mirror of
https://github.com/Helsinki-NLP/OPUS-MT-train.git
synced 2024-10-05 16:47:21 +03:00
317 lines
8.9 KiB
Makefile
317 lines
8.9 KiB
Makefile
#
|
|
# forward translation to be used for
|
|
# knowledge distillation
|
|
#
|
|
# only works with sentencepiece models!
|
|
#
|
|
# TODO's
|
|
#
|
|
# - forward-translate monolingual data (re-use bt-data)
|
|
# - reconstruction filtering (score translation in opposite direction)
|
|
# (use weights? normalise-script from bergamot/students)
|
|
# - other kind of data filtering / selection?
|
|
# - create lexical shortlists (see bergamot)
|
|
# - finetune alphas in intgemm8 models (see bergamot)
|
|
# - benchmark distilled models
|
|
#
|
|
|
|
PWD := ${shell pwd}
|
|
REPOHOME := ${PWD}/../
|
|
|
|
include ${REPOHOME}lib/env.mk
|
|
include ${REPOHOME}lib/config.mk
|
|
include ${REPOHOME}lib/slurm.mk
|
|
|
|
|
|
SRC = fin
|
|
TRG = eng
|
|
|
|
|
|
## change decoder settings
|
|
## TODO: do we need this?
|
|
|
|
MARIAN_BEAM_SIZE=1
|
|
MARIAN_MINI_BATCH=100
|
|
MARIAN_MAXI_BATCH=100
|
|
MARIAN_MAX_LENGTH=200
|
|
MARIAN_WORKSPACE=12000
|
|
|
|
|
|
TATOEBA_VERSION ?= v2021-08-07
|
|
TATOEBA_VERSION_NOHYPHEN ?= $(subst -,,${TATOEBA_VERSION})
|
|
|
|
TATOEBA_GITRAW = https://raw.githubusercontent.com/Helsinki-NLP/Tatoeba-Challenge/master
|
|
TATOEBA_RELEASED = ${TATOEBA_GITRAW}/models/released-model-results-all.txt
|
|
TATOEBA_RELEASED_BT = https://object.pouta.csc.fi/Tatoeba-MT-bt/released-data.txt
|
|
TATOEBA_MODEL_STORAGE = https://object.pouta.csc.fi/Tatoeba-MT-models
|
|
|
|
## container for storing backtranslations
|
|
BT_CONTAINER = Tatoeba-MT-bt
|
|
BT_CWORK_ONTAINER = project-Tatoeba-MT-bt
|
|
|
|
## split size in nr-of-lines
|
|
## default part to be selected = aa
|
|
SPLIT_SIZE ?= 1000000
|
|
|
|
## maximum input length (number sentence piece segments)
|
|
## maximum number of sentences to be translated (top N lines)
|
|
MAX_LENGTH ?= 200
|
|
MAX_SENTENCES ?= ${SPLIT_SIZE}
|
|
|
|
SORTLANGS = $(sort ${SRC} ${TRG})
|
|
LANGPAIR = ${SRC}-${TRG}
|
|
SORTED_LANGPAIR = ${firstword ${SORTLANGS}}-${lastword ${SORTLANGS}}
|
|
|
|
PWD := $(shell pwd)
|
|
|
|
|
|
|
|
# MODELZIP = https://object.pouta.csc.fi/Tatoeba-Challenge/ang.tar
|
|
MODELZIP := ${shell wget -qq -O - ${TATOEBA_RELEASED} | grep '^${LANGPAIR}' | head -1 | cut -f4}
|
|
MODELINFO = ${patsubst ${TATOEBA_MODEL_STORAGE}/%.zip,${TATOEBA_GITRAW}/models/%.yml,${MODELZIP}}
|
|
MODELNAME = ${patsubst %.zip,%,${notdir ${MODELZIP}}}
|
|
|
|
MULTI_TARGET_MODEL := ${shell wget -qq -O - ${MODELINFO} | grep 'use-target-labels' | wc -l}
|
|
ifneq (${MULTI_TARGET_MODEL},0)
|
|
TARGET_LANG_LABEL := ${shell wget -qq -O - ${MODELINFO} | grep -o '>>${TRG}.*<<'}
|
|
endif
|
|
|
|
RELEASED_BITEXTS := $(patsubst %.tar,%,${shell wget -qq -O - ${TATOEBA_GITRAW}/Wiki.md | \
|
|
grep -o 'WikiShuffled/...\.tar' | cut -f2 -d'/'})
|
|
|
|
RELEASED_BITEXTS_REV = ${shell (for d in ${RELEASED_BITEXTS}; do echo $$d; done) | tac}
|
|
|
|
|
|
PART ?= aa
|
|
OUTPUT_DIR ?= ${LANGPAIR}
|
|
|
|
BITEXT_DATADIR = ${PWD}/../work-tatoeba/data/simple
|
|
MODEL_WORKDIR = ${PWD}/../work-tatoeba/${LANGPAIR}
|
|
BITEXT_SRCRAW = ${BITEXT_DATADIR}/Tatoeba-train-${TATOEBA_VERSION}.${SORTED_LANGPAIR}.clean.${SRC}.gz
|
|
BITEXT_SRCPRE = ${wildcard ${MODEL_WORKDIR}/train/opusTC${TATOEBA_VERSION_NOHYPHEN}.src.clean.spm*.gz}
|
|
|
|
BITEXT_BASE = ${OUTPUT_DIR}/Tatoeba-train.${MODELNAME}.${LANGPAIR}
|
|
BITEXT_SRC = ${BITEXT_BASE}.${SRC}.${PART}.gz
|
|
BITEXT_PRE = ${BITEXT_BASE}.${SRC}.spm.${PART}.gz
|
|
BITEXT_TRG = ${BITEXT_BASE}.${TRG}.${PART}.gz
|
|
|
|
BITEXT_LATEST_SRC = ${OUTPUT_DIR}/latest/Tatoeba-train.${PART}.${LANGPAIR}.${SRC}.gz
|
|
BITEXT_LATEST_TRG = ${OUTPUT_DIR}/latest/Tatoeba-train.${PART}.${LANGPAIR}.${TRG}.gz
|
|
BITEXT_LATEST_README = ${OUTPUT_DIR}/latest/README.md
|
|
|
|
|
|
## all parts of the bitext
|
|
PARTS = $(subst .,,${suffix ${basename ${wildcard ${BITEXT_PRE:${PART}.gz=}??.gz}}})
|
|
ALL_BITEXT_LATEST_SRC = ${patsubst %,${OUTPUT_DIR}/latest/Tatoeba-train.%.${LANGPAIR}.${SRC}.gz,${PARTS}}
|
|
ALL_BITEXT_LATEST_TRG = ${patsubst %,${OUTPUT_DIR}/latest/Tatoeba-train.%.${LANGPAIR}.${TRG}.gz,${PARTS}}
|
|
|
|
|
|
## don't delete translated text even if the process crashes
|
|
.PRECIOUS: ${BITEXT_BASE}.${TRG}.%.gz
|
|
|
|
.PHONY: all
|
|
all: translate
|
|
|
|
.PHONY: prepare
|
|
prepare: ${LANGPAIR}/${MODELNAME}/decoder.yml ${BITEXT_PRE}
|
|
|
|
.PHONY: translate
|
|
translate: ${BITEXT_LATEST_README} ${BITEXT_LATEST_TRG}
|
|
${MAKE} ${BITEXT_LATEST_SRC}
|
|
|
|
## translate all parts
|
|
.PHONY: translate-all-parts
|
|
translate-all-parts: ${ALL_BITEXT_LATEST_TRG}
|
|
${MAKE} source-all-parts
|
|
|
|
.PHONY: source-all-parts
|
|
source-all-parts: ${ALL_BITEXT_LATEST_SRC}
|
|
|
|
|
|
.PHONY: print-modelinfo
|
|
print-modelinfo:
|
|
@echo ${MODELNAME}
|
|
@echo ${MODELZIP}
|
|
@echo ${MODELINFO}
|
|
@echo "multi-target model: ${MULTI_TARGET_MODEL}"
|
|
@echo "target language label: ${TARGET_LANG_LABEL}"
|
|
|
|
## fetch the latest model
|
|
|
|
${LANGPAIR}/${MODELNAME}/decoder.yml:
|
|
ifneq (${MODELZIP},)
|
|
mkdir -p ${dir $@}
|
|
wget -O ${dir $@}/model.zip ${MODELZIP}
|
|
cd ${dir $@} && unzip model.zip
|
|
rm -f ${dir $@}/model.zip
|
|
mv ${dir $@}/preprocess.sh ${dir $@}/preprocess-old.sh
|
|
sed 's#perl -C -pe.*$$#perl -C -pe "s/(?!\\n)\\p{C}/ /g;" |#' \
|
|
< ${dir $@}/preprocess-old.sh > ${dir $@}/preprocess.sh
|
|
chmod +x ${dir $@}/preprocess.sh
|
|
endif
|
|
|
|
|
|
## pre-process data
|
|
|
|
ifeq (${MULTI_TARGET_MODEL},1)
|
|
PREPROCESS_ARGS = ${SRC} ${TRG} ${LANGPAIR}/${MODELNAME}/source.spm
|
|
else
|
|
PREPROCESS_ARGS = ${SRC} ${LANGPAIR}/${MODELNAME}/source.spm
|
|
endif
|
|
|
|
|
|
ifeq (${BITEXT_SRCPRE},)
|
|
|
|
${BITEXT_SRCRAW}:
|
|
${MAKE} -C .. SRCLANGS=${SRC} TRGLANGS=${TRG} clean-data-tatoeba
|
|
|
|
${BITEXT_PRE}: ${BITEXT_SRCRAW}
|
|
ifneq (${MODELZIP},)
|
|
mkdir -p ${dir $@}
|
|
${MAKE} ${LANGPAIR}/${MODELNAME}/decoder.yml
|
|
${GZCAT} $< |\
|
|
grep -v '[<>{}]' |\
|
|
${LANGPAIR}/${MODELNAME}/preprocess.sh ${PREPROCESS_ARGS} |\
|
|
perl -e 'while (<>){next if (split(/\s+/)>${MAX_LENGTH});print;}' |\
|
|
split -l ${SPLIT_SIZE} - ${patsubst %${PART}.gz,%,$@} |\
|
|
gzip -f > $@
|
|
endif
|
|
|
|
else
|
|
|
|
${BITEXT_PRE}: ${BITEXT_SRCPRE}
|
|
${GZCAT} $< |\
|
|
perl -e 'while (<>){next if (split(/\s+/)>${MAX_LENGTH});print;}' |\
|
|
split -l ${SPLIT_SIZE} - ${patsubst %${PART}.gz,%,$@}
|
|
${GZIP} -f ${patsubst %${PART}.gz,%,$@}??
|
|
|
|
endif
|
|
|
|
|
|
|
|
## merge SentencePiece segments in the source text
|
|
## (Why? because we filter out some data from the original wiki text, see above)
|
|
|
|
${BITEXT_BASE}.${SRC}.%.gz: ${BITEXT_BASE}.${SRC}.spm.%.gz
|
|
if [ -e ${patsubst ${BITEXT_BASE}.${SRC}.%.gz,${BITEXT_BASE}.${TRG}.%.gz,$@} ]; then \
|
|
mkdir -p ${dir $@}; \
|
|
${GZCAT} $< |\
|
|
sed 's/ //g;s/▁/ /g' | \
|
|
sed 's/^ *//;s/ *$$//' |\
|
|
sed 's/^>>[a-z]*<< //' |\
|
|
gzip -c > $@; \
|
|
fi
|
|
|
|
|
|
## overwrite the file with the latest translations
|
|
## --> this allows multiple translation iterations
|
|
## without duplicating the data we want to use in MT training
|
|
|
|
${OUTPUT_DIR}/latest/Tatoeba-train.%.${LANGPAIR}.${SRC}.gz: ${BITEXT_BASE}.${SRC}.%.gz
|
|
mkdir -p ${dir $@}
|
|
cp $< $@
|
|
|
|
${OUTPUT_DIR}/latest/Tatoeba-train.%.${LANGPAIR}.${TRG}.gz: ${BITEXT_BASE}.${TRG}.%.gz
|
|
mkdir -p ${dir $@}
|
|
cp $< $@
|
|
|
|
${BITEXT_LATEST_README}: ${LANGPAIR}/${MODELNAME}/README.md
|
|
mkdir -p ${dir $@}
|
|
cp $< $@
|
|
|
|
|
|
## translate
|
|
|
|
${BITEXT_BASE}.${TRG}.%.gz: ${BITEXT_BASE}.${SRC}.spm.%.gz
|
|
ifneq (${MODELZIP},)
|
|
mkdir -p ${dir $@}
|
|
${MAKE} ${LANGPAIR}/${MODELNAME}/decoder.yml
|
|
${LOAD_ENV} && cd ${LANGPAIR}/${MODELNAME} && \
|
|
${MARIAN_DECODER} \
|
|
-c decoder.yml \
|
|
-i ${PWD}/$< \
|
|
-d ${MARIAN_GPUS} \
|
|
${MARIAN_DECODER_FLAGS} |\
|
|
sed 's/ //g;s/▁/ /g' | sed 's/^ *//;s/ *$$//' |\
|
|
gzip -c > ${PWD}/$@
|
|
endif
|
|
|
|
|
|
|
|
check-latest:
|
|
@if [ -d ${LANGPAIR}/latest ]; then \
|
|
for S in `ls ${LANGPAIR}/latest/*.${SRC}.gz`; do \
|
|
T=`echo $$S | sed 's/.${SRC}.gz/.${TRG}.gz/'`; \
|
|
a=`${GZCAT} $$S | wc -l`; \
|
|
b=`${GZCAT} $$T | wc -l`; \
|
|
if [ $$a != $$b ]; then \
|
|
echo "$$a != $$b $$S $$T"; \
|
|
else \
|
|
echo "$$a $$S $$T"; \
|
|
fi \
|
|
done \
|
|
fi
|
|
|
|
check-translated:
|
|
@for S in `ls ${LANGPAIR}/*.${SRC}.spm.gz`; do \
|
|
T=`echo $$S | sed 's/.${SRC}.gz/.${TRG}.gz/'`; \
|
|
a=`${GZCAT} $$S | wc -l`; \
|
|
b=`${GZCAT} $$T | wc -l`; \
|
|
if [ $$a != $$b ]; then \
|
|
echo "$$a != $$b $$S $$T"; \
|
|
else \
|
|
echo "$$a $$S $$T"; \
|
|
fi \
|
|
done
|
|
|
|
check-length:
|
|
@echo "check ${LANGPAIR}"
|
|
@${MAKE} check-translated
|
|
@${MAKE} check-latest
|
|
|
|
|
|
remove-%-all check-%-all:
|
|
for d in `find . -maxdepth 1 -type d -name '*-*' -printf "%f "`; do \
|
|
s=`echo $$d | cut -f1 -d'-'`; \
|
|
t=`echo $$d | cut -f2 -d'-'`; \
|
|
make SRC=$$s TRG=$$t ${@:-all=}; \
|
|
done
|
|
|
|
|
|
|
|
remove-incomplete:
|
|
${MAKE} remove-incomplete-translated
|
|
${MAKE} remove-incomplete-latest
|
|
|
|
remove-incomplete-translated:
|
|
@echo "check ${LANGPAIR}"
|
|
@mkdir -p ${LANGPAIR}/incomplete
|
|
@for S in `ls ${LANGPAIR}/*.${SRC}.gz`; do \
|
|
T=`echo $$S | sed 's/.${SRC}.gz/.${TRG}.gz/'`; \
|
|
a=`${GZCAT} $$S | wc -l`; \
|
|
b=`${GZCAT} $$T | wc -l`; \
|
|
if [ $$a != $$b ]; then \
|
|
echo "$$a != $$b $$S $$T"; \
|
|
mv $$S ${LANGPAIR}/incomplete/; \
|
|
mv $$T ${LANGPAIR}/incomplete/; \
|
|
fi \
|
|
done
|
|
|
|
|
|
remove-incomplete-latest:
|
|
@echo "check ${LANGPAIR}"
|
|
@mkdir -p ${LANGPAIR}/incomplete/latest
|
|
@if [ -d ${LANGPAIR}/latest ]; then \
|
|
for S in `ls ${LANGPAIR}/latest/*.${SRC}.gz`; do \
|
|
T=`echo $$S | sed 's/.${SRC}.gz/.${TRG}.gz/'`; \
|
|
a=`${GZCAT} $$S | wc -l`; \
|
|
b=`${GZCAT} $$T | wc -l`; \
|
|
if [ $$a != $$b ]; then \
|
|
echo "$$a != $$b $$S $$T"; \
|
|
mv $$S ${LANGPAIR}/incomplete/latest/; \
|
|
mv $$T ${LANGPAIR}/incomplete/latest/; \
|
|
fi \
|
|
done \
|
|
fi
|
|
|