Merge branch 'master' of https://github.com/kaldi-asr/kaldi

* 'master' of https://github.com/kaldi-asr/kaldi:
  [egs] Add chain recipe for Fisher English (kaldi-asr#1803)
  [src] Cosmetic fix to usage message (kaldi-asr#1800)
  [egs] Fix bug RE xent_regularize in Aspire chain recipes. (kaldi-asr#1797)

egs/aspire/s5/local/chain/tuning/run_blstm_7b.sh

@@ -20,6 +20,7 @@ num_data_reps=3
 
 
 min_seg_len=
+xent_regularize=0.1
 chunk_width=150
 chunk_left_context=40
 chunk_right_context=40
@@ -197,7 +198,7 @@ if [ $stage -le 12 ]; then
     --cmd "$decode_cmd" \
     --feat.online-ivector-dir exp/nnet3/ivectors_train_min${min_seg_len} \
     --feat.cmvn-opts "--norm-means=false --norm-vars=false" \
-    --chain.xent-regularize 0.1 \
+    --chain.xent-regularize $xent_regularize \
     --chain.leaky-hmm-coefficient 0.1 \
     --chain.l2-regularize 0.00005 \
     --chain.apply-deriv-weights false \

egs/aspire/s5/local/chain/tuning/run_tdnn_7b.sh

@@ -20,6 +20,7 @@ num_data_reps=3
 
 
 min_seg_len=
+xent_regularize=0.1
 frames_per_eg=150
 # End configuration section.
 echo "$0 $@"  # Print the command line for logging
@@ -191,7 +192,7 @@ if [ $stage -le 12 ]; then
     --cmd "$decode_cmd" \
     --feat.online-ivector-dir exp/nnet3/ivectors_train_min${min_seg_len} \
     --feat.cmvn-opts "--norm-means=false --norm-vars=false" \
-    --chain.xent-regularize 0.1 \
+    --chain.xent-regularize $xent_regularize \
     --chain.leaky-hmm-coefficient 0.1 \
     --chain.l2-regularize 0.00005 \
     --chain.apply-deriv-weights false \

egs/fisher_english/s5/local/chain/compare_wer_general.sh

@@ -0,0 +1,106 @@
+#!/bin/bash
+
+# this script is used for comparing decoding results between systems.
+# e.g. local/chain/compare_wer_general.sh exp/chain_cleaned/tdnn_{c,d}_sp
+# For use with discriminatively trained systems you specify the epochs after a colon:
+# for instance,
+# local/chain/compare_wer_general.sh exp/chain_cleaned/tdnn_c_sp exp/chain_cleaned/tdnn_c_sp_smbr:{1,2,3}
+
+
+echo "# $0 $*"
+
+include_looped=false
+if [ "$1" == "--looped" ]; then
+  include_looped=true
+  shift
+fi
+
+used_epochs=false
+
+# this function set_names is used to separate the epoch-related parts of the name
+# [for discriminative training] and the regular parts of the name.
+# If called with a colon-free directory name, like:
+#  set_names exp/chain_cleaned/tdnn_lstm1e_sp_bi_smbr
+# it will set dir=exp/chain_cleaned/tdnn_lstm1e_sp_bi_smbr and epoch_infix=""
+# If called with something like:
+#  set_names exp/chain_cleaned/tdnn_d_sp_smbr:3
+# it will set dir=exp/chain_cleaned/tdnn_d_sp_smbr and epoch_infix="_epoch3"
+
+
+set_names() {
+  if [ $# != 1 ]; then
+    echo "compare_wer_general.sh: internal error"
+    exit 1  # exit the program
+  fi
+  dirname=$(echo $1 | cut -d: -f1)
+  epoch=$(echo $1 | cut -s -d: -f2)
+  if [ -z $epoch ]; then
+    epoch_infix=""
+  else
+    used_epochs=true
+    epoch_infix=_epoch${epoch}
+  fi
+}
+
+
+
+echo -n "# System               "
+for x in $*; do   printf "% 10s" " $(basename $x)";   done
+echo
+
+strings=("# WER on dev    "  "# WER on test    ")
+
+for n in 0 1; do
+   echo -n "${strings[$n]}"
+   for x in $*; do
+     set_names $x  # sets $dirname and $epoch_infix
+     decode_names=(dev${epoch_infix} test${epoch_infix})
+     wer=$(grep WER $dirname/decode_${decode_names[$n]}/wer* | utils/best_wer.sh | awk '{print $2}')
+     printf "% 10s" $wer
+   done
+   echo
+   if $include_looped; then
+     echo -n "#         [looped:]    "
+     for x in $*; do
+       set_names $x  # sets $dirname and $epoch_infix
+       decode_names=(dev${epoch_infix} test${epoch_infix})
+       wer=$(grep WER $dirname/decode_looped_${decode_names[$n]}/wer* | utils/best_wer.sh | awk '{print $2}')
+       printf "% 10s" $wer
+     done
+     echo
+   fi
+done
+
+
+if $used_epochs; then
+  exit 0;  # the diagnostics aren't comparable between regular and discriminatively trained systems.
+fi
+
+echo -n "# Final train prob     "
+for x in $*; do
+  prob=$(grep Overall $x/log/compute_prob_train.final.log | grep -v xent | awk '{printf("%.4f", $8)}')
+  printf "% 10s" $prob
+done
+echo
+
+echo -n "# Final valid prob     "
+for x in $*; do
+  prob=$(grep Overall $x/log/compute_prob_valid.final.log | grep -v xent | awk '{printf("%.4f", $8)}')
+  printf "% 10s" $prob
+done
+echo
+
+echo -n "# Final train prob (xent)"
+for x in $*; do
+  prob=$(grep Overall $x/log/compute_prob_train.final.log | grep -w xent | awk '{printf("%.4f", $8)}')
+  printf "% 10s" $prob
+done
+echo
+
+echo -n "# Final valid prob (xent)"
+for x in $*; do
+  prob=$(grep Overall $x/log/compute_prob_valid.final.log | grep -w xent | awk '{printf("%.4f", $8)}')
+  printf "% 10s" $prob
+done
+
+echo

egs/fisher_english/s5/local/chain/run_tdnn.sh

@@ -0,0 +1,217 @@
+#!/bin/bash
+set -e
+
+# Based on run_tdnn_7b.sh in the fisher swbd recipe
+
+# Results on a 350hr random subset of fisher english:
+# local/chain/compare_wer_general.sh exp/chain_350k/tdnn7b_sp
+# System                tdnn7b_sp
+# WER on dev         17.74
+# WER on test         17.57
+# Final train prob        -0.1128
+# Final valid prob        -0.1251
+# Final train prob (xent)   -1.7908
+# Final valid prob (xent)   -1.7712
+
+# steps/info/nnet3_dir_info.pl exp/chain_350k/tdnn7b_sp
+# exp/chain_350k/tdnn7b_sp: num-iters=319 nj=3..16 num-params=22.1M dim=40+100->8617 combine=-0.14->-0.13
+
+
+
+# configs for 'chain'
+stage=0
+tdnn_affix=7b
+train_stage=-10
+get_egs_stage=-10
+decode_iter=
+train_set=train
+tree_affix=
+nnet3_affix=
+xent_regularize=0.1
+hidden_dim=725
+num_leaves=11000
+
+# training options
+num_epochs=4
+remove_egs=false
+common_egs_dir=
+minibatch_size=128
+num_jobs_initial=3
+num_jobs_final=16
+initial_effective_lrate=0.001
+final_effective_lrate=0.0001
+frames_per_iter=1500000
+
+gmm=tri5a
+build_tree_ali_dir=exp/tri4a_ali  # used to make a new tree for chain topology, should match train data
+# End configuration section.
+echo "$0 $@"  # Print the command line for logging
+
+. ./cmd.sh
+. ./path.sh
+. ./utils/parse_options.sh
+
+if ! cuda-compiled; then
+  cat <<EOF && exit 1
+This script is intended to be used with GPUs but you have not compiled Kaldi with CUDA
+If you want to use GPUs (and have them), go to src/, and configure and make on a machine
+where "nvcc" is installed.
+EOF
+fi
+
+
+gmm_dir=exp/$gmm   # used to get training lattices (for chain supervision)
+treedir=exp/chain${nnet3_affix}/tree_${tree_affix}
+lat_dir=exp/chain${nnet3_affix}/tri5a_${train_set}_sp_lats  # training lattices directory
+dir=exp/chain${nnet3_affix}/tdnn${tdnn_affix}_sp
+train_data_dir=data/${train_set}_sp_hires
+lores_train_data_dir=data/${train_set}_sp
+build_tree_train_data_dir=data/${train_set}
+train_ivector_dir=exp/nnet3${nnet3_affix}/ivectors_${train_set}_sp_hires
+lang=data/lang_chain
+
+
+# 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.
+local/nnet3/run_ivector_common.sh --stage $stage \
+                                  --speed-perturb true \
+                                  --train-set $train_set \
+                                  --nnet3-affix $nnet3_affix \
+                                  --generate-alignments false || exit 1;
+
+if [ $stage -le 9 ]; then
+  # Get the alignments as lattices (gives the chain training more freedom).
+  # use the same num-jobs as the alignments
+  nj=$(cat $build_tree_ali_dir/num_jobs) || exit 1;
+  steps/align_fmllr_lats.sh --nj $nj --cmd "$train_cmd" $lores_train_data_dir \
+    data/lang $gmm_dir $lat_dir || exit 1;
+  rm $lat_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.]
+  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 [ $stage -le 11 ]; then
+  # Build a tree using our new topology.
+  steps/nnet3/chain/build_tree.sh --frame-subsampling-factor 3 \
+      --leftmost-questions-truncate -1 \
+      --cmd "$train_cmd" $num_leaves $build_tree_train_data_dir $lang $build_tree_ali_dir $treedir || exit 1;
+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)
+
+  mkdir -p $dir/configs
+  cat <<EOF > $dir/configs/network.xconfig
+  input dim=100 name=ivector
+  input dim=40 name=input
+
+  # please note that it is important to have input layer with the name=input
+  # as the layer immediately preceding the fixed-affine-layer to enable
+  # the use of short notation for the descriptor
+  fixed-affine-layer name=lda input=Append(-1,0,1,ReplaceIndex(ivector, t, 0)) affine-transform-file=$dir/configs/lda.mat
+
+  # the first splicing is moved before the lda layer, so no splicing here
+  relu-batchnorm-layer name=tdnn1 dim=$hidden_dim
+  relu-batchnorm-layer name=tdnn2 input=Append(-1,0,1,2) dim=$hidden_dim
+  relu-batchnorm-layer name=tdnn3 input=Append(-3,0,3) dim=$hidden_dim
+  relu-batchnorm-layer name=tdnn4 input=Append(-3,0,3) dim=$hidden_dim
+  relu-batchnorm-layer name=tdnn5 input=Append(-3,0,3) dim=$hidden_dim
+  relu-batchnorm-layer name=tdnn6 input=Append(-6,-3,0) dim=$hidden_dim
+
+  ## adding the layers for chain branch
+  relu-batchnorm-layer name=prefinal-chain input=tdnn6 dim=$hidden_dim target-rms=0.5
+  output-layer name=output include-log-softmax=false dim=$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=$hidden_dim target-rms=0.5
+  output-layer name=output-xent dim=$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 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/fisher_english-$(date +'%m_%d_%H_%M')/s5c/$dir/egs/storage $dir/egs/storage
+  fi
+
+  touch $dir/egs/.nodelete # keep egs around when that run dies.
+
+  steps/nnet3/chain/train.py --stage $train_stage \
+    --egs.dir "$common_egs_dir" \
+    --cmd "$decode_cmd" \
+    --feat.online-ivector-dir $train_ivector_dir \
+    --feat.cmvn-opts "--norm-means=false --norm-vars=false" \
+    --chain.xent-regularize 0.1 \
+    --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.stage $get_egs_stage \
+    --egs.opts "--frames-overlap-per-eg 0" \
+    --egs.chunk-width 150 \
+    --trainer.num-chunk-per-minibatch $minibatch_size \
+    --trainer.frames-per-iter $frames_per_iter \
+    --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 2.0 \
+    --cleanup.remove-egs $remove_egs \
+    --feat-dir $train_data_dir \
+    --tree-dir $treedir \
+    --lat-dir $lat_dir \
+    --dir $dir  || exit 1;
+fi
+
+graph_dir=$dir/graph
+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_test $dir $graph_dir
+fi
+
+decode_suff=
+if [ $stage -le 15 ]; then
+  iter_opts=
+  if [ ! -z $decode_iter ]; then
+    iter_opts=" --iter $decode_iter "
+  fi
+  for decode_set in dev test; do
+      (
+      num_jobs=`cat data/${decode_set}_hires/utt2spk|cut -d' ' -f2|sort -u|wc -l`
+      steps/nnet3/decode.sh --acwt 1.0 --post-decode-acwt 10.0 \
+          --nj $num_jobs --cmd "$decode_cmd" $iter_opts \
+          --online-ivector-dir exp/nnet3${nnet3_affix}/ivectors_${decode_set}_hires \
+          $graph_dir data/${decode_set}_hires $dir/decode_${decode_set}${decode_iter:+_$decode_iter}${decode_suff} || exit 1;
+      ) &
+  done
+fi
+wait;
+exit 0;