[src,scripts,egs] Speedups to GRU-based networks (special components) (…

…#2712)

egs/swbd/s5c/local/chain/tuning/run_tdnn_opgru_1a.sh

@@ -4,31 +4,36 @@
 
 # This is based on TDNN_LSTM_1b, but using the NormOPGRU to replace the LSTMP,
 # and adding chunk-{left,right}-context-initial=0
+# For the details of OPGRU structure, please check the paper
+# "Output-Gate Projected Gated Recurrent Unit for Speech Recognition"
+# by Gaofeng Cheng et al,
+# http://www.danielpovey.com/files/2018_interspeech_opgru.pdf
+
 # Different from the vanilla OPGRU, Norm-OPGRU adds batchnorm in its output (forward direction)
 # and renorm in its recurrence. Experiments show that the TDNN-NormOPGRU could achieve similar
 # results than TDNN-LSTMP and BLSTMP in both large or small data sets (80 ~ 2300 Hrs).
 
 # ./local/chain/compare_wer_general.sh --looped tdnn_lstm_1e_sp tdnn_opgru_1a_sp
 # System                tdnn_lstm_1e_sp tdnn_opgru_1a_sp
-# WER on train_dev(tg)      12.81     12.39
-#           [looped:]       12.93     12.32
-# WER on train_dev(fg)      11.92     11.39
-#           [looped:]       12.07     11.35
+# WER on train_dev(tg)      12.81     12.31
+#           [looped:]       12.93     12.26
+# WER on train_dev(fg)      11.92     11.60
+#           [looped:]       12.07     11.65
 # WER on eval2000(tg)        15.6      15.1
 #           [looped:]        16.0      15.1
-# WER on eval2000(fg)        14.1      13.6
+# WER on eval2000(fg)        14.1      13.5
 #           [looped:]        14.5      13.5
-# Final train prob         -0.065    -0.066
-# Final valid prob         -0.087    -0.085
-# Final train prob (xent)        -0.918    -0.889
-# Final valid prob (xent)       -1.0309   -0.9837
+# Final train prob         -0.065    -0.068
+# Final valid prob         -0.087    -0.091
+# Final train prob (xent)        -0.918    -0.879
+# Final valid prob (xent)       -1.0309   -0.9667
 
 
 
 set -e
 
 # configs for 'chain'
-stage=12
+stage=0
 train_stage=-10
 get_egs_stage=-10
 speed_perturb=true

egs/swbd/s5c/local/chain/tuning/run_tdnn_opgru_1b.sh

@@ -0,0 +1,315 @@
+#!/bin/bash
+# Apache 2.0
+
+# This is based on TDNN_OPGRU_1A, but using the FastNormOPGRU to replace the NormPGRU.
+# For the details of OPGRU structure, please check the paper
+# "Output-Gate Projected Gated Recurrent Unit for Speech Recognition"
+# by Gaofeng Cheng et al,
+# http://www.danielpovey.com/files/2018_interspeech_opgru.pdf
+
+# Different from the vanilla OPGRU, Norm-OPGRU adds batchnorm in its output (forward direction)
+# and renorm in its recurrence. Experiments show that the TDNN-NormOPGRU could achieve similar
+# results than TDNN-LSTMP and BLSTMP in both large or small data sets (80 ~ 2300 Hrs).
+
+# ./local/chain/compare_wer_general.sh --looped tdnn_opgru_1a_sp tdnn_opgru_1b_sp
+# System                tdnn_opgru_1a_sp tdnn_opgru_1b_sp
+# WER on train_dev(tg)      12.31     12.41
+#           [looped:]       12.26     12.38
+# WER on train_dev(fg)      11.49     11.60
+#           [looped:]       11.43     11.65
+# WER on eval2000(tg)        14.9      15.1
+#           [looped:]        15.0      15.1
+# WER on eval2000(fg)        13.5      13.7
+#           [looped:]        13.5      13.7
+# Final train prob         -0.068    -0.070
+# Final valid prob         -0.091    -0.092
+# Final train prob (xent)        -0.879    -0.889
+# Final valid prob (xent)       -0.9667   -0.9723
+
+
+
+set -e
+
+# configs for 'chain'
+stage=0
+train_stage=-10
+get_egs_stage=-10
+speed_perturb=true
+dir=exp/chain/tdnn_opgru_1b # 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=0
+xent_regularize=0.025
+self_repair_scale=0.00001
+label_delay=5
+dropout_schedule='0,0@0.20,0.2@0.50,0'
+# decode options
+extra_left_context=50
+extra_right_context=0
+frames_per_chunk=
+test_online_decoding=
+
+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:+_$affix}
+dir=${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 $2}')
+  learning_rate_factor=$(echo "print 0.5/$xent_regularize" | python)
+  gru_opts="dropout-per-frame=true dropout-proportion=0.0 gru-nonlinearity-options=\"max-change=0.75\""
+
+  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=1024
+  relu-batchnorm-layer name=tdnn2 input=Append(-1,0,1) dim=1024
+  relu-batchnorm-layer name=tdnn3 input=Append(-1,0,1) dim=1024
+
+  # check steps/libs/nnet3/xconfig/gru.py for the other options and defaults
+  fast-norm-opgru-layer name=opgru1 cell-dim=1024 recurrent-projection-dim=256 non-recurrent-projection-dim=256 delay=-3 $gru_opts
+  relu-batchnorm-layer name=tdnn4 input=Append(-3,0,3) dim=1024
+  relu-batchnorm-layer name=tdnn5 input=Append(-3,0,3) dim=1024
+  fast-norm-opgru-layer name=opgru2 cell-dim=1024 recurrent-projection-dim=256 non-recurrent-projection-dim=256 delay=-3 $gru_opts
+  relu-batchnorm-layer name=tdnn6 input=Append(-3,0,3) dim=1024
+  relu-batchnorm-layer name=tdnn7 input=Append(-3,0,3) dim=1024
+  fast-norm-opgru-layer name=opgru3 cell-dim=1024 recurrent-projection-dim=256 non-recurrent-projection-dim=256 delay=-3 $gru_opts
+
+  ## adding the layers for chain branch
+  output-layer name=output input=opgru3 output-delay=$label_delay include-log-softmax=false dim=$num_targets max-change=1.5
+
+  # adding the layers for xent branch
+  # This block prints the configs for a separate output that will be
+  # trained with a cross-entropy objective in the 'chain' models... this
+  # has the effect of regularizing the hidden parts of the model.  we use
+  # 0.5 / args.xent_regularize as the learning rate factor- the factor of
+  # 0.5 / args.xent_regularize is suitable as it means the xent
+  # final-layer learns at a rate independent of the regularization
+  # constant; and the 0.5 was tuned so as to make the relative progress
+  # similar in the xent and regular final layers.
+  output-layer name=output-xent input=opgru3 output-delay=$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 \
+    --trainer.dropout-schedule $dropout_schedule \
+    --egs.chunk-left-context-initial 0 \
+    --egs.chunk-right-context-final 0 \
+    --egs.dir "$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 $dir/graph_sw1_tg
+fi
+
+decode_suff=sw1_tg
+graph_dir=$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 $extra_right_context  \
+          --extra-left-context-initial 0 \
+          --extra-right-context-final 0 \
+          --frames-per-chunk "$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
+
+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 50 --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
+
+if [ $stage -le 17 ]; then
+  rm $dir/.error 2>/dev/null || true
+  for decode_set in train_dev eval2000; do
+    (
+      steps/nnet3/decode_looped.sh \
+         --acwt 1.0 --post-decode-acwt 10.0 \
+         --nj 50 --cmd "$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
+
+wait;
+exit 0;