OPUS-MT-train/lib/generic.mk
Joerg Tiedemann 1fe26b4b0b elg work
2022-03-23 12:08:51 +02:00

483 lines
14 KiB
Makefile

# -*-makefile-*-
#
# generic implic targets that make our life a bit easier
unidirectional:
${MAKE} data
${MAKE} WALLTIME=72 HPC_MEM=4g HPC_CORES=1 train.submit-multigpu
bilingual:
${MAKE} data
${MAKE} WALLTIME=72 HPC_MEM=4g HPC_CORES=1 train.submit-multigpu
${MAKE} reverse-data
${MAKE} WALLTIME=72 HPC_MEM=4g HPC_CORES=1 SRCLANGS="${TRGLANGS}" TRGLANGS="${SRCLANGS}" \
train.submit-multigpu
bilingual-big:
${MAKE} data
${MAKE} WALLTIME=72 HPC_MEM=8g HPC_CORES=1 train.submit-multigpu
${MAKE} reverse-data
${MAKE} WALLTIME=72 HPC_MEM=8g HPC_CORES=1 SRCLANGS="${TRGLANGS}" TRGLANGS="${SRCLANGS}" \
train.submit-multigpu
bilingual-medium:
${MAKE} data
${MAKE} WALLTIME=72 HPC_MEM=4g HPC_CORES=1 \
MARIAN_VALID_FREQ=5000 MARIAN_WORKSPACE=10000 train.submit
${MAKE} reverse-data
${MAKE} WALLTIME=72 HPC_MEM=4g HPC_CORES=1 SRCLANGS="${TRGLANGS}" TRGLANGS="${SRCLANGS}" \
MARIAN_VALID_FREQ=5000 MARIAN_WORKSPACE=10000 train.submit
bilingual-small:
${MAKE} data
${MAKE} WALLTIME=72 HPC_MEM=4g HPC_CORES=1 \
MARIAN_WORKSPACE=5000 MARIAN_VALID_FREQ=2500 train.submit
${MAKE} reverse-data
${MAKE} WALLTIME=72 HPC_MEM=4g HPC_CORES=1 SRCLANGS="${TRGLANGS}" TRGLANGS="${SRCLANGS}" \
MARIAN_WORKSPACE=5000 MARIAN_VALID_FREQ=2500 train.submit
multilingual:
${MAKE} SRCLANGS="${LANGS}" TRGLANGS="${LANGS}" data
${MAKE} SRCLANGS="${LANGS}" TRGLANGS="${LANGS}" \
WALLTIME=72 HPC_CORES=1 HPC_MEM=4g train.submit-multigpu
multilingual-big:
${MAKE} SRCLANGS="${LANGS}" TRGLANGS="${LANGS}" data
${MAKE} SRCLANGS="${LANGS}" TRGLANGS="${LANGS}" \
WALLTIME=72 HPC_CORES=1 HPC_MEM=8g train.submit-multigpu
multilingual-medium:
${MAKE} SRCLANGS="${LANGS}" TRGLANGS="${LANGS}" data
${MAKE} SRCLANGS="${LANGS}" TRGLANGS="${LANGS}" \
MARIAN_VALID_FREQ=5000 MARIAN_WORKSPACE=10000 \
WALLTIME=72 HPC_CORES=1 HPC_MEM=4g train.submit-multigpu
all2pivot:
for l in ${filter-out ${PIVOT},${LANGS}}; do \
${MAKE} SRCLANGS="$$l" TRGLANGS="${PIVOT}" data; \
${MAKE} SRCLANGS="$$l" TRGLANGS="${PIVOT}" HPC_CORES=1 HPC_MEM=4g train.submit-multigpu; \
${MAKE} SRCLANGS="$$l" TRGLANGS="${PIVOT}" reverse-data; \
${MAKE} SRCLANGS="${PIVOT}" TRGLANGS="$$l" HPC_CORES=1 HPC_MEM=4g train.submit-multigpu; \
done
## submit train jobs with settings that depend on the size of the training data
## --> change WORKSPACE, MEM, nr of GPUs, validation frequency, stopping criterion
train-dynamic:
if [ ! -e "${WORKHOME}/${LANGPAIRSTR}/train.submit" ]; then \
${MAKE} data; \
s=`${ZCAT} ${TRAIN_SRC}.clean.${PRE_SRC}${TRAINSIZE}.gz | head -10000001 | wc -l`; \
if [ $$s -gt 10000000 ]; then \
echo "${LANGPAIRSTR} bigger than 10 million"; \
${MAKE} HPC_CORES=1 HPC_MEM=8g train.submit-multigpu; \
elif [ $$s -gt 1000000 ]; then \
echo "${LANGPAIRSTR} bigger than 1 million"; \
${MAKE} \
MARIAN_VALID_FREQ=2500 \
HPC_CORES=1 HPC_MEM=8g train.submit; \
elif [ $$s -gt 100000 ]; then \
echo "${LANGPAIRSTR} bigger than 100k"; \
${MAKE} \
MARIAN_VALID_FREQ=1000 \
MARIAN_WORKSPACE=5000 \
MARIAN_VALID_MINI_BATCH=8 \
HPC_CORES=1 HPC_MEM=4g train.submit; \
elif [ $$s -gt 10000 ]; then \
echo "${LANGPAIRSTR} bigger than 10k"; \
${MAKE} \
MARIAN_WORKSPACE=3500 \
MARIAN_VALID_MINI_BATCH=4 \
MARIAN_DROPOUT=0.5 \
MARIAN_VALID_FREQ=1000 \
HPC_CORES=1 HPC_MEM=4g train.submit; \
else \
echo "${LANGPAIRSTR} too small"; \
fi \
fi
# MARIAN_EARLY_STOPPING=5 \
bilingual-dynamic: train-dynamic
if [ "${SRCLANGS}" != "${TRGLANGS}" ]; then \
${MAKE} reverse-data; \
${MAKE} TRGLANGS="${SRCLANGS}" SRCLANGS='${TRGLANGS}' train-dynamic; \
fi
test-skip:
for s in ${SRCLANGS}; do \
for t in ${TRGLANGS}; do \
if [ `echo "$$s-$$t $$t-$$s" | egrep '${SKIP_LANGPAIRS}' | wc -l` -gt 0 ]; then \
echo "skip $$s-$$t"; \
fi \
done \
done
## extension -all-langpairs: run something over all language pairs, e.g.
## make wordalign-all-langpairs
## this goes sequentially over all language pairs
## for the parallelizable version of this: look at %-all-parallel
%-all-langpairs:
for l in ${ALL_LANG_PAIRS}; do \
${MAKE} SRCLANGS="`echo $$l | cut -f1 -d'-' | sed 's/\\+/ /g'`" \
TRGLANGS="`echo $$l | cut -f2 -d'-' | sed 's/\\+/ /g'`" ${@:-all-langpairs=}; \
done
# run something over all language pairs that have trained models
## - make eval-allmodels
## - make dist-allmodels
%-allmodels:
for l in ${ALL_LANG_PAIRS}; do \
m=`find ${WORKHOME}/$$l -maxdepth 1 -name '*.best-perplexity.npz' -printf "%f\n"`; \
for i in $$m; do \
s=`echo $$l | cut -f1 -d'-' | sed 's/\\+/ /g'`; \
t=`echo $$l | cut -f2 -d'-' | sed 's/\\+/ /g'`; \
d=`echo $$i | cut -f1 -d.`; \
x=`echo $$i | cut -f2 -d. | cut -f1 -d-`; \
y=`echo $$i | cut -f2 -d. | cut -f2 -d-`; \
v=`echo $$i | cut -f3 -d.`; \
echo "model = $$i"; \
echo "dataset = $$d"; \
echo "src-lang = $$s"; \
echo "trg-lang = $$t"; \
echo "pre-src = $$x"; \
echo "pre-trg = $$y"; \
echo "type = $$v"; \
${MAKE} \
SRCLANGS="$$s" TRGLANGS="$$t" \
DATASET=$$d \
PRE_SRC=$$x PRE_TRG=$$y \
MODELTYPE=$$v ${@:-allmodels=}; \
done \
done
## OLD: doesn't work for different model variants
##
# %-allmodels:
# for l in ${ALL_LANG_PAIRS}; do \
# if [ `find ${WORKHOME}/$$l -name '*.${PRE_SRC}-${PRE_TRG}.*.best-perplexity.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'`" ${@:-allmodels=}; \
# fi \
# done
listallmodels:
@m=`find ${WORKDIR} -maxdepth 1 -name '*.best-perplexity.npz' -printf "%f\n"`; \
for i in $$m; do \
d=`echo $$i | cut -f1 -d.`; \
s=`echo $$i | cut -f2 -d. | cut -f1 -d-`; \
t=`echo $$i | cut -f2 -d. | cut -f1 -d-`; \
v=`echo $$i | cut -f3 -d.`; \
echo "model = $$i"; \
echo "dataset = $$d"; \
echo "pre-src = $$s"; \
echo "pre-trg = $$t"; \
echo "type = $$v"; \
done
## only bilingual models
%-allbilingual:
for l in ${ALL_BILINGUAL_MODELS}; do \
if [ `find ${WORKHOME}/$$l -name '*.${PRE_SRC}-${PRE_TRG}.*.best-perplexity.npz' | wc -l` -gt 0 ]; then \
${MAKE} SRCLANGS="`echo $$l | cut -f1 -d'-' | sed 's/\\+/ /g'`" \
TRGLANGS="`echo $$l | cut -f2 -d'-' | sed 's/\\+/ /g'`" ${@:-allbilingual=}; \
fi \
done
## only bilingual models
%-allmultilingual:
for l in ${ALL_MULTILINGUAL_MODELS}; do \
if [ `find ${WORKHOME}/$$l -name '*.${PRE_SRC}-${PRE_TRG}.*.best-perplexity.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'`" ${@:-allmultilingual=}; \
fi \
done
## run something over all language pairs but make it possible to do it in parallel, for example
## - make dist-all-parallel
%-all-parallel:
${MAKE} $(subst -all-parallel,,${patsubst %,$@__%-run-for-langpair,${ALL_LANG_PAIRS}})
## run a command that includes the langpair, for example
## make wordalign__en-da+sv-run-for-langpair ...... runs wordalign with SRCLANGS="en" TRGLANGS="da sv"
## What is this good for?
## ---> can run many lang-pairs in parallel instead of having a for loop and run sequencetially
%-run-for-langpair:
${MAKE} SRCLANGS='$(subst +, ,$(firstword $(subst -, ,${lastword ${subst __, ,${@:-run-for-langpair=}}})))' \
TRGLANGS='$(subst +, ,$(lastword $(subst -, ,${lastword ${subst __, ,${@:-run-for-langpair=}}})))' \
${shell echo $@ | sed 's/__.*$$//'}
## right-to-left model
%-RL:
${MAKE} MODEL=${MODEL}-RL \
MARIAN_EXTRA="${MARIAN_EXTRA} --right-left" \
${@:-RL=}
## include all backtranslation data as well in training
%-bt:
${MAKE} DATASET=${DATASET}+bt USE_BACKTRANS=1 SHUFFLE_TRAINING_DATA=1 ${@:-bt=}
## adding a pivot language to the model
## --> add pivot language to each side (source and target)
## --> only start the task if the pivot language adds anything on either side
## --> make a new fresh vocabulary (don't copy an existing one)
## --> make a new BPE/sentencepiece model
## --> make a new config file
DEFAULT_PIVOT_LANG ?= en
PIVOT_LANG ?= ${DEFAULT_PIVOT_LANG}
%-pivotlang:
if [ "$(sort ${SRCLANGS} ${TRGLANGS} ${PIVOT_LANG})" != "$(sort ${SRCLANGS} ${TRGLANGS})" ]; then \
${MAKE} DATASET=${DATASET}+${PIVOT_LANG} \
SRCLANGS="$(sort ${SRCLANGS} ${PIVOT_LANG})" \
TRGLANGS="$(sort ${TRGLANGS} ${PIVOT_LANG})" \
SKIP_LANGPAIRS=${PIVOT_LANG}-${PIVOT_LANG} \
MODEL_LATEST_VOCAB= \
PIVOT=${PIVOT_LANG} \
SUBWORD_MODEL_NAME=opus+${PIVOT_LANG} \
${@:-pivotlang=}; \
fi
## add forward translations
%-ft:
${MAKE} DATASET=${DATASET}+ft USE_FORWARDTRANS=1 SHUFFLE_TRAINING_DATA=1 ${@:-ft=}
# use a selected set of forward translation
# TODO: only call the translation extraction recipe if there is no file that matches
# the extension of that kind of extraction type (e.g. *.best95.gz)
# --> this can go wrong
# default for ce-filter
FT_SELECTED ?= 95
%-ftbest:
@for s in ${SRCLANGS}; do \
for t in ${TRGLANGS}; do \
if [ -e ${FORWARDTRANS_HOME}/$$s-$$t/latest ]; then \
if [ ! -e `ls ${FORWARDTRANS_HOME}/$$s-$$t/latest/*.best${FT_SELECTED}.gz | head -1` ]; then \
${MAKE} -C ${FORWARDTRANS_HOME} SRC=$$s TRG=$$t \
RETAIN=${FT_SELECTED} extract-best-translations; \
fi \
fi \
done \
done
${MAKE} DATASET=${DATASET}+ft${FT_SELECTED} \
USE_FORWARDTRANS_SELECTED=${FT_SELECTED} \
SHUFFLE_TRAINING_DATA=1 \
${@:-ftbest=}
%-ftrawbest:
@for s in ${SRCLANGS}; do \
for t in ${TRGLANGS}; do \
if [ -e ${FORWARDTRANS_HOME}/$$s-$$t/latest ]; then \
${MAKE} -C ${FORWARDTRANS_HOME} SRC=$$s TRG=$$t \
RETAIN=${FT_SELECTED} extract-rawbest-translations; \
fi \
done \
done
${MAKE} DATASET=${DATASET}+ftraw${FT_SELECTED} \
USE_FORWARDTRANS_SELECTED_RAW=${FT_SELECTED} \
SHUFFLE_TRAINING_DATA=1 \
${@:-ftrawbest=}
## add forward translation of monolingual data
%-ftmono:
${MAKE} DATASET=${DATASET}+ftmono USE_FORWARDTRANSMONO=1 SHUFFLE_TRAINING_DATA=1 ${@:-ftmono=}
## pivot-based backward translation
%-pbt:
${MAKE} DATASET=${DATASET}+pbt USE_BACKWARD_PIVOTING=1 SHUFFLE_TRAINING_DATA=1 ${@:-pbt=}
## pivot-based backward translation
%-pft:
${MAKE} DATASET=${DATASET}+pft USE_FORWARD_PIVOTING=1 SHUFFLE_TRAINING_DATA=1 ${@:-pft=}
## pivot-based backward translation
%-pivot:
${MAKE} DATASET=${DATASET}+pivot USE_PIVOTING=1 ${@:-pivot=}
## don't use the regular parallel training data
## (only makes sense if bt, ft, or pivot-based data are activated)
%-nopar:
${MAKE} DATASET=${DATASET}+nopar TRAINSET= ${@:-nopar=}
# TATOEBA_TRAINSET=
##-------------------------------------------------------------
## default: make separate sentencepiece models
## but create a joint vocabulary for training with
## tied embeddings
## alternatives:
## * joint-spm: create a joint sentencepiece model
## * separate-spm: create separate SPMs and use them directly
##-------------------------------------------------------------
## joint sentencepiece model and joint vocabulary
%-joint-spm:
${MAKE} USE_JOINT_SUBWORD_MODEL=1 USE_SPM_VOCAB=1 MODEL_VARIANT=-jointvoc ${@:-joint-spm=}
## use sentenceopiece models directly as separate vocabularies
%-separate-spm:
${MAKE} USE_SPM_VOCAB=1 MODEL_VARIANT=-sepvoc ${@:-separate-spm=}
## for compatibility reasons: train on forward translated data and no regular train data
%-ftonly:
${MAKE} DATASET=${DATASET}+ftonly \
USE_FORWARDTRANS=1 \
CONTINUE_EXISTING=1 \
MARIAN_EARLY_STOPPING=${FT_MARIAN_EARLY_STOPPING} \
TRAINSET= \
${@:-ftonly=}
# TATOEBA_TRAINSET= \
# MODELCONFIG=config-ft.mk \
%-big-align:
${MAKE} MODELTYPE=transformer-big-align ${@:-big-align=}
## run a multigpu job (2 or 4 GPUs)
%-multigpu %-gpu0123:
${MAKE} NR_GPUS=4 MARIAN_GPUS='0 1 2 3' $(subst -gpu0123,,${@:-multigpu=})
%-twogpu %-gpu01:
${MAKE} NR_GPUS=2 MARIAN_GPUS='0 1' $(subst -gpu01,,${@:-twogpu=})
%-gpu23:
${MAKE} NR_GPUS=2 MARIAN_GPUS='2 3' ${@:-gpu23=}
## document level models
## devtest data should not be shuffled
%-doc:
${MAKE} PRE_SRC=${SUBWORDS}${SUBWORD_SRCVOCAB_SIZE:000=}k.doc${CONTEXT_SIZE} \
PRE_TRG=${SUBWORDS}${SUBWORD_TRGVOCAB_SIZE:000=}k.doc${CONTEXT_SIZE} \
SHUFFLE_DEVDATA=0 \
${@:-doc=}
## sentence-piece models
%-spm:
${MAKE} WORKHOME=${shell realpath ${PWD}/work-spm} \
PRE=simple SUBWORDS=spm \
${@:-spm=}
## sentence-piece models with space-separated strings
%-nospace:
${MAKE} WORKHOME=${shell realpath ${PWD}/work-nospace} \
PRE=simple \
SPMEXTRA=--split_by_whitespace=false \
${@:-nospace=}
## with SPM models trained on monolingual data
%-monospm: ${SPMSRCMONO} ${SPMTRGMONO}
${MAKE} WORKHOME=${shell realpath ${PWD}/work-monospm} \
SPMSRCMODEL=${SPMSRCMONO} \
SPMTRGMODEL=${SPMTRGMONO} \
${@:-monospm=}
## BPE models
%-bpe:
${MAKE} WORKHOME=${shell realpath ${PWD}/work-bpe} \
PRE=tok SUBWORDS=bpe \
MODELTYPE=transformer \
${@:-bpe=}
# ## sentence-piece models with langid-filtering (new default)
# %-langid:
# ${MAKE} WORKHOME=${shell realpath ${PWD}/work-langid} \
# PRE=simple \
# ${@:-langid=}
# %-bpe-memad:
# ${MAKE} WORKHOME=${shell realpath ${PWD}/work-bpe-memad} \
# PRE=tok SUBWORDS=bpe \
# MODELTYPE=transformer \
# ${@:-bpe-memad=}
## NEW: don't continue from existing models when including pivot data
## TODO: should we do that?
#
# PIVOT_MODEL = ${MODEL_SUBDIR}${DATASET}+pivot${TRAINSIZE}.${PRE_SRC}-${PRE_TRG}
# PIVOT_MODEL_BASE = ${PIVOT_MODEL}.${MODELTYPE}.model${NR}
# PIVOT_MODEL_START = ${WORKDIR}/${PIVOT_MODEL_BASE}.npz
# PIVOT_MODEL_VOCAB = ${WORKDIR}/${PIVOT_MODEL}.vocab.yml
# %-pivot:
# ifneq (${wildcard ${MODEL_FINAL}},)
# ifeq (${wildcard ${PIVOT_MODEL_START}},)
# cp ${MODEL_FINAL} ${PIVOT_MODEL_START}
# cp ${MODEL_VOCAB} ${PIVOT_MODEL_VOCAB}
# endif
# endif
# rm -f ${WORKHOME}/${LANGPAIRSTR}/train.submit
# ${MAKE} DATASET=${DATASET}+pivot \
# USE_PIVOTING=1 \
# CONTINUE_EXISTING=1 \
# MARIAN_EARLY_STOPPING=10 \
# ${@:-pivot=}
# OLD: start with parameters from the standard transformer model
# NOW: big-align is not compatible with the dimensions settings anymore
#
# %-big-align:
# ifneq (${wildcard ${MODEL_FINAL}},)
# ${MAKE} PRE_TRAINED_MODEL=${MODEL_FINAL} MODELTYPE=transformer-big-align ${@:-big-align=}
# else ifneq (${wildcard ${MODEL_START}},)
# ${MAKE} PRE_TRAINED_MODEL=${MODEL_START} MODELTYPE=transformer-big-align ${@:-big-align=}
# else
# ${MAKE} MODELTYPE=transformer-big-align ${@:-big-align=}
# endif