mirror of
https://github.com/Helsinki-NLP/OPUS-MT-train.git
synced 2024-12-12 07:03:01 +03:00
108 lines
2.9 KiB
Makefile
108 lines
2.9 KiB
Makefile
|
|
include ../Makefile.env
|
|
include ../Makefile.config
|
|
include ../Makefile.slurm
|
|
|
|
|
|
SRC = en
|
|
TRG = de
|
|
LANGPAIR = ${SRC}-${TRG}
|
|
|
|
MODELHOME = ../models/${LANGPAIR}
|
|
MODELZIP = ${lastword ${sort ${wildcard ${MODELHOME}/*-20*.zip}}}
|
|
MODELNAME = ${patsubst %.zip,%,${notdir ${MODELZIP}}}
|
|
MODELSTORE = OPUS-MT-models
|
|
MODELURL = https://object.pouta.csc.fi/${MODELSTORE}/${LANGPAIR}
|
|
|
|
TESTSETDIR = ../testsets/${LANGPAIR}
|
|
TESTSETS = ${sort ${wildcard ${TESTSETDIR}/*.${SRC}.gz}}
|
|
TESTSET = ${lastword ${TESTSETS}}
|
|
|
|
WORKDIR = ${PWD}/${LANGPAIR}/${MODELNAME}
|
|
TEST_PRE = ${WORKDIR}/${patsubst %.gz,%.pre,${notdir ${TESTSET}}}
|
|
TEST_TRANS = ${TEST_PRE}.${TRG}
|
|
TEST_EVAL = ${TEST_TRANS}.eval
|
|
|
|
|
|
|
|
all: ${WORKDIR}/model/decoder.yml
|
|
${MAKE} ${TEST_EVAL}
|
|
${MAKE} cleanup
|
|
|
|
|
|
|
|
|
|
|
|
.INTERMEDIATE: ${WORKDIR}/model/decoder.yml
|
|
|
|
prepare-model: ${WORKDIR}/model/decoder.yml
|
|
${WORKDIR}/model/decoder.yml:
|
|
mkdir -p ${dir $@}
|
|
ifneq (${wildcard ${MODELZIP}},)
|
|
cp ${MODELZIP} ${dir $@}
|
|
else
|
|
cd ${dir $@} && wget ${MODELURL}/${notdir ${MODELZIP}}
|
|
endif
|
|
cd ${dir $@} && unzip -u *.zip
|
|
|
|
|
|
SUBWORD_MODEL = ${filter-out ${WORKDIR}/model/source.tcmodel,${wildcard ${WORKDIR}/model/source.*}}
|
|
|
|
ifneq (${wildcard ${WORKDIR}/model/preprocess.sh},)
|
|
PREPROCESS = ${WORKDIR}/model/preprocess.sh ${SRC} ${SUBWORD_MODEL}
|
|
else ifeq (${SUBWORD_MODEL},${WORKDIR}/model/source.spm)
|
|
PREPROCESS = ${TOKENIZER}/replace-unicode-punctuation.perl |\
|
|
${TOKENIZER}/remove-non-printing-char.perl |\
|
|
${TOKENIZER}/normalize-punctuation.perl -l ${SRC} |\
|
|
sed 's/ */ /g;s/^ *//g;s/ *$$//g' |\
|
|
${SPM_HOME}/spm_encode --model ${SUBWORD_MODEL}
|
|
else
|
|
PREPROCESS = ${TOKENIZER}/replace-unicode-punctuation.perl |\
|
|
${TOKENIZER}/remove-non-printing-char.perl |\
|
|
${TOKENIZER}/normalize-punctuation.perl -l ${SRC} |\
|
|
${TOKENIZER}/tokenizer.perl -a -l ${SRC} |\
|
|
sed 's/ */ /g;s/^ *//g;s/ *$$//g' |\
|
|
python3 ${SNMTPATH}/apply_bpe.py -c ${SUBWORD_MODEL}
|
|
endif
|
|
|
|
ifneq (${wildcard ${WORKDIR}/model/postprocess.sh},)
|
|
POSTPROCESS = ${WORKDIR}/model/postprocess.sh
|
|
else ifeq (${SUBWORD_MODEL},${WORKDIR}/model/source.spm)
|
|
POSTPROCESS = sed 's/ //g;s/▁/ /g' | sed 's/^ *//;s/ *$$//'
|
|
else
|
|
POSTPROCESS = sed 's/\@\@ //g;s/ \@\@//g;s/ \@\-\@ /-/g' |\
|
|
$(TOKENIZER)/detokenizer.perl -l ${TRG}
|
|
endif
|
|
|
|
|
|
prepare-data: ${TEST_PRE}
|
|
${TEST_PRE}: ${TESTSET} ${WORKDIR}/model/decoder.yml
|
|
zcat $< | ${PREPROCESS} > $@
|
|
|
|
|
|
translate: ${TEST_TRANS}
|
|
|
|
## translate test set
|
|
${TEST_TRANS}: ${TEST_PRE} ${WORKDIR}/model/decoder.yml
|
|
mkdir -p ${dir $@}
|
|
cd ${dir ${word 2,$^}}; \
|
|
${LOADMODS} && ${MARIAN}/marian-decoder -i $< \
|
|
-c decoder.yml \
|
|
-d ${MARIAN_GPUS} \
|
|
${MARIAN_DECODER_FLAGS} |\
|
|
${POSTPROCESS} |\
|
|
sed 's/^ *//;s/ *$$//' > $@
|
|
|
|
|
|
eval: ${TEST_EVAL}
|
|
|
|
${TEST_EVAL}: ${TEST_TRANS}
|
|
zcat ${patsubst %.${SRC}.gz,%.${TRG}.gz,${TESTSET}} > $@.ref
|
|
cat $< | sacrebleu $@.ref > $@
|
|
cat $< | sacrebleu --metrics=chrf --width=3 $@.ref >> $@
|
|
rm -f $@.ref
|
|
|
|
|
|
cleanup:
|
|
rm -fr ${WORKDIR}/model
|