Add a BPE-based recipe for IAM

egs/iam/v2/cmd.sh

@@ -0,0 +1,13 @@
+# you can change cmd.sh depending on what type of queue you are using.
+# If you have no queueing system and want to run on a local machine, you
+# can change all instances 'queue.pl' to run.pl (but be careful and run
+# commands one by one: most recipes will exhaust the memory on your
+# machine).  queue.pl works with GridEngine (qsub).  slurm.pl works
+# with slurm.  Different queues are configured differently, with different
+# queue names and different ways of specifying things like memory;
+# to account for these differences you can create and edit the file
+# conf/queue.conf to match your queue's configuration.  Search for
+# conf/queue.conf in http://kaldi-asr.org/doc/queue.html for more information,
+# or search for the string 'default_config' in utils/queue.pl or utils/slurm.pl.
+
+export cmd="queue.pl"

egs/iam/v2/image

@@ -0,0 +1 @@
+../../cifar/v1/image/

egs/iam/v2/local/chain/compare_wer.sh

@@ -0,0 +1,90 @@
+#!/bin/bash
+
+# this script is used for comparing decoding results between systems.
+# e.g. local/chain/compare_wer.sh exp/chain/cnn{1a,1b}
+
+# Copyright      2017  Chun Chieh Chang
+#                2017  Ashish Arora
+
+if [ $# == 0 ]; then
+  echo "Usage: $0: <dir1> [<dir2> ... ]"
+  echo "e.g.: $0 exp/chain/cnn{1a,1b}"
+  exit 1
+fi
+. ./path.sh
+
+echo "# $0 $*"
+used_epochs=false
+
+echo -n "# System                     "
+for x in $*; do   printf "% 10s" " $(basename $x)";   done
+echo
+
+echo -n "# WER                        "
+for x in $*; do
+  wer=$(cat $x/decode_test/scoring_kaldi/best_wer | awk '{print $2}')
+  printf "% 10s" $wer
+done
+echo
+
+echo -n "# WER (rescored)             "
+for x in $*; do
+  wer="--"
+  [ -d $x/decode_test_rescored ] && wer=$(cat $x/decode_test_rescored/scoring_kaldi/best_wer | awk '{print $2}')
+  printf "% 10s" $wer
+done
+echo
+
+echo -n "# CER                        "
+for x in $*; do
+  cer=$(cat $x/decode_test/scoring_kaldi/best_cer | awk '{print $2}')
+  printf "% 10s" $cer
+done
+echo
+
+echo -n "# CER (rescored)             "
+for x in $*; do
+  cer="--"
+  [ -d $x/decode_test_rescored ] && cer=$(cat $x/decode_test_rescored/scoring_kaldi/best_cer | awk '{print $2}')
+  printf "% 10s" $cer
+done
+echo
+
+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
+
+echo -n "# Parameters                 "
+for x in $*; do
+  params=$(nnet3-info $x/final.mdl 2>/dev/null | grep num-parameters | cut -d' ' -f2 | awk '{printf "%0.2fM\n",$1/1000000}')
+  printf "% 10s" $params
+done
+echo

egs/iam/v2/local/chain/run_cnn_e2eali.sh

@@ -0,0 +1 @@
+tuning/run_cnn_e2eali_1c.sh

egs/iam/v2/local/chain/run_e2e_cnn.sh

@@ -0,0 +1,174 @@
+#!/bin/bash
+# Copyright    2017  Hossein Hadian
+
+# This script does end2end chain training (i.e. from scratch)
+
+# local/chain/compare_wer.sh exp/chain/cnn_1a exp/chain/cnn_chainali_1c exp/chain/e2e_cnn_1a
+# System                         cnn_1a cnn_chainali_1c e2e_cnn_1a
+# WER                             18.52     12.72     12.15
+# CER                             10.07      5.99      6.03
+# Final train prob              -0.0077   -0.0291   -0.0371
+# Final valid prob              -0.0970   -0.0359   -0.0636
+# Final train prob (xent)       -0.5484   -0.9781
+# Final valid prob (xent)       -0.9643   -1.1544
+# Parameters                      4.36M     3.96M     9.13M
+
+# steps/info/chain_dir_info.pl exp/chain/e2e_cnn_1a
+# exp/chain/e2e_cnn_1a: num-iters=21 nj=2..4 num-params=9.1M dim=40->12640 combine=-0.033->-0.033 (over 1) logprob:train/valid[13,20,final]=(-0.058,-0.042,-0.035/-0.070,-0.064,-0.059)
+
+set -e
+
+# configs for 'chain'
+stage=0
+train_stage=-10
+get_egs_stage=-10
+affix=1a
+
+# training options
+tdnn_dim=450
+num_epochs=4
+num_jobs_initial=2
+num_jobs_final=4
+minibatch_size=150=100,64/300=50,32/600=25,16/1200=16,8
+common_egs_dir=
+l2_regularize=0.00005
+frames_per_iter=1000000
+cmvn_opts="--norm-means=true --norm-vars=true"
+train_set=train
+lang_decode=data/lang
+lang_rescore=data/lang_rescore_6g
+
+# End configuration section.
+echo "$0 $@"  # Print the command line for logging
+
+. ./cmd.sh
+. ./path.sh
+. ./utils/parse_options.sh
+
+if ! cuda-compiled; then
+  cat <<EOF && exit 1
+This script is intended to be used with GPUs but you have not compiled Kaldi with CUDA
+If you want to use GPUs (and have them), go to src/, and configure and make on a machine
+where "nvcc" is installed.
+EOF
+fi
+
+lang=data/lang_e2e
+treedir=exp/chain/e2e_bitree  # it's actually just a trivial tree (no tree building)
+dir=exp/chain/e2e_cnn_${affix}
+
+if [ $stage -le 0 ]; 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 [ $stage -le 1 ]; then
+  steps/nnet3/chain/e2e/prepare_e2e.sh --nj 30 --cmd "$cmd" \
+                                       --shared-phones true \
+                                       --type biphone \
+                                       data/$train_set $lang $treedir
+  $cmd $treedir/log/make_phone_lm.log \
+  cat data/$train_set/text \| \
+    steps/nnet3/chain/e2e/text_to_phones.py data/lang \| \
+    utils/sym2int.pl -f 2- data/lang/phones.txt \| \
+    chain-est-phone-lm --num-extra-lm-states=500 \
+                       ark:- $treedir/phone_lm.fst
+fi
+
+if [ $stage -le 2 ]; then
+  echo "$0: creating neural net configs using the xconfig parser";
+  num_targets=$(tree-info $treedir/tree | grep num-pdfs | awk '{print $2}')
+
+  cnn_opts="l2-regularize=0.075"
+  tdnn_opts="l2-regularize=0.075"
+  output_opts="l2-regularize=0.1"
+  common1="$cnn_opts required-time-offsets= height-offsets=-2,-1,0,1,2 num-filters-out=36"
+  common2="$cnn_opts required-time-offsets= height-offsets=-2,-1,0,1,2 num-filters-out=70"
+  common3="$cnn_opts required-time-offsets= height-offsets=-1,0,1 num-filters-out=70"
+  mkdir -p $dir/configs
+  cat <<EOF > $dir/configs/network.xconfig
+  input dim=40 name=input
+
+  conv-relu-batchnorm-layer name=cnn1 height-in=40 height-out=40 time-offsets=-3,-2,-1,0,1,2,3 $common1
+  conv-relu-batchnorm-layer name=cnn2 height-in=40 height-out=20 time-offsets=-2,-1,0,1,2 $common1 height-subsample-out=2
+  conv-relu-batchnorm-layer name=cnn3 height-in=20 height-out=20 time-offsets=-4,-2,0,2,4 $common2
+  conv-relu-batchnorm-layer name=cnn4 height-in=20 height-out=20 time-offsets=-4,-2,0,2,4 $common2
+  conv-relu-batchnorm-layer name=cnn5 height-in=20 height-out=10 time-offsets=-4,-2,0,2,4 $common2 height-subsample-out=2
+  conv-relu-batchnorm-layer name=cnn6 height-in=10 height-out=10 time-offsets=-1,0,1 $common3
+  conv-relu-batchnorm-layer name=cnn7 height-in=10 height-out=10 time-offsets=-1,0,1 $common3
+  relu-batchnorm-layer name=tdnn1 input=Append(-4,-2,0,2,4) dim=$tdnn_dim $tdnn_opts
+  relu-batchnorm-layer name=tdnn2 input=Append(-4,0,4) dim=$tdnn_dim $tdnn_opts
+  relu-batchnorm-layer name=tdnn3 input=Append(-4,0,4) dim=$tdnn_dim $tdnn_opts
+
+  ## adding the layers for chain branch
+  relu-batchnorm-layer name=prefinal-chain dim=$tdnn_dim target-rms=0.5 $output_opts
+  output-layer name=output include-log-softmax=false dim=$num_targets max-change=1.5 $output_opts
+EOF
+
+  steps/nnet3/xconfig_to_configs.py --xconfig-file $dir/configs/network.xconfig --config-dir $dir/configs
+fi
+
+if [ $stage -le 3 ]; then
+  # no need to store the egs in a shared storage because we always
+  # remove them. Anyway, it takes only 5 minutes to generate them.
+
+  steps/nnet3/chain/e2e/train_e2e.py --stage $train_stage \
+    --cmd "$cmd" \
+    --feat.cmvn-opts "$cmvn_opts" \
+    --chain.leaky-hmm-coefficient 0.1 \
+    --chain.l2-regularize $l2_regularize \
+    --chain.apply-deriv-weights false \
+    --egs.dir "$common_egs_dir" \
+    --egs.stage $get_egs_stage \
+    --egs.opts "--num_egs_diagnostic 100 --num_utts_subset 400" \
+    --chain.frame-subsampling-factor 4 \
+    --chain.alignment-subsampling-factor 4 \
+    --trainer.num-chunk-per-minibatch $minibatch_size \
+    --trainer.frames-per-iter $frames_per_iter \
+    --trainer.num-epochs $num_epochs \
+    --trainer.optimization.momentum 0 \
+    --trainer.optimization.num-jobs-initial $num_jobs_initial \
+    --trainer.optimization.num-jobs-final $num_jobs_final \
+    --trainer.optimization.initial-effective-lrate 0.001 \
+    --trainer.optimization.final-effective-lrate 0.0001 \
+    --trainer.optimization.shrink-value 1.0 \
+    --trainer.max-param-change 2.0 \
+    --cleanup.remove-egs true \
+    --feat-dir data/${train_set} \
+    --tree-dir $treedir \
+    --dir $dir  || exit 1;
+fi
+
+if [ $stage -le 4 ]; then
+  # The reason we are using data/lang here, instead of $lang, is just to
+  # emphasize that it's not actually important to give mkgraph.sh the
+  # lang directory with the matched topology (since it gets the
+  # topology file from the model).  So you could give it a different
+  # lang directory, one that contained a wordlist and LM of your choice,
+  # as long as phones.txt was compatible.
+
+  utils/mkgraph.sh \
+    --self-loop-scale 1.0 $lang_decode \
+    $dir $dir/graph || exit 1;
+fi
+
+if [ $stage -le 5 ]; then
+  frames_per_chunk=$(echo $chunk_width | cut -d, -f1)
+  steps/nnet3/decode.sh --acwt 1.0 --post-decode-acwt 10.0 \
+    --nj 30 --cmd "$cmd" \
+    $dir/graph data/test $dir/decode_test || exit 1;
+
+  steps/lmrescore_const_arpa.sh --cmd "$cmd" $lang_decode $lang_rescore \
+                                data/test $dir/decode_test{,_rescored} || exit 1
+fi
+
+echo "Done. Date: $(date). Results:"
+local/chain/compare_wer.sh $dir