def __init__(self,
en_file,
ru_file,
en_tokenizer_file='en_tokenizer.model',
ru_tokenizer_file='ru_tokenizer.model',
en_vocab_size=5000,
ru_vocab_size=5000):
if not (os.path.exists(en_tokenizer_file)
and os.path.exists(ru_tokenizer_file)):
self.en_tokenizer, self.ru_tokenizer = init_tokenizers(
en_file, ru_file, en_vocab_size, ru_vocab_size)
else:
self.en_tokenizer = yttm.BPE(model=en_tokenizer_file)
self.ru_tokenizer = yttm.BPE(model=ru_tokenizer_file)
self.en_corpus, self.ru_corpus = read_corupuses(en_file, ru_file)
def construct(self, booster: "xgboost.core.Booster",
params: Mapping[str, int], bpe_model_path: str):
self._booster = booster
self._params = params
self._bpe_model_path = bpe_model_path
self._bpe_model = youtokentome.BPE(bpe_model_path)
return self
def init_tokenizers(
en_file,
ru_file,
en_vocab_size=5000,
ru_vocab_size=5000
) -> Tuple[yttm.youtokentome.BPE, yttm.youtokentome.BPE]:
if not (os.path.exists(en_file) and os.path.exists(ru_file)):
raise FileNotFoundError('Couldn\'t find corpus files')
yttm.BPE.train(data=en_file,
vocab_size=en_vocab_size,
model='en_tokenizer.model')
yttm.BPE.train(data=ru_file,
vocab_size=ru_vocab_size,
model='ru_tokenizer.model')
return yttm.BPE(model='en_tokenizer.model'), yttm.BPE(
model='ru_tokenizer.model')
def __init__(self, api_url):
self.path = os.path.dirname(os.path.abspath(__file__))
self.api_url = api_url
self.bpe = yttm.BPE(model=os.path.join(self.path, 'yttm.model'))
self.vocab_size = self.bpe.vocab_size()
self.sequence_length = 20
self.newline_token = 88
def __init__(self,
vocab_file,
errors='replace',
bos_token="<s>",
eos_token="</s>",
sep_token="</s>",
cls_token="<s>",
unk_token="<unk>",
pad_token='<pad>',
mask_token='<mask>',
**kwargs):
super(RubertaTokenizer, self).__init__(bos_token=bos_token,
eos_token=eos_token,
unk_token=unk_token,
**kwargs)
self.max_len_single_sentence = self.max_len - 2 # take into account special tokens
self.max_len_sentences_pair = self.max_len - 3 # take into account special tokens
self.vocab_file = vocab_file
self.bpe = yttm.BPE(model=vocab_file)
self.pat = re.compile(
r""" ?\p{L}+| ?\p{N}+| ?[^\s\p{L}\p{N}]+|\s+(?!\S)|\s+""")
self.encoder = {
self.bpe.id_to_subword(i): i
for i in range(self.bpe.vocab_size())
}
self.encoder['<|endoftext|>'] = self.bpe.vocab_size()
self.decoder = {v: k for k, v in self.encoder.items()}
def __init__(self, bpe_path = None):
bpe_path = Path(bpe_path)
assert bpe_path.exists(), f'BPE json path {str(bpe_path)} does not exist'
tokenizer = yttm.BPE(model = str(bpe_path))
self.tokenizer = tokenizer
self.vocab_size = tokenizer.vocab_size()
def test_encode_decode():
generate_artifacts()
os.remove(BASE_MODEL_FILE)
yttm.BPE.train(
data=TRAIN_FILE,
vocab_size=16000,
model=BASE_MODEL_FILE,
bos_id=BOS_ID,
eos_id=EOS_ID,
)
bpe = yttm.BPE(BASE_MODEL_FILE)
text_in = [
" ".join("".join([random.choice("abcd ") for _ in range(50)]).split())
]
ids = bpe.encode(text_in, yttm.OutputType.ID)
# It is necessary to add first empty line, since everything in BPE starts from a new line
text_in[0] = "\n" + text_in[0]
assert text_in == bpe.decode(ids)
ids_bos_eos = bpe.encode(text_in, yttm.OutputType.ID, bos=True, eos=True)
assert text_in == bpe.decode(ids_bos_eos, ignore_ids=[BOS_ID, EOS_ID])
assert bpe.decode(ids,
ignore_ids=[]) == bpe.decode(ids_bos_eos,
ignore_ids=[BOS_ID, EOS_ID])
def _load_model(self):
path_saved_model = os.path.join(self.path_cache, self.config_hash)
if not os.path.isdir(path_saved_model):
if not globals_vars.TRAINING:
raise Exception(
f"Embedding: While running in TEST mode: Model is not trained with this config yet \n({path_saved_model})"
)
else:
print(
f"Embedding: While running in TRAINING mode: Model is not trained with this config yet -> Now training the model with this config \n({path_saved_model})"
)
train_path = os.path.join(
os.path.dirname(os.path.abspath(__file__)), "train.py",
)
training_return = os.system(f"python {train_path}")
if training_return != 0:
raise Exception(
f"EMBEDDING TRAINING HAS FAILED OUCH\npython {train_path} failed"
)
else:
print(
f"Embedding: loading model already trained with config: \n({path_saved_model})\n"
)
self.model = yttm.BPE(
model=os.path.join(self.path_cache, self.config_hash, "model.bin")
)
# already contain '<PAD>', '<UNK>', '<EOS>', '<BOS>' (no +1 necessary)
self.vocab_size = self.model.vocab_size()
print(
"\nvocab_size from lang {}: \n{}".format(self.lang, self.model.vocab_size())
)
print("\nvocab from lang {}: \n{}".format(self.lang, self.model.vocab()))
def save_vocab(self):
bpe = yttm.BPE(model=self.bpe_model_path)
vocab = bpe.vocab()
with open(os.path.join(self.config.data_dir, 'vocab.txt'),
mode='w') as file_object:
file_object.write('\n'.join(vocab))
def train_tokenizer(
self,
corpus_file=CORPUS_FILE,
model_file=TOKENIZER_MODEL_FILE,
vocab_sz=50000,
dump_labels=True,
):
assert self.built
import youtokentome as yttm
# first we need to dump labels
if dump_labels:
self.dump_labels(corpus_file)
# train model
print("Training yttm model...")
yttm.BPE.train(data=corpus_file, vocab_size=vocab_sz, model=model_file)
print("Done.")
# load model (for testing)
print("Testing yttm model...")
bpe = yttm.BPE(model=model_file)
# Two types of tokenization
test_text = "Are you freakin' crazy?"
encoded1 = bpe.encode([test_text], output_type=yttm.OutputType.ID)
encoded2 = bpe.encode([test_text], output_type=yttm.OutputType.SUBWORD)
decoded = bpe.decode(encoded1)
print(encoded1)
print(encoded2)
print(decoded)
def create_tokenizer(tokenizer_path,
datasets,
vocab_size,
tokens,
temp_file_path='tokenizer_text.temp'):
# Load tokenizer
if os.path.exists(tokenizer_path):
print('Loading pretrained tokenizer...')
tokenizer = yttm.BPE(model=tokenizer_path)
else:
print('Creating new tokenizer...')
# Create the corresponding folder (if needed)
os.makedirs(os.path.dirname(tokenizer_path), exist_ok=True)
# Create temp file with data to train tokenizer.
with open(temp_file_path, 'w', encoding='utf8') as out_file:
for data in datasets:
out_file.write('\n'.join(map(str.lower, data)))
# Train tokenizer.
tokenizer = yttm.BPE.train(data=temp_file_path,
vocab_size=vocab_size,
model=tokenizer_path,
n_threads=-1,
**tokens)
# Delete temp file.
os.remove(temp_file_path)
return tokenizer
def __init__(self, model_path, bpe_dropout=0.0):
model_path = Path(model_path).expanduser()
self.tokenizer = yttm.BPE(model=str(model_path))
self.vocab_size = len(self.tokenizer.vocab())
self.special_tokens = self.tokens_to_ids(
["<PAD>", "<UNK>", "<BOS>", "<EOS>"])
self.bpe_dropout = bpe_dropout
def test(self, mode, bpe_model_path=None):
while True:
file_path = input("File path: ").strip()
file_path = r"C:\Users\lezgy\OneDrive\Рабочий стол\Data_summ\data.txt"
if file_path == "q":
break
try:
with open(file_path, "r", encoding="utf-8") as r:
article = r.read().strip().split("\n")
article = " ".join(article)
if mode in ["lemm", "stem", "gram", "base"]:
article = article.lower()
article = word_tokenize(article)
article = " ".join(article)
print(f"real_text : {article}")
if mode == "lemm":
lemmatizer = mystem.Mystem()
article = preprocess_lemm(article, lemmatizer)
elif mode == "stem":
stemmer = RussianStemmer(False)
article = preprocess_stemm(article, stemmer)
elif mode == "gram":
token_model = youtokentome.BPE(model=bpe_model_path)
article = preprocess_gramm(article, token_model)
self.test_calc(article)
except Exception as e:
print(e)
print("File not found")
def __init__(self, path_to_bpe: str, path_to_model: str,
model_params: Dict[str, int]):
self.bpe_model = yttm.BPE(path_to_bpe)
self.categories: List[str] = [
"Алкоголь",
"Бытовая техника",
"Воды, соки, напитки",
"Дача и гриль",
"Другое",
"Замороженные продукты",
"Зоотовары",
"Красота, гигиена, бытовая химия",
"Макароны, крупы, специи",
"Молоко, сыр, яйца",
"Овощи, фрукты, ягоды",
"Подборки и готовые блюда",
"Постные продукты",
"Посуда",
"Птица, мясо, деликатесы",
"Рыба, икра",
"Соусы, орехи, консервы",
"Товары для дома и дачи",
"Товары для мам и детей",
"Хлеб, сладости, снеки",
"Чай, кофе, сахар",
]
self.device = torch.device("cpu")
self.model = CategoryClassifier(**model_params)
self.model.load_state_dict(
torch.load(path_to_model, map_location=self.device))
self.model.eval()
def __init__(self, vocab_size: int=10000,
train_fname: str='train_texts.txt',
bpe_path: str=''):
self.bpe_path = bpe_path if len(bpe_path) else 'yttm_bpe.bin'
self.bpe_model = yttm.BPE(bpe_path) if len(bpe_path) else None
self.train_fname = train_fname
self.vocab_size = vocab_size
def compute_features(data, model_path="vocab.model", max_len=20):
bpe = yttm.BPE(model=model_path)
features_ids = bpe.encode(data.feature_string.values.tolist(),
output_type=yttm.OutputType.ID)
features_ids = [
f[:max_len] + [0] * (max_len - len(f)) for f in features_ids
]
return np.array(features_ids)
def main(args):
training_args = TrainingArguments(
output_dir=args.output_dir,
overwrite_output_dir=True,
num_train_epochs=args.epoch,
per_device_train_batch_size=args.batch_size,
save_steps=args.save_steps,
save_total_limit=10,
gradient_accumulation_steps=args.gradient_accumulation_steps,
dataloader_num_workers=args.n_workers,
)
print("Loading tokenizer...")
tokenizer = yttm.BPE(args.tokenizer)
print("Loading model...")
model = get_model(vocab_size=tokenizer.vocab_size())
print("List training files...")
train_paths = get_training_files(args.train_dir)
if args.check:
train_paths = train_paths[:10000]
print("Loading train texts...")
train_data = []
for p in tqdm(train_paths):
(keys, notes) = read_abc(p)
if keys is None:
continue
keys_tokens = tokenizer.encode(keys)
bars = notes.split(" | ")
notes_tokens = [tokenizer.encode(i + " | ") for i in bars]
## To avoid OOM
sequence_len = sum(len(i) for i in notes_tokens)
if not (args.min_sequence_lenght < sequence_len <
args.max_sequence_lenght):
print("Skip", p)
continue
train_data.append((keys_tokens, notes_tokens))
print("Making dataset...")
train_dataset = ABCDataset(train_data)
if args.checkpoint:
state_dict = torch.load(args.checkpoint)
model.load_state_dict(state_dict)
trainer = Trainer(model=model,
args=training_args,
data_collator=collate_function,
train_dataset=train_dataset)
print("Start training...")
trainer.train()
def test_encode_decode():
generate_artifacts()
os.remove(BASE_MODEL_FILE)
yttm.BPE.train(data=TRAIN_FILE, vocab_size=16000, model=BASE_MODEL_FILE)
bpe = yttm.BPE(BASE_MODEL_FILE)
text_in = [" ".join("".join([random.choice("abcd ") for _ in range(50)]).split())]
ids = bpe.encode(text_in, yttm.OutputType.ID)
assert text_in == bpe.decode(ids)
def get_bpe_tokenizer(train_texts, train_txt_path, bpe_model_name, vocab_size):
_save_text(train_texts, train_txt_path)
yttm.BPE.train(data=train_txt_path,
vocab_size=vocab_size,
model=bpe_model_name)
tokenizer = yttm.BPE(bpe_model_name)
return tokenizer
def main():
parser = argparse.ArgumentParser(description="Compute BLEU.")
parser.add_argument('ckpt', type=str, help="Checkpoint to restore.")
parser.add_argument('--dir',
type=str,
default="./wmt14",
help="Directory of dataset.")
parser.add_argument('--split',
default='test',
type=str,
help="Specify which split of data to evaluate.")
parser.add_argument(
'--gpu_id',
default=0,
type=int,
help="CUDA visible GPU ID. Currently only support single GPU.")
parser.add_argument('--beams',
default=1,
type=int,
help="Beam Search width.")
args = parser.parse_args()
os.environ["CUDA_VISIBLE_DEVICES"] = str(args.gpu_id)
assert torch.cuda.is_available()
import data
import build_model
# Restore checkpoint
info = torch.load(args.ckpt)
cfg = info['cfg']
# Build model
bpe_model = yttm.BPE(model=cfg['bpe'])
model = build_model.Seq2Seq(bpe_model.vocab_size(),
bpe_model.vocab_size(),
hidden_size=cfg['model']['hidden_size'],
encoder_layers=cfg['model']['encoder_layers'],
decoder_layers=cfg['model']['decoder_layers'],
use_bn=cfg['model']['use_bn'])
model.load_state_dict(info['weights'])
model.eval()
model = model.cuda()
# Create dataset
if args.beams == 1:
batch_size = cfg['train']['batch_size']
else:
batch_size = 1
loader = data.load(args.dir,
split=args.split,
batch_size=batch_size,
bpe_model=bpe_model)
# Evaluate
_, bleu = utils.eval_dataset(loader, model, bpe_model, args.beams)
print("BLEU on %s set = %.4f" % (args.split, error))
def youtoken():
bpe = yttm.BPE(model="statementPrediction.model")
with open("../results/all_projects.json") as jsonf:
logs = json.load(jsonf)
with open("tokenizedExample.txt", "w") as output:
for log in logs:
msg = log["msg"]
tokenizedMsg = bpe.encode(
[msg], output_type=yttm.OutputType.SUBWORD)[0]
output.write(f"{msg} ----> {str(tokenizedMsg)}\n")
def __init__(self, filename, *inputs, **kwargs):
super().__init__(*inputs, **kwargs)
if os.path.isdir(filename):
filename = os.path.join(filename, self.def_name)
self.bpe = yttm.BPE(filename)
self.hash = hashlib.sha512(open(filename,
'rb').read()).hexdigest()[:10]
self.filename = filename
def train(self, train_fname: str='', vocab_size: int=10000):
if len(train_fname):
self.train_fname = train_fname
if vocab_size > 0:
self.vocab_size = vocab_size
yttm.BPE.train(data=self.train_fname, vocab_size=self.vocab_size, model=self.bpe_path)
self.bpe_model = yttm.BPE(self.bpe_path)
def __init__(self, data_folder, source_suffix, target_suffix, split,
tokens_in_batch):
"""
:param data_folder: folder containing the source and target language data files
:param source_suffix: the filename suffix for the source language files
:param target_suffix: the filename suffix for the target language files
:param split: train, or val, or test?
:param tokens_in_batch: the number of target language tokens in each batch
"""
self.tokens_in_batch = tokens_in_batch
self.source_suffix = source_suffix
self.target_suffix = target_suffix
assert split.lower() in {
"train", "val", "test"
}, "'split' must be one of 'train', 'val', 'test'! (case-insensitive)"
self.split = split.lower()
# Is this for training?
self.for_training = self.split == "train"
# Load BPE model
self.bpe_model = youtokentome.BPE(
model=os.path.join(data_folder, "bpe.model"))
# Load data
with codecs.open(os.path.join(data_folder,
".".join([split, source_suffix])),
"r",
encoding="utf-8") as f:
source_data = f.read().split("\n")[:-1]
with codecs.open(os.path.join(data_folder,
".".join([split, target_suffix])),
"r",
encoding="utf-8") as f:
target_data = f.read().split("\n")[:-1]
assert len(source_data) == len(
target_data
), "There are a different number of source or target sequences!"
source_lengths = [
len(s)
for s in self.bpe_model.encode(source_data, bos=False, eos=False)
]
target_lengths = [
len(t)
for t in self.bpe_model.encode(target_data, bos=True, eos=True)
] # target language sequences have <BOS> and <EOS> tokens
self.data = list(
zip(source_data, target_data, source_lengths, target_lengths))
# If for training, pre-sort by target lengths - required for itertools.groupby() later
if self.for_training:
self.data.sort(key=lambda x: x[3])
# Create batches
self.create_batches()
def __init__(self, path):
super().__init__()
data = pickle.loads(Path(path).read_bytes())
self.samples = data['samples']
self.seq_length = data['seq_length']
self.bpe = yttm.BPE(data['bpe_path'])
self.vocab = self.bpe.vocab()
self.vocab_size = len(self.vocab)
self.pad_idx, self.unk_idx, self.bos_idx, self.eos_idx = list(range(4))
def bpe_tokenize(self):
self.split_files(self.file_path)
if self.force or not os.path.isfile(
self.src_bpe_file) or not os.path.isfile(self.src_bpe_file):
self._remove_file(self.src_bpe_file)
self._remove_file(self.trg_bpe_file)
yttm.BPE.train(data=self.src_file,
vocab_size=self.src_vocab_size,
model=self.src_bpe_file)
yttm.BPE.train(data=self.trg_file,
vocab_size=self.trg_vocab_size,
model=self.trg_bpe_file)
# Loading model
self.src_bpe: yttm.BPE = yttm.BPE(model=self.src_bpe_file)
self.trg_bpe: yttm.BPE = yttm.BPE(model=self.trg_bpe_file)
return
def make_yttm_tokenizer(train_conll: List[Instance], vocab_size=400):
tokens = []
for instance in train_conll:
tokens += [token.text for token in instance['tokens']]
text = ' '.join(tokens)
with open('train_chunks.txt', 'w') as fobj:
fobj.write(text)
yttm.BPE.train(data='train_chunks.txt',
vocab_size=vocab_size,
model='conll_model.yttm')
return yttm.BPE('conll_model.yttm')
def test_vocabulary_consistency():
generate_artifacts()
os.remove(BASE_MODEL_FILE)
yttm.BPE.train(data=TRAIN_FILE, vocab_size=16000, model=BASE_MODEL_FILE)
bpe = yttm.BPE(BASE_MODEL_FILE)
assert bpe.vocab_size() == len(bpe.vocab())
assert bpe.vocab_size() == len(set(bpe.vocab()))
vc = bpe.vocab()
for i, subword in enumerate(vc):
assert i == bpe.subword_to_id(subword)
assert subword == bpe.id_to_subword(i)
def __init__(self, filename, *inputs, **kwargs):
super().__init__(*inputs, **kwargs)
#self.max_len_single_sentence = 1024 # no default special tokens - you can update this value if you add special tokens
#self.max_len_sentences_pair = 1024 # no default special tokens - you can update this value if you add special tokens
if os.path.isdir(filename):
filename = os.path.join(filename, self.def_name)
self.bpe = yttm.BPE(filename)
self.hash = hashlib.sha512(open(filename,
'rb').read()).hexdigest()[:10]
self.filename = filename
def __init__(self, bpe_path=None):
bpe_path = Path(bpe_path)
print(f"************ bpe_path : {bpe_path}")
assert bpe_path.exists(
), f'BPE json path {str(bpe_path)} does not exist'
tokenizer = yttm.BPE(model=str(bpe_path))
self.tokenizer = tokenizer
self.vocab_size = tokenizer.vocab_size()
print(f"************ self.tokenizer : {self.tokenizer}")
print(f"************ self.vocab_size : {self.vocab_size}")
