OCR: Decomposition for CASIA and YOMDLE_ZH datasets

egs/casia_hwdb/v1/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/casia_hwdb/v1/image

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

egs/casia_hwdb/v1/local/augment_data.sh

@@ -0,0 +1,37 @@
+#!/bin/bash
+# Copyright   2018 Hossein Hadian
+#             2018 Ashish Arora
+
+# Apache 2.0
+# This script performs data augmentation.
+
+nj=4
+cmd=run.pl
+feat_dim=40
+fliplr=false
+verticle_shift=0
+echo "$0 $@"
+
+. ./cmd.sh
+. ./path.sh
+. ./utils/parse_options.sh || exit 1;
+
+srcdir=$1
+outdir=$2
+datadir=$3
+
+mkdir -p $datadir/augmentations
+echo "copying $srcdir to $datadir/augmentations/aug1, allowed length, creating feats.scp"
+
+for set in aug1; do
+  image/copy_data_dir.sh --spk-prefix $set- --utt-prefix $set- \
+    $srcdir $datadir/augmentations/$set
+  cat $srcdir/allowed_lengths.txt > $datadir/augmentations/$set/allowed_lengths.txt
+  local/extract_features.sh --nj $nj --cmd "$cmd" --feat-dim $feat_dim \
+    --vertical-shift $verticle_shift \
+    --fliplr $fliplr --augment 'random_scale' $datadir/augmentations/$set
+done
+
+echo " combine original data and data from different augmentations"
+utils/combine_data.sh --extra-files images.scp $outdir $srcdir $datadir/augmentations/aug1
+cat $srcdir/allowed_lengths.txt > $outdir/allowed_lengths.txt

egs/casia_hwdb/v1/local/chain/compare_wer.sh

@@ -0,0 +1,67 @@
+#!/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
+
+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 "# CER                        "
+for x in $*; do
+  cer=$(cat $x/decode_test/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

egs/casia_hwdb/v1/local/chain/run_cnn_e2eali_1b.sh

@@ -0,0 +1,231 @@
+#!/bin/bash
+
+# e2eali_1b is the same as chainali_1a but uses the e2e chain model to get the
+# lattice alignments and to build a tree
+
+# ./local/chain/compare_wer.sh exp_yomdle_chinese/chain/e2e_cnn_1a exp_yomdle_chinese/chain/cnn_e2eali_1b
+# System                      e2e_cnn_1a cnn_e2eali_1b
+# CER                             15.44     13.57
+# Final train prob               0.0616   -0.0512
+# Final valid prob               0.0390   -0.0718
+# Final train prob (xent)                 -0.6199
+# Final valid prob (xent)                 -0.7448
+
+set -e -o pipefail
+
+data_dir=data
+exp_dir=exp
+
+stage=0
+
+nj=30
+train_set=train
+nnet3_affix=    # affix for exp dirs, e.g. it was _cleaned in tedlium.
+affix=_1b  #affix for TDNN+LSTM directory e.g. "1a" or "1b", in case we change the configuration.
+common_egs_dir=
+reporting_email=
+
+# chain options
+train_stage=-10
+xent_regularize=0.1
+frame_subsampling_factor=4
+# training chunk-options
+chunk_width=340,300,200,100
+num_leaves=2000
+# we don't need extra left/right context for TDNN systems.
+chunk_left_context=0
+chunk_right_context=0
+tdnn_dim=450
+# training options
+srand=0
+remove_egs=true
+lang_test=lang_test
+# End configuration section.
+echo "$0 $@"  # Print the command line for logging
+
+
+. ./cmd.sh
+. ./path.sh
+. ./utils/parse_options.sh
+
+
+if ! cuda-compiled; then
+  cat <<EOF && exit 1
+This script is intended to be used with GPUs but you have not compiled Kaldi with CUDA
+If you want to use GPUs (and have them), go to src/, and configure and make on a machine
+where "nvcc" is installed.
+EOF
+fi
+
+e2echain_model_dir=$exp_dir/chain/e2e_cnn_1a
+ali_dir=$exp_dir/chain/e2e_ali_train
+lat_dir=$exp_dir/chain${nnet3_affix}/e2e_${train_set}_lats
+dir=$exp_dir/chain${nnet3_affix}/cnn_e2eali${affix}
+train_data_dir=$data_dir/${train_set}
+tree_dir=$exp_dir/chain${nnet3_affix}/tree_e2e
+
+# the 'lang' directory is created by this script.
+# If you create such a directory with a non-standard topology
+# you should probably name it differently.
+lang=$data_dir/lang_chain
+for f in $train_data_dir/feats.scp $ali_dir/ali.1.gz $ali_dir/final.mdl; do
+  [ ! -f $f ] && echo "$0: expected file $f to exist" && exit 1
+done
+
+
+if [ $stage -le 1 ]; then
+  echo "$0: creating lang directory $lang with chain-type topology"
+  cp -r $data_dir/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 2 ]; then
+  # Get the alignments as lattices (gives the chain training more freedom).
+  # use the same num-jobs as the alignments
+  steps/nnet3/align_lats.sh --nj $nj --cmd "$cmd" \
+                            --acoustic-scale 1.0 \
+                            --scale-opts '--transition-scale=1.0 --self-loop-scale=1.0' \
+                            ${train_data_dir} $data_dir/lang $e2echain_model_dir $lat_dir
+  echo "" >$lat_dir/splice_opts
+
+fi
+
+if [ $stage -le 3 ]; then
+  # Build a tree using our new topology.  We know we have alignments for the
+  # speed-perturbed data (local/nnet3/run_ivector_common.sh made them), so use
+  # those.  The num-leaves is always somewhat less than the num-leaves from
+  # the GMM baseline.
+  if [ -f $tree_dir/final.mdl ]; then
+    echo "$0: $tree_dir/final.mdl already exists, refusing to overwrite it."
+  fi
+
+  steps/nnet3/chain/build_tree.sh \
+    --frame-subsampling-factor $frame_subsampling_factor \
+    --alignment-subsampling-factor 1 \
+    --context-opts "--context-width=2 --central-position=1" \
+    --cmd "$cmd" $num_leaves ${train_data_dir} \
+    $lang $ali_dir $tree_dir
+fi
+
+
+if [ $stage -le 4 ]; then
+  mkdir -p $dir
+  echo "$0: creating neural net configs using the xconfig parser";
+  num_targets=$(tree-info $tree_dir/tree | grep num-pdfs | awk '{print $2}')
+  learning_rate_factor=$(echo "print 0.5/$xent_regularize" | python)
+  cnn_opts="l2-regularize=0.075"
+  tdnn_opts="l2-regularize=0.075"
+  output_opts="l2-regularize=0.1"
+  common1="$cnn_opts required-time-offsets= height-offsets=-2,-1,0,1,2 num-filters-out=32"
+  common2="$cnn_opts required-time-offsets= height-offsets=-2,-1,0,1,2 num-filters-out=128"
+  common3="$cnn_opts required-time-offsets= height-offsets=-1,0,1 num-filters-out=512"
+  mkdir -p $dir/configs
+  cat <<EOF > $dir/configs/network.xconfig
+  input dim=180 name=input
+
+  conv-relu-batchnorm-layer name=cnn1 height-in=60 height-out=60 time-offsets=-3,-2,-1,0,1,2,3 $common1
+  conv-relu-batchnorm-layer name=cnn2 height-in=60 height-out=60 time-offsets=-3,-2,-1,0,1,2,3 $common1
+  conv-relu-batchnorm-layer name=cnn3 height-in=60 height-out=30 time-offsets=-2,-1,0,1,2 $common1 height-subsample-out=2
+  conv-relu-batchnorm-layer name=cnn4 height-in=30 height-out=30 time-offsets=-4,-2,0,2,4 $common2
+  conv-relu-batchnorm-layer name=cnn5 height-in=30 height-out=30 time-offsets=-4,-2,0,2,4 $common2
+  conv-relu-batchnorm-layer name=cnn6 height-in=30 height-out=15 time-offsets=-4,-2,0,2,4 $common2 height-subsample-out=2
+  conv-relu-batchnorm-layer name=cnn7 height-in=15 height-out=15 time-offsets=-4,0,4 $common3
+  conv-relu-batchnorm-layer name=cnn8 height-in=15 height-out=15 time-offsets=-4,0,4 $common3
+  conv-relu-batchnorm-layer name=cnn9 height-in=15 height-out=15 time-offsets=-4,0,4 $common3
+  relu-batchnorm-layer name=tdnn1 input=Append(-8,-4,0,4,8) dim=$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 $tdnn_opts
+  output-layer name=output include-log-softmax=false dim=$num_targets max-change=1.5 $output_opts
+
+  # adding the layers for xent branch
+  # This block prints the configs for a separate output that will be
+  # trained with a cross-entropy objective in the 'chain' mod?els... this
+  # has the effect of regularizing the hidden parts of the model.  we use
+  # 0.5 / args.xent_regularize as the learning rate factor- the factor of
+  # 0.5 / args.xent_regularize is suitable as it means the xent
+  # final-layer learns at a rate independent of the regularization
+  # constant; and the 0.5 was tuned so as to make the relative progress
+  # similar in the xent and regular final layers.
+  relu-batchnorm-layer name=prefinal-xent input=tdnn3 dim=$tdnn_dim target-rms=0.5 $tdnn_opts
+  output-layer name=output-xent dim=$num_targets learning-rate-factor=$learning_rate_factor 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 5 ]; then
+  if [[ $(hostname -f) == *.clsp.jhu.edu ]] && [ ! -d $dir/egs/storage ]; then
+    utils/create_split_dir.pl \
+     /export/b0{3,4,5,6}/$USER/kaldi-data/egs/iam-$(date +'%m_%d_%H_%M')/s5/$dir/egs/storage $dir/egs/storage
+  fi
+
+  steps/nnet3/chain/train.py --stage=$train_stage \
+    --cmd="$cmd" \
+    --feat.cmvn-opts="--norm-means=false --norm-vars=false" \
+    --chain.xent-regularize $xent_regularize \
+    --chain.leaky-hmm-coefficient=0.1 \
+    --chain.l2-regularize=0.00005 \
+    --chain.apply-deriv-weights=false \
+    --chain.lm-opts="--ngram-order=2 --no-prune-ngram-order=1 --num-extra-lm-states=1500" \
+    --chain.frame-subsampling-factor=$frame_subsampling_factor \
+    --chain.alignment-subsampling-factor=1 \
+    --chain.left-tolerance 3 \
+    --chain.right-tolerance 3 \
+    --trainer.srand=$srand \
+    --trainer.max-param-change=2.0 \
+    --trainer.num-epochs=6 \
+    --trainer.frames-per-iter=1000000 \
+    --trainer.optimization.num-jobs-initial=4 \
+    --trainer.optimization.num-jobs-final=8 \
+    --trainer.optimization.initial-effective-lrate=0.001 \
+    --trainer.optimization.final-effective-lrate=0.0001 \
+    --trainer.optimization.shrink-value=1.0 \
+    --trainer.num-chunk-per-minibatch=16,8 \
+    --trainer.optimization.momentum=0.0 \
+    --egs.chunk-width=$chunk_width \
+    --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="$common_egs_dir" \
+    --egs.opts="--frames-overlap-per-eg 0 --constrained false" \
+    --cleanup.remove-egs=$remove_egs \
+    --use-gpu=wait \
+    --reporting.email="$reporting_email" \
+    --feat-dir=$train_data_dir \
+    --tree-dir=$tree_dir \
+    --lat-dir=$lat_dir \
+    --dir=$dir  || exit 1;
+fi
+
+if [ $stage -le 6 ]; 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 $data_dir/$lang_test \
+    $dir $dir/graph || exit 1;
+fi
+
+if [ $stage -le 7 ]; then
+  frames_per_chunk=$(echo $chunk_width | cut -d, -f1)
+  steps/nnet3/decode.sh --acwt 1.0 --post-decode-acwt 10.0 \
+    --extra-left-context $chunk_left_context \
+    --extra-right-context $chunk_right_context \
+    --extra-left-context-initial 0 \
+    --extra-right-context-final 0 \
+    --frames-per-chunk $frames_per_chunk \
+    --nj $nj --cmd "$cmd" \
+    $dir/graph $data_dir/test $dir/decode_test || exit 1;
+fi