[egs] swbd/s5c, added 5 layer (b)lstm recipes (#1759)
authorVijayaditya Peddinti <vijayaditya@users.noreply.github.com>
Wed, 12 Jul 2017 03:49:55 +0000 (20:49 -0700)
committerDaniel Povey <dpovey@gmail.com>
Wed, 12 Jul 2017 03:49:55 +0000 (23:49 -0400)
egs/swbd/s5c/local/chain/tuning/run_blstm_6m.sh [new file with mode: 0755]
egs/swbd/s5c/local/chain/tuning/run_blstm_6n.sh [new file with mode: 0755]
egs/swbd/s5c/local/chain/tuning/run_blstm_6o.sh [new file with mode: 0755]
egs/swbd/s5c/local/chain/tuning/run_lstm_6l.sh [new file with mode: 0755]

diff --git a/egs/swbd/s5c/local/chain/tuning/run_blstm_6m.sh b/egs/swbd/s5c/local/chain/tuning/run_blstm_6m.sh
new file mode 100755 (executable)
index 0000000..4668aac
--- /dev/null
@@ -0,0 +1,246 @@
+#!/bin/bash
+
+# Copyright 2015  Johns Hopkins University (Author: Daniel Povey).
+#           2015  Vijayaditya Peddinti
+#           2016  Yiming Wang
+#           2017  Google  Inc. (vpeddinti@google.com)
+# Apache 2.0.
+
+# 6m is same as 6k, but with delay of -1,-1 at lower blstm layer
+# local/chain/compare_wer_general.sh blstm_6k_sp blstm_6m_sp
+# System                blstm_6k_sp blstm_6m_sp
+# WER on train_dev(tg)      12.95     12.79
+# WER on train_dev(fg)      11.98     11.81
+# WER on eval2000(tg)        15.5      15.0
+# WER on eval2000(fg)        14.1      13.6
+# Final train prob         -0.041    -0.050
+# Final valid prob         -0.072    -0.075
+# Final train prob (xent)        -0.629    -0.713
+# Final valid prob (xent)       -0.8091   -0.8568
+
+
+set -e
+
+# configs for 'chain'
+stage=12
+train_stage=-10
+get_egs_stage=-10
+speed_perturb=true
+dir=exp/chain/blstm_6m  # Note: _sp will get added to this if $speed_perturb == true.
+decode_iter=
+decode_dir_affix=
+
+# training options
+leftmost_questions_truncate=-1
+chunk_width=150
+chunk_left_context=40
+chunk_right_context=40
+xent_regularize=0.025
+self_repair_scale=0.00001
+label_delay=0
+
+# decode options
+extra_left_context=50
+extra_right_context=50
+frames_per_chunk=
+
+remove_egs=false
+common_egs_dir=
+
+affix=
+# End configuration section.
+echo "$0 $@"  # Print the command line for logging
+
+. ./cmd.sh
+. ./path.sh
+. ./utils/parse_options.sh
+
+if ! cuda-compiled; then
+  cat <<EOF && exit 1
+This script is intended to be used with GPUs but you have not compiled Kaldi with CUDA
+If you want to use GPUs (and have them), go to src/, and configure and make on a machine
+where "nvcc" is installed.
+EOF
+fi
+
+# The iVector-extraction and feature-dumping parts are the same as the standard
+# nnet3 setup, and you can skip them by setting "--stage 8" if you have already
+# run those things.
+
+suffix=
+if [ "$speed_perturb" == "true" ]; then
+  suffix=_sp
+fi
+
+dir=$dir${affix:+_$affix}
+if [ $label_delay -gt 0 ]; then dir=${dir}_ld$label_delay; fi
+dir=${dir}$suffix
+train_set=train_nodup$suffix
+ali_dir=exp/tri4_ali_nodup$suffix
+treedir=exp/chain/tri5_7d_tree$suffix
+lang=data/lang_chain_2y
+
+
+# if we are using the speed-perturbed data we need to generate
+# alignments for it.
+local/nnet3/run_ivector_common.sh --stage $stage \
+  --speed-perturb $speed_perturb \
+  --generate-alignments $speed_perturb || exit 1;
+
+
+if [ $stage -le 9 ]; then
+  # Get the alignments as lattices (gives the CTC training more freedom).
+  # use the same num-jobs as the alignments
+  nj=$(cat exp/tri4_ali_nodup$suffix/num_jobs) || exit 1;
+  steps/align_fmllr_lats.sh --nj $nj --cmd "$train_cmd" data/$train_set \
+    data/lang exp/tri4 exp/tri4_lats_nodup$suffix
+  rm exp/tri4_lats_nodup$suffix/fsts.*.gz # save space
+fi
+
+
+if [ $stage -le 10 ]; 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 11 ]; then
+  # Build a tree using our new topology.
+  steps/nnet3/chain/build_tree.sh --frame-subsampling-factor 3 \
+      --leftmost-questions-truncate $leftmost_questions_truncate \
+      --context-opts "--context-width=2 --central-position=1" \
+      --cmd "$train_cmd" 7000 data/$train_set $lang $ali_dir $treedir
+fi
+
+if [ $stage -le 12 ]; then
+  echo "$0: creating neural net configs using the xconfig parser";
+
+  num_targets=$(tree-info $treedir/tree |grep num-pdfs|awk '{print $2}')
+  [ -z $num_targets ] && { echo "$0: error getting num-targets"; exit 1; }
+  learning_rate_factor=$(echo "print 0.5/$xent_regularize" | python)
+
+  lstm_opts="decay-time=20"
+
+  mkdir -p $dir/configs
+  cat <<EOF > $dir/configs/network.xconfig
+  input dim=100 name=ivector
+  input dim=40 name=input
+
+  # please note that it is important to have input layer with the name=input
+  # as the layer immediately preceding the fixed-affine-layer to enable
+  # the use of short notation for the descriptor
+  fixed-affine-layer name=lda input=Append(-2,-1,0,1,2,ReplaceIndex(ivector, t, 0)) affine-transform-file=$dir/configs/lda.mat
+
+  # the first splicing is moved before the lda layer, so no splicing here
+
+  # check steps/libs/nnet3/xconfig/lstm.py for the other options and defaults
+  fast-lstmp-layer name=blstm1-forward input=lda cell-dim=1024 recurrent-projection-dim=256 non-recurrent-projection-dim=256 delay=-1 $lstm_opts
+  fast-lstmp-layer name=blstm1-backward input=lda cell-dim=1024 recurrent-projection-dim=256 non-recurrent-projection-dim=256 delay=1 $lstm_opts
+
+  fast-lstmp-layer name=blstm2-forward input=Append(blstm1-forward, blstm1-backward) cell-dim=1024 recurrent-projection-dim=256 non-recurrent-projection-dim=256 delay=-3 $lstm_opts
+  fast-lstmp-layer name=blstm2-backward input=Append(blstm1-forward, blstm1-backward) cell-dim=1024 recurrent-projection-dim=256 non-recurrent-projection-dim=256 delay=3 $lstm_opts
+
+  fast-lstmp-layer name=blstm3-forward input=Append(blstm2-forward, blstm2-backward) cell-dim=1024 recurrent-projection-dim=256 non-recurrent-projection-dim=256 delay=-3 $lstm_opts
+  fast-lstmp-layer name=blstm3-backward input=Append(blstm2-forward, blstm2-backward) cell-dim=1024 recurrent-projection-dim=256 non-recurrent-projection-dim=256 delay=3 $lstm_opts
+
+  ## adding the layers for chain branch
+  output-layer name=output input=Append(blstm3-forward, blstm3-backward) output-delay=$label_delay include-log-softmax=false dim=$num_targets max-change=1.5
+
+  # adding the layers for xent branch
+  # This block prints the configs for a separate output that will be
+  # trained with a cross-entropy objective in the 'chain' models... this
+  # has the effect of regularizing the hidden parts of the model.  we use
+  # 0.5 / args.xent_regularize as the learning rate factor- the factor of
+  # 0.5 / args.xent_regularize is suitable as it means the xent
+  # final-layer learns at a rate independent of the regularization
+  # constant; and the 0.5 was tuned so as to make the relative progress
+  # similar in the xent and regular final layers.
+  output-layer name=output-xent input=Append(blstm3-forward, blstm3-backward) output-delay=$label_delay 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 13 ]; 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/swbd-$(date +'%m_%d_%H_%M')/s5c/$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" \
+    --trainer.num-chunk-per-minibatch 64 \
+    --trainer.frames-per-iter 1200000 \
+    --trainer.max-param-change 2.0 \
+    --trainer.num-epochs 4 \
+    --trainer.optimization.shrink-value 0.99 \
+    --trainer.optimization.num-jobs-initial 3 \
+    --trainer.optimization.num-jobs-final 16 \
+    --trainer.optimization.initial-effective-lrate 0.001 \
+    --trainer.optimization.final-effective-lrate 0.0001 \
+    --trainer.optimization.momentum 0.0 \
+    --trainer.deriv-truncate-margin 8 \
+    --egs.stage $get_egs_stage \
+    --egs.opts "--frames-overlap-per-eg 0" \
+    --egs.chunk-width $chunk_width \
+    --egs.chunk-left-context $chunk_left_context \
+    --egs.chunk-right-context $chunk_right_context \
+    --egs.dir "$common_egs_dir" \
+    --cleanup.remove-egs $remove_egs \
+    --feat-dir data/${train_set}_hires \
+    --tree-dir $treedir \
+    --lat-dir exp/tri4_lats_nodup$suffix \
+    --dir $dir  || exit 1;
+fi
+
+if [ $stage -le 14 ]; 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_sw1_tg $dir $dir/graph_sw1_tg
+fi
+
+decode_suff=sw1_tg
+graph_dir=$dir/graph_sw1_tg
+if [ $stage -le 15 ]; then
+  [ -z $extra_left_context ] && extra_left_context=$chunk_left_context;
+  [ -z $extra_right_context ] && extra_right_context=$chunk_right_context;
+  [ -z $frames_per_chunk ] && frames_per_chunk=$chunk_width;
+  iter_opts=
+  if [ ! -z $decode_iter ]; then
+    iter_opts=" --iter $decode_iter "
+  fi
+  for decode_set in train_dev eval2000; do
+      (
+      steps/nnet3/decode.sh --acwt 1.0 --post-decode-acwt 10.0 \
+          --nj 50 --cmd "$decode_cmd" $iter_opts \
+          --extra-left-context $extra_left_context  \
+          --extra-right-context $extra_right_context  \
+          --frames-per-chunk "$frames_per_chunk" \
+          --online-ivector-dir exp/nnet3/ivectors_${decode_set} \
+         $graph_dir data/${decode_set}_hires \
+         $dir/decode_${decode_set}${decode_dir_affix:+_$decode_dir_affix}_${decode_suff} || exit 1;
+      if $has_fisher; then
+          steps/lmrescore_const_arpa.sh --cmd "$decode_cmd" \
+            data/lang_sw1_{tg,fsh_fg} data/${decode_set}_hires \
+            $dir/decode_${decode_set}${decode_dir_affix:+_$decode_dir_affix}_sw1_{tg,fsh_fg} || exit 1;
+      fi
+      ) &
+  done
+fi
+wait;
+exit 0;
diff --git a/egs/swbd/s5c/local/chain/tuning/run_blstm_6n.sh b/egs/swbd/s5c/local/chain/tuning/run_blstm_6n.sh
new file mode 100755 (executable)
index 0000000..22316d5
--- /dev/null
@@ -0,0 +1,252 @@
+#!/bin/bash
+
+# Copyright 2015  Johns Hopkins University (Author: Daniel Povey).
+#           2015  Vijayaditya Peddinti
+#           2016  Yiming Wang
+#           2017  Google  Inc. (vpeddinti@google.com)
+# Apache 2.0.
+
+# 6n is same as 6k, but with two additional BLSTM layers
+# local/chain/compare_wer_general.sh blstm_6k_sp blstm_6n_sp
+# System                blstm_6k_sp blstm_6n_sp
+# WER on train_dev(tg)      12.95     12.69
+# WER on train_dev(fg)      11.98     11.90
+# WER on eval2000(tg)        15.5      15.2
+# WER on eval2000(fg)        14.1      13.9
+# Final train prob         -0.041    -0.046
+# Final valid prob         -0.072    -0.072
+# Final train prob (xent)        -0.629    -0.666
+# Final valid prob (xent)       -0.8091   -0.8047
+
+set -e
+
+# configs for 'chain'
+stage=12
+train_stage=-10
+get_egs_stage=-10
+speed_perturb=true
+dir=exp/chain/blstm_6n  # Note: _sp will get added to this if $speed_perturb == true.
+decode_iter=
+decode_dir_affix=
+
+# training options
+leftmost_questions_truncate=-1
+chunk_width=150
+chunk_left_context=40
+chunk_right_context=40
+xent_regularize=0.025
+self_repair_scale=0.00001
+label_delay=0
+
+# decode options
+extra_left_context=50
+extra_right_context=50
+frames_per_chunk=
+
+remove_egs=false
+common_egs_dir=
+
+affix=
+# End configuration section.
+echo "$0 $@"  # Print the command line for logging
+
+. ./cmd.sh
+. ./path.sh
+. ./utils/parse_options.sh
+
+if ! cuda-compiled; then
+  cat <<EOF && exit 1
+This script is intended to be used with GPUs but you have not compiled Kaldi with CUDA
+If you want to use GPUs (and have them), go to src/, and configure and make on a machine
+where "nvcc" is installed.
+EOF
+fi
+
+# The iVector-extraction and feature-dumping parts are the same as the standard
+# nnet3 setup, and you can skip them by setting "--stage 8" if you have already
+# run those things.
+
+suffix=
+if [ "$speed_perturb" == "true" ]; then
+  suffix=_sp
+fi
+
+dir=$dir${affix:+_$affix}
+if [ $label_delay -gt 0 ]; then dir=${dir}_ld$label_delay; fi
+dir=${dir}$suffix
+train_set=train_nodup$suffix
+ali_dir=exp/tri4_ali_nodup$suffix
+treedir=exp/chain/tri5_7d_tree$suffix
+lang=data/lang_chain_2y
+
+
+# if we are using the speed-perturbed data we need to generate
+# alignments for it.
+local/nnet3/run_ivector_common.sh --stage $stage \
+  --speed-perturb $speed_perturb \
+  --generate-alignments $speed_perturb || exit 1;
+
+
+if [ $stage -le 9 ]; then
+  # Get the alignments as lattices (gives the CTC training more freedom).
+  # use the same num-jobs as the alignments
+  nj=$(cat exp/tri4_ali_nodup$suffix/num_jobs) || exit 1;
+  steps/align_fmllr_lats.sh --nj $nj --cmd "$train_cmd" data/$train_set \
+    data/lang exp/tri4 exp/tri4_lats_nodup$suffix
+  rm exp/tri4_lats_nodup$suffix/fsts.*.gz # save space
+fi
+
+
+if [ $stage -le 10 ]; 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 11 ]; then
+  # Build a tree using our new topology.
+  steps/nnet3/chain/build_tree.sh --frame-subsampling-factor 3 \
+      --leftmost-questions-truncate $leftmost_questions_truncate \
+      --context-opts "--context-width=2 --central-position=1" \
+      --cmd "$train_cmd" 7000 data/$train_set $lang $ali_dir $treedir
+fi
+
+if [ $stage -le 12 ]; then
+  echo "$0: creating neural net configs using the xconfig parser";
+
+  num_targets=$(tree-info $treedir/tree |grep num-pdfs|awk '{print $2}')
+  [ -z $num_targets ] && { echo "$0: error getting num-targets"; exit 1; }
+  learning_rate_factor=$(echo "print 0.5/$xent_regularize" | python)
+
+  lstm_opts="decay-time=20"
+
+  mkdir -p $dir/configs
+  cat <<EOF > $dir/configs/network.xconfig
+  input dim=100 name=ivector
+  input dim=40 name=input
+
+  # please note that it is important to have input layer with the name=input
+  # as the layer immediately preceding the fixed-affine-layer to enable
+  # the use of short notation for the descriptor
+  fixed-affine-layer name=lda input=Append(-2,-1,0,1,2,ReplaceIndex(ivector, t, 0)) affine-transform-file=$dir/configs/lda.mat
+
+  # the first splicing is moved before the lda layer, so no splicing here
+
+  # check steps/libs/nnet3/xconfig/lstm.py for the other options and defaults
+  fast-lstmp-layer name=blstm1-forward input=lda cell-dim=1024 recurrent-projection-dim=256 non-recurrent-projection-dim=256 delay=-3 $lstm_opts
+  fast-lstmp-layer name=blstm1-backward input=lda cell-dim=1024 recurrent-projection-dim=256 non-recurrent-projection-dim=256 delay=3 $lstm_opts
+
+  fast-lstmp-layer name=blstm2-forward input=Append(blstm1-forward, blstm1-backward) cell-dim=1024 recurrent-projection-dim=256 non-recurrent-projection-dim=256 delay=-3 $lstm_opts
+  fast-lstmp-layer name=blstm2-backward input=Append(blstm1-forward, blstm1-backward) cell-dim=1024 recurrent-projection-dim=256 non-recurrent-projection-dim=256 delay=3 $lstm_opts
+
+  fast-lstmp-layer name=blstm3-forward input=Append(blstm2-forward, blstm2-backward) cell-dim=1024 recurrent-projection-dim=256 non-recurrent-projection-dim=256 delay=-3 $lstm_opts
+  fast-lstmp-layer name=blstm3-backward input=Append(blstm2-forward, blstm2-backward) cell-dim=1024 recurrent-projection-dim=256 non-recurrent-projection-dim=256 delay=3 $lstm_opts
+
+
+  fast-lstmp-layer name=blstm4-forward input=Append(blstm3-forward, blstm3-backward) cell-dim=1024 recurrent-projection-dim=256 non-recurrent-projection-dim=256 delay=-3 $lstm_opts
+  fast-lstmp-layer name=blstm4-backward input=Append(blstm3-forward, blstm3-backward) cell-dim=1024 recurrent-projection-dim=256 non-recurrent-projection-dim=256 delay=3 $lstm_opts
+
+  fast-lstmp-layer name=blstm5-forward input=Append(blstm4-forward, blstm4-backward) cell-dim=1024 recurrent-projection-dim=256 non-recurrent-projection-dim=256 delay=-3 $lstm_opts
+  fast-lstmp-layer name=blstm5-backward input=Append(blstm4-forward, blstm4-backward) cell-dim=1024 recurrent-projection-dim=256 non-recurrent-projection-dim=256 delay=3 $lstm_opts
+
+  ## adding the layers for chain branch
+  output-layer name=output input=Append(blstm5-forward, blstm5-backward) output-delay=$label_delay include-log-softmax=false dim=$num_targets max-change=1.5
+
+  # adding the layers for xent branch
+  # This block prints the configs for a separate output that will be
+  # trained with a cross-entropy objective in the 'chain' models... this
+  # has the effect of regularizing the hidden parts of the model.  we use
+  # 0.5 / args.xent_regularize as the learning rate factor- the factor of
+  # 0.5 / args.xent_regularize is suitable as it means the xent
+  # final-layer learns at a rate independent of the regularization
+  # constant; and the 0.5 was tuned so as to make the relative progress
+  # similar in the xent and regular final layers.
+  output-layer name=output-xent input=Append(blstm5-forward, blstm5-backward) output-delay=$label_delay 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 13 ]; 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/swbd-$(date +'%m_%d_%H_%M')/s5c/$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" \
+    --trainer.num-chunk-per-minibatch 64 \
+    --trainer.frames-per-iter 1200000 \
+    --trainer.max-param-change 2.0 \
+    --trainer.num-epochs 4 \
+    --trainer.optimization.shrink-value 0.99 \
+    --trainer.optimization.num-jobs-initial 3 \
+    --trainer.optimization.num-jobs-final 16 \
+    --trainer.optimization.initial-effective-lrate 0.001 \
+    --trainer.optimization.final-effective-lrate 0.0001 \
+    --trainer.optimization.momentum 0.0 \
+    --trainer.deriv-truncate-margin 8 \
+    --egs.stage $get_egs_stage \
+    --egs.opts "--frames-overlap-per-eg 0" \
+    --egs.chunk-width $chunk_width \
+    --egs.chunk-left-context $chunk_left_context \
+    --egs.chunk-right-context $chunk_right_context \
+    --egs.dir "$common_egs_dir" \
+    --cleanup.remove-egs $remove_egs \
+    --feat-dir data/${train_set}_hires \
+    --tree-dir $treedir \
+    --lat-dir exp/tri4_lats_nodup$suffix \
+    --dir $dir  || exit 1;
+fi
+
+if [ $stage -le 14 ]; 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_sw1_tg $dir $dir/graph_sw1_tg
+fi
+
+decode_suff=sw1_tg
+graph_dir=$dir/graph_sw1_tg
+if [ $stage -le 15 ]; then
+  [ -z $extra_left_context ] && extra_left_context=$chunk_left_context;
+  [ -z $extra_right_context ] && extra_right_context=$chunk_right_context;
+  [ -z $frames_per_chunk ] && frames_per_chunk=$chunk_width;
+  iter_opts=
+  if [ ! -z $decode_iter ]; then
+    iter_opts=" --iter $decode_iter "
+  fi
+  for decode_set in train_dev eval2000; do
+      (
+      steps/nnet3/decode.sh --acwt 1.0 --post-decode-acwt 10.0 \
+          --nj 50 --cmd "$decode_cmd" $iter_opts \
+          --extra-left-context $extra_left_context  \
+          --extra-right-context $extra_right_context  \
+          --frames-per-chunk "$frames_per_chunk" \
+          --online-ivector-dir exp/nnet3/ivectors_${decode_set} \
+         $graph_dir data/${decode_set}_hires \
+         $dir/decode_${decode_set}${decode_dir_affix:+_$decode_dir_affix}_${decode_suff} || exit 1;
+      if $has_fisher; then
+          steps/lmrescore_const_arpa.sh --cmd "$decode_cmd" \
+            data/lang_sw1_{tg,fsh_fg} data/${decode_set}_hires \
+            $dir/decode_${decode_set}${decode_dir_affix:+_$decode_dir_affix}_sw1_{tg,fsh_fg} || exit 1;
+      fi
+      ) &
+  done
+fi
+wait;
+exit 0;
diff --git a/egs/swbd/s5c/local/chain/tuning/run_blstm_6o.sh b/egs/swbd/s5c/local/chain/tuning/run_blstm_6o.sh
new file mode 100755 (executable)
index 0000000..ad2ac4b
--- /dev/null
@@ -0,0 +1,254 @@
+#!/bin/bash
+
+# Copyright 2015  Johns Hopkins University (Author: Daniel Povey).
+#           2015  Vijayaditya Peddinti
+#           2016  Yiming Wang
+#           2017  Google  Inc. (vpeddinti@google.com)
+# Apache 2.0.
+
+
+# 6o is same as 6k, but with two additional BLSTM layers
+# and delay of -1 for the first blstm layer
+# local/chain/compare_wer_general.sh blstm_6k_sp blstm_6o_sp
+# System                blstm_6k_sp blstm_6o_sp
+# WER on train_dev(tg)      12.95     12.60
+# WER on train_dev(fg)      11.98     11.75
+# WER on eval2000(tg)        15.5      14.7
+# WER on eval2000(fg)        14.1      13.4
+# Final train prob         -0.041    -0.041
+# Final valid prob         -0.072    -0.069
+# Final train prob (xent)        -0.629    -0.636
+# Final valid prob (xent)       -0.8091   -0.7854
+
+set -e
+
+# configs for 'chain'
+stage=12
+train_stage=-10
+get_egs_stage=-10
+speed_perturb=true
+dir=exp/chain/blstm_6o  # Note: _sp will get added to this if $speed_perturb == true.
+decode_iter=
+decode_dir_affix=
+
+# training options
+leftmost_questions_truncate=-1
+chunk_width=150
+chunk_left_context=40
+chunk_right_context=40
+xent_regularize=0.025
+self_repair_scale=0.00001
+label_delay=0
+
+# decode options
+extra_left_context=50
+extra_right_context=50
+frames_per_chunk=
+
+remove_egs=false
+common_egs_dir=
+
+affix=
+# End configuration section.
+echo "$0 $@"  # Print the command line for logging
+
+. ./cmd.sh
+. ./path.sh
+. ./utils/parse_options.sh
+
+if ! cuda-compiled; then
+  cat <<EOF && exit 1
+This script is intended to be used with GPUs but you have not compiled Kaldi with CUDA
+If you want to use GPUs (and have them), go to src/, and configure and make on a machine
+where "nvcc" is installed.
+EOF
+fi
+
+# The iVector-extraction and feature-dumping parts are the same as the standard
+# nnet3 setup, and you can skip them by setting "--stage 8" if you have already
+# run those things.
+
+suffix=
+if [ "$speed_perturb" == "true" ]; then
+  suffix=_sp
+fi
+
+dir=$dir${affix:+_$affix}
+if [ $label_delay -gt 0 ]; then dir=${dir}_ld$label_delay; fi
+dir=${dir}$suffix
+train_set=train_nodup$suffix
+ali_dir=exp/tri4_ali_nodup$suffix
+treedir=exp/chain/tri5_7d_tree$suffix
+lang=data/lang_chain_2y
+
+
+# if we are using the speed-perturbed data we need to generate
+# alignments for it.
+local/nnet3/run_ivector_common.sh --stage $stage \
+  --speed-perturb $speed_perturb \
+  --generate-alignments $speed_perturb || exit 1;
+
+
+if [ $stage -le 9 ]; then
+  # Get the alignments as lattices (gives the CTC training more freedom).
+  # use the same num-jobs as the alignments
+  nj=$(cat exp/tri4_ali_nodup$suffix/num_jobs) || exit 1;
+  steps/align_fmllr_lats.sh --nj $nj --cmd "$train_cmd" data/$train_set \
+    data/lang exp/tri4 exp/tri4_lats_nodup$suffix
+  rm exp/tri4_lats_nodup$suffix/fsts.*.gz # save space
+fi
+
+
+if [ $stage -le 10 ]; 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 11 ]; then
+  # Build a tree using our new topology.
+  steps/nnet3/chain/build_tree.sh --frame-subsampling-factor 3 \
+      --leftmost-questions-truncate $leftmost_questions_truncate \
+      --context-opts "--context-width=2 --central-position=1" \
+      --cmd "$train_cmd" 7000 data/$train_set $lang $ali_dir $treedir
+fi
+
+if [ $stage -le 12 ]; then
+  echo "$0: creating neural net configs using the xconfig parser";
+
+  num_targets=$(tree-info $treedir/tree |grep num-pdfs|awk '{print $2}')
+  [ -z $num_targets ] && { echo "$0: error getting num-targets"; exit 1; }
+  learning_rate_factor=$(echo "print 0.5/$xent_regularize" | python)
+
+  lstm_opts="decay-time=20"
+
+  mkdir -p $dir/configs
+  cat <<EOF > $dir/configs/network.xconfig
+  input dim=100 name=ivector
+  input dim=40 name=input
+
+  # please note that it is important to have input layer with the name=input
+  # as the layer immediately preceding the fixed-affine-layer to enable
+  # the use of short notation for the descriptor
+  fixed-affine-layer name=lda input=Append(-2,-1,0,1,2,ReplaceIndex(ivector, t, 0)) affine-transform-file=$dir/configs/lda.mat
+
+  # the first splicing is moved before the lda layer, so no splicing here
+
+  # check steps/libs/nnet3/xconfig/lstm.py for the other options and defaults
+  fast-lstmp-layer name=blstm1-forward input=lda cell-dim=1024 recurrent-projection-dim=256 non-recurrent-projection-dim=256 delay=-1 $lstm_opts
+  fast-lstmp-layer name=blstm1-backward input=lda cell-dim=1024 recurrent-projection-dim=256 non-recurrent-projection-dim=256 delay=1 $lstm_opts
+
+  fast-lstmp-layer name=blstm2-forward input=Append(blstm1-forward, blstm1-backward) cell-dim=1024 recurrent-projection-dim=256 non-recurrent-projection-dim=256 delay=-3 $lstm_opts
+  fast-lstmp-layer name=blstm2-backward input=Append(blstm1-forward, blstm1-backward) cell-dim=1024 recurrent-projection-dim=256 non-recurrent-projection-dim=256 delay=3 $lstm_opts
+
+  fast-lstmp-layer name=blstm3-forward input=Append(blstm2-forward, blstm2-backward) cell-dim=1024 recurrent-projection-dim=256 non-recurrent-projection-dim=256 delay=-3 $lstm_opts
+  fast-lstmp-layer name=blstm3-backward input=Append(blstm2-forward, blstm2-backward) cell-dim=1024 recurrent-projection-dim=256 non-recurrent-projection-dim=256 delay=3 $lstm_opts
+
+
+  fast-lstmp-layer name=blstm4-forward input=Append(blstm3-forward, blstm3-backward) cell-dim=1024 recurrent-projection-dim=256 non-recurrent-projection-dim=256 delay=-3 $lstm_opts
+  fast-lstmp-layer name=blstm4-backward input=Append(blstm3-forward, blstm3-backward) cell-dim=1024 recurrent-projection-dim=256 non-recurrent-projection-dim=256 delay=3 $lstm_opts
+
+  fast-lstmp-layer name=blstm5-forward input=Append(blstm4-forward, blstm4-backward) cell-dim=1024 recurrent-projection-dim=256 non-recurrent-projection-dim=256 delay=-3 $lstm_opts
+  fast-lstmp-layer name=blstm5-backward input=Append(blstm4-forward, blstm4-backward) cell-dim=1024 recurrent-projection-dim=256 non-recurrent-projection-dim=256 delay=3 $lstm_opts
+
+  ## adding the layers for chain branch
+  output-layer name=output input=Append(blstm5-forward, blstm5-backward) output-delay=$label_delay include-log-softmax=false dim=$num_targets max-change=1.5
+
+  # adding the layers for xent branch
+  # This block prints the configs for a separate output that will be
+  # trained with a cross-entropy objective in the 'chain' models... this
+  # has the effect of regularizing the hidden parts of the model.  we use
+  # 0.5 / args.xent_regularize as the learning rate factor- the factor of
+  # 0.5 / args.xent_regularize is suitable as it means the xent
+  # final-layer learns at a rate independent of the regularization
+  # constant; and the 0.5 was tuned so as to make the relative progress
+  # similar in the xent and regular final layers.
+  output-layer name=output-xent input=Append(blstm5-forward, blstm5-backward) output-delay=$label_delay 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 13 ]; 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/swbd-$(date +'%m_%d_%H_%M')/s5c/$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" \
+    --trainer.num-chunk-per-minibatch 32 \
+    --trainer.frames-per-iter 1200000 \
+    --trainer.max-param-change 2.0 \
+    --trainer.num-epochs 4 \
+    --trainer.optimization.shrink-value 0.99 \
+    --trainer.optimization.num-jobs-initial 3 \
+    --trainer.optimization.num-jobs-final 16 \
+    --trainer.optimization.initial-effective-lrate 0.001 \
+    --trainer.optimization.final-effective-lrate 0.0001 \
+    --trainer.optimization.momentum 0.0 \
+    --trainer.deriv-truncate-margin 8 \
+    --egs.stage $get_egs_stage \
+    --egs.opts "--frames-overlap-per-eg 0" \
+    --egs.chunk-width $chunk_width \
+    --egs.chunk-left-context $chunk_left_context \
+    --egs.chunk-right-context $chunk_right_context \
+    --egs.dir "$common_egs_dir" \
+    --cleanup.remove-egs $remove_egs \
+    --feat-dir data/${train_set}_hires \
+    --tree-dir $treedir \
+    --lat-dir exp/tri4_lats_nodup$suffix \
+    --dir $dir  || exit 1;
+fi
+
+if [ $stage -le 14 ]; 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_sw1_tg $dir $dir/graph_sw1_tg
+fi
+
+decode_suff=sw1_tg
+graph_dir=$dir/graph_sw1_tg
+if [ $stage -le 15 ]; then
+  [ -z $extra_left_context ] && extra_left_context=$chunk_left_context;
+  [ -z $extra_right_context ] && extra_right_context=$chunk_right_context;
+  [ -z $frames_per_chunk ] && frames_per_chunk=$chunk_width;
+  iter_opts=
+  if [ ! -z $decode_iter ]; then
+    iter_opts=" --iter $decode_iter "
+  fi
+  for decode_set in train_dev eval2000; do
+      (
+      steps/nnet3/decode.sh --acwt 1.0 --post-decode-acwt 10.0 \
+          --nj 50 --cmd "$decode_cmd" $iter_opts \
+          --extra-left-context $extra_left_context  \
+          --extra-right-context $extra_right_context  \
+          --frames-per-chunk "$frames_per_chunk" \
+          --online-ivector-dir exp/nnet3/ivectors_${decode_set} \
+         $graph_dir data/${decode_set}_hires \
+         $dir/decode_${decode_set}${decode_dir_affix:+_$decode_dir_affix}_${decode_suff} || exit 1;
+      if $has_fisher; then
+          steps/lmrescore_const_arpa.sh --cmd "$decode_cmd" \
+            data/lang_sw1_{tg,fsh_fg} data/${decode_set}_hires \
+            $dir/decode_${decode_set}${decode_dir_affix:+_$decode_dir_affix}_sw1_{tg,fsh_fg} || exit 1;
+      fi
+      ) &
+  done
+fi
+wait;
+exit 0;
diff --git a/egs/swbd/s5c/local/chain/tuning/run_lstm_6l.sh b/egs/swbd/s5c/local/chain/tuning/run_lstm_6l.sh
new file mode 100755 (executable)
index 0000000..12564c4
--- /dev/null
@@ -0,0 +1,315 @@
+#!/bin/bash
+
+# Copyright 2015  Johns Hopkins University (Author: Daniel Povey).
+#           2015  Vijayaditya Peddinti
+#           2015  Xingyu Na
+#           2015  Pegah Ghahrmani
+#           2017  Google  Inc. (vpeddinti@google.com)
+# Apache 2.0.
+
+
+
+
+# 6l is same as 6k but with two additional lstm layers
+
+# local/chain/compare_wer_general.sh lstm_6k_sp lstm_6l_sp
+# System                lstm_6k_sp lstm_6l_sp
+# WER on train_dev(tg)      14.50     14.41
+# WER on train_dev(fg)      13.36     13.11
+# WER on eval2000(tg)        17.1      17.3
+# WER on eval2000(fg)        15.6      15.6
+# Final train prob         -0.095    -0.093
+# Final valid prob         -0.113    -0.113
+# Final train prob (xent)        -1.271    -1.258
+# Final valid prob (xent)       -1.3479   -1.3307
+
+set -e
+
+# configs for 'chain'
+stage=12
+train_stage=-10
+get_egs_stage=-10
+speed_perturb=true
+dir=exp/chain/lstm_6l # Note: _sp will get added to this if $speed_perturb == true.
+decode_iter=
+decode_nj=50
+
+# training options
+xent_regularize=0.01
+self_repair_scale=0.00001
+label_delay=5
+
+chunk_left_context=40
+chunk_right_context=0
+# we'll put chunk-left-context-initial=0 and chunk-right-context-final=0
+# directly without variables.
+frames_per_chunk=140,100,160
+
+# (non-looped) decoding options
+frames_per_chunk_primary=$(echo $frames_per_chunk | cut -d, -f1)
+extra_left_context=50
+extra_right_context=0
+# we'll put extra-left-context-initial=0 and extra-right-context-final=0
+# directly without variables.
+
+
+remove_egs=false
+common_egs_dir=
+
+test_online_decoding=false  # if true, it will run the last decoding stage.
+
+# End configuration section.
+echo "$0 $@"  # Print the command line for logging
+
+. ./cmd.sh
+. ./path.sh
+. ./utils/parse_options.sh
+
+if ! cuda-compiled; then
+  cat <<EOF && exit 1
+This script is intended to be used with GPUs but you have not compiled Kaldi with CUDA
+If you want to use GPUs (and have them), go to src/, and configure and make on a machine
+where "nvcc" is installed.
+EOF
+fi
+
+# The iVector-extraction and feature-dumping parts are the same as the standard
+# nnet3 setup, and you can skip them by setting "--stage 8" if you have already
+# run those things.
+
+suffix=
+if [ "$speed_perturb" == "true" ]; then
+  suffix=_sp
+fi
+
+dir=${dir}$suffix
+train_set=train_nodup$suffix
+ali_dir=exp/tri4_ali_nodup$suffix
+treedir=exp/chain/tri5_7d_tree$suffix
+lang=data/lang_chain_2y
+
+
+# if we are using the speed-perturbed data we need to generate
+# alignments for it.
+local/nnet3/run_ivector_common.sh --stage $stage \
+  --speed-perturb $speed_perturb \
+  --generate-alignments $speed_perturb || exit 1;
+
+
+if [ $stage -le 9 ]; then
+  # Get the alignments as lattices (gives the CTC training more freedom).
+  # use the same num-jobs as the alignments
+  nj=$(cat exp/tri4_ali_nodup$suffix/num_jobs) || exit 1;
+  steps/align_fmllr_lats.sh --nj $nj --cmd "$train_cmd" data/$train_set \
+    data/lang exp/tri4 exp/tri4_lats_nodup$suffix
+  rm exp/tri4_lats_nodup$suffix/fsts.*.gz # save space
+fi
+
+
+if [ $stage -le 10 ]; 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 [ $stage -le 11 ]; then
+  # Build a tree using our new topology.
+  steps/nnet3/chain/build_tree.sh --frame-subsampling-factor 3 \
+      --context-opts "--context-width=2 --central-position=1" \
+      --cmd "$train_cmd" 7000 data/$train_set $lang $ali_dir $treedir
+fi
+
+if [ $stage -le 12 ]; then
+  echo "$0: creating neural net configs using the xconfig parser";
+
+  num_targets=$(tree-info $treedir/tree |grep num-pdfs|awk '{print $2}')
+  [ -z $num_targets ] && { echo "$0: error getting num-targets"; exit 1; }
+  learning_rate_factor=$(echo "print 0.5/$xent_regularize" | python)
+
+  lstm_opts="decay-time=20"
+
+  mkdir -p $dir/configs
+  cat <<EOF > $dir/configs/network.xconfig
+  input dim=100 name=ivector
+  input dim=40 name=input
+
+  # please note that it is important to have input layer with the name=input
+  # as the layer immediately preceding the fixed-affine-layer to enable
+  # the use of short notation for the descriptor
+  fixed-affine-layer name=lda input=Append(-2,-1,0,1,2,ReplaceIndex(ivector, t, 0)) affine-transform-file=$dir/configs/lda.mat
+
+  # check steps/libs/nnet3/xconfig/lstm.py for the other options and defaults
+  fast-lstmp-layer name=fastlstm1 cell-dim=768 recurrent-projection-dim=256 non-recurrent-projection-dim=256 delay=-3 $lstm_opts
+  fast-lstmp-layer name=fastlstm2 cell-dim=768 recurrent-projection-dim=256 non-recurrent-projection-dim=256 delay=-3 $lstm_opts
+  fast-lstmp-layer name=fastlstm3 cell-dim=768 recurrent-projection-dim=256 non-recurrent-projection-dim=256 delay=-3 $lstm_opts
+  fast-lstmp-layer name=fastlstm4 cell-dim=768 recurrent-projection-dim=256 non-recurrent-projection-dim=256 delay=-3 $lstm_opts
+  fast-lstmp-layer name=fastlstm5 cell-dim=768 recurrent-projection-dim=256 non-recurrent-projection-dim=256 delay=-3 $lstm_opts
+
+  ## adding the layers for chain branch
+  output-layer name=output input=fastlstm5 output-delay=$label_delay include-log-softmax=false dim=$num_targets max-change=1.5
+
+  # adding the layers for xent branch
+  # This block prints the configs for a separate output that will be
+  # trained with a cross-entropy objective in the 'chain' models... this
+  # has the effect of regularizing the hidden parts of the model.  we use
+  # 0.5 / args.xent_regularize as the learning rate factor- the factor of
+  # 0.5 / args.xent_regularize is suitable as it means the xent
+  # final-layer learns at a rate independent of the regularization
+  # constant; and the 0.5 was tuned so as to make the relative progress
+  # similar in the xent and regular final layers.
+  output-layer name=output-xent input=fastlstm5 output-delay=$label_delay 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 13 ]; 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/swbd-$(date +'%m_%d_%H_%M')/s5c/$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" \
+    --trainer.num-chunk-per-minibatch 64,32 \
+    --trainer.frames-per-iter 1500000 \
+    --trainer.max-param-change 2.0 \
+    --trainer.num-epochs 4 \
+    --trainer.optimization.shrink-value 0.99 \
+    --trainer.optimization.num-jobs-initial 3 \
+    --trainer.optimization.num-jobs-final 16 \
+    --trainer.optimization.initial-effective-lrate 0.001 \
+    --trainer.optimization.final-effective-lrate 0.0001 \
+    --trainer.optimization.momentum 0.0 \
+    --trainer.deriv-truncate-margin 8 \
+    --egs.stage $get_egs_stage \
+    --egs.opts "--frames-overlap-per-eg 0" \
+    --egs.chunk-width $frames_per_chunk \
+    --egs.chunk-left-context $chunk_left_context \
+    --egs.chunk-right-context $chunk_right_context \
+    --egs.chunk-left-context-initial 0 \
+    --egs.chunk-right-context-final 0 \
+    --egs.dir "$common_egs_dir" \
+    --cleanup.remove-egs $remove_egs \
+    --feat-dir data/${train_set}_hires \
+    --tree-dir $treedir \
+    --lat-dir exp/tri4_lats_nodup$suffix \
+    --dir $dir  || exit 1;
+fi
+
+if [ $stage -le 14 ]; 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_sw1_tg $dir $dir/graph_sw1_tg
+fi
+
+
+graph_dir=$dir/graph_sw1_tg
+iter_opts=
+if [ ! -z $decode_iter ]; then
+  iter_opts=" --iter $decode_iter "
+fi
+
+if [ $stage -le 15 ]; then
+  rm $dir/.error 2>/dev/null || true
+  for decode_set in train_dev eval2000; do
+      (
+        steps/nnet3/decode.sh --num-threads 4 \
+          --acwt 1.0 --post-decode-acwt 10.0 \
+          --nj 25 --cmd "$decode_cmd" $iter_opts \
+          --extra-left-context $extra_left_context  \
+          --extra-right-context $extra_right_context  \
+          --extra-left-context-initial 0 \
+          --extra-right-context-final 0 \
+          --frames-per-chunk "$frames_per_chunk_primary" \
+          --online-ivector-dir exp/nnet3/ivectors_${decode_set} \
+         $graph_dir data/${decode_set}_hires \
+         $dir/decode_${decode_set}${decode_iter:+_$decode_iter}_sw1_tg || exit 1;
+      if $has_fisher; then
+          steps/lmrescore_const_arpa.sh --cmd "$decode_cmd" \
+            data/lang_sw1_{tg,fsh_fg} data/${decode_set}_hires \
+            $dir/decode_${decode_set}${decode_iter:+_$decode_iter}_sw1_{tg,fsh_fg} || exit 1;
+      fi
+      ) &
+  done
+  wait
+  if [ -f $dir/.error ]; then
+    echo "$0: something went wrong in decoding"
+    exit 1
+  fi
+fi
+
+if [ $stage -le 16 ]; then
+  # looped decoding.  Note: this does not make sense for BLSTMs or other
+  # backward-recurrent setups, and for TDNNs and other non-recurrent there is no
+  # point doing it because it would give identical results to regular decoding.
+  rm $dir/.error 2>/dev/null || true
+  for decode_set in train_dev eval2000; do
+    (
+      steps/nnet3/decode_looped.sh \
+         --acwt 1.0 --post-decode-acwt 10.0 \
+         --nj $decode_nj --cmd "$decode_cmd" $iter_opts \
+         --online-ivector-dir exp/nnet3/ivectors_${decode_set} \
+         $graph_dir data/${decode_set}_hires \
+         $dir/decode_${decode_set}${decode_iter:+_$decode_iter}_sw1_tg_looped || exit 1;
+      if $has_fisher; then
+          steps/lmrescore_const_arpa.sh --cmd "$decode_cmd" \
+            data/lang_sw1_{tg,fsh_fg} data/${decode_set}_hires \
+            $dir/decode_${decode_set}${decode_iter:+_$decode_iter}_sw1_{tg,fsh_fg}_looped || exit 1;
+      fi
+      ) &
+  done
+  wait
+  if [ -f $dir/.error ]; then
+    echo "$0: something went wrong in looped decoding"
+    exit 1
+  fi
+fi
+
+if $test_online_decoding && [ $stage -le 17 ]; then
+  # note: if the features change (e.g. you add pitch features), you will have to
+  # change the options of the following command line.
+  steps/online/nnet3/prepare_online_decoding.sh \
+       --mfcc-config conf/mfcc_hires.conf \
+       $lang exp/nnet3/extractor $dir ${dir}_online
+
+  rm $dir/.error 2>/dev/null || true
+  for decode_set in train_dev eval2000; do
+    (
+      # note: we just give it "$decode_set" as it only uses the wav.scp, the
+      # feature type does not matter.
+
+      steps/online/nnet3/decode.sh --nj $decode_nj --cmd "$decode_cmd" $iter_opts \
+          --acwt 1.0 --post-decode-acwt 10.0 \
+         $graph_dir data/${decode_set}_hires \
+         ${dir}_online/decode_${decode_set}${decode_iter:+_$decode_iter}_sw1_tg || exit 1;
+      if $has_fisher; then
+          steps/lmrescore_const_arpa.sh --cmd "$decode_cmd" \
+            data/lang_sw1_{tg,fsh_fg} data/${decode_set}_hires \
+            ${dir}_online/decode_${decode_set}${decode_iter:+_$decode_iter}_sw1_{tg,fsh_fg} || exit 1;
+      fi
+    ) || touch $dir/.error &
+  done
+  wait
+  if [ -f $dir/.error ]; then
+    echo "$0: something went wrong in online decoding"
+    exit 1
+  fi
+fi
+
+exit 0;