run_tdnn_a.sh 3.76 KB
#!/bin/bash

# Copyright 2012  Johns Hopkins University (author: Daniel Povey)  Tony Robinson
#           2015  Guoguo Chen

# This script trains LMs on the WSJ LM-training data.
# It requires that you have already run wsj_extend_dict.sh,
# to get the larger-size dictionary including all of CMUdict
# plus any OOVs and possible acronyms that we could easily
# derive pronunciations for.

# rnnlm/train_rnnlm.sh: best iteration (out of 40) was 39, linking it to final iteration.
# Train objf: -1225.00 -5.50 -5.29 -5.17 -5.09 -5.02 -4.97 -4.92 -4.88 -4.84 -4.79 -4.78 -4.76 -4.74 -4.73 -4.72 -4.71 -4.70 -4.68 -4.67 -4.66 -4.65 -4.64 -4.63 -4.62 -4.61 -4.60 -4.59 -4.58 -4.58 -4.57 -4.56 -4.56 -4.56 -4.56 -4.55 -4.55 -4.55 -4.54 -4.50
# Dev objf:   -11.89 -5.69 -5.40 -5.26 -5.19 -5.14 -5.10 -5.08 -5.06 -5.04 -4.95 -5.03 -4.93 -4.91 -4.90 -4.89 -4.88 -4.88 -4.87 -4.86 -4.85 -4.85 -4.84 -4.84 -4.84 -4.83 -4.83 -4.83 -4.83 -4.83 -4.83 -4.83 -4.79 -4.78 -4.78 -4.77 -4.76 -4.76 -4.76 -4.76 -4.75


# This script takes no command-line arguments but takes the --cmd option.

# Begin configuration section.
cmd=run.pl
dir=exp/rnnlm_tdnn_a
embedding_dim=600
stage=0
train_stage=0

. utils/parse_options.sh


text=data/local/dict_nosp_larger/cleaned.gz
lexicon=data/local/dict_nosp_larger/lexiconp.txt
text_dir=data/rnnlm/text_nosp
mkdir -p $dir/config
set -e

for f in $text $lexicon; do
  [ ! -f $f ] && \
    echo "$0: expected file $f to exist; search for local/wsj_extend_dict.sh in run.sh" && exit 1
done

if [ $stage -le 0 ]; then
  mkdir -p $text_dir
  echo -n >$text_dir/dev.txt
  # hold out one in every 500 lines as dev data.
  gunzip -c data/local/dict_nosp_larger/cleaned.gz  | awk -v text_dir=$text_dir '{if(NR%500 == 0) { print >text_dir"/dev.txt"; } else {print;}}' >$text_dir/wsj.txt
fi

if [ $stage -le 1 ]; then
  # the training scripts require that <s>, </s> and <brk> be present in a particular
  # order.
  awk '{print $1}' $lexicon | sort | uniq | \
    awk 'BEGIN{print "<eps> 0";print "<s> 1"; print "</s> 2"; print "<brk> 3";n=4;} {print $1, n++}' \
        >$dir/config/words.txt
  # words that are not present in words.txt but are in the training or dev data, will be
  # mapped to <SPOKEN_NOISE> during training.
  echo "<SPOKEN_NOISE>" >$dir/config/oov.txt

  cat > $dir/config/data_weights.txt <<EOF
wsj   1   1.0
EOF

  rnnlm/get_unigram_probs.py --vocab-file=$dir/config/words.txt \
                             --unk-word="<SPOKEN_NOISE>" \
                             --data-weights-file=$dir/config/data_weights.txt \
                             $text_dir >$dir/config/unigram_probs.txt

  # choose features
  rnnlm/choose_features.py --unigram-probs=$dir/config/unigram_probs.txt \
                           --use-constant-feature=true \
                           --special-words='<s>,</s>,<brk>,<SPOKEN_NOISE>' \
                           $dir/config/words.txt > $dir/config/features.txt

  cat >$dir/config/xconfig <<EOF
input dim=$embedding_dim name=input
relu-renorm-layer name=tdnn1 dim=600 input=Append(0, IfDefined(-1))
relu-renorm-layer name=tdnn2 dim=600 input=Append(0, IfDefined(-1))
relu-renorm-layer name=tdnn3 dim=600 input=Append(0, IfDefined(-2))
output-layer name=output include-log-softmax=false dim=$embedding_dim
EOF
  rnnlm/validate_config_dir.sh $text_dir $dir/config
fi

if [ $stage -le 2 ]; then
  # the --unigram-factor option is set larger than the default (100)
  # in order to reduce the size of the sampling LM, because rnnlm-get-egs
  # was taking up too much CPU (as much as 10 cores).
  rnnlm/prepare_rnnlm_dir.sh --unigram-factor 200.0 \
                             $text_dir $dir/config $dir
fi

if [ $stage -le 3 ]; then
  rnnlm/train_rnnlm.sh --num-jobs-initial 1 --num-jobs-final 3 \
                  --stage $train_stage --num-epochs 10 --cmd "queue.pl" $dir
fi

exit 0