Vietnamese Diacritic Restoration

Vietnamese Diacritic Restoration using Transformer Sequence-to-Sequence Moddel

Requirement

tensorflow-gpu==1.14.0
tensor2tensor==1.14.1

Generate data

Generate data for default problem translate_vndt

./gen_data.sh

Generate data for custom problem A

./gen_data.sh A

Train model

Define problem

@registry.register_problem
class TranslateVndt(translate.TranslateProblem):
    @property
    def approx_vocab_size(self):
        return 2**15  # 32768

    def source_data_files(self, dataset_split):
        train = dataset_split == problem.DatasetSplit.TRAIN
        return _VNDT_TRAIN_DATASETS if train else _VNDT_DEV_DATASETS

Define hyperparams

@registry.register_hparams
def transformer_base_h256():
    hparams = transformer_base()
    hparams.hidden_size = 256
    return hparams

On problem translate_vndt, to train model transformer with hparams transformer_base on GPUs 0,1

./train.sh 0,1 transformer_base transformer translate_vndt

Predict

Similar to train.sh

./predict.sh 0,1 transformer_base transformer translate_vndt

The output is stored in sub-translate_vndt-transformer-transformer_base.csv