genz-tokenize 1.2.7

Vietnamese tokenization, preprocess and models NLP

Genz Tokenize

Using for tokenize

    from genz_tokenize import Tokenize
    # using vocab from lib
    tokenize = Tokenize()
    print(tokenize('sinh_viên công_nghệ', 'hello', max_len = 10, padding = True, truncation = True))
    # {'input_ids': [1, 770, 1444, 2, 2, 30469, 2, 0, 0, 0], 'attention_mask': [1, 1, 1, 1, 1, 1, 1, 0, 0, 0], 'sequence_id': [None, 0, 0, None, None, 1, None]}

    print(tokenize.decode([1, 770, 2]))
    # <s> sinh_viên </s>

    # from your vocab
    tokenize = Tokenize.fromFile('vocab.txt','bpe.codes')

Preprocessing data

    from genz_tokenize.preprocess import remove_punctuations,  convert_unicode, remove_emoji, vncore_tokenize

Model

1. Seq2Seq with Bahdanau Attention
2. Transformer classification
3. Transformer
4. BERT

Trainer

    from genz_tokenize.base_model.utils import Config
    from genz_tokenize.base_model.models import Seq2Seq, Transformer, TransformerClassification
    from genz_tokenize.base_model.training import TrainArgument, Trainer
    # create config hyper parameter
    config = Config()
    config.vocab_size = 100
    config.target_vocab_size = 120
    config.units = 16
    config.maxlen = 20
    # initial model
    model = Seq2Seq(config)
    x = tf.zeros(shape=(10, config.maxlen))
    y = tf.zeros(shape=(10, config.maxlen))
    # create dataset
    BUFFER_SIZE = len(x)
    dataset_train = tf.data.Dataset.from_tensor_slices((x, y))
    dataset_train = dataset_train.shuffle(BUFFER_SIZE)
    dataset_train = dataset_train.batch(2)
    dataset_train = dataset_train.prefetch(tf.data.experimental.AUTOTUNE)

    args = TrainArgument(batch_size=2, epochs=2)
    trainer = Trainer(model=model, args=args, data_train=dataset_train)
    trainer.train()

    from genz_tokenize.models.bert import DataCollection
    from genz_tokenize.models.bert.training import TrainArg, Trainner
    from genz_tokenize.models.bert.roberta import RoBertaClassification, RobertaConfig
    import tensorflow as tf

    x = tf.zeros(shape=(10, 10), dtype=tf.int32)
    mask = tf.zeros(shape=(10, 10), dtype=tf.int32)
    y = tf.zeros(shape=(10, 2), dtype=tf.int32)

    dataset = DataCollection(
                    input_ids=x,
                    attention_mask=mask,
                    token_type_ids=None,
                    dec_input_ids=None,
                    dec_attention_mask=None,
                    dec_token_type_ids=None,
                    y=y
                )
    tf_dataset = dataset.to_tf_dataset(batch_size=2)

    config = RobertaConfig()
    config.num_class = 2
    model = RoBertaQAEncoderDecoder(config)
    arg = TrainArg(epochs=2, batch_size=2, learning_rate=1e-2)
    trainer = Trainner(model, arg, tf_dataset)
    trainer.train()