tesseract/training/tesstrain_utils.sh

570 lines
20 KiB
Bash
Executable File

#!/bin/bash
# (C) Copyright 2014, Google Inc.
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
# http://www.apache.org/licenses/LICENSE-2.0
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
#
# This script defines functions that are used by tesstrain.sh
# For a detailed description of the phases, see
# https://github.com/tesseract-ocr/tesseract/wiki/TrainingTesseract
#
# USAGE: source tesstrain_utils.sh
if [ "$(uname)" == "Darwin" ];then
FONTS_DIR="/Library/Fonts/"
else
FONTS_DIR="/usr/share/fonts/"
fi
OUTPUT_DIR="/tmp/tesstrain/tessdata"
OVERWRITE=0
LINEDATA=0
RUN_SHAPE_CLUSTERING=0
EXTRACT_FONT_PROPERTIES=1
WORKSPACE_DIR=$(mktemp -d)
# Logging helper functions.
tlog() {
echo -e $* 2>&1 1>&2 | tee -a ${LOG_FILE}
}
err_exit() {
echo -e "ERROR: "$* 2>&1 1>&2 | tee -a ${LOG_FILE}
exit 1
}
# Helper function to run a command and append its output to a log. Aborts early
# if the program file is not found.
# Usage: run_command CMD ARG1 ARG2...
run_command() {
local cmd=$(which $1)
if [[ -z ${cmd} ]]; then
for d in api training; do
cmd=$(which $d/$1)
if [[ ! -z ${cmd} ]]; then
break
fi
done
if [[ -z ${cmd} ]]; then
err_exit "$1 not found"
fi
fi
shift
tlog "[$(date)] ${cmd} $@"
"${cmd}" "$@" 2>&1 1>&2 | tee -a ${LOG_FILE}
# check completion status
if [[ $? -gt 0 ]]; then
err_exit "Program $(basename ${cmd}) failed. Abort."
fi
}
# Check if all the given files exist, or exit otherwise.
# Used to check required input files and produced output files in each phase.
# Usage: check_file_readable FILE1 FILE2...
check_file_readable() {
for file in $@; do
if [[ ! -r ${file} ]]; then
err_exit "${file} does not exist or is not readable"
fi
done
}
# Sets the named variable to given value. Aborts if the value is missing or
# if it looks like a flag.
# Usage: parse_value VAR_NAME VALUE
parse_value() {
local val="$2"
if [[ -z $val ]]; then
err_exit "Missing value for variable $1"
exit
fi
if [[ ${val:0:2} == "--" ]]; then
err_exit "Invalid value $val passed for variable $1"
exit
fi
eval $1=\"$val\"
}
# Does simple command-line parsing and initialization.
parse_flags() {
local i=0
while test $i -lt ${#ARGV[@]}; do
local j=$((i+1))
case ${ARGV[$i]} in
--)
break;;
--fontlist)
fn=0
FONTS=""
while test $j -lt ${#ARGV[@]}; do
test -z "${ARGV[$j]}" && break
test $(echo ${ARGV[$j]} | cut -c -2) = "--" && break
FONTS[$fn]="${ARGV[$j]}"
fn=$((fn+1))
j=$((j+1))
done
i=$((j-1)) ;;
--exposures)
exp=""
while test $j -lt ${#ARGV[@]}; do
test -z "${ARGV[$j]}" && break
test $(echo ${ARGV[$j]} | cut -c -2) = "--" && break
exp="$exp ${ARGV[$j]}"
j=$((j+1))
done
parse_value "EXPOSURES" "$exp"
i=$((j-1)) ;;
--fonts_dir)
parse_value "FONTS_DIR" ${ARGV[$j]}
i=$j ;;
--lang)
parse_value "LANG_CODE" ${ARGV[$j]}
i=$j ;;
--langdata_dir)
parse_value "LANGDATA_ROOT" ${ARGV[$j]}
i=$j ;;
--output_dir)
parse_value "OUTPUT_DIR" ${ARGV[$j]}
i=$j ;;
--overwrite)
OVERWRITE=1 ;;
--linedata_only)
LINEDATA=1 ;;
--extract_font_properties)
EXTRACT_FONT_PROPERTIES=1 ;;
--noextract_font_properties)
EXTRACT_FONT_PROPERTIES=0 ;;
--tessdata_dir)
parse_value "TESSDATA_DIR" ${ARGV[$j]}
i=$j ;;
--training_text)
parse_value "TRAINING_TEXT" "${ARGV[$j]}"
i=$j ;;
--wordlist)
parse_value "WORDLIST_FILE" ${ARGV[$j]}
i=$j ;;
*)
err_exit "Unrecognized argument ${ARGV[$i]}" ;;
esac
i=$((i+1))
done
if [[ -z ${LANG_CODE} ]]; then
err_exit "Need to specify a language --lang"
fi
if [[ -z ${LANGDATA_ROOT} ]]; then
err_exit "Need to specify path to language files --langdata_dir"
fi
if [[ -z ${TESSDATA_DIR} ]]; then
if [[ -z ${TESSDATA_PREFIX} ]]; then
err_exit "Need to specify a --tessdata_dir or have a "\
"TESSDATA_PREFIX variable defined in your environment"
else
TESSDATA_DIR="${TESSDATA_PREFIX}"
fi
fi
# Location where intermediate files will be created.
TRAINING_DIR=${WORKSPACE_DIR}/${LANG_CODE}
# Location of log file for the whole run.
LOG_FILE=${TRAINING_DIR}/tesstrain.log
# Take training text and wordlist from the langdata directory if not
# specified in the command-line.
if [[ -z ${TRAINING_TEXT} ]]; then
TRAINING_TEXT=${LANGDATA_ROOT}/${LANG_CODE}/${LANG_CODE}.training_text
fi
if [[ -z ${WORDLIST_FILE} ]]; then
WORDLIST_FILE=${LANGDATA_ROOT}/${LANG_CODE}/${LANG_CODE}.wordlist
fi
WORD_BIGRAMS_FILE=${LANGDATA_ROOT}/${LANG_CODE}/${LANG_CODE}.word.bigrams
NUMBERS_FILE=${LANGDATA_ROOT}/${LANG_CODE}/${LANG_CODE}.numbers
PUNC_FILE=${LANGDATA_ROOT}/${LANG_CODE}/${LANG_CODE}.punc
BIGRAM_FREQS_FILE=${TRAINING_TEXT}.bigram_freqs
UNIGRAM_FREQS_FILE=${TRAINING_TEXT}.unigram_freqs
TRAIN_NGRAMS_FILE=${TRAINING_TEXT}.train_ngrams
GENERATE_DAWGS=1
}
# Function initializes font config with a unique font cache dir.
initialize_fontconfig() {
export FONT_CONFIG_CACHE=$(mktemp -d --tmpdir font_tmp.XXXXXXXXXX)
local sample_path=${FONT_CONFIG_CACHE}/sample_text.txt
echo "Text" >${sample_path}
run_command text2image --fonts_dir=${FONTS_DIR} \
--font="${FONTS[0]}" --outputbase=${sample_path} --text=${sample_path} \
--fontconfig_tmpdir=${FONT_CONFIG_CACHE}
}
# Helper function for phaseI_generate_image. Generates the image for a single
# language/font combination in a way that can be run in parallel.
generate_font_image() {
local font="$1"
tlog "Rendering using ${font}"
local fontname=$(echo ${font} | tr ' ' '_' | sed 's/,//g')
local outbase=${TRAINING_DIR}/${LANG_CODE}.${fontname}.exp${EXPOSURE}
local common_args="--fontconfig_tmpdir=${FONT_CONFIG_CACHE}"
common_args+=" --fonts_dir=${FONTS_DIR} --strip_unrenderable_words"
common_args+=" --leading=${LEADING}"
common_args+=" --char_spacing=${CHAR_SPACING} --exposure=${EXPOSURE}"
common_args+=" --outputbase=${outbase} --max_pages=3"
# add --writing_mode=vertical-upright to common_args if the font is
# specified to be rendered vertically.
for vfont in "${VERTICAL_FONTS[@]}"; do
if [[ "${font}" == "${vfont}" ]]; then
common_args+=" --writing_mode=vertical-upright "
break
fi
done
run_command text2image ${common_args} --font="${font}" \
--text=${TRAINING_TEXT} ${TEXT2IMAGE_EXTRA_ARGS}
check_file_readable ${outbase}.box ${outbase}.tif
if ((EXTRACT_FONT_PROPERTIES)) &&
[[ -r ${TRAIN_NGRAMS_FILE} ]]; then
tlog "Extracting font properties of ${font}"
run_command text2image ${common_args} --font="${font}" \
--ligatures=false --text=${TRAIN_NGRAMS_FILE} \
--only_extract_font_properties --ptsize=32
check_file_readable ${outbase}.fontinfo
fi
}
# Phase I : Generate (I)mages from training text for each font.
phase_I_generate_image() {
local par_factor=$1
if [[ -z ${par_factor} || ${par_factor} -le 0 ]]; then
par_factor=1
fi
tlog "\n=== Phase I: Generating training images ==="
if [[ -z ${TRAINING_TEXT} ]] || [[ ! -r ${TRAINING_TEXT} ]]; then
err_exit "Could not find training text file ${TRAINING_TEXT}"
fi
CHAR_SPACING="0.0"
for EXPOSURE in $EXPOSURES; do
if ((EXTRACT_FONT_PROPERTIES)) && [[ -r ${BIGRAM_FREQS_FILE} ]]; then
# Parse .bigram_freqs file and compose a .train_ngrams file with text
# for tesseract to recognize during training. Take only the ngrams whose
# combined weight accounts for 95% of all the bigrams in the language.
NGRAM_FRAC=$(cat ${BIGRAM_FREQS_FILE} \
| awk '{s=s+$2}; END {print (s/100)*p}' p=99)
cat ${BIGRAM_FREQS_FILE} | sort -rnk2 \
| awk '{s=s+$2; if (s <= x) {printf "%s ", $1; } }' \
x=${NGRAM_FRAC} > ${TRAIN_NGRAMS_FILE}
check_file_readable ${TRAIN_NGRAMS_FILE}
fi
local counter=0
for font in "${FONTS[@]}"; do
generate_font_image "${font}" &
let counter=counter+1
let rem=counter%par_factor
if [[ "${rem}" -eq 0 ]]; then
wait
fi
done
wait
# Check that each process was successful.
for font in "${FONTS[@]}"; do
local fontname=$(echo ${font} | tr ' ' '_' | sed 's/,//g')
local outbase=${TRAINING_DIR}/${LANG_CODE}.${fontname}.exp${EXPOSURE}
check_file_readable ${outbase}.box ${outbase}.tif
done
done
}
# Phase UP : Generate (U)nicharset and (P)roperties file.
phase_UP_generate_unicharset() {
tlog "\n=== Phase UP: Generating unicharset and unichar properties files ==="
local box_files=$(ls ${TRAINING_DIR}/*.box)
UNICHARSET_FILE="${TRAINING_DIR}/${LANG_CODE}.unicharset"
run_command unicharset_extractor --output_unicharset "${UNICHARSET_FILE}" \
--norm_mode "${NORM_MODE}" ${box_files}
check_file_readable ${UNICHARSET_FILE}
XHEIGHTS_FILE="${TRAINING_DIR}/${LANG_CODE}.xheights"
run_command set_unicharset_properties \
-U ${UNICHARSET_FILE} -O ${UNICHARSET_FILE} -X ${XHEIGHTS_FILE} \
--script_dir=${LANGDATA_ROOT}
check_file_readable ${XHEIGHTS_FILE}
}
# Phase D : Generate (D)awg files from unicharset file and wordlist files
phase_D_generate_dawg() {
tlog "\n=== Phase D: Generating Dawg files ==="
# Skip if requested
if [[ ${GENERATE_DAWGS} -eq 0 ]]; then
tlog "Skipping ${phase_name}"
return
fi
# Output files
WORD_DAWG=${TRAINING_DIR}/${LANG_CODE}.word-dawg
FREQ_DAWG=${TRAINING_DIR}/${LANG_CODE}.freq-dawg
PUNC_DAWG=${TRAINING_DIR}/${LANG_CODE}.punc-dawg
NUMBER_DAWG=${TRAINING_DIR}/${LANG_CODE}.number-dawg
BIGRAM_DAWG=${TRAINING_DIR}/${LANG_CODE}.bigram-dawg
# Word DAWG
local freq_wordlist_file=${TRAINING_DIR}/${LANG_CODE}.wordlist.clean.freq
if [[ -s ${WORDLIST_FILE} ]]; then
tlog "Generating word Dawg"
check_file_readable ${UNICHARSET_FILE}
run_command wordlist2dawg -r 1 ${WORDLIST_FILE} ${WORD_DAWG} \
${UNICHARSET_FILE}
check_file_readable ${WORD_DAWG}
FREQ_DAWG_SIZE=100
head -n ${FREQ_DAWG_SIZE} ${WORDLIST_FILE} > ${freq_wordlist_file}
fi
# Freq-word DAWG
if [[ -s ${freq_wordlist_file} ]]; then
check_file_readable ${UNICHARSET_FILE}
tlog "Generating frequent-word Dawg"
run_command wordlist2dawg -r 1 ${freq_wordlist_file} \
${FREQ_DAWG} ${UNICHARSET_FILE}
check_file_readable ${FREQ_DAWG}
fi
# Punctuation DAWG
# -r arguments to wordlist2dawg denote RTL reverse policy
# (see Trie::RTLReversePolicy enum in third_party/tesseract/dict/trie.h).
# We specify 0/RRP_DO_NO_REVERSE when generating number DAWG,
# 1/RRP_REVERSE_IF_HAS_RTL for freq and word DAWGS,
# 2/RRP_FORCE_REVERSE for the punctuation DAWG.
local punc_reverse_policy=0;
if [[ "${LANG_IS_RTL}" == "1" ]]; then
punc_reverse_policy=2
fi
if [[ ! -s ${PUNC_FILE} ]]; then
PUNC_FILE="${LANGDATA_ROOT}/common.punc"
fi
check_file_readable ${PUNC_FILE}
run_command wordlist2dawg -r ${punc_reverse_policy} \
${PUNC_FILE} ${PUNC_DAWG} ${UNICHARSET_FILE}
check_file_readable ${PUNC_DAWG}
# Numbers DAWG
if [[ -s ${NUMBERS_FILE} ]]; then
run_command wordlist2dawg -r 0 \
${NUMBERS_FILE} ${NUMBER_DAWG} ${UNICHARSET_FILE}
check_file_readable ${NUMBER_DAWG}
fi
# Bigram dawg
if [[ -s ${WORD_BIGRAMS_FILE} ]]; then
run_command wordlist2dawg -r 1 \
${WORD_BIGRAMS_FILE} ${BIGRAM_DAWG} ${UNICHARSET_FILE}
check_file_readable ${BIGRAM_DAWG}
fi
}
# Phase E : (E)xtract .tr feature files from .tif/.box files
phase_E_extract_features() {
local box_config=$1
local par_factor=$2
local ext=$3
if [[ -z ${par_factor} || ${par_factor} -le 0 ]]; then
par_factor=1
fi
tlog "\n=== Phase E: Generating ${ext} files ==="
local img_files=""
for exposure in ${EXPOSURES}; do
img_files=${img_files}' '$(ls ${TRAINING_DIR}/*.exp${exposure}.tif)
done
# Use any available language-specific configs.
local config=""
if [[ -r ${LANGDATA_ROOT}/${LANG_CODE}/${LANG_CODE}.config ]]; then
config=${LANGDATA_ROOT}/${LANG_CODE}/${LANG_CODE}.config
fi
OLD_TESSDATA_PREFIX=${TESSDATA_PREFIX}
export TESSDATA_PREFIX=${TESSDATA_DIR}
tlog "Using TESSDATA_PREFIX=${TESSDATA_PREFIX}"
local counter=0
for img_file in ${img_files}; do
run_command tesseract ${img_file} ${img_file%.*} \
${box_config} ${config} &
let counter=counter+1
let rem=counter%par_factor
if [[ "${rem}" -eq 0 ]]; then
wait
fi
done
wait
export TESSDATA_PREFIX=${OLD_TESSDATA_PREFIX}
# Check that all the output files were produced.
for img_file in ${img_files}; do
check_file_readable "${img_file%.*}.${ext}"
done
}
# Phase C : (C)luster feature prototypes in .tr into normproto file (cnTraining)
# phaseC_cluster_prototypes ${TRAINING_DIR}/${LANG_CODE}.normproto
phase_C_cluster_prototypes() {
tlog "\n=== Phase C: Clustering feature prototypes (cnTraining) ==="
local out_normproto=$1
run_command cntraining -D "${TRAINING_DIR}/" \
$(ls ${TRAINING_DIR}/*.tr)
check_file_readable ${TRAINING_DIR}/normproto
mv ${TRAINING_DIR}/normproto ${out_normproto}
}
# Phase S : (S)hape clustering
phase_S_cluster_shapes() {
if ((! RUN_SHAPE_CLUSTERING)); then
tlog "\n=== Shape Clustering disabled ==="
return
fi
check_file_readable ${LANGDATA_ROOT}/font_properties
local font_props="-F ${LANGDATA_ROOT}/font_properties"
if [[ -r ${TRAINING_DIR}/${LANG_CODE}.xheights ]] &&\
[[ -s ${TRAINING_DIR}/${LANG_CODE}.xheights ]]; then
font_props=${font_props}" -X ${TRAINING_DIR}/${LANG_CODE}.xheights"
fi
run_command shapeclustering \
-D "${TRAINING_DIR}/" \
-U ${TRAINING_DIR}/${LANG_CODE}.unicharset \
-O ${TRAINING_DIR}/${LANG_CODE}.mfunicharset \
${font_props} \
$(ls ${TRAINING_DIR}/*.tr)
check_file_readable ${TRAINING_DIR}/shapetable \
${TRAINING_DIR}/${LANG_CODE}.mfunicharset
}
# Phase M : Clustering microfeatures (mfTraining)
phase_M_cluster_microfeatures() {
tlog "\n=== Phase M : Clustering microfeatures (mfTraining) ==="
check_file_readable ${LANGDATA_ROOT}/font_properties
font_props="-F ${LANGDATA_ROOT}/font_properties"
if [[ -r ${TRAINING_DIR}/${LANG_CODE}.xheights ]] && \
[[ -s ${TRAINING_DIR}/${LANG_CODE}.xheights ]]; then
font_props=${font_props}" -X ${TRAINING_DIR}/${LANG_CODE}.xheights"
fi
run_command mftraining \
-D "${TRAINING_DIR}/" \
-U ${TRAINING_DIR}/${LANG_CODE}.unicharset \
-O ${TRAINING_DIR}/${LANG_CODE}.mfunicharset \
${font_props} \
$(ls ${TRAINING_DIR}/*.tr)
check_file_readable ${TRAINING_DIR}/inttemp ${TRAINING_DIR}/shapetable \
${TRAINING_DIR}/pffmtable ${TRAINING_DIR}/${LANG_CODE}.mfunicharset
mv ${TRAINING_DIR}/inttemp ${TRAINING_DIR}/${LANG_CODE}.inttemp
mv ${TRAINING_DIR}/shapetable ${TRAINING_DIR}/${LANG_CODE}.shapetable
mv ${TRAINING_DIR}/pffmtable ${TRAINING_DIR}/${LANG_CODE}.pffmtable
mv ${TRAINING_DIR}/${LANG_CODE}.mfunicharset ${TRAINING_DIR}/${LANG_CODE}.unicharset
}
phase_B_generate_ambiguities() {
tlog "\n=== Phase B : ambiguities training ==="
# Check for manually created ambiguities data.
if [[ -r ${LANGDATA_ROOT}/${LANG_CODE}/${LANG_CODE}.unicharambigs ]]; then
tlog "Found file ${LANGDATA_ROOT}/${LANG_CODE}/${LANG_CODE}.unicharambigs"
cp ${LANGDATA_ROOT}/${LANG_CODE}/${LANG_CODE}.unicharambigs \
${TRAINING_DIR}/${LANG_CODE}.unicharambigs
# Make it writable, as it may be read-only in the client.
chmod u+w ${TRAINING_DIR}/${LANG_CODE}.unicharambigs
return
else
tlog "No unicharambigs file found!"
fi
# TODO: Add support for generating ambiguities automatically.
}
make__lstmdata() {
tlog "\n=== Constructing LSTM training data ==="
local lang_prefix="${LANGDATA_ROOT}/${LANG_CODE}/${LANG_CODE}"
if [[ ! -d "${OUTPUT_DIR}" ]]; then
tlog "Creating new directory ${OUTPUT_DIR}"
mkdir -p "${OUTPUT_DIR}"
fi
local lang_is_rtl=""
if [[ "${LANG_IS_RTL}" == "1" ]]; then
lang_is_rtl="--lang_is_rtl"
fi
local pass_through=""
if [[ "${NORM_MODE}" -ge "2" ]]; then
pass_through="--pass_through_recoder"
fi
# Build the starter traineddata from the inputs.
run_command combine_lang_model \
--input_unicharset "${TRAINING_DIR}/${LANG_CODE}.unicharset" \
--script_dir "${LANGDATA_ROOT}" \
--words "${lang_prefix}.wordlist" \
--numbers "${lang_prefix}.numbers" \
--puncs "${lang_prefix}.punc" \
--output_dir "${OUTPUT_DIR}" --lang "${LANG_CODE}" \
"${pass_through}" "${lang_is_rtl}"
for f in "${TRAINING_DIR}/${LANG_CODE}".*.lstmf; do
tlog "Moving ${f} to ${OUTPUT_DIR}"
mv "${f}" "${OUTPUT_DIR}"
done
local lstm_list="${OUTPUT_DIR}/${LANG_CODE}.training_files.txt"
ls -1 "${OUTPUT_DIR}/${LANG_CODE}".*.lstmf > "${lstm_list}"
}
make__traineddata() {
tlog "\n=== Making final traineddata file ==="
local lang_prefix=${LANGDATA_ROOT}/${LANG_CODE}/${LANG_CODE}
# Combine available files for this language from the langdata dir.
if [[ -r ${lang_prefix}.config ]]; then
tlog "Copying ${lang_prefix}.config to ${TRAINING_DIR}"
cp ${lang_prefix}.config ${TRAINING_DIR}
chmod u+w ${TRAINING_DIR}/${LANG_CODE}.config
fi
if [[ -r ${lang_prefix}.cube-unicharset ]]; then
tlog "Copying ${lang_prefix}.cube-unicharset to ${TRAINING_DIR}"
cp ${lang_prefix}.cube-unicharset ${TRAINING_DIR}
chmod u+w ${TRAINING_DIR}/${LANG_CODE}.cube-unicharset
fi
if [[ -r ${lang_prefix}.cube-word-dawg ]]; then
tlog "Copying ${lang_prefix}.cube-word-dawg to ${TRAINING_DIR}"
cp ${lang_prefix}.cube-word-dawg ${TRAINING_DIR}
chmod u+w ${TRAINING_DIR}/${LANG_CODE}.cube-word-dawg
fi
if [[ -r ${lang_prefix}.params-model ]]; then
tlog "Copying ${lang_prefix}.params-model to ${TRAINING_DIR}"
cp ${lang_prefix}.params-model ${TRAINING_DIR}
chmod u+w ${TRAINING_DIR}/${LANG_CODE}.params-model
fi
# Compose the traineddata file.
run_command combine_tessdata ${TRAINING_DIR}/${LANG_CODE}.
# Copy it to the output dir, overwriting only if allowed by the cmdline flag.
if [[ ! -d ${OUTPUT_DIR} ]]; then
tlog "Creating new directory ${OUTPUT_DIR}"
mkdir -p ${OUTPUT_DIR}
fi
local destfile=${OUTPUT_DIR}/${LANG_CODE}.traineddata;
if [[ -f ${destfile} ]] && ((! OVERWRITE)); then
err_exit "File ${destfile} exists and no --overwrite specified";
fi
tlog "Moving ${TRAINING_DIR}/${LANG_CODE}.traineddata to ${OUTPUT_DIR}"
cp -f ${TRAINING_DIR}/${LANG_CODE}.traineddata ${destfile}
}