mirror of
https://github.com/Helsinki-NLP/OPUS-MT-train.git
synced 2024-09-11 20:27:19 +03:00
285 lines
7.6 KiB
Makefile
285 lines
7.6 KiB
Makefile
# -*-makefile-*-
|
|
#
|
|
# model configurations
|
|
#
|
|
|
|
|
|
# SRCLANGS = da no sv
|
|
# TRGLANGS = fi
|
|
|
|
SRCLANGS = sv
|
|
TRGLANGS = fi
|
|
|
|
ifndef SRC
|
|
SRC := ${firstword ${SRCLANGS}}
|
|
endif
|
|
ifndef TRG
|
|
TRG := ${lastword ${TRGLANGS}}
|
|
endif
|
|
|
|
|
|
# sorted languages and langpair used to match resources in OPUS
|
|
SORTLANGS = $(sort ${SRC} ${TRG})
|
|
SPACE = $(empty) $(empty)
|
|
LANGPAIR = ${firstword ${SORTLANGS}}-${lastword ${SORTLANGS}}
|
|
LANGSTR = ${subst ${SPACE},+,$(SRCLANGS)}-${subst ${SPACE},+,$(TRGLANGS)}
|
|
|
|
|
|
## for same language pairs: add numeric extension
|
|
ifeq (${SRC},$(TRG))
|
|
SRCEXT = ${SRC}1
|
|
TRGEXT = ${SRC}2
|
|
else
|
|
SRCEXT = ${SRC}
|
|
TRGEXT = ${TRG}
|
|
endif
|
|
|
|
|
|
## all of OPUS (NEW: don't require MOSES format)
|
|
# OPUSCORPORA = ${patsubst %/latest/moses/${LANGPAIR}.txt.zip,%,\
|
|
# ${patsubst ${OPUSHOME}/%,%,\
|
|
# ${shell ls ${OPUSHOME}/*/latest/moses/${LANGPAIR}.txt.zip}}}
|
|
OPUSCORPORA = ${patsubst %/latest/xml/${LANGPAIR}.xml.gz,%,\
|
|
${patsubst ${OPUSHOME}/%,%,\
|
|
${shell ls ${OPUSHOME}/*/latest/xml/${LANGPAIR}.xml.gz}}}
|
|
|
|
|
|
ALL_LANG_PAIRS = ${shell ls ${WORKHOME} | grep -- '-' | grep -v old}
|
|
ALL_BILINGUAL_MODELS = ${shell echo '${ALL_LANG_PAIRS}' | tr ' ' "\n" | grep -v -- '\+'}
|
|
ALL_MULTILINGUAL_MODELS = ${shell echo '${ALL_LANG_PAIRS}' | tr ' ' "\n" | grep -- '\+'}
|
|
|
|
# ALL_BILINGUAL_MODELS = ${shell ls ${WORKHOME} | grep -- '-' | grep -v old | grep -v -- '\+'}
|
|
# ALL_MULTILINGUAL_MODELS = ${shell ls ${WORKHOME} | grep -- '-' | grep -v old | grep -- '\+'}
|
|
|
|
|
|
## size of dev data, test data and BPE merge operations
|
|
|
|
DEVSIZE = 5000
|
|
TESTSIZE = 5000
|
|
|
|
## NEW: significantly reduce devminsize
|
|
## (= absolute minimum we need as devdata)
|
|
## NEW: define an alternative small size for DEV and TEST
|
|
## OLD DEVMINSIZE:
|
|
# DEVMINSIZE = 1000
|
|
|
|
DEVSMALLSIZE = 1000
|
|
TESTSMALLSIZE = 1000
|
|
DEVMINSIZE = 250
|
|
|
|
## size of heldout data for each sub-corpus
|
|
## (only if there is at least twice as many examples in the corpus)
|
|
HELDOUTSIZE = ${DEVSIZE}
|
|
|
|
##----------------------------------------------------------------------------
|
|
## train/dev/test data
|
|
##----------------------------------------------------------------------------
|
|
|
|
## dev/test data: default = Tatoeba otherwise, GlobalVoices, JW300, GNOME or bibl-uedin
|
|
## - check that data exist
|
|
## - check that there are at least 2 x DEVMINSIZE examples
|
|
## TODO: this does not work well for multilingual models!
|
|
|
|
ifneq ($(wildcard ${OPUSHOME}/Tatoeba/latest/moses/${LANGPAIR}.txt.zip),)
|
|
ifeq ($(shell if (( `head -1 ${OPUSHOME}/Tatoeba/latest/info/${LANGPAIR}.txt.info` \
|
|
> $$((${DEVMINSIZE} + ${DEVMINSIZE})) )); then echo "ok"; fi),ok)
|
|
DEVSET = Tatoeba
|
|
endif
|
|
endif
|
|
|
|
## backoff to GlobalVoices
|
|
ifndef DEVSET
|
|
ifneq ($(wildcard ${OPUSHOME}/GlobalVoices/latest/moses/${LANGPAIR}.txt.zip),)
|
|
ifeq ($(shell if (( `head -1 ${OPUSHOME}/GlobalVoices/latest/info/${LANGPAIR}.txt.info` \
|
|
> $$((${DEVMINSIZE} + ${DEVMINSIZE})) )); then echo "ok"; fi),ok)
|
|
DEVSET = GlobalVoices
|
|
endif
|
|
endif
|
|
endif
|
|
|
|
## backoff to JW300
|
|
ifndef DEVSET
|
|
ifneq ($(wildcard ${OPUSHOME}/JW300/latest/xml/${LANGPAIR}.xml.gz),)
|
|
ifeq ($(shell if (( `sed -n 2p ${OPUSHOME}/JW300/latest/info/${LANGPAIR}.info` \
|
|
> $$((${DEVMINSIZE} + ${DEVMINSIZE})) )); then echo "ok"; fi),ok)
|
|
DEVSET = JW300
|
|
endif
|
|
endif
|
|
endif
|
|
|
|
## otherwise: bible-uedin
|
|
ifndef DEVSET
|
|
DEVSET = bible-uedin
|
|
endif
|
|
|
|
## in case we want to use some additional data sets
|
|
EXTRA_TRAINSET =
|
|
|
|
## TESTSET= DEVSET, TRAINSET = OPUS - WMT-News,DEVSET.TESTSET
|
|
TESTSET = ${DEVSET}
|
|
TRAINSET = $(filter-out WMT-News ${DEVSET} ${TESTSET},${OPUSCORPORA} ${EXTRA_TRAINSET})
|
|
TUNESET = OpenSubtitles
|
|
|
|
|
|
## 1 = use remaining data from dev/test data for training
|
|
USE_REST_DEVDATA = 1
|
|
|
|
|
|
##----------------------------------------------------------------------------
|
|
## pre-processing and vocabulary
|
|
##----------------------------------------------------------------------------
|
|
|
|
BPESIZE = 32000
|
|
SRCBPESIZE = ${BPESIZE}
|
|
TRGBPESIZE = ${BPESIZE}
|
|
|
|
ifndef VOCABSIZE
|
|
VOCABSIZE = $$((${SRCBPESIZE} + ${TRGBPESIZE} + 1000))
|
|
endif
|
|
|
|
## for document-level models
|
|
CONTEXT_SIZE = 100
|
|
|
|
## pre-processing type
|
|
PRE = norm
|
|
PRE_SRC = spm${SRCBPESIZE:000=}k
|
|
PRE_TRG = spm${TRGBPESIZE:000=}k
|
|
|
|
|
|
##-------------------------------------
|
|
## default name of the data set (and the model)
|
|
##-------------------------------------
|
|
|
|
ifndef DATASET
|
|
DATASET = opus
|
|
endif
|
|
|
|
##-------------------------------------
|
|
## OLD OLD OLD
|
|
## name of the data set (and the model)
|
|
## - single corpus = use that name
|
|
## - multiple corpora = opus
|
|
## add also vocab size to the name
|
|
##-------------------------------------
|
|
|
|
# ifndef DATASET
|
|
# ifeq (${words ${TRAINSET}},1)
|
|
# DATASET = ${TRAINSET}
|
|
# else
|
|
# DATASET = opus
|
|
# endif
|
|
# endif
|
|
|
|
|
|
|
|
## DATADIR = directory where the train/dev/test data are
|
|
## WORKDIR = directory used for training
|
|
|
|
DATADIR = ${WORKHOME}/data
|
|
WORKDIR = ${WORKHOME}/${LANGSTR}
|
|
|
|
|
|
## data sets
|
|
TRAIN_BASE = ${WORKDIR}/train/${DATASET}
|
|
TRAIN_SRC = ${TRAIN_BASE}.src
|
|
TRAIN_TRG = ${TRAIN_BASE}.trg
|
|
TRAIN_ALG = ${TRAIN_BASE}${TRAINSIZE}.${PRE_SRC}-${PRE_TRG}.src-trg.alg.gz
|
|
|
|
## training data in local space
|
|
LOCAL_TRAIN_SRC = ${TMPDIR}/${LANGSTR}/train/${DATASET}.src
|
|
LOCAL_TRAIN_TRG = ${TMPDIR}/${LANGSTR}/train/${DATASET}.trg
|
|
|
|
TUNE_SRC = ${WORKDIR}/tune/${TUNESET}.src
|
|
TUNE_TRG = ${WORKDIR}/tune/${TUNESET}.trg
|
|
|
|
DEV_SRC = ${WORKDIR}/val/${DEVSET}.src
|
|
DEV_TRG = ${WORKDIR}/val/${DEVSET}.trg
|
|
|
|
TEST_SRC = ${WORKDIR}/test/${TESTSET}.src
|
|
TEST_TRG = ${WORKDIR}/test/${TESTSET}.trg
|
|
|
|
|
|
## heldout data directory (keep one set per data set)
|
|
HELDOUT_DIR = ${WORKDIR}/heldout
|
|
|
|
MODEL_SUBDIR =
|
|
MODEL = ${MODEL_SUBDIR}${DATASET}${TRAINSIZE}.${PRE_SRC}-${PRE_TRG}
|
|
MODELTYPE = transformer-align
|
|
NR = 1
|
|
|
|
MODEL_BASENAME = ${MODEL}.${MODELTYPE}.model${NR}
|
|
MODEL_VALIDLOG = ${MODEL}.${MODELTYPE}.valid${NR}.log
|
|
MODEL_TRAINLOG = ${MODEL}.${MODELTYPE}.train${NR}.log
|
|
MODEL_FINAL = ${WORKDIR}/${MODEL_BASENAME}.npz.best-perplexity.npz
|
|
MODEL_VOCABTYPE = yml
|
|
MODEL_VOCAB = ${WORKDIR}/${MODEL}.vocab.${MODEL_VOCABTYPE}
|
|
MODEL_DECODER = ${MODEL_FINAL}.decoder.yml
|
|
|
|
|
|
## test set translation and scores
|
|
|
|
TEST_TRANSLATION = ${WORKDIR}/${TESTSET}.${MODEL}${NR}.${MODELTYPE}.${SRC}.${TRG}
|
|
TEST_EVALUATION = ${TEST_TRANSLATION}.eval
|
|
TEST_COMPARISON = ${TEST_TRANSLATION}.compare
|
|
|
|
|
|
|
|
## parameters for running Marian NMT
|
|
|
|
MARIAN_GPUS = 0
|
|
MARIAN_EXTRA =
|
|
MARIAN_VALID_FREQ = 10000
|
|
MARIAN_SAVE_FREQ = ${MARIAN_VALID_FREQ}
|
|
MARIAN_DISP_FREQ = ${MARIAN_VALID_FREQ}
|
|
MARIAN_EARLY_STOPPING = 10
|
|
MARIAN_VALID_MINI_BATCH = 16
|
|
MARIAN_MAXI_BATCH = 500
|
|
MARIAN_DROPOUT = 0.1
|
|
|
|
MARIAN_DECODER_GPU = -b 12 -n1 -d ${MARIAN_GPUS} --mini-batch 8 --maxi-batch 32 --maxi-batch-sort src
|
|
MARIAN_DECODER_CPU = -b 12 -n1 --cpu-threads ${HPC_CORES} --mini-batch 8 --maxi-batch 32 --maxi-batch-sort src
|
|
MARIAN_DECODER_FLAGS = ${MARIAN_DECODER_GPU}
|
|
|
|
## TODO: currently marianNMT crashes with workspace > 26000
|
|
ifeq (${GPU},p100)
|
|
MARIAN_WORKSPACE = 13000
|
|
else ifeq (${GPU},v100)
|
|
# MARIAN_WORKSPACE = 30000
|
|
# MARIAN_WORKSPACE = 26000
|
|
# MARIAN_WORKSPACE = 24000
|
|
# MARIAN_WORKSPACE = 18000
|
|
MARIAN_WORKSPACE = 16000
|
|
else
|
|
MARIAN_WORKSPACE = 10000
|
|
endif
|
|
|
|
|
|
ifeq (${shell nvidia-smi | grep failed | wc -l},1)
|
|
MARIAN = ${MARIANCPU}
|
|
MARIAN_DECODER_FLAGS = ${MARIAN_DECODER_CPU}
|
|
MARIAN_EXTRA = --cpu-threads ${HPC_CORES}
|
|
endif
|
|
|
|
|
|
|
|
|
|
ifneq ("$(wildcard ${TRAIN_WEIGHTS})","")
|
|
MARIAN_TRAIN_WEIGHTS = --data-weighting ${TRAIN_WEIGHTS}
|
|
endif
|
|
|
|
|
|
### training a model with Marian NMT
|
|
##
|
|
## NR allows to train several models for proper ensembling
|
|
## (with shared vocab)
|
|
##
|
|
## DANGER: if several models are started at the same time
|
|
## then there is some racing issue with creating the vocab!
|
|
|
|
ifdef NR
|
|
SEED=${NR}${NR}${NR}${NR}
|
|
else
|
|
SEED=1234
|
|
endif
|
|
|