 ]; then
+  # Create a version of the lang/ directory that has one state per phone in the
+  # topo file. [note, it really has two states.. the first one is only repeated
+  # once, the second one has zero or more repeats.]
+  rm -rf $lang
+  cp -r data/lang $lang
+  silphonelist=$(cat $lang/phones/silence.csl) || exit 1;
+  nonsilphonelist=$(cat $lang/phones/nonsilence.csl) || exit 1;
+  # Use our special topology... note that later on may have to tune this
+  # topology.
+  steps/nnet3/chain/gen_topo.py $nonsilphonelist $silphonelist >$lang/topo
+fi
+
+if [ $stage -le 9 ]; then
+  # Build a tree using our new topology. This is the critically different
+  # step compared with other recipes.
+  steps/nnet3/chain/build_tree.sh --frame-subsampling-factor 3 \
+      --context-opts "--context-width=2 --central-position=1" \
+      --cmd "$train_cmd" 5000 data/$train_set $lang $ali_dir $treedir
+fi
+
+if [ $stage -le 10 ]; then
+  echo "$0: creating neural net configs using the xconfig parser";
+
+  num_targets=$(tree-info $treedir/tree |grep num-pdfs|awk '{print $2}')
+  learning_rate_factor=$(echo "print (0.5/$xent_regularize)" | python)
+
+  mkdir -p $dir/configs
+  cat <<EOF > $dir/configs/network.xconfig
+  input dim=100 name=ivector
+  input dim=43 name=input
+
+  # please note that it is important to have input layer with the name=input
+  # as the layer immediately preceding the fixed-affine-layer to enable
+  # the use of short notation for the descriptor
+  fixed-affine-layer name=lda input=Append(-1,0,1,ReplaceIndex(ivector, t, 0)) affine-transform-file=$dir/configs/lda.mat
+
+  # the first splicing is moved before the lda layer, so no splicing here
+  relu-batchnorm-layer name=tdnn1 dim=625
+  relu-batchnorm-layer name=tdnn2 input=Append(-1,0,1) dim=625
+  relu-batchnorm-layer name=tdnn3 input=Append(-1,0,1) dim=625
+  relu-batchnorm-layer name=tdnn4 input=Append(-3,0,3) dim=625
+  relu-batchnorm-layer name=tdnn5 input=Append(-3,0,3) dim=625
+  relu-batchnorm-layer name=tdnn6 input=Append(-3,0,3) dim=625
+
+  ## adding the layers for chain branch
+  relu-batchnorm-layer name=prefinal-chain input=tdnn6 dim=625 target-rms=0.5
+  output-layer name=output include-log-softmax=false dim=$num_targets max-change=1.5
+
+  # adding the layers for xent branch
+  # This block prints the configs for a separate output that will be
+  # trained with a cross-entropy objective in the 'chain' models... this
+  # has the effect of regularizing the hidden parts of the model.  we use
+  # 0.5 / args.xent_regularize as the learning rate factor- the factor of
+  # 0.5 / args.xent_regularize is suitable as it means the xent
+  # final-layer learns at a rate independent of the regularization
+  # constant; and the 0.5 was tuned so as to make the relative progress
+  # similar in the xent and regular final layers.
+  relu-batchnorm-layer name=prefinal-xent input=tdnn6 dim=625 target-rms=0.5
+  output-layer name=output-xent dim=$num_targets learning-rate-factor=$learning_rate_factor max-change=1.5
+
+EOF
+  steps/nnet3/xconfig_to_configs.py --xconfig-file $dir/configs/network.xconfig --config-dir $dir/configs/
+fi
+
+if [ $stage -le 11 ]; then
+  if [[ $(hostname -f) == *.clsp.jhu.edu ]] && [ ! -d $dir/egs/storage ]; then
+    utils/create_split_dir.pl \
+     /export/b0{5,6,7,8}/$USER/kaldi-data/egs/aishell-$(date +'%m_%d_%H_%M')/s5c/$dir/egs/storage $dir/egs/storage
+  fi
+
+  steps/nnet3/chain/train.py --stage $train_stage \
+    --cmd "$decode_cmd" \
+    --feat.online-ivector-dir exp/nnet3/ivectors_${train_set} \
+    --feat.cmvn-opts "--norm-means=false --norm-vars=false" \
+    --chain.xent-regularize $xent_regularize \
+    --chain.leaky-hmm-coefficient 0.1 \
+    --chain.l2-regularize 0.00005 \
+    --chain.apply-deriv-weights false \
+    --chain.lm-opts="--num-extra-lm-states=2000" \
+    --egs.dir "$common_egs_dir" \
+    --egs.stage $get_egs_stage \
+    --egs.opts "--frames-overlap-per-eg 0" \
+    --egs.chunk-width $frames_per_eg \
+    --trainer.num-chunk-per-minibatch $minibatch_size \
+    --trainer.frames-per-iter 1500000 \
+    --trainer.num-epochs $num_epochs \
+    --trainer.optimization.num-jobs-initial $num_jobs_initial \
+    --trainer.optimization.num-jobs-final $num_jobs_final \
+    --trainer.optimization.initial-effective-lrate $initial_effective_lrate \
+    --trainer.optimization.final-effective-lrate $final_effective_lrate \
+    --trainer.max-param-change $max_param_change \
+    --cleanup.remove-egs $remove_egs \
+    --feat-dir data/${train_set}_hires_online \
+    --tree-dir $treedir \
+    --lat-dir exp/tri5a_sp_lats \
+    --dir $dir  || exit 1;
+fi
+
+if [ $stage -le 12 ]; then
+  # Note: it might appear that this $lang directory is mismatched, and it is as
+  # far as the 'topo' is concerned, but this script doesn't read the 'topo' from
+  # the lang directory.
+  utils/mkgraph.sh --self-loop-scale 1.0 data/lang_test $dir $dir/graph
+fi
+
+graph_dir=$dir/graph
+if [ $stage -le 13 ]; then
+  for test_set in dev test; do
+    steps/nnet3/decode.sh --acwt 1.0 --post-decode-acwt 10.0 \
+      --nj 10 --cmd "$decode_cmd" \
+      --online-ivector-dir exp/nnet3/ivectors_$test_set \
+      $graph_dir data/${test_set}_hires_online $dir/decode_${test_set} || exit 1;
+  done
+fi
+
+if [ $stage -le 14 ]; then
+  steps/online/nnet3/prepare_online_decoding.sh --mfcc-config conf/mfcc_hires.conf \
+    --add-pitch true \
+    $lang exp/nnet3/extractor "$dir" ${dir}_online || exit 1;
+fi
+
+dir=${dir}_online
+if [ $stage -le 15 ]; then
+  for test_set in dev test; do
+    steps/online/nnet3/decode.sh --acwt 1.0 --post-decode-acwt 10.0 \
+      --nj 10 --cmd "$decode_cmd" \
+      --config conf/decode.config \
+      $graph_dir data/${test_set}_hires_online $dir/decode_${test_set} || exit 1;
+  done
+fi
+
+if [ $stage -le 16 ]; then
+  for test_set in dev test; do
+    steps/online/nnet3/decode.sh --acwt 1.0 --post-decode-acwt 10.0 \
+      --nj 10 --cmd "$decode_cmd" --per-utt true \
+      --config conf/decode.config \
+      $graph_dir data/${test_set}_hires_online $dir/decode_${test_set}_per_utt || exit 1;
+  done
+fi
+
+exit;

egs/aishell/s5/local/nnet3/run_ivector_common.sh

@@ -14,7 +14,7 @@ stage=0
 train_set=train
 test_sets="dev test"
 gmm=tri5a
-
+online=false
 nnet3_affix=
 
 . ./cmd.sh
@@ -31,6 +31,11 @@ for f in data/${train_set}/feats.scp ${gmm_dir}/final.mdl; do
   fi
 done
 
+online_affix=
+if [ $online = true ]; then
+  online_affix=_online
+fi
+
 if [ $stage -le 1 ]; then
   # Although the nnet will be trained by high resolution data, we still have to
   # perturb the normal data to get the alignment _sp stands for speed-perturbed
@@ -54,26 +59,26 @@ if [ $stage -le 3 ]; then
   # Create high-resolution MFCC features (with 40 cepstra instead of 13).
   # this shows how you can split across multiple file-systems.
   echo "$0: creating high-resolution MFCC features"
-  mfccdir=mfcc_perturbed_hires
+  mfccdir=mfcc_perturbed_hires$online_affix
   if [[ $(hostname -f) == *.clsp.jhu.edu ]] && [ ! -d $mfccdir/storage ]; then
     utils/create_split_dir.pl /export/b0{5,6,7,8}/$USER/kaldi-data/mfcc/aishell-$(date +'%m_%d_%H_%M')/s5/$mfccdir/storage $mfccdir/storage
   fi
 
   for datadir in ${train_set}_sp ${test_sets}; do
-    utils/copy_data_dir.sh data/$datadir data/${datadir}_hires
+    utils/copy_data_dir.sh data/$datadir data/${datadir}_hires$online_affix
   done
 
   # do volume-perturbation on the training data prior to extracting hires
   # features; this helps make trained nnets more invariant to test data volume.
-  utils/data/perturb_data_dir_volume.sh data/${train_set}_sp_hires || exit 1;
+  utils/data/perturb_data_dir_volume.sh data/${train_set}_sp_hires$online_affix || exit 1;
 
   for datadir in ${train_set}_sp ${test_sets}; do
-    steps/make_mfcc_pitch.sh --nj 10 --mfcc-config conf/mfcc_hires.conf \
-      --cmd "$train_cmd" data/${datadir}_hires exp/make_hires/$datadir $mfccdir || exit 1;
-    steps/compute_cmvn_stats.sh data/${datadir}_hires exp/make_hires/$datadir $mfccdir || exit 1;
-    utils/fix_data_dir.sh data/${datadir}_hires || exit 1;
+    steps/make_mfcc_pitch$online_affix.sh --nj 10 --mfcc-config conf/mfcc_hires.conf \
+      --cmd "$train_cmd" data/${datadir}_hires$online_affix exp/make_hires/$datadir $mfccdir || exit 1;
+    steps/compute_cmvn_stats.sh data/${datadir}_hires$online_affix exp/make_hires/$datadir $mfccdir || exit 1;
+    utils/fix_data_dir.sh data/${datadir}_hires$online_affix || exit 1;
     # create MFCC data dir without pitch to extract iVector
-    utils/data/limit_feature_dim.sh 0:39 data/${datadir}_hires data/${datadir}_hires_nopitch || exit 1;
+    utils/data/limit_feature_dim.sh 0:39 data/${datadir}_hires$online_affix data/${datadir}_hires_nopitch || exit 1;
     steps/compute_cmvn_stats.sh data/${datadir}_hires_nopitch exp/make_hires/$datadir $mfccdir || exit 1;
   done
 fi