-
Notifications
You must be signed in to change notification settings - Fork 39
/
Makefile
347 lines (265 loc) · 11 KB
/
Makefile
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
#
# pivoting - translate training data via pivot models
#
#
# TODO: exclude certain corpora like GNOME Ubuntu, (bible-uedin) ...
# ---> this will basically be the same for all languages
#
PWD := ${shell pwd}
REPOHOME := ${PWD}/../
TOOLSDIR := ${REPOHOME}tools
## language (SRC->TRG) pair we need
SRC = fi
TRG = se
## pivot language
PIVOT = nb
## always include those data sets even if they have the same target language pair
INCLUDE = OpenSubtitles
## specify additional data sets to exclude
# EXCLUDE = bible-uedin DGT GlobalVoices GNOME infopankki KDE4 KDEdoc Tanzil Ubuntu
## langpair (sorted lang id's) of the original data
## to be translated from PIVOT to SRC
ORIGINAL_LANGPAIR = ${firstword ${sort ${PIVOT} ${TRG}}}-${lastword ${sort ${PIVOT} ${TRG}}}
PIVOT_LANGPAIR = ${PIVOT}-${SRC}
include ${REPOHOME}lib/env.mk
include ${REPOHOME}lib/config.mk
include ${REPOHOME}lib/slurm.mk
include ${REPOHOME}lib/dist.mk
include lib/models.mk
## max number of sentences to translate
## TODO: should also support splitting and translating subsets in parallel
MAX_PIVOT_SENTENCES ?= 1000000
NMT_WORKDIR ?= ../work
ORIGINAL_DATADIR ?= ${PWD}/${NMT_WORKDIR}/data
EXCLUDE_DATASETS = ${sort \
${filter-out \
${patsubst %,${ORIGINAL_DATADIR}/${PRE}/%.${ORIGINAL_LANGPAIR}.clean.${TRG}.gz,${INCLUDE}},\
${patsubst %.${LANGPAIR}.clean.${TRG}.gz,%.${ORIGINAL_LANGPAIR}.clean.${TRG}.gz,\
${wildcard ${ORIGINAL_DATADIR}/${PRE}/*.${LANGPAIR}.clean.${TRG}.gz}}} \
${patsubst %,${ORIGINAL_DATADIR}/${PRE}/%.${ORIGINAL_LANGPAIR}.clean.${TRG}.gz,${EXCLUDE}}}
ORIGINAL_DATASETS_TRG = ${filter-out ${EXCLUDE_DATASETS},\
${wildcard ${ORIGINAL_DATADIR}/${PRE}/*.${ORIGINAL_LANGPAIR}.clean.${TRG}.gz}}
ORIGINAL_DATASETS_SRC = ${patsubst %.${TRG}.gz,%.${PIVOT}.gz,${ORIGINAL_DATASETS_TRG}}
ORIGINAL_DATASRC ?= ${firstword ${ORIGINAL_DATASETS_SRC}}
ORIGINAL_DATATRG ?= ${firstword ${ORIGINAL_DATASETS_TRG}}
DATASET_NAMES = $(patsubst %.clean,%,$(patsubst %.${PIVOT}.gz,%,${notdir ${ORIGINAL_DATASETS_SRC}}))
DATASET_NAME = $(patsubst %.clean,%,$(patsubst %.${PIVOT}.gz,%,${notdir ${ORIGINAL_DATASRC}}))
## output dir
OUTPUT_DIR = ${SRC}-${TRG}
PIVOT_MODEL_ZIP = ${firstword ${call get-model-distro,../models,${PIVOT_LANGPAIR}}}
ifeq (${PIVOT_MODEL_ZIP},)
PIVOT_MODEL_ZIP = ${firstword ${call get-model-distro,../models,.*${PIVOT}.*-${SRC}}}
endif
ifeq (${PIVOT_MODEL_ZIP},)
PIVOT_MODEL_ZIP = ${firstword ${call get-model-distro,../models,${PIVOT}-.*${SRC}.*}}
endif
ifeq (${PIVOT_MODEL_ZIP},)
PIVOT_MODEL_ZIP = ${firstword ${call get-model-distro,../models,.*${PIVOT}.*-.*${SRC}.*}}
endif
PIVOT_MODEL_DIR ?= ${patsubst %/,%,${dir ${PIVOT_MODEL_ZIP}}}
PIVOT_MODEL_NAME ?= ${patsubst %.zip,%,${notdir ${PIVOT_MODEL_ZIP}}}
## set to 1 if the model for backtranslation is a multi-target model
## --> need to use pre-processing script differently
ifeq ($(words $(subst +, ,$(lastword $(subst -, ,$(notdir ${PIVOT_MODEL_DIR}))))),1)
MULTI_TARGET_MODEL = 0
else
MULTI_TARGET_MODEL = 1
endif
test-model:
@echo ${call get-model-distro,../models,fr-en}
@echo ${call get-model-release,fr-en}
# #--------------------------------------------------------------------------
# # find a model to translate the original data
# #--------------------------------------------------------------------------
# ## if there is no model for the pivot translation pair
# ## ---> look for a multilingual model that includes both languages
# ## TODO: make sure that PIVOT and SRC do not match other lang-IDs!!!
# ## TODO: what do we do if there is more than one multilingual model?
# ## --> need to define preference mechanism
# ## TODO: this should better come from some API or at least ObjectStorage
# ## (not local disk)
# MODELSDIR ?= ../models
# ifeq ($(wildcard ${MODELSDIR}/${PIVOT_LANGPAIR}),)
# ifeq ($(wildcard ${MODELSDIR}/*${PIVOT}*-${SRC}),)
# ifeq ($(wildcard ${MODELSDIR}/${PIVOT}-*${SRC}*),)
# MODELHOME = $(firstword $(wildcard ${MODELSDIR}/*${PIVOT}*-*${SRC}*))
# else
# MODELHOME = $(firstword $(wildcard ${MODELSDIR}/${PIVOT}-*${SRC}*))
# endif
# else
# MODELHOME = $(firstword $(wildcard ${MODELSDIR}/*${PIVOT}*-${SRC}))
# endif
# else
# MODELHOME = ${MODELSDIR}/${PIVOT_LANGPAIR}
# endif
# ## select the latest NMT model (assume this is the best one)
# ## standard sort is different from UTF8-based sort
# ## --> prefer models with augmented data sets (separated by +)
# ## we need the UTF8 sort order
# ## --> use bash sort and UTF8 locale
# # MODELZIP = ${lastword ${sort ${wildcard ${MODELHOME}/*-20*.zip}}}
# MODELZIP = ${lastword ${shell ls ${MODELHOME}/*-20*.zip | LANG=en_US.UTF-8 sort}}
# MODELNAME = ${patsubst %.zip,%,${notdir ${MODELZIP}}}
# ## no released model found?
# ## ---> find a local one in the work directory
# ifeq (${MODELNAME},)
# ifeq ($(wildcard ${NMT_WORKDIR}/models/${PIVOT_LANGPAIR}),)
# ifeq ($(wildcard ${NMT_WORKDIR}/models/*${PIVOT}*-${SRC}),)
# ifeq ($(wildcard ${NMT_WORKDIR}/models/${PIVOT}-*${SRC}*),)
# MODELHOME = $(firstword $(wildcard ${NMT_WORKDIR}/models/*${PIVOT}*-*${SRC}*))
# else
# MODELHOME = $(firstword $(wildcard ${NMT_WORKDIR}/models/${PIVOT}-*${SRC}*))
# endif
# else
# MODELHOME = $(firstword $(wildcard ${NMT_WORKDIR}/models/*${PIVOT}*-${SRC}))
# endif
# else
# MODELHOME = ${NMT_WORKDIR}/models/${PIVOT_LANGPAIR}
# endif
# # MODELZIP = ${lastword ${sort ${wildcard ${MODELHOME}/*-20*.zip}}}
# MODELZIP = ${lastword ${sort ${wildcard ${MODELHOME}/opus-20*.zip}}}
# MODELNAME = ${patsubst %.zip,%,${notdir ${MODELZIP}}}
# endif
# ## set to 1 if the model for backtranslation is a multi-target model
# ## --> need to use pre-processing script differently
# ifeq ($(words $(subst +, ,$(lastword $(subst -, ,$(notdir ${MODELHOME}))))),1)
# MULTI_TARGET_MODEL = 0
# else
# MULTI_TARGET_MODEL = 1
# endif
## decoder config
DECODER = ${OUTPUT_DIR}/${PIVOT_MODEL_NAME}/decoder.yml
ifdef LOCAL_SCRATCH
TMPDIR = ${LOCAL_SCRATCH}
endif
#--------------------------------------------------------------------------
# target files to be created
#--------------------------------------------------------------------------
## just one data set
TRANSLATED_SRC = ${OUTPUT_DIR}/${DATASET_NAME}.${PIVOT_MODEL_NAME}.${LANGPAIR}.${PIVOT}.gz
TRANSLATED_PRE = ${OUTPUT_DIR}/${DATASET_NAME}.${PIVOT_MODEL_NAME}.${LANGPAIR}.${PIVOT}.spm.gz
TRANSLATED_TRG = ${OUTPUT_DIR}/${DATASET_NAME}.${PIVOT_MODEL_NAME}.${LANGPAIR}.${SRC}.gz
TRANSLATED_LATEST_SRC = ${OUTPUT_DIR}/latest/${DATASET_NAME}.${LANGPAIR}.${SRC}.gz
TRANSLATED_LATEST_TRG = ${OUTPUT_DIR}/latest/${DATASET_NAME}.${LANGPAIR}.${TRG}.gz
## all data sets
ALL_TRANSLATED_SRC = $(patsubst %,${OUTPUT_DIR}/%.${PIVOT_MODEL_NAME}.${LANGPAIR}.${PIVOT}.gz,${DATASET_NAMES})
ALL_TRANSLATED_TRG = $(patsubst %,${OUTPUT_DIR}/%.${PIVOT_MODEL_NAME}.${LANGPAIR}.${SRC}.gz,${DATASET_NAMES})
ALL_TRANSLATED_LATEST_SRC = $(patsubst %,${OUTPUT_DIR}/latest/%.${LANGPAIR}.${SRC}.gz,${DATASET_NAMES})
ALL_TRANSLATED_LATEST_TRG = $(patsubst %,${OUTPUT_DIR}/latest/%.${LANGPAIR}.${TRG}.gz,${DATASET_NAMES})
# don't delete even if uncomplete
.PRECIOUS: ${ALL_TRANSLATED_SRC} ${ALL_TRANSLATED_TRG}
# .INTERMEDIATE: ${TRANSLATED_PRE}
.PHONY: all prepare translate
all: ${DECODER} ${ALL_TRANSLATED_LATEST_SRC} ${ALL_TRANSLATED_LATEST_TRG}
prepare: ${TRANSLATED_PRE}
translate: ${TRANSLATED_TRG}
## tatoeba MT challenge ...
%-tatoeba:
${MAKE} NMT_WORKDIR=../work-tatoeba \
MODELSDIR=../models-tatoeba \
MODEL_CONTAINER=Tatoeba-MT-models \
INCLUDE=Tatoeba-train \
${@:-tatoeba=}
## aux function to print the selected modelname and data sets
.PHONY: print-modelname
print-modelname:
@echo ${PIVOT_MODEL_NAME}
@echo ${PIVOT_MODEL_ZIP}
.PHONY: print-data
print-data:
# @echo ${ORIGINAL_DATASRC}
@echo ${DATASET_NAME}
# @echo ${TRANSLATED_SRC}
# @echo ${TRANSLATED_TRG}
# @echo ${TRANSLATED_LATEST_SRC}
# @echo ${TRANSLATED_LATEST_TRG}
.PHONY: print-all-data
print-all-data:
# @echo "${ORIGINAL_DATASETS_TRG}"
@echo "${DATASET_NAMES}"
# @echo "${ALL_TRANSLATED_SRC}"
# @echo "${ALL_TRANSLATED_TRG}"
# @echo "${ALL_TRANSLATED_LATEST_SRC}"
# @echo "${ALL_TRANSLATED_LATEST_TRG}"
print-excludes:
@echo ${patsubst %,%.${ORIGINAL_LANGPAIR}.clean.${TRG}.gz,${INCLUDE}}
@echo "${EXCLUDE_PATTERN}"
@echo "${EXCLUDE_DATASETS}"
## fetch the latest model
## ---> TODO: should we fetch from ObjectStorage instead?
## ---> could have as fallback to fetch frpm ObjectStore?
${OUTPUT_DIR}/${PIVOT_MODEL_NAME}/decoder.yml:
ifneq (${PIVOT_MODEL_ZIP},)
mkdir -p ${dir $@}
# ${WGET} -O ${dir $@}${PIVOT_MODEL_NAME}.zip ${OBJECTSTORAGE}/${MODEL_CONTAINER}/${PIVOT_MODEL_ZIP}
cp ${PIVOT_MODEL_ZIP} ${dir $@}
cd ${dir $@} && unzip *.zip
rm -f ${dir $@}*.zip
endif
#--------------------------------------------------------------------------
## pre-process data
#--------------------------------------------------------------------------
ifeq (${MULTI_TARGET_MODEL},1)
PREPROCESS_ARGS = ${PIVOT} ${SRC} ${OUTPUT_DIR}/${PIVOT_MODEL_NAME}/source.spm
else
PREPROCESS_ARGS = ${PIVOT} ${OUTPUT_DIR}/${PIVOT_MODEL_NAME}/source.spm
endif
${TRANSLATED_PRE}: ${ORIGINAL_DATASRC}
ifneq (${PIVOT_MODEL_ZIP},)
mkdir -p ${dir $@}
${MAKE} ${DECODER}
${GZCAT} $< |\
head -${MAX_PIVOT_SENTENCES} |\
${OUTPUT_DIR}/${PIVOT_MODEL_NAME}/preprocess.sh ${PREPROCESS_ARGS} |\
gzip -c > $@
endif
# the same in an implict rule (makes it possible to run things in parallel)
${OUTPUT_DIR}/%.${PIVOT_MODEL_NAME}.${LANGPAIR}.${PIVOT}.spm.gz: ${ORIGINAL_DATADIR}/${PRE}/%.clean.${PIVOT}.gz
ifneq (${PIVOT_MODEL_ZIP},)
mkdir -p ${dir $@}
${MAKE} ${DECODER}
${GZCAT} $< |\
${OUTPUT_DIR}/${PIVOT_MODEL_NAME}/preprocess.sh ${PREPROCESS_ARGS} |\
gzip -c > $@
endif
## overwrite the file with the latest translations
## --> this allows multiple translation iterations
## without duplicating the data we want to use in MT training
${OUTPUT_DIR}/latest/%.${LANGPAIR}.${SRC}.gz: ${OUTPUT_DIR}/%.${PIVOT_MODEL_NAME}.${LANGPAIR}.${SRC}.gz
mkdir -p ${dir $@}
cp $< $@
${OUTPUT_DIR}/latest/%.${LANGPAIR}.${TRG}.gz: ${ORIGINAL_DATADIR}/${PRE}/%.clean.${TRG}.gz \
${OUTPUT_DIR}/latest/%.${LANGPAIR}.${SRC}.gz
mkdir -p ${dir $@}
cp $< $@
${TRANSLATED_LATEST_TRG}: ${ORIGINAL_DATATRG} ${TRANSLATED_LATEST_SRC}
mkdir -p ${dir $@}
cp $< $@
## translate
%.${LANGPAIR}.${SRC}.gz: %.${LANGPAIR}.${PIVOT}.spm.gz
ifneq (${PIVOT_MODEL_ZIP},)
mkdir -p ${dir $@}
${MAKE} ${OUTPUT_DIR}/${PIVOT_MODEL_NAME}/decoder.yml
${LOAD_ENV} && cd ${OUTPUT_DIR}/${PIVOT_MODEL_NAME} && ${MARIAN_DECODER} \
-i ${PWD}/$< \
-c decoder.yml \
-d ${MARIAN_GPUS} \
${MARIAN_DECODER_FLAGS} |\
sed 's/ //g;s/▁/ /g' | sed 's/^ *//;s/ *$$//' |\
gzip -c > ${PWD}/$@
endif
check-length:
for d in `find . -maxdepth 1 -type d -name '*-*' -printf "%f "`; do \
s=`echo $$d | cut -f1 -d'-'`; \
t=`echo $$d | cut -f2 -d'-'`; \
echo "check $$d"; \
for S in `ls $$d/*.$$s.gz`; do \
T=`echo $$S | sed 's/.$$s.gz/.$$t.gz/'`; \
echo "$$S -- $$T"; \
${GZCAT} $$S | wc -l; \
${GZCAT} $$T | wc -l; \
if [ `${GZCAT} $$S | wc -l` != `${GZCAT} $$T | wc -l` ]; then \
echo "$$S != $$T"; \
fi \
done \
done