egs/ptb/LM/prepare.sh

#!/usr/bin/env bash

# fix segmentation fault reported in https://github.com/k2-fsa/icefall/issues/674
export PROTOCOL_BUFFERS_PYTHON_IMPLEMENTATION=python

set -eou pipefail

nj=15
stage=-1
stop_stage=100

dl_dir=$PWD/download
# The following files will be downloaded to $dl_dir
#  - ptb.train.txt
#  - ptb.valid.txt
#  - ptb.test.txt

. shared/parse_options.sh || exit 1

# vocab size for sentence piece models.
# It will generate data/bpe_xxx, data/bpe_yyy
# if the array contains xxx, yyy
vocab_sizes=(
  500
  # 1000
  # 2000
  # 5000
)

# All files generated by this script are saved in "data".
# You can safely remove "data" and rerun this script to regenerate it.
mkdir -p data
mkdir -p $dl_dir

log() {
  # This function is from espnet
  local fname=${BASH_SOURCE[1]##*/}
  echo -e "$(date '+%Y-%m-%d %H:%M:%S') (${fname}:${BASH_LINENO[0]}:${FUNCNAME[1]}) $*"
}

log "dl_dir: $dl_dir"

if [ $stage -le -1 ] && [ $stop_stage -ge -1 ]; then
  log "Stage -1: Download data"

  # Caution: The downloaded data has already been normalized for LM training.

  if [ ! -f $dl_dir/.complete ]; then
    url=http://raw.githubusercontent.com/townie/PTB-dataset-from-Tomas-Mikolov-s-webpage/master/data
    wget --directory-prefix $dl_dir $url/ptb.train.txt
    wget --directory-prefix $dl_dir $url/ptb.valid.txt
    wget --directory-prefix $dl_dir $url/ptb.test.txt
    touch $dl_dir/.complete
  fi
fi

if [ $stage -le 0 ] && [ $stop_stage -ge 0 ]; then
  log "Stage 0: Train BPE model"

  # Caution: You have to use the same bpe model for training your acoustic model
  # Caution: You have to use the same bpe model for training your acoustic model
  # Caution: You have to use the same bpe model for training your acoustic model

  for vocab_size in ${vocab_sizes[@]}; do
    lang_dir=data/lang_bpe_${vocab_size}
    mkdir -p $lang_dir
    ./local/train_bpe_model.py \
      --lang-dir $lang_dir \
      --vocab-size $vocab_size \
      --transcript $dl_dir/ptb.train.txt
  done
fi

if [ $stage -le 1 ] && [ $stop_stage -ge 1 ]; then
  log "Stage 1: Generate LM training data"
  # Note: ptb.train.txt has already been normalized

  for vocab_size in ${vocab_sizes[@]}; do
    lang_dir=data/lang_bpe_${vocab_size}
    out_dir=data/lm_training_bpe_${vocab_size}
    mkdir -p $out_dir
    ./local/prepare_lm_training_data.py \
      --bpe-model $lang_dir/bpe.model \
      --lm-data $dl_dir/ptb.train.txt \
      --lm-archive $out_dir/lm_data.pt

    ./local/prepare_lm_training_data.py \
      --bpe-model $lang_dir/bpe.model \
      --lm-data $dl_dir/ptb.valid.txt \
      --lm-archive $out_dir/lm_data-valid.pt

    ./local/prepare_lm_training_data.py \
      --bpe-model $lang_dir/bpe.model \
      --lm-data $dl_dir/ptb.test.txt \
      --lm-archive $out_dir/lm_data-test.pt
  done
fi

if [ $stage -le 2 ] && [ $stop_stage -ge 2 ]; then
  log "Stage 2: Sort LM training data"
  # Sort LM training data generated in stage 1
  # by sentence length in descending order
  # for ease of training.
  #
  # Sentence length equals to the number of BPE tokens
  # in a sentence.

  for vocab_size in ${vocab_sizes[@]}; do
    out_dir=data/lm_training_bpe_${vocab_size}
    mkdir -p $out_dir
    ./local/sort_lm_training_data.py \
      --in-lm-data $out_dir/lm_data.pt \
      --out-lm-data $out_dir/sorted_lm_data.pt \
      --out-statistics $out_dir/statistics.txt

    ./local/sort_lm_training_data.py \
      --in-lm-data $out_dir/lm_data-valid.pt \
      --out-lm-data $out_dir/sorted_lm_data-valid.pt \
      --out-statistics $out_dir/statistics-valid.txt

    ./local/sort_lm_training_data.py \
      --in-lm-data $out_dir/lm_data-test.pt \
      --out-lm-data $out_dir/sorted_lm_data-test.pt \
      --out-statistics $out_dir/statistics-test.txt
  done
fi