Helsinki-NLP/OPUS-MT-train
1
1
Fork 0
mirror of https://github.com/Helsinki-NLP/OPUS-MT-train.git synced 2024-09-11 20:27:19 +03:00
Code Issues Packages Projects Releases Wiki Activity
4b7ae1a39b
OPUS-MT-train/Makefile.dist

418 lines
15 KiB
Makefile
Raw Blame History

# -*-makefile-*-
#
# make distribution packages
# and upload them to cPouta ObjectStorage
#
MODELSHOME = ${WORKHOME}/models
DIST_PACKAGE = ${MODELSHOME}/${LANGPAIRSTR}/${DATASET}.zip
## minimum BLEU score for models to be accepted as distribution package
MIN_BLEU_SCORE = 20
.PHONY: dist
dist: ${DIST_PACKAGE}
.PHONY: scores
scores:
${MAKE} FIND_EVAL_FILES=1 ${WORKHOME}/eval/scores.txt
## get the best model from all kind of alternative setups
## in the following sub directories (add prefix work-)
# ALT_MODEL_DIR = bpe-old bpe-memad bpe spm-noalign bpe-align spm
ALT_MODEL_DIR = spm langid
best_dist_all:
for l in $(sort ${shell ls work* | grep -- '-' | grep -v old | grep -v work}); do \
if [ `find work*/$$l -name '*.npz' | wc -l` -gt 0 ]; then \
d=`find work-spm/$$l -name '*.best-perplexity.npz' -exec basename {} \; | cut -f1 -d.`; \
${MAKE} SRCLANGS="`echo $$l | cut -f1 -d'-' | sed 's/\\+/ /g'`" \
TRGLANGS="`echo $$l | cut -f2 -d'-' | sed 's/\\+/ /g'`" \
DATASET=$$d best_dist; \
fi \
done
# best_dist_all:
# for l in $(sort ${shell ls work* | grep -- '-' | grep -v old | grep -v work}); do \
# if [ `find work*/$$l -name '${DATASET}${TRAINSIZE}.*.npz' | wc -l` -gt 0 ]; then \
# ${MAKE} SRCLANGS="`echo $$l | cut -f1 -d'-' | sed 's/\\+/ /g'`" \
# TRGLANGS="`echo $$l | cut -f2 -d'-' | sed 's/\\+/ /g'`" best_dist; \
# fi \
# done
## find the best model according to test set scores
## and make a distribution package from that model
## (BLEU needs to be above MIN_BLEU_SCORE)
## NEW: don't trust models tested with GNOME test sets!
best_dist:
@m=0;\
s=''; \
echo "------------------------------------------------"; \
echo "search best model for ${LANGPAIRSTR}"; \
for d in ${ALT_MODEL_DIR}; do \
e=`ls work-$$d/${LANGPAIRSTR}/val/*.trg | xargs basename | sed 's/\.trg//'`; \
echo "evaldata = $$e"; \
if [ "$$e" != "GNOME" ]; then \
if ls work-$$d/${LANGPAIRSTR}/$$e*.eval 1> /dev/null 2>&1; then \
b=`grep 'BLEU+' work-$$d/${LANGPAIRSTR}/$$e*.eval | cut -f3 -d' '`; \
if (( $$(echo "$$m-$$b < 1" |bc -l) )); then \
echo "$$d ($$b) is better or not much worse than $$s ($$m)!"; \
m=$$b; \
s=$$d; \
else \
echo "$$d ($$b) is worse than $$s ($$m)!"; \
fi \
fi \
fi \
done; \
echo "------------------------------------------------"; \
if [ "$$s" != "" ]; then \
if (( $$(echo "$$m > ${MIN_BLEU_SCORE}" |bc -l) )); then \
${MAKE} MODELSHOME=${PWD}/models \
MODELS_URL=https://object.pouta.csc.fi/OPUS-MT-models dist-$$s; \
fi; \
fi
## make a package for distribution
## old: only accept models with a certain evaluation score:
# if [ `grep BLEU $(TEST_EVALUATION) | cut -f3 -d ' ' | cut -f1 -d '.'` -ge ${MIN_BLEU_SCORE} ]; then \
DATE = ${shell date +%F}
MODELS_URL = https://object.pouta.csc.fi/OPUS-MT-dev
SKIP_DIST_EVAL = 0
## determine pre-processing type
ifneq ("$(wildcard ${BPESRCMODEL})","")
PREPROCESS_TYPE = bpe
PREPROCESS_SRCMODEL = ${BPESRCMODEL}
PREPROCESS_TRGMODEL = ${BPETRGMODEL}
PREPROCESS_DESCRIPTION = normalization + tokenization + BPE
else
PREPROCESS_TYPE = spm
PREPROCESS_SRCMODEL = ${SPMSRCMODEL}
PREPROCESS_TRGMODEL = ${SPMTRGMODEL}
PREPROCESS_DESCRIPTION = normalization + SentencePiece
endif
ifneq (${words ${TRGLANGS}},1)
PREPROCESS_SCRIPT = preprocess-${PREPROCESS_TYPE}-multi-target.sh
else
PREPROCESS_SCRIPT = preprocess-${PREPROCESS_TYPE}.sh
endif
POSTPROCESS_SCRIPT = postprocess-${PREPROCESS_TYPE}.sh
${DIST_PACKAGE}: ${MODEL_FINAL}
ifneq (${SKIP_DIST_EVAL},1)
@${MAKE} $(TEST_EVALUATION)
@${MAKE} $(TEST_COMPARISON)
endif
@mkdir -p ${dir $@}
@touch ${WORKDIR}/source.tcmodel
@echo "# $(notdir ${@:.zip=})-${DATE}.zip" > ${WORKDIR}/README.md
@echo '' >> ${WORKDIR}/README.md
@echo "* dataset: ${DATASET}" >> ${WORKDIR}/README.md
@echo "* model: ${MODELTYPE}" >> ${WORKDIR}/README.md
@echo "* pre-processing: ${PREPROCESS_DESCRIPTION}" >> ${WORKDIR}/README.md
@cp ${PREPROCESS_SRCMODEL} ${WORKDIR}/source.${PREPROCESS_TYPE}
@cp ${PREPROCESS_TRGMODEL} ${WORKDIR}/target.${PREPROCESS_TYPE}
@cp ${PREPROCESS_SCRIPT} ${WORKDIR}/preprocess.sh
@cp ${POSTPROCESS_SCRIPT} ${WORKDIR}/postprocess.sh
@if [ ${words ${TRGLANGS}} -gt 1 ]; then \
echo '* a sentence initial language token is required in the form of `>>id<<` (id = valid target language ID)' \
>> ${WORKDIR}/README.md; \
fi
@echo "* download: [$(notdir ${@:.zip=})-${DATE}.zip](${MODELS_URL}/${LANGPAIRSTR}/$(notdir ${@:.zip=})-${DATE}.zip)" >> ${WORKDIR}/README.md
if [ -e $(TEST_EVALUATION) ]; then \
echo "* test set translations: [$(notdir ${@:.zip=})-${DATE}.test.txt](${MODELS_URL}/${LANGPAIRSTR}/$(notdir ${@:.zip=})-${DATE}.test.txt)" >> ${WORKDIR}/README.md; \
echo "* test set scores: [$(notdir ${@:.zip=})-${DATE}.eval.txt](${MODELS_URL}/${LANGPAIRSTR}/$(notdir ${@:.zip=})-${DATE}.eval.txt)" >> ${WORKDIR}/README.md; \
echo '' >> ${WORKDIR}/README.md; \
echo '## Benchmarks' >> ${WORKDIR}/README.md; \
echo '' >> ${WORKDIR}/README.md; \
cd ${WORKDIR}; \
grep -H BLEU *k${NR}.*eval | \
tr '.' '/' | cut -f1,5,6 -d '/' | tr '/' "." > $@.1; \
grep BLEU *k${NR}.*eval | cut -f3 -d ' ' > $@.2; \
grep chrF *k${NR}.*eval | cut -f3 -d ' ' > $@.3; \
echo '| testset | BLEU | chr-F |' >> README.md; \
echo '|-----------------------|-------|-------|' >> README.md; \
paste $@.1 $@.2 $@.3 | sed "s/\t/ | /g;s/^/| /;s/$$/ |/" >> README.md; \
rm -f $@.1 $@.2 $@.3; \
fi
@cat ${WORKDIR}/README.md >> ${dir $@}README.md
@echo '' >> ${dir $@}README.md
@cp models/LICENSE ${WORKDIR}/
@chmod +x ${WORKDIR}/preprocess.sh
@sed -e 's# - /.*/\([^/]*\)$$# - \1#' \
-e 's/beam-size: [0-9]*$$/beam-size: 6/' \
-e 's/mini-batch: [0-9]*$$/mini-batch: 1/' \
-e 's/maxi-batch: [0-9]*$$/maxi-batch: 1/' \
-e 's/relative-paths: false/relative-paths: true/' \
< ${MODEL_DECODER} > ${WORKDIR}/decoder.yml
@cd ${WORKDIR} && zip ${notdir $@} \
README.md LICENSE \
${notdir ${MODEL_FINAL}} \
${notdir ${MODEL_VOCAB}} \
${notdir ${MODEL_VALIDLOG}} \
${notdir ${MODEL_TRAINLOG}} \
source.* target.* decoder.yml preprocess.sh postprocess.sh
@mkdir -p ${dir $@}
@mv -f ${WORKDIR}/${notdir $@} ${@:.zip=}-${DATE}.zip
if [ -e $(TEST_EVALUATION) ]; then \
cp $(TEST_EVALUATION) ${@:.zip=}-${DATE}.eval.txt; \
cp $(TEST_COMPARISON) ${@:.zip=}-${DATE}.test.txt; \
fi
@rm -f $@
@cd ${dir $@} && ln -s $(notdir ${@:.zip=})-${DATE}.zip ${notdir $@}
@rm -f ${WORKDIR}/decoder.yml ${WORKDIR}/source.* ${WORKDIR}/target.*
@rm -f ${WORKDIR}/preprocess.sh ${WORKDIR}/postprocess.sh
## do this only if the flag is set
## --> avoid expensive wildcard searches each time make is called
ifeq (${FIND_EVAL_FILES},1)
EVALSCORES = ${patsubst ${WORKHOME}/%.eval,${WORKHOME}/eval/%.eval.txt,${wildcard ${WORKHOME}/*/*.eval}}
EVALTRANSL = ${patsubst ${WORKHOME}/%.compare,${WORKHOME}/eval/%.test.txt,${wildcard ${WORKHOME}/*/*.compare}}
endif
## upload to Object Storage
## Don't forget to run this before uploading!
# source project_2000661-openrc.sh
#
# - make upload ......... released models = all sub-dirs in models/
# - make upload-models .. trained models in current WORKHOME to OPUS-MT-dev
# - make upload-scores .. score file with benchmark results to OPUS-MT-eval
# - make upload-eval .... benchmark tests from models in WORKHOME
# - make upload-images .. images of VMs that run OPUS-MT
upload:
find models/ -type l | tar -cf models-links.tar -T -
find models/ -type l -delete
cd models && swift upload OPUS-MT-models --changed --skip-identical *
tar -xf models-links.tar
rm -f models-links.tar
swift post OPUS-MT-models --read-acl ".r:*"
swift list OPUS-MT-models > index.txt
swift upload OPUS-MT-models index.txt
rm -f index.txt
upload-models:
find ${WORKHOME}/models -type l | tar -cf dev-models-links.tar -T -
find ${WORKHOME}/models -type l -delete
cd ${WORKHOME} && swift upload OPUS-MT-dev --changed --skip-identical models
tar -xf dev-models-links.tar
rm -f dev-models-links.tar
swift post OPUS-MT-dev --read-acl ".r:*"
swift list OPUS-MT-dev > index.txt
swift upload OPUS-MT-dev index.txt
rm -f index.txt
upload-scores: scores
cd ${WORKHOME} && swift upload OPUS-MT-eval --changed --skip-identical eval/scores.txt
swift post OPUS-MT-eval --read-acl ".r:*"
upload-eval: scores
cd ${WORKHOME} && swift upload OPUS-MT-eval --changed --skip-identical eval
swift post OPUS-MT-eval --read-acl ".r:*"
upload-images:
cd ${WORKHOME} && swift upload OPUS-MT --changed --skip-identical \
--use-slo --segment-size 5G opusMT-images
swift post OPUS-MT-images --read-acl ".r:*"
## this is for the multeval scores
# ${WORKHOME}/eval/scores.txt: ${EVALSCORES}
# cd ${WORKHOME} && \
# grep base */*eval | cut -f1,2- -d '/' | cut -f1,6- -d '.' | \
# sed 's/-/ /' | sed 's/\// /' | sed 's/ ([^)]*)//g' |\
# sed 's/.eval:baseline//' | sed "s/ */\t/g" | sort > $@
${WORKHOME}/eval/scores.txt: ${EVALSCORES} ${EVALTRANSL}
cd ${WORKHOME} && grep BLEU */*k${NR}.*eval | cut -f1 -d '/' | tr '-' "\t" > $@.1
cd ${WORKHOME} && grep BLEU */*k${NR}.*eval | tr '.' '/' | cut -f2,6,7 -d '/' | tr '/' "." > $@.2
cd ${WORKHOME} && grep BLEU */*k${NR}.*eval | cut -f3 -d ' ' > $@.3
cd ${WORKHOME} && grep chrF */*k${NR}.*eval | cut -f3 -d ' ' > $@.4
paste $@.1 $@.2 $@.3 $@.4 > $@
rm -f $@.1 $@.2 $@.3 $@.4
${EVALSCORES}: # ${WORKHOME}/eval/%.eval.txt: ${WORKHOME}/models/%.eval
mkdir -p ${dir $@}
cp ${patsubst ${WORKHOME}/eval/%.eval.txt,${WORKHOME}/%.eval,$@} $@
# cp $< $@
${EVALTRANSL}: # ${WORKHOME}/eval/%.test.txt: ${WORKHOME}/models/%.compare
mkdir -p ${dir $@}
cp ${patsubst ${WORKHOME}/eval/%.test.txt,${WORKHOME}/%.compare,$@} $@
# cp $< $@
# ## dangerous area ....
# delete-eval:
# swift delete OPUS-MT eval
######################################################################
## handle old models in previous work directories
## obsolete now?
######################################################################
##-----------------------------------
## make packages from trained models
## check old-models as well!
TRAINED_NEW_MODELS = ${patsubst ${WORKHOME}/%/,%,${dir ${wildcard ${WORKHOME}/*/*.best-perplexity.npz}}}
# TRAINED_OLD_MODELS = ${patsubst ${WORKHOME}/old-models/%/,%,${dir ${wildcard ${WORKHOME}/old-models/*/*.best-perplexity.npz}}}
TRAINED_OLD_MODELS = ${patsubst ${WORKHOME}/old-models/%/,%,${dir ${wildcard ${WORKHOME}/old-models/??-??/*.best-perplexity.npz}}}
TRAINED_OLD_ONLY_MODELS = ${filter-out ${TRAINED_NEW_MODELS},${TRAINED_OLD_MODELS}}
TRAINED_NEW_ONLY_MODELS = ${filter-out ${TRAINED_OLD_MODELS},${TRAINED_NEW_MODELS}}
TRAINED_DOUBLE_MODELS = ${filter ${TRAINED_NEW_MODELS},${TRAINED_OLD_MODELS}}
## make packages of all new models
## unless there are better models in old-models
new-models-dist:
@echo "nr of extra models: ${words ${TRAINED_NEW_ONLY_MODELS}}"
for l in ${TRAINED_NEW_ONLY_MODELS}; do \
${MAKE} SRCLANGS="`echo $$l | cut -f1 -d'-' | sed 's/\\+/ /g'`" \
TRGLANGS="`echo $$l | cut -f2 -d'-' | sed 's/\\+/ /g'`" dist; \
done
@echo "trained double ${words ${TRAINED_DOUBLE_MODELS}}"
for l in ${TRAINED_DOUBLE_MODELS}; do \
n=`grep 'new best' work/$$l/*.valid1.log | tail -1 | cut -f12 -d ' '`; \
o=`grep 'new best' work/old-models/$$l/*.valid1.log | tail -1 | cut -f12 -d ' '`; \
if (( $$(echo "$$n < $$o" |bc -l) )); then \
${MAKE} SRCLANGS="`echo $$l | cut -f1 -d'-' | sed 's/\\+/ /g'`" \
TRGLANGS="`echo $$l | cut -f2 -d'-' | sed 's/\\+/ /g'`" dist; \
fi \
done
## fix decoder path in old-models (to run evaluations
fix-decoder-path:
for l in ${wildcard ${WORKHOME}/old-models/*/*.best-perplexity.npz.decoder.yml}; do \
sed --in-place=.backup 's#/\(..-..\)/opus#/old-models/\1/opus#' $$l; \
sed --in-place=.backup2 's#/old-models/old-models/#/old-models/#' $$l; \
sed --in-place=.backup2 's#/old-models/old-models/#/old-models/#' $$l; \
done
## make packages of all old models from old-models
## unless there are better models in work (new models)
old-models-dist:
@echo "nr of extra models: ${words ${TRAINED_OLD_ONLY_MODELS}}"
for l in ${TRAINED_OLD_ONLY_MODELS}; do \
${MAKE} SRCLANGS="`echo $$l | cut -f1 -d'-' | sed 's/\\+/ /g'`" \
TRGLANGS="`echo $$l | cut -f2 -d'-' | sed 's/\\+/ /g'`" \
WORKHOME=${WORKHOME}/old-models \
MODELSHOME=${WORKHOME}/models dist; \
done
@echo "trained double ${words ${TRAINED_DOUBLE_MODELS}}"
for l in ${TRAINED_DOUBLE_MODELS}; do \
n=`grep 'new best' work/$$l/*.valid1.log | tail -1 | cut -f12 -d ' '`; \
o=`grep 'new best' work/old-models/$$l/*.valid1.log | tail -1 | cut -f12 -d ' '`; \
if (( $$(echo "$$o < $$n" |bc -l) )); then \
${MAKE} SRCLANGS="`echo $$l | cut -f1 -d'-' | sed 's/\\+/ /g'`" \
TRGLANGS="`echo $$l | cut -f2 -d'-' | sed 's/\\+/ /g'`" \
WORKHOME=${WORKHOME}/old-models \
MODELSHOME=${WORKHOME}/models dist; \
else \
echo "$$l: new better than old"; \
fi \
done
## old models had slightly different naming conventions
LASTSRC = ${lastword ${SRCLANGS}}
LASTTRG = ${lastword ${TRGLANGS}}
MODEL_OLD = ${MODEL_SUBDIR}${DATASET}${TRAINSIZE}.${PRE_SRC}-${PRE_TRG}.${LASTSRC}${LASTTRG}
MODEL_OLD_BASENAME = ${MODEL_OLD}.${MODELTYPE}.model${NR}
MODEL_OLD_FINAL = ${WORKDIR}/${MODEL_OLD_BASENAME}.npz.best-perplexity.npz
MODEL_OLD_VOCAB = ${WORKDIR}/${MODEL_OLD}.vocab.${MODEL_VOCABTYPE}
MODEL_OLD_DECODER = ${MODEL_OLD_FINAL}.decoder.yml
MODEL_TRANSLATE = ${WORKDIR}/${TESTSET}.${MODEL}${NR}.${MODELTYPE}.${SRC}.${TRG}
MODEL_OLD_TRANSLATE = ${WORKDIR}/${TESTSET}.${MODEL_OLD}${NR}.${MODELTYPE}.${SRC}.${TRG}
MODEL_OLD_VALIDLOG = ${MODEL_OLD}.${MODELTYPE}.valid${NR}.log
MODEL_OLD_TRAINLOG = ${MODEL_OLD}.${MODELTYPE}.train${NR}.log
link-old-models:
if [ ! -e ${MODEL_FINAL} ]; then \
if [ -e ${MODEL_OLD_FINAL} ]; then \
ln -s ${MODEL_OLD_FINAL} ${MODEL_FINAL}; \
ln -s ${MODEL_OLD_VOCAB} ${MODEL_VOCAB}; \
ln -s ${MODEL_OLD_DECODER} ${MODEL_DECODER}; \
fi \
fi
if [ ! -e ${MODEL_TRANSLATE} ]; then \
if [ -e ${MODEL_OLD_TRANSLATE} ]; then \
ln -s ${MODEL_OLD_TRANSLATE} ${MODEL_TRANSLATE}; \
fi \
fi
if [ ! -e ${WORKDIR}/${MODEL_VALIDLOG} ]; then \
if [ -e ${WORKDIR}/${MODEL_OLD_VALIDLOG} ]; then \
ln -s ${WORKDIR}/${MODEL_OLD_VALIDLOG} ${WORKDIR}/${MODEL_VALIDLOG}; \
ln -s ${WORKDIR}/${MODEL_OLD_TRAINLOG} ${WORKDIR}/${MODEL_TRAINLOG}; \
fi \
fi
rm -f ${MODEL_TRANSLATE}.eval
rm -f ${MODEL_TRANSLATE}.compare
ifneq (${DATASET},${OLDDATASET})
TRAINFILES = ${wildcard ${WORKDIR}/train/*${OLDDATASET}*.*}
MODELFILES = ${wildcard ${WORKDIR}/*${OLDDATASET}*.*}
DECODERFILES = ${wildcard ${WORKDIR}/*${OLDDATASET}*.decoder.yml}
endif
## fix model names from old style
## where models trained on a single corpus got the name
## of that corpus
## Now: always use 'opus' as the name of the default dataset
fix-model-names:
ifneq (${DATASET},${OLDDATASET})
for f in ${DECODERFILES}; do \
perl -i.bak -pe 's/${OLDDATASET}/${DATASET}/' $$f; \
done
for f in ${TRAINFILES}; do \
mv -f $$f `echo $$f | sed 's/${OLDDATASET}/${DATASET}/'`; \
ln -s `echo $$f | sed 's/${OLDDATASET}/${DATASET}/'` $$f; \
done
for f in ${MODELFILES}; do \
mv -f $$f `echo $$f | sed 's/${OLDDATASET}/${DATASET}/'`; \
done
endif
Reference in New Issue View Git Blame Copy Permalink