def test_compute_type():
model_path = _get_model_path()
with pytest.raises(ValueError):
ctranslate2.Translator(model_path, compute_type="float64")
with pytest.raises(TypeError):
ctranslate2.Translator(model_path, compute_type=["int8", "int16"])
ctranslate2.Translator(model_path, compute_type="int8")
ctranslate2.Translator(model_path, compute_type={"cuda": "float16", "cpu": "int8"})
def test_compute_type():
model_path = _get_model_path()
with pytest.raises(ValueError, match="compute type"):
ctranslate2.Translator(model_path, compute_type="float64")
with pytest.raises(TypeError, match="incompatible constructor arguments"):
ctranslate2.Translator(model_path, compute_type=["int8", "int16"])
ctranslate2.Translator(model_path, compute_type="int8")
ctranslate2.Translator(model_path,
compute_type={
"cuda": "float16",
"cpu": "int8"
})
def test_raw_file_translation(tmpdir):
input_path = str(tmpdir.join("input.txt"))
output_path = str(tmpdir.join("output.txt"))
with open(input_path, "w", encoding="utf-8") as input_file:
input_file.write("آتزمون")
input_file.write("\n")
input_file.write("آتشيسون")
input_file.write("\n")
translator = ctranslate2.Translator(_get_model_path())
tokenize_fn = lambda text: list(text)
detokenize_fn = lambda tokens: "".join(tokens)
with pytest.raises(ValueError, match="target_detokenize_fn"):
translator.translate_file(input_path,
output_path,
source_tokenize_fn=tokenize_fn)
with pytest.raises(ValueError, match="source_tokenize_fn"):
translator.translate_file(input_path,
output_path,
target_detokenize_fn=detokenize_fn)
translator.translate_file(
input_path,
output_path,
source_tokenize_fn=tokenize_fn,
target_detokenize_fn=detokenize_fn,
)
with open(output_path, encoding="utf-8") as output_file:
lines = output_file.readlines()
assert lines[0].strip() == "atzmon"
assert lines[1].strip() == "achison"
def test_raw_file_translation(tmpdir):
input_path = str(tmpdir.join("input.txt"))
output_path = str(tmpdir.join("output.txt"))
with open(input_path, "w") as input_file:
input_file.write("آتزمون")
input_file.write("\n")
input_file.write("آتشيسون")
input_file.write("\n")
translator = ctranslate2.Translator(_get_model_path())
tokenize_fn = lambda text: list(text)
detokenize_fn = lambda tokens: "".join(tokens)
max_batch_size = 4
with pytest.raises(ValueError):
translator.translate_file(input_path,
output_path,
max_batch_size,
tokenize_fn=tokenize_fn)
with pytest.raises(ValueError):
translator.translate_file(input_path,
output_path,
max_batch_size,
detokenize_fn=detokenize_fn)
translator.translate_file(input_path,
output_path,
max_batch_size,
tokenize_fn=tokenize_fn,
detokenize_fn=detokenize_fn)
with open(output_path) as output_file:
lines = output_file.readlines()
assert lines[0].strip() == "atzmon"
assert lines[1].strip() == "achison"
def return_loaded_models(self, model_paths, ids):
loaded_models = {}
for i, path in enumerate(model_paths):
translator = ctranslate2.Translator(path, device="auto")
loaded_models[ids[i]] = translator
log_info("Model Loaded: {}".format(ids[i]), MODULE_CONTEXT)
return loaded_models
def __init__(self, model_path):
self.model_path = model_path
self.tokenizer_source = None
self.tokenizer_target = None
if self.INTER_THREADS in os.environ:
inter_threads = int(os.environ[self.INTER_THREADS])
else:
inter_threads = 1
if self.INTRA_THREADS in os.environ:
intra_threads = int(os.environ[self.INTRA_THREADS])
else:
intra_threads = 4
if self.BEAM_SIZE in os.environ:
self.beam_size = int(os.environ[self.BEAM_SIZE])
else:
self.beam_size = 2
if self.USE_VMAP in os.environ:
self.use_vmap = True
else:
self.use_vmap = False
print(
f"inter_threads: {inter_threads}, intra_threads: {intra_threads}, beam_size {self.beam_size}, use_vmap {self.use_vmap}"
)
self.translator = ctranslate2.Translator(model_path,
inter_threads=inter_threads,
intra_threads=intra_threads)
def hypotheses(self, input_text, num_hypotheses):
if self.translator == None:
model_path = str(self.pkg.package_path / 'model')
self.translator = ctranslate2.Translator(model_path)
paragraphs = ITranslation.split_into_paragraphs(input_text)
info("paragraphs", paragraphs)
translated_paragraphs = []
for paragraph in paragraphs:
translated_paragraphs.append(
apply_packaged_translation(self.pkg, paragraph,
self.translator, num_hypotheses))
info("translated_paragraphs", translated_paragraphs)
# Construct new hypotheses using all paragraphs
hypotheses_to_return = [
Hypothesis('', 0) for i in range(num_hypotheses)
]
for i in range(num_hypotheses):
for translated_paragraph in translated_paragraphs:
value = ITranslation.combine_paragraphs([
hypotheses_to_return[i].value,
translated_paragraph[i].value
])
score = hypotheses_to_return[i].score + translated_paragraph[
i].score
hypotheses_to_return[i] = Hypothesis(value, score)
hypotheses_to_return[i].value = hypotheses_to_return[
i].value.lstrip('\n')
info('hypotheses_to_return', hypotheses_to_return)
return hypotheses_to_return
def __init__(self, models_path, model_name, tokenizer_source = None, tokenizer_target = None, translator = None):
inter_threads, intra_threads = self._init_read_env_vars()
model_path = os.path.join(models_path, model_name)
if tokenizer_source:
self.tokenizer_source = tokenizer_target
else:
src_model_path = self.get_source_tokenizer_file(model_path, model_name)
self.tokenizer_source = pyonmttok.Tokenizer(mode="none", sp_model_path = src_model_path)
if tokenizer_target:
self.tokenizer_target = tokenizer_target
else:
tgt_model_path = self.get_target_tokenizer_file(model_path, model_name)
self.tokenizer_target = pyonmttok.Tokenizer(mode="none", sp_model_path = tgt_model_path)
self.tokenizer_source_language = self._get_sentence_tokenizer_source_language(model_name)
print(f"inter_threads: {inter_threads}, intra_threads: {intra_threads}, beam_size {self.beam_size}, use_vmap {self.use_vmap}")
if translator is None:
self.model_path = model_path
ctranslate_model_path = os.path.join(model_path, "ctranslate2")
self.translator = ctranslate2.Translator(ctranslate_model_path, inter_threads = inter_threads, intra_threads = intra_threads)
else:
self.translator = translator
self.model_name = model_name
def __init__(self,
model_path,
device,
device_index,
batch_size,
beam_size,
n_best,
preload=False):
import ctranslate2
self.translator = ctranslate2.Translator(model_path,
device=device,
device_index=device_index,
inter_threads=1,
intra_threads=1,
compute_type="default")
self.batch_size = batch_size
self.beam_size = beam_size
self.n_best = n_best
if preload:
# perform a first request to initialize everything
dummy_translation = self.translate(["a"])
print("Performed a dummy translation to initialize the model",
dummy_translation)
time.sleep(1)
self.translator.unload_model(to_cpu=True)
def test_fairseq_user_start_token(tmpdir):
class _CustomFairseqConverter(ctranslate2.converters.FairseqConverter):
def _load(self):
model_spec = super()._load()
model_spec.user_decoder_start_tokens = True
return model_spec
data_dir = os.path.join(
_TEST_DATA_DIR,
"models",
"transliteration-aren-all",
"fairseq",
)
converter = _CustomFairseqConverter(os.path.join(data_dir, "model.pt"),
data_dir)
output_dir = str(tmpdir.join("ctranslate2_model"))
converter.convert(output_dir)
translator = ctranslate2.Translator(output_dir)
tokens = [["آ", "ت", "ز", "م", "و", "ن"]]
with pytest.raises(ValueError, match="start token"):
translator.translate_batch(tokens)
output = translator.translate_batch(tokens, target_prefix=[["</s>"]])
assert output[0].hypotheses[0] == ["a", "t", "z", "m", "o", "n"]
def test_opennmt_tf_variables_conversion(tmpdir):
import opennmt
model_path = os.path.join(
_TEST_DATA_DIR,
"models",
"transliteration-aren-all",
"opennmt_tf",
"v2",
"checkpoint",
)
src_vocab = opennmt.data.Vocab.from_file(
os.path.join(model_path, "ar.vocab"))
tgt_vocab = opennmt.data.Vocab.from_file(
os.path.join(model_path, "en.vocab"))
_, variables = opennmt_tf.load_model(model_path)
converter = ctranslate2.converters.OpenNMTTFConverter(
ctranslate2.specs.TransformerSpec(6, 8),
src_vocab,
tgt_vocab,
variables=variables,
)
output_dir = str(tmpdir.join("ctranslate2_model"))
converter.convert(output_dir)
translator = ctranslate2.Translator(output_dir)
output = translator.translate_batch([["آ", "ت", "ز", "م", "و", "ن"]])
assert output[0].hypotheses[0] == ["a", "t", "z", "m", "o", "n"]
def test_translator_properties():
translator = ctranslate2.Translator(_get_model_path(), inter_threads=2)
assert translator.model_is_loaded
assert translator.device == "cpu"
assert translator.device_index == 0
assert translator.num_translators == 2
assert translator.num_queued_batches == 0
def test_opennmt_py_model_conversion(tmpdir):
model_path = os.path.join(
_TEST_DATA_DIR, "models", "transliteration-aren-all", "opennmt_py", "aren_7000.pt")
converter = ctranslate2.converters.OpenNMTPyConverter(model_path)
output_dir = str(tmpdir.join("ctranslate2_model"))
converter.convert(output_dir, ctranslate2.specs.TransformerBase())
translator = ctranslate2.Translator(output_dir)
output = translator.translate_batch([["آ" ,"ت" ,"ز" ,"م" ,"و" ,"ن"]])
assert output[0][0]["tokens"] == ["a", "t", "z", "m", "o", "n"]
def translate(self, input_text):
if self.translator == None:
model_path = str(self.pkg.package_path / 'model')
self.translator = ctranslate2.Translator(model_path)
paragraphs = self.split_into_paragraphs(input_text)
translated_paragraphs = []
for paragraph in paragraphs:
translated_paragraphs.append(
apply_packaged_translation(self.pkg, paragraph, self.translator))
return self.combine_paragraphs(translated_paragraphs)
def load_translator(self, inter_threads, intra_threads):
import ctranslate2
translator = ctranslate2.Translator('./models/multi_lang_v3',
device='auto',
device_index=0,
inter_threads=inter_threads,
intra_threads=intra_threads,
compute_type="default")
print('Using device:', translator.device)
return translator
def testCTranslate2Export(self, variant):
try:
import ctranslate2
except ImportError:
self.skipTest("ctranslate2 module is not available")
export_dir = os.path.join(self.get_temp_dir(), "export")
runner = self._getTransliterationRunner()
runner.export(export_dir, exporter=exporters.make_exporter(variant))
self.assertTrue(ctranslate2.contains_model(export_dir))
translator = ctranslate2.Translator(export_dir)
output = translator.translate_batch([["آ", "ت", "ز", "م", "و", "ن"]])
self.assertListEqual(output[0][0]["tokens"], ["a", "t", "z", "m", "o", "n"])
def test_opennmt_tf_model_quantization(tmpdir, quantization):
model_path = os.path.join(
_TEST_DATA_DIR, "models", "transliteration-aren-all", "opennmt_tf", "v2", "checkpoint")
converter = ctranslate2.converters.OpenNMTTFConverter(
model_path,
src_vocab=os.path.join(model_path, "ar.vocab"),
tgt_vocab=os.path.join(model_path, "en.vocab"))
output_dir = str(tmpdir.join("ctranslate2_model"))
converter.convert(output_dir, ctranslate2.specs.TransformerBase(), quantization=quantization)
translator = ctranslate2.Translator(output_dir)
output = translator.translate_batch([["آ" ,"ت" ,"ز" ,"م" ,"و" ,"ن"]])
assert output[0][0]["tokens"] == ["a", "t", "z", "m", "o", "n"]
def test_opennmt_py_source_features(tmpdir, filename):
model_path = os.path.join(
_TEST_DATA_DIR,
"models",
"transliteration-aren-all",
"opennmt_py",
filename,
)
converter = ctranslate2.converters.OpenNMTPyConverter(model_path)
output_dir = str(tmpdir.join("ctranslate2_model"))
converter.convert(output_dir)
assert os.path.isfile(os.path.join(output_dir, "source_1_vocabulary.txt"))
assert os.path.isfile(os.path.join(output_dir, "source_2_vocabulary.txt"))
source = [
["آ", "ت", "ز", "م", "و", "ن"],
["آ", "ت", "ش", "ي", "س", "و", "ن"],
]
source_features = [
["0", "1", "2", "3", "4", "5"],
["0", "1", "2", "3", "4", "5", "6"],
]
expected_target = [
["a", "t", "z", "m", "o", "n"],
["a", "c", "h", "i", "s", "o", "n"],
]
source_w_features = []
for tokens, features in zip(source, source_features):
source_w_features.append(
["%s￨%s" % pair for pair in zip(tokens, features)])
translator = ctranslate2.Translator(output_dir)
with pytest.raises(ValueError, match="features"):
translator.translate_batch(source)
outputs = translator.translate_batch(source_w_features)
for output, expected_hypothesis in zip(outputs, expected_target):
assert output.hypotheses[0] == expected_hypothesis
input_path = str(tmpdir.join("input.txt"))
output_path = str(tmpdir.join("output.txt"))
_write_tokens(source, input_path)
with pytest.raises(ValueError, match="features"):
translator.translate_file(input_path, output_path)
_write_tokens(source_w_features, input_path)
translator.translate_file(input_path, output_path)
with open(output_path) as output_file:
for line, expected_hypothesis in zip(output_file, expected_target):
assert line.strip().split() == expected_hypothesis
def test_fairseq_model_conversion(tmpdir):
data_dir = os.path.join(
_TEST_DATA_DIR,
"models",
"transliteration-aren-all",
"fairseq",
)
converter = ctranslate2.converters.FairseqConverter(
os.path.join(data_dir, "model.pt"), data_dir)
output_dir = str(tmpdir.join("ctranslate2_model"))
converter.convert(output_dir)
translator = ctranslate2.Translator(output_dir)
output = translator.translate_batch([["آ", "ت", "ز", "م", "و", "ن"]])
assert output[0].hypotheses[0] == ["a", "t", "z", "m", "o", "n"]
def test_opennmt_tf_model_conversion(tmpdir, model_path, src_vocab, tgt_vocab, model_spec):
model_path = os.path.join(
_TEST_DATA_DIR, "models", "transliteration-aren-all", "opennmt_tf", model_path)
if src_vocab is not None:
src_vocab = os.path.join(model_path, src_vocab)
if tgt_vocab is not None:
tgt_vocab = os.path.join(model_path, tgt_vocab)
converter = ctranslate2.converters.OpenNMTTFConverter(
model_path, src_vocab=src_vocab, tgt_vocab=tgt_vocab)
output_dir = str(tmpdir.join("ctranslate2_model"))
converter.convert(output_dir, model_spec)
translator = ctranslate2.Translator(output_dir)
output = translator.translate_batch([["آ" ,"ت" ,"ز" ,"م" ,"و" ,"ن"]])
assert output[0][0]["tokens"] == ["a", "t", "z", "m", "o", "n"]
def test_hard_target_prefix_with_vmap(tmpdir, beam_size):
model_dir = str(tmpdir.join("model"))
shutil.copytree(_get_model_path(), model_dir)
with open(os.path.join(model_dir, "vmap.txt"), "w",
encoding="utf-8") as vmap:
vmap.write("ن\tt z m o n\n")
translator = ctranslate2.Translator(model_dir)
output = translator.translate_batch(
[["آ", "ت", "ز", "م", "و", "ن"]],
target_prefix=[["a", "t", "z"]],
beam_size=beam_size,
use_vmap=True,
)
assert output[0].hypotheses[0] == ["a", "t", "z", "m", "o", "n"]
def quickstart_example():
# Convert a model trained with OpenNMT-py or OpenNMT-tf.
# REF [site] >> https://github.com/OpenNMT/CTranslate2
#--------------------
translator = ctranslate2.Translator('ende_ctranslate2/')
print('Start translating...')
start_time = time.time()
# Translate tokenized inputs.
translated = translator.translate_batch([['▁H', 'ello', '▁world', '!']])
print('End translating: {} secs.'.format(time.time() - start_time))
print('Tokens: {}.'.format(translated[0][0]['tokens']))
print('Score = {}.'.format(translated[0][0]['score']))
def test_opennmt_py_relative_transformer(tmpdir):
model_path = os.path.join(
_TEST_DATA_DIR,
"models",
"transliteration-aren-all",
"opennmt_py",
"aren_relative_6000.pt",
)
converter = ctranslate2.converters.OpenNMTPyConverter(model_path)
output_dir = str(tmpdir.join("ctranslate2_model"))
converter.convert(output_dir)
translator = ctranslate2.Translator(output_dir)
output = translator.translate_batch([["آ", "ت", "ز", "م", "و", "ن"],
["آ", "ر", "ث", "ر"]])
assert output[0].hypotheses[0] == ["a", "t", "z", "o", "m", "o", "n"]
assert output[1].hypotheses[0] == ["a", "r", "t", "h", "e", "r"]
def multi_translate(self, input_text, nresults=4):
if self.translator == None:
model_path = str(self.pkg.package_path / 'model')
self.translator = ctranslate2.Translator(model_path)
paragraphs = self.split_into_paragraphs(input_text)
info("paragraphs", paragraphs)
translated_paragraphs = []
for paragraph in paragraphs:
translated_paragraphs.append(
apply_packaged_translation(self.pkg, paragraph,
self.translator, nresults))
info("translated_paragraphs", translated_paragraphs)
pre_combine_paragraphs = [[s[i] for s in translated_paragraphs]
for i in range(nresults)]
info("pre_combine_paragraphs", pre_combine_paragraphs)
return self.combine_paragraphs(pre_combine_paragraphs, nresults)
def test_raw_file_translation_with_prefix(tmpdir):
source_path = str(tmpdir.join("input.txt"))
target_path = str(tmpdir.join("target.txt"))
output_path = str(tmpdir.join("output.txt"))
with open(source_path, "w") as source_file:
source_file.write("آتزمون")
source_file.write("\n")
source_file.write("آتشيسون")
source_file.write("\n")
with open(target_path, "w") as target_file:
# Write target in reverse to use a different tokenization.
target_file.write("sta\n")
target_file.write("\n")
translator = ctranslate2.Translator(_get_model_path())
source_tokenize_fn = lambda text: list(text)
target_tokenize_fn = lambda text: list(reversed(list(text)))
detokenize_fn = lambda tokens: "".join(tokens)
max_batch_size = 4
with pytest.raises(ValueError):
# Target tokenization is missing.
translator.translate_file(
source_path,
output_path,
max_batch_size,
tokenize_fn=source_tokenize_fn,
detokenize_fn=detokenize_fn,
target_path=target_path,
)
translator.translate_file(
source_path,
output_path,
max_batch_size,
tokenize_fn=source_tokenize_fn,
detokenize_fn=detokenize_fn,
target_path=target_path,
target_tokenize_fn=target_tokenize_fn,
)
with open(output_path) as output_file:
lines = output_file.readlines()
assert lines[0].strip() == "atsumon"
assert lines[1].strip() == "achison"
def test_opennmt_tf_shared_embeddings_conversion(tmpdir):
# Issue https://github.com/OpenNMT/CTranslate2/issues/118
import tensorflow as tf
import opennmt
vocab = opennmt.data.Vocab()
for i in range(10):
vocab.add(str(i))
vocab_path = str(tmpdir.join("vocab.txt"))
vocab.serialize(vocab_path)
num_layers = 3
num_heads = 4
model = opennmt.models.Transformer(
opennmt.inputters.WordEmbedder(32),
opennmt.inputters.WordEmbedder(32),
num_layers,
num_units=32,
num_heads=num_heads,
ffn_inner_dim=64,
share_embeddings=opennmt.models.EmbeddingsSharingLevel.ALL,
)
model.initialize({
"source_vocabulary": vocab_path,
"target_vocabulary": vocab_path
})
model.create_variables()
checkpoint_prefix = str(tmpdir.join("ckpt"))
checkpoint = tf.train.Checkpoint(model=model)
checkpoint.write(checkpoint_prefix)
converter = ctranslate2.converters.OpenNMTTFConverter(
model_path=checkpoint_prefix,
src_vocab=vocab_path,
tgt_vocab=vocab_path)
output_dir = str(tmpdir.join("ctranslate2_model"))
converter.convert(output_dir,
ctranslate2.specs.TransformerSpec(num_layers, num_heads))
assert os.path.isfile(os.path.join(output_dir, "shared_vocabulary.txt"))
# Check that the translation runs.
translator = ctranslate2.Translator(output_dir)
translator.translate_batch([["1", "2", "3"]], max_decoding_length=10)
def __init__(self,
model_path,
ct2_translator_args,
ct2_translate_batch_args,
target_prefix=False,
preload=False):
import ctranslate2
self.translator = ctranslate2.Translator(model_path,
**ct2_translator_args)
self.ct2_translate_batch_args = ct2_translate_batch_args
self.target_prefix = target_prefix
if preload:
# perform a first request to initialize everything
dummy_translation = self.translate(["a"])
print("Performed a dummy translation to initialize the model",
dummy_translation)
time.sleep(1)
self.translator.unload_model(to_cpu=True)
def translate(src_list, sp_path_src, sp_path_tgt, ct_path):
tokenize = MosesTokenizer('ru')
sp_src = spm.SentencePieceProcessor()
sp_src.load(sp_path_src)
lengths = []
temp = []
p_big = re.compile('Ҧ')
g_big = re.compile('Ҕ')
p_small = re.compile('ҧ')
g_small = re.compile('ҕ')
for text in src_list:
text = p_big.sub('Ԥ', text)
text = g_big.sub('Ӷ', text)
text = p_small.sub('ԥ', text)
text = g_small.sub('ӷ', text)
if text != '':
with MosesSentenceSplitter('ru') as splitsents:
text = splitsents([text])
lengths.append(len(text))
temp.extend(text)
src_list = temp
for i, text in enumerate(src_list):
text = ' '.join(tokenize(text)).lower()
text = sp_src.encode(text, out_type=str)
src_list[i] = text
translator = ctranslate2.Translator(ct_path)
tgt_list = translator.translate_batch(src_list)
for i, text in enumerate(tgt_list):
detokenize = MosesDetokenizer('ru')
sp_tgt = spm.SentencePieceProcessor()
sp_tgt.load(sp_path_tgt)
text = sp_tgt.decode(text[0]['tokens'])
text = detokenize(text.split(' '))
tgt_list[i] = text
temp = []
i = 0
for length in lengths:
text = ''
for jw in range(length):
text = text + tgt_list[i + jw] + ' '
temp.append(text.strip())
i = i + length
tgt_list = temp
return tgt_list
def create_app():
import ctranslate2
import pathlib
path = pathlib.Path(__file__).absolute().parents[2] / 'model_store'
logger.info("model directory path = '{}'".format(path))
# start packing
net_models = {}
# pack arabic model
model_ara = ctranslate2.Translator(
str(path / 'arabic' / 'ctranslate2_released')
)
net_models["ara"] = model_ara
logger.info("packed arabic model")
# pack chiense model
model_chi = ctranslate2.Translator(
str(path / 'chinese' / 'ctranslate2_released')
)
net_models["chi"] = model_chi
logger.info("packed chinese model")
# pack hebrew model
model_heb = ctranslate2.Translator(
str(path / 'hebrew' / 'ctranslate2_released')
)
net_models["heb"] = model_heb
logger.info("packed hebrew model")
# pack japanese model
model_jpn = ctranslate2.Translator(
str(path / 'katakana' / 'ctranslate2_released')
)
net_models["jpn"] = model_jpn
logger.info("packed japanese model")
# pack korean model
model_kor = ctranslate2.Translator(
str(path / 'korean' / 'ctranslate2_released')
)
net_models["kor"] = model_kor
logger.info("packed korean model")
# pack russian model
model_rus = ctranslate2.Translator(
str(path / 'russian' / 'ctranslate2_released')
)
net_models["rus"] = model_rus
logger.info("packed russian model")
# init flask objects
app = Flask(__name__)
api = Api(app)
api.app.config['RESTFUL_JSON'] = {'ensure_ascii': False}
api.add_resource(TransformerNETransliterator,
'/predict',
resource_class_kwargs={"net_models": net_models})
return app
def encode_itranslate_decode(i,
sp_encoder,
sp_decoder,
num_map,
tp_tokenizer,
num_hypotheses=3):
try:
logger.info("Inside encode_itranslate_decode function")
model_path = get_model_path(i['id'])
translator = ctranslate2.Translator(model_path)
i['src'] = str(sp.encode_line(sp_encoder, i['src']))
i_final = format_converter(i['src'])
if 'target_prefix' in i and len(
i['target_prefix']) > 0 and i['target_prefix'].isspace(
) == False:
logger.info("target prefix: {}".format(i['target_prefix']))
i['target_prefix'] = i['target_prefix']
i['target_prefix'] = replace_num_target_prefix(i, num_map)
if tp_tokenizer is not None:
i['target_prefix'] = tp_tokenizer(i['target_prefix'])
i['target_prefix'] = str(
sp.encode_line(sp_decoder, i['target_prefix']))
tp_final = format_converter(i['target_prefix'])
tp_final[-1] = tp_final[-1].replace(']', ",")
m_out = translator.translate_batch([i_final],
beam_size=5,
target_prefix=[tp_final],
num_hypotheses=num_hypotheses)
else:
m_out = translator.translate_batch([i_final],
beam_size=5,
num_hypotheses=num_hypotheses)
translation = multiple_hypothesis_decoding(m_out[0], sp_decoder)
return translation
except Exception as e:
logger.error(
"Unexpexcted error in encode_itranslate_decode: {} and {}".format(
e,
sys.exc_info()[0]))
raise
