[egs] Improvements to multi_en tdnn-opgru/lstm recipes (#2824)

Author: xiaohui-zhang

egs/multi_en/s5/local/chain/run_tdnn_lstm.sh

@@ -0,0 +1 @@
+tuning/run_tdnn_lstm_1a.sh

egs/multi_en/s5/local/chain/run_tdnn_opgru.sh

@@ -1 +1 @@
-tuning/run_tdnn_opgru_1a.sh
+tuning/run_tdnn_opgru_1b.sh

egs/multi_en/s5/local/chain/tuning/run_tdnn_lstm_1a.sh

@@ -0,0 +1,319 @@
+#!/bin/bash
+# Copyright 2017 University of Chinese Academy of Sciences (UCAS) Gaofeng Cheng
+#           2018 Xiaohui Zhang
+#           2018 Vimal Manohar
+# Apache 2.0
+
+# This recipe is similar with tdnn_lstm_1b recipefrom fisher_swbd/s5, and is currently
+# the best performing multi-en recipe.
+
+# System             tdnn_opgru_1b_sp  tdnn_lstm_1a_sp
+# WER on eval2000(tg)           11.4     11.4
+# WER on eval2000(fg)           11.2     11.2
+# WER on rt03(tg)               11.1     10.7
+# WER on rt03(fg)               10.8     10.5
+# Final train prob            -0.091    -0.095
+# Final valid prob             -0.091   -0.089
+# Final train prob (xent)     -0.990    -0.970
+# Final valid prob (xent)     -0.091    -0.9638
+# Num-parameters            34976320    39704128
+
+# ./steps/info/chain_dir_info.pl exp/multi_a/chain/tdnn_lstm_1a_sp
+# exp/multi_a/chain/tdnn_lstm_1a_sp: num-iters=2096 nj=3..16 num-params=39.7M dim=40+100->6176 combine=-0.088->-0.087 (over 3) 
+# xent:train/valid[1395,2095,final]=(-1.38,-0.960,-0.970/-1.39,-0.964,-0.964) 
+# logprob:train/valid[1395,2095,final]=(-0.117,-0.091,-0.095/-0.109,-0.087,-0.089)
+
+# online results
+# Eval2000
+# %WER 14.2 | 2628 21594 | 87.8 8.6 3.5 2.1 14.2 49.1 | exp/multi_a/chain/tdnn_lstm_1a_sp_online/decode_eval2000/score_8_0.0/eval2000_hires.ctm.callhm.filt.sys
+# %WER 11.4 | 4459 42989 | 90.3 7.0 2.7 1.7 11.4 46.1 | exp/multi_a/chain/tdnn_lstm_1a_sp_online/decode_eval2000/score_8_0.0/eval2000_hires.ctm.filt.sys
+# %WER 8.4 | 1831 21395 | 92.8 5.3 2.0 1.2 8.4 41.2 | exp/multi_a/chain/tdnn_lstm_1a_sp_online/decode_eval2000/score_9_0.0/eval2000_hires.ctm.swbd.filt.sys
+# %WER 14.0 | 2628 21594 | 88.0 8.5 3.4 2.1 14.0 48.6 | exp/multi_a/chain/tdnn_lstm_1a_sp_online/decode_eval2000_fg/score_8_0.0/eval2000_hires.ctm.callhm.filt.sys
+# %WER 11.2 | 4459 42989 | 90.5 6.9 2.6 1.7 11.2 45.4 | exp/multi_a/chain/tdnn_lstm_1a_sp_online/decode_eval2000_fg/score_8_0.0/eval2000_hires.ctm.filt.sys
+# %WER 8.1 | 1831 21395 | 93.1 5.1 1.8 1.2 8.1 40.6 | exp/multi_a/chain/tdnn_lstm_1a_sp_online/decode_eval2000_fg/score_9_0.0/eval2000_hires.ctm.swbd.filt.sys
+
+# RT03
+# %WER 8.7 | 3970 36721 | 92.2 5.3 2.5 1.0 8.7 37.3 | exp/multi_a/chain/tdnn_lstm_1a_sp_online/decode_rt03/score_7_0.0/rt03_hires.ctm.fsh.filt.sys
+# %WER 10.8 | 8420 76157 | 90.4 6.5 3.2 1.2 10.8 40.1 | exp/multi_a/chain/tdnn_lstm_1a_sp_online/decode_rt03/score_8_0.0/rt03_hires.ctm.filt.sys
+# %WER 12.7 | 4450 39436 | 88.7 7.7 3.6 1.4 12.7 42.5 | exp/multi_a/chain/tdnn_lstm_1a_sp_online/decode_rt03/score_8_0.0/rt03_hires.ctm.swbd.filt.sys
+# %WER 8.5 | 3970 36721 | 92.4 5.1 2.5 0.9 8.5 37.2 | exp/multi_a/chain/tdnn_lstm_1a_sp_online/decode_rt03_fg/score_7_1.0/rt03_hires.ctm.fsh.filt.sys
+# %WER 10.5 | 8420 76157 | 90.6 6.3 3.1 1.2 10.5 40.1 | exp/multi_a/chain/tdnn_lstm_1a_sp_online/decode_rt03_fg/score_8_0.0/rt03_hires.ctm.filt.sys
+# %WER 12.4 | 4450 39436 | 88.9 7.2 3.9 1.3 12.4 42.7 | exp/multi_a/chain/tdnn_lstm_1a_sp_online/decode_rt03_fg/score_9_0.0/rt03_hires.ctm.swbd.filt.sys
+
+set -e
+
+# configs for 'chain'
+stage=-10
+train_stage=-10
+get_egs_stage=-10
+speed_perturb=true
+multi=multi_a
+gmm=tri5a
+decode_iter=
+decode_dir_affix=
+decode_nj=50
+
+# training options
+frames_per_chunk=140,100,160
+frames_per_chunk_primary=$(echo $frames_per_chunk | cut -d, -f1)
+chunk_left_context=40
+chunk_right_context=0
+xent_regularize=0.025
+self_repair_scale=0.00001
+label_delay=5
+# decode options
+extra_left_context=50
+extra_right_context=0
+dropout_schedule='0,0@0.20,0.3@0.50,0'
+num_epochs=4
+
+remove_egs=false
+common_egs_dir=
+
+test_online_decoding=true  # if true, it will run the last decoding stage.
+
+nnet3_affix=
+tdnn_affix=_1a
+
+# 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=exp/$multi/chain/tdnn_lstm${tdnn_affix}${suffix}
+train_set=${multi}/${gmm}${suffix}
+lats_dir=exp/${multi}/${gmm}_lats_nodup${suffix}
+treedir=exp/$multi/chain/${gmm}_tree
+lang=data/${multi}/lang_${gmm}_chain
+lang_dir=data/lang_${multi}_${gmm}_fsh_sw1_tg
+rescore_lang_dir=data/lang_${multi}_${gmm}_fsh_sw1_fg
+
+local/nnet3/run_ivector_common.sh --stage $stage --nnet3-affix "$nnet3_affix" \
+  --multi $multi \
+  --gmm $gmm \
+  --speed-perturb $speed_perturb || exit 1
+
+online_ivector_dir=exp/$multi/nnet3${nnet3_affix}/ivectors_${train_set}
+
+if [ $stage -le 9 ]; then
+  steps/align_fmllr_lats.sh --nj 100 --cmd "$train_cmd" \
+    --generate-ali-from-lats true data/$train_set \
+    data/lang_${multi}_${gmm} exp/${multi}/$gmm $lats_dir
+  rm ${lats_dir}/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.]
+  if [ -d $lang ]; then 
+    echo "$lang exists. Remove it or skip this stage."
+    exit 1
+  fi
+
+  cp -r data/lang_${multi}_${gmm} $lang
+  silphonelist=$(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.
+
+  if [ -f $treedir/final.mdl ]; then 
+    echo "$treedir exists. Remove it or skip this stage."
+    exit 1
+  fi
+
+  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 $lats_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 $2}')
+  learning_rate_factor=$(echo "print 0.5/$xent_regularize" | python)
+  lstm_opts="dropout-proportion=0.0 decay-time=40"
+
+  relu_dim=1024
+  cell_dim=1024
+  projection_dim=256
+
+  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
+  relu-batchnorm-layer name=tdnn1 dim=$relu_dim
+  relu-batchnorm-layer name=tdnn2 input=Append(-1,0,1) dim=$relu_dim
+  relu-batchnorm-layer name=tdnn3 input=Append(-1,0,1) dim=$relu_dim
+
+  # check steps/libs/nnet3/xconfig/lstm.py for the other options and defaults
+  fast-lstmp-layer name=lstm1 cell-dim=$cell_dim recurrent-projection-dim=$projection_dim non-recurrent-projection-dim=$projection_dim delay=-3 $lstm_opts
+  relu-batchnorm-layer name=tdnn4 input=Append(-3,0,3) dim=$relu_dim
+  relu-batchnorm-layer name=tdnn5 input=Append(-3,0,3) dim=$relu_dim
+  fast-lstmp-layer name=lstm2 cell-dim=$cell_dim recurrent-projection-dim=$projection_dim non-recurrent-projection-dim=$projection_dim delay=-3 $lstm_opts
+  relu-batchnorm-layer name=tdnn6 input=Append(-3,0,3) dim=$relu_dim
+  relu-batchnorm-layer name=tdnn7 input=Append(-3,0,3) dim=$relu_dim
+  fast-lstmp-layer name=lstm3 cell-dim=$cell_dim recurrent-projection-dim=$projection_dim non-recurrent-projection-dim=$projection_dim delay=-3 $lstm_opts
+
+  ## adding the layers for chain branch
+  output-layer name=output input=lstm3 output-delay=$label_delay include-log-softmax=false dim=$num_targets max-change=1.5
+
+  # adding the layers for xent branch
+  # This block prints the configs for a separate output that will be
+  # trained with a cross-entropy objective in the 'chain' models... this
+  # has the effect of regularizing the hidden parts of the model.  we use
+  # 0.5 / args.xent_regularize as the learning rate factor- the factor of
+  # 0.5 / args.xent_regularize is suitable as it means the xent
+  # final-layer learns at a rate independent of the regularization
+  # constant; and the 0.5 was tuned so as to make the relative progress
+  # similar in the xent and regular final layers.
+  output-layer name=output-xent input=lstm3 output-delay=$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/multi-en-$(date +'%m_%d_%H_%M')/s5/$dir/egs/storage $dir/egs/storage
+  fi
+
+  steps/nnet3/chain/train.py --stage $train_stage \
+    --cmd "$decode_cmd" \
+    --feat.online-ivector-dir exp/$multi/nnet3${nnet3_affix}/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 $num_epochs \
+    --trainer.optimization.shrink-value 0.99 \
+    --trainer.optimization.num-jobs-initial 3 \
+    --trainer.optimization.num-jobs-final 16 \
+    --trainer.optimization.initial-effective-lrate 0.001 \
+    --trainer.optimization.final-effective-lrate 0.0001 \
+    --trainer.dropout-schedule=$dropout_schedule \
+    --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 $lats_dir \
+    --dir $dir  || exit 1;
+fi
+
+lang_suffix=${lang_dir##*lang}
+
+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 $lang_dir \
+    $dir $dir/graph${lang_suffix}
+fi
+
+graph_dir=$dir/graph${lang_suffix}
+if [ $stage -le 15 ]; then
+  iter_opts=
+  [ -z $extra_left_context ] && extra_left_context=$chunk_left_context;
+  [ -z $extra_right_context ] && extra_right_context=$chunk_right_context;
+  if [ ! -z $decode_iter ]; then
+    iter_opts=" --iter $decode_iter "
+  fi
+  for decode_set in eval2000 rt03; do
+      (
+      steps/nnet3/decode.sh --acwt 1.0 --post-decode-acwt 10.0 \
+        --nj 50 --cmd "$decode_cmd" $iter_opts \
+        --extra-left-context $extra_left_context \
+        --extra-right-context $extra_right_context \
+        --extra-left-context-initial 0 \
+        --extra-right-context-final 0 \
+        --frames-per-chunk "$frames_per_chunk_primary" \
+        --online-ivector-dir exp/$multi/nnet3${nnet3_affix}/ivectors_${decode_set} \
+        $graph_dir data/${decode_set}_hires \
+         $dir/decode${lang_suffix}_${decode_set}${decode_dir_affix:+_$decode_dir_affix}${decode_iter:+_iter$decode_iter} 
+      
+      steps/lmrescore_const_arpa.sh --cmd "$decode_cmd" \
+          $lang_dir $rescore_lang_dir data/${decode_set}_hires \
+          $dir/decode${lang_suffix}_${decode_set}${decode_dir_affix:+_$decode_dir_affix}{,_fg}${decode_iter:+_iter$decode_iter} || exit 1;
+      ) &
+  done
+fi
+wait;
+
+if $test_online_decoding && [ $stage -le 16 ]; 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;