def main_sweep(argv):
parser = argparse.ArgumentParser(
description=
'Sweep SA sample size and measure BLEU scores at various settings.')
parser.add_argument(
'-e',
'--engine',
dest='engine',
help='the engine name, \'default\' will be used if absent',
default=None)
parser.add_argument(
'--path',
dest='corpora_path',
metavar='CORPORA',
default=None,
help=
'the path to the test corpora (default is the automatically splitted sample)'
)
args = parser.parse_args(argv)
samples = [
int(e) for e in
'10 20 50 70 80 90 100 110 120 150 200 350 500 800 1000 2000 5000'.
split()
]
node = ConfiguredClusterNode(args.engine)
# more or less copy-pasted from mmt evaluate:
evaluator = Evaluator(node.engine, node)
corpora = ParallelCorpus.list(args.corpora_path) if args.corpora_path is not None \
else ParallelCorpus.list(os.path.join(node.engine.data_path, TrainingPreprocessor.TEST_FOLDER_NAME))
lines = 0
for corpus in corpora:
lines += corpus.count_lines()
# end copy-paste
print('sample bleu')
for sample in samples:
node.set('suffixarrays', 'sample', sample)
node.apply_configs()
scores = evaluator.evaluate(corpora=corpora,
google_key='1234',
heval_output=None,
use_sessions=True,
debug=False)
engine_scores = scores['MMT']
if isinstance(engine_scores, str):
raise RuntimeError(engine_scores)
bleu = engine_scores['bleu']
print(sample, '%.2f' % (bleu * 100))
def clean(self, corpora, dest_folder, langs=None):
if langs is None and len(corpora) > 0:
langs = (corpora[0].langs[0], corpora[0].langs[1])
self._pool_exec(self._clean_file,
[(corpus, ParallelCorpus(corpus.name, dest_folder, corpus.langs), langs) for corpus in corpora])
return ParallelCorpus.list(dest_folder)
def encode(self, corpora, dest_folder):
if not os.path.isdir(dest_folder):
fileutils.makedirs(dest_folder, exist_ok=True)
for corpus in corpora:
for lang in corpus.langs:
source = corpus.get_file(lang)
dest = ParallelCorpus(corpus.name, dest_folder,
[lang]).get_file(lang)
self.encode_file(source, dest)
return ParallelCorpus.list(dest_folder)
def process(self,
corpora,
dest_folder,
print_tags=True,
print_placeholders=False,
original_spacing=False):
for corpus in corpora:
for lang in corpus.langs:
source = corpus.get_file(lang)
dest = ParallelCorpus(corpus.name, dest_folder,
[lang]).get_file(lang)
self.__process_file(source, dest, lang, print_tags,
print_placeholders, original_spacing)
return ParallelCorpus.list(dest_folder)
def process(self,
source,
target,
input_paths,
output_path,
data_path=None):
args = ['-s', source, '-t', target, '--output', output_path, '--input']
for root in input_paths:
args.append(root)
if data_path is not None:
args.append('--dev')
args.append(
os.path.join(data_path, TrainingPreprocessor.DEV_FOLDER_NAME))
args.append('--test')
args.append(
os.path.join(data_path, TrainingPreprocessor.TEST_FOLDER_NAME))
command = mmt_javamain(self._java_mainclass, args)
shell.execute(command,
stdin=shell.DEVNULL,
stdout=shell.DEVNULL,
stderr=shell.DEVNULL)
return ParallelCorpus.splitlist(source, target, roots=output_path)
def translate(self, corpora, output):
"""
Translate the given corpora in parallel processing fashion.
:param corpora: list of ParallelCorpus
:param output: path to output directory
:return: ([ParallelCorpus, ...], time_per_sentence, parallelism)
"""
pool = multithread.Pool(self._threads)
try:
translations = []
start_time = datetime.now()
for corpus in corpora:
self._before_translate(corpus)
with open(corpus.get_file(self.source_lang)) as source:
output_path = os.path.join(output, corpus.name + '.' + self.target_lang)
for line in source:
translation = pool.apply_async(self._get_translation, (line, corpus))
translations.append((translation, output_path))
self._after_translate(corpus)
elapsed_time = 0
translation_count = 0
path = None
stream = None
for translation_job, output_path in translations:
translation, elapsed = translation_job.get()
if output_path != path:
if stream is not None:
stream.close()
stream = open(output_path, 'wb')
path = output_path
stream.write(translation.encode('utf-8'))
stream.write('\n')
elapsed_time += elapsed
translation_count += 1
if stream is not None:
stream.close()
end_time = datetime.now()
total_time = end_time - start_time
return ParallelCorpus.list(output), (elapsed_time / translation_count), (
elapsed_time / total_time.total_seconds())
finally:
pool.terminate()
def main_sweep(argv):
parser = argparse.ArgumentParser(description='Sweep SA sample size and measure BLEU scores at various settings.')
parser.add_argument('-e', '--engine', dest='engine', help='the engine name, \'default\' will be used if absent',
default=None)
parser.add_argument('--path', dest='corpora_path', metavar='CORPORA', default=None,
help='the path to the test corpora (default is the automatically splitted sample)')
args = parser.parse_args(argv)
samples = [int(e) for e in '10 20 50 70 80 90 100 110 120 150 200 350 500 800 1000 2000 5000'.split()]
node = ConfiguredClusterNode(args.engine)
# more or less copy-pasted from mmt evaluate:
evaluator = Evaluator(node.engine, node)
corpora = ParallelCorpus.list(args.corpora_path) if args.corpora_path is not None \
else ParallelCorpus.list(os.path.join(node.engine.data_path, TrainingPreprocessor.TEST_FOLDER_NAME))
lines = 0
for corpus in corpora:
lines += corpus.count_lines()
# end copy-paste
print('sample bleu')
for sample in samples:
node.set('suffixarrays', 'sample', sample)
node.apply_configs()
scores = evaluator.evaluate(corpora=corpora, google_key='1234', heval_output=None,
use_sessions=True, debug=False)
engine_scores = scores['MMT']
if isinstance(engine_scores, str):
raise RuntimeError(engine_scores)
bleu = engine_scores['bleu']
print(sample, '%.2f' % (bleu * 100))
def clean(self, source, target, input_paths, output_path):
args = ['-s', source, '-t', target, '--output', output_path, '--input']
for root in input_paths:
args.append(root)
command = mmt_javamain(self._java_mainclass, args)
shell.execute(command,
stdin=shell.DEVNULL,
stdout=shell.DEVNULL,
stderr=shell.DEVNULL)
return ParallelCorpus.splitlist(source, target, roots=output_path)[0]
def train(self, corpora, aligner, working_dir='.', log_file=None):
if os.path.isdir(self._model) and len(os.listdir(self._model)) > 0:
raise Exception('Model already exists at ' + self._model)
if not os.path.isdir(self._model):
fileutils.makedirs(self._model, exist_ok=True)
if not os.path.isdir(working_dir):
fileutils.makedirs(working_dir, exist_ok=True)
l1 = self._source_lang
l2 = self._target_lang
langs = (l1, l2)
langs_suffix = l1 + '-' + l2
mct_base = self._get_model_basename()
dmp_file = mct_base + '.dmp'
mam_file = mct_base + '.' + langs_suffix + '.mam'
lex_file = mct_base + '.' + langs_suffix + '.lex'
log = shell.DEVNULL
try:
if log_file is not None:
log = open(log_file, 'a')
# Clean corpus for training
clean_output = os.path.join(working_dir, 'clean_corpora')
fileutils.makedirs(clean_output, exist_ok=True)
corpora = self._cleaner.clean(corpora, clean_output, (self._source_lang, self._target_lang))
# Create merged corpus and domains list file (dmp)
merged_corpus = ParallelCorpus(os.path.basename(mct_base), working_dir, langs)
fileutils.merge([corpus.get_file(l1) for corpus in corpora], merged_corpus.get_file(l1))
fileutils.merge([corpus.get_file(l2) for corpus in corpora], merged_corpus.get_file(l2))
with open(dmp_file, 'w') as dmp:
for corpus in corpora:
dmp.write(str(corpus.name) + ' ' + str(corpus.count_lines()) + '\n')
# Create alignments in 'bal' file and symmetrize
bal_file = aligner.align(merged_corpus, langs, self._model, working_dir, log_file)
symal_file = os.path.join(working_dir, 'alignments.' + langs_suffix + '.symal')
symal_command = [self._symal_bin, '-a=g', '-d=yes', '-f=yes', '-b=yes']
with open(bal_file) as stdin:
with open(symal_file, 'w') as stdout:
shell.execute(symal_command, stdin=stdin, stdout=stdout, stderr=log)
# Execute mtt-build
mttbuild_command = self._get_mttbuild_command(mct_base, dmp_file, l1)
with open(merged_corpus.get_file(l1)) as stdin:
shell.execute(mttbuild_command, stdin=stdin, stdout=log, stderr=log)
mttbuild_command = self._get_mttbuild_command(mct_base, dmp_file, l2)
with open(merged_corpus.get_file(l2)) as stdin:
shell.execute(mttbuild_command, stdin=stdin, stdout=log, stderr=log)
# Create 'mam' file
mam_command = [self._symal2mam_bin, mam_file]
with open(symal_file) as stdin:
shell.execute(mam_command, stdin=stdin, stdout=log, stderr=log)
# Create 'lex' file
lex_command = [self._mmlexbuild_bin, mct_base + '.', l1, l2, '-o', lex_file]
shell.execute(lex_command, stdout=log, stderr=log)
finally:
if log_file is not None:
log.close()
def tune(self, corpora=None, tokenize=True, debug=False, context_enabled=True):
if corpora is None:
corpora = ParallelCorpus.list(os.path.join(self.engine.data_path, TrainingPreprocessor.DEV_FOLDER_NAME))
if len(corpora) == 0:
raise IllegalArgumentException('empty corpora')
if not self.is_running():
raise IllegalStateException('No MMT Server running, start the engine first')
target_lang = self.engine.target_lang
source_lang = self.engine.source_lang
cmdlogger = _tuning_logger(4 if tokenize else 3)
cmdlogger.start(self, corpora)
working_dir = self.engine.get_tempdir('tuning')
mert_wd = os.path.join(working_dir, 'mert')
try:
original_corpora = corpora
# Tokenization
tokenized_corpora = original_corpora
if tokenize:
tokenizer_output = os.path.join(working_dir, 'tokenized_corpora')
fileutils.makedirs(tokenizer_output, exist_ok=True)
with cmdlogger.step('Corpus tokenization') as _:
tokenized_corpora = self.engine.preprocessor.process(corpora, tokenizer_output, print_tags=False,
print_placeholders=True,
original_spacing=False)
# Create merged corpus
with cmdlogger.step('Merging corpus') as _:
source_merged_corpus = os.path.join(working_dir, 'corpus.' + source_lang)
with open(source_merged_corpus, 'wb') as out:
original_root = original_corpora[0].root
for corpus in tokenized_corpora:
tokenized = corpus.get_file(source_lang)
original = os.path.join(original_root, corpus.name + '.' + source_lang)
out.write(tokenized + ':' + original + '\n')
target_merged_corpus = os.path.join(working_dir, 'corpus.' + target_lang)
fileutils.merge([corpus.get_file(target_lang) for corpus in tokenized_corpora], target_merged_corpus)
# Run MERT algorithm
with cmdlogger.step('Tuning') as _:
# Start MERT
decoder_flags = ['--port', str(self.api.port)]
if not context_enabled:
decoder_flags.append('--skip-context-analysis')
decoder_flags.append('1')
fileutils.makedirs(mert_wd, exist_ok=True)
with tempfile.NamedTemporaryFile() as runtime_moses_ini:
command = [self._mert_script, source_merged_corpus, target_merged_corpus,
self._mert_i_script, runtime_moses_ini.name, '--threads',
str(multiprocessing.cpu_count()), '--mertdir', os.path.join(Moses.bin_path, 'bin'),
'--mertargs', '\'--binary --sctype BLEU\'', '--working-dir', mert_wd, '--nbest', '100',
'--decoder-flags', '"' + ' '.join(decoder_flags) + '"', '--nonorm', '--closest',
'--no-filter-phrase-table']
with open(self.engine.get_logfile('mert'), 'wb') as log:
shell.execute(' '.join(command), stdout=log, stderr=log)
# Read optimized configuration
with cmdlogger.step('Applying changes') as _:
bleu_score = 0
weights = {}
found_weights = False
with open(os.path.join(mert_wd, 'moses.ini')) as moses_ini:
for line in moses_ini:
line = line.strip()
if len(line) == 0:
continue
elif found_weights:
tokens = line.split()
weights[tokens[0].rstrip('=')] = [float(val) for val in tokens[1:]]
elif line.startswith('# BLEU'):
bleu_score = float(line.split()[2])
elif line == '[weight]':
found_weights = True
_ = self.api.update_features(weights)
cmdlogger.completed(bleu_score)
finally:
if not debug:
self.engine.clear_tempdir()
def tune(self,
corpora=None,
tokenize=True,
debug=False,
context_enabled=True):
if corpora is None:
corpora = ParallelCorpus.list(
os.path.join(self.engine.data_path,
TrainingPreprocessor.DEV_FOLDER_NAME))
if len(corpora) == 0:
raise IllegalArgumentException('empty corpora')
if not self.is_running():
raise IllegalStateException(
'No MMT Server running, start the engine first')
target_lang = self.engine.target_lang
source_lang = self.engine.source_lang
cmdlogger = _tuning_logger(4 if tokenize else 3)
cmdlogger.start(self, corpora)
working_dir = self.engine.get_tempdir('tuning')
mert_wd = os.path.join(working_dir, 'mert')
try:
original_corpora = corpora
# Tokenization
tokenized_corpora = original_corpora
if tokenize:
tokenizer_output = os.path.join(working_dir,
'tokenized_corpora')
fileutils.makedirs(tokenizer_output, exist_ok=True)
with cmdlogger.step('Corpus tokenization') as _:
tokenized_corpora = self.engine.preprocessor.process(
corpora,
tokenizer_output,
print_tags=False,
print_placeholders=True,
original_spacing=False)
# Create merged corpus
with cmdlogger.step('Merging corpus') as _:
source_merged_corpus = os.path.join(working_dir,
'corpus.' + source_lang)
with open(source_merged_corpus, 'wb') as out:
original_root = original_corpora[0].root
for corpus in tokenized_corpora:
tokenized = corpus.get_file(source_lang)
original = os.path.join(
original_root, corpus.name + '.' + source_lang)
out.write(tokenized + ':' + original + '\n')
target_merged_corpus = os.path.join(working_dir,
'corpus.' + target_lang)
fileutils.merge([
corpus.get_file(target_lang)
for corpus in tokenized_corpora
], target_merged_corpus)
# Run MERT algorithm
with cmdlogger.step('Tuning') as _:
# Start MERT
decoder_flags = ['--port', str(self.api.port)]
if not context_enabled:
decoder_flags.append('--skip-context-analysis')
decoder_flags.append('1')
fileutils.makedirs(mert_wd, exist_ok=True)
with tempfile.NamedTemporaryFile() as runtime_moses_ini:
command = [
self._mert_script, source_merged_corpus,
target_merged_corpus, self._mert_i_script,
runtime_moses_ini.name, '--threads',
str(multiprocessing.cpu_count()), '--mertdir',
os.path.join(Moses.bin_path, 'bin'), '--mertargs',
'\'--binary --sctype BLEU\'', '--working-dir', mert_wd,
'--nbest', '100', '--decoder-flags',
'"' + ' '.join(decoder_flags) + '"', '--nonorm',
'--closest', '--no-filter-phrase-table'
]
with open(self.engine.get_logfile('mert'), 'wb') as log:
shell.execute(' '.join(command),
stdout=log,
stderr=log)
# Read optimized configuration
with cmdlogger.step('Applying changes') as _:
bleu_score = 0
weights = {}
found_weights = False
with open(os.path.join(mert_wd, 'moses.ini')) as moses_ini:
for line in moses_ini:
line = line.strip()
if len(line) == 0:
continue
elif found_weights:
tokens = line.split()
weights[tokens[0].rstrip('=')] = [
float(val) for val in tokens[1:]
]
elif line.startswith('# BLEU'):
bleu_score = float(line.split()[2])
elif line == '[weight]':
found_weights = True
_ = self.api.update_features(weights)
cmdlogger.completed(bleu_score)
finally:
if not debug:
self.engine.clear_tempdir()
def build(self, roots, debug=False, steps=None, split_trainingset=True):
self._temp_dir = self._engine.get_tempdir('training', ensure=True)
source_lang = self._engine.source_lang
target_lang = self._engine.target_lang
bilingual_corpora, monolingual_corpora = ParallelCorpus.splitlist(source_lang, target_lang, roots=roots)
if len(bilingual_corpora) == 0:
raise IllegalArgumentException(
'you project does not include %s-%s data.' % (source_lang.upper(), target_lang.upper()))
if steps is None:
steps = self._engine.training_steps
else:
unknown_steps = [step for step in steps if step not in self._engine.training_steps]
if len(unknown_steps) > 0:
raise IllegalArgumentException('Unknown training steps: ' + str(unknown_steps))
cmdlogger = _builder_logger(len(steps) + 1)
cmdlogger.start(self._engine, bilingual_corpora, monolingual_corpora)
shutil.rmtree(self._engine.path, ignore_errors=True)
os.makedirs(self._engine.path)
# Check disk space constraints
free_space_on_disk = fileutils.df(self._engine.path)[2]
corpus_size_on_disk = 0
for root in roots:
corpus_size_on_disk += fileutils.du(root)
free_memory = fileutils.free()
recommended_mem = self.__GB * corpus_size_on_disk / (350 * self.__MB) # 1G RAM every 350M on disk
recommended_disk = 10 * corpus_size_on_disk
if free_memory < recommended_mem or free_space_on_disk < recommended_disk:
if free_memory < recommended_mem:
print '> WARNING: more than %.fG of RAM recommended, only %.fG available' % \
(recommended_mem / self.__GB, free_memory / self.__GB)
if free_space_on_disk < recommended_disk:
print '> WARNING: more than %.fG of storage recommended, only %.fG available' % \
(recommended_disk / self.__GB, free_space_on_disk / self.__GB)
print
try:
corpora_roots = roots
unprocessed_bicorpora = bilingual_corpora
unprocessed_mocorpora = monolingual_corpora
# TM cleanup
if 'tm_cleanup' in steps:
with cmdlogger.step('TMs clean-up') as _:
cleaned_output = self._get_tempdir('clean_tms')
self._engine.cleaner.clean(source_lang, target_lang, roots, cleaned_output)
for corpus in monolingual_corpora:
cfile = corpus.get_file(target_lang)
link = os.path.join(cleaned_output, os.path.basename(cfile))
os.symlink(cfile, link)
corpora_roots = [cleaned_output]
unprocessed_bicorpora, unprocessed_mocorpora = ParallelCorpus.splitlist(source_lang, target_lang,
roots=corpora_roots)
# Preprocessing
processed_bicorpora = unprocessed_bicorpora
processed_mocorpora = unprocessed_mocorpora
if 'preprocess' in steps:
with cmdlogger.step('Corpora preprocessing') as _:
preprocessor_output = self._get_tempdir('preprocessed')
processed_bicorpora, processed_mocorpora = self._engine.training_preprocessor.process(
source_lang, target_lang, corpora_roots, preprocessor_output,
(self._engine.data_path if split_trainingset else None)
)
# Training Context Analyzer
if 'context_analyzer' in steps:
with cmdlogger.step('Context Analyzer training') as _:
log_file = self._engine.get_logfile('training.context')
self._engine.analyzer.create_index(unprocessed_bicorpora, source_lang, log_file=log_file)
# Training Adaptive Language Model (on the target side of all bilingual corpora)
if 'lm' in steps:
with cmdlogger.step('Language Model training') as _:
working_dir = self._get_tempdir('lm')
log_file = self._engine.get_logfile('training.lm')
self._engine.lm.train(processed_bicorpora + processed_mocorpora, target_lang, working_dir, log_file)
# Training Translation Model
if 'tm' in steps:
with cmdlogger.step('Translation Model training') as _:
working_dir = self._get_tempdir('tm')
log_file = self._engine.get_logfile('training.tm')
self._engine.pt.train(processed_bicorpora, self._engine.aligner, working_dir, log_file)
# Writing config file
with cmdlogger.step('Writing config files') as _:
self._engine.write_configs()
cmdlogger.completed()
finally:
if not debug:
self._engine.clear_tempdir('training')
def build(self, roots, debug=False, steps=None, split_trainingset=True):
self._temp_dir = self._engine.get_tempdir('training', ensure=True)
source_lang = self._engine.source_lang
target_lang = self._engine.target_lang
bilingual_corpora, monolingual_corpora = ParallelCorpus.splitlist(
source_lang, target_lang, roots=roots)
if len(bilingual_corpora) == 0:
raise IllegalArgumentException(
'you project does not include %s-%s data.' %
(source_lang.upper(), target_lang.upper()))
if steps is None:
steps = self._engine.training_steps
else:
unknown_steps = [
step for step in steps
if step not in self._engine.training_steps
]
if len(unknown_steps) > 0:
raise IllegalArgumentException('Unknown training steps: ' +
str(unknown_steps))
cmdlogger = _builder_logger(len(steps) + 1)
cmdlogger.start(self._engine, bilingual_corpora, monolingual_corpora)
shutil.rmtree(self._engine.path, ignore_errors=True)
os.makedirs(self._engine.path)
# Check disk space constraints
free_space_on_disk = fileutils.df(self._engine.path)[2]
corpus_size_on_disk = 0
for root in roots:
corpus_size_on_disk += fileutils.du(root)
free_memory = fileutils.free()
recommended_mem = self.__GB * corpus_size_on_disk / (
350 * self.__MB) # 1G RAM every 350M on disk
recommended_disk = 10 * corpus_size_on_disk
if free_memory < recommended_mem or free_space_on_disk < recommended_disk:
if free_memory < recommended_mem:
print '> WARNING: more than %.fG of RAM recommended, only %.fG available' % \
(recommended_mem / self.__GB, free_memory / self.__GB)
if free_space_on_disk < recommended_disk:
print '> WARNING: more than %.fG of storage recommended, only %.fG available' % \
(recommended_disk / self.__GB, free_space_on_disk / self.__GB)
print
try:
corpora_roots = roots
unprocessed_bicorpora = bilingual_corpora
unprocessed_mocorpora = monolingual_corpora
# TM cleanup
if 'tm_cleanup' in steps:
with cmdlogger.step('TMs clean-up') as _:
cleaned_output = self._get_tempdir('clean_tms')
self._engine.cleaner.clean(source_lang, target_lang, roots,
cleaned_output)
for corpus in monolingual_corpora:
cfile = corpus.get_file(target_lang)
link = os.path.join(cleaned_output,
os.path.basename(cfile))
os.symlink(cfile, link)
corpora_roots = [cleaned_output]
unprocessed_bicorpora, unprocessed_mocorpora = ParallelCorpus.splitlist(
source_lang, target_lang, roots=corpora_roots)
# Preprocessing
processed_bicorpora = unprocessed_bicorpora
processed_mocorpora = unprocessed_mocorpora
if 'preprocess' in steps:
with cmdlogger.step('Corpora preprocessing') as _:
preprocessor_output = self._get_tempdir('preprocessed')
processed_bicorpora, processed_mocorpora = self._engine.training_preprocessor.process(
source_lang, target_lang, corpora_roots,
preprocessor_output, (self._engine.data_path
if split_trainingset else None))
# Training Context Analyzer
if 'context_analyzer' in steps:
with cmdlogger.step('Context Analyzer training') as _:
log_file = self._engine.get_logfile('training.context')
self._engine.analyzer.create_index(unprocessed_bicorpora,
source_lang,
log_file=log_file)
# Training Adaptive Language Model (on the target side of all bilingual corpora)
if 'lm' in steps:
with cmdlogger.step('Language Model training') as _:
working_dir = self._get_tempdir('lm')
log_file = self._engine.get_logfile('training.lm')
self._engine.lm.train(
processed_bicorpora + processed_mocorpora, target_lang,
working_dir, log_file)
# Training Translation Model
if 'tm' in steps:
with cmdlogger.step('Translation Model training') as _:
working_dir = self._get_tempdir('tm')
log_file = self._engine.get_logfile('training.tm')
self._engine.pt.train(processed_bicorpora,
self._engine.aligner, working_dir,
log_file)
# Writing config file
with cmdlogger.step('Writing config files') as _:
self._engine.write_configs()
cmdlogger.completed()
finally:
if not debug:
self._engine.clear_tempdir('training')