README.md

metadata

library_name: pytorch
pipeline_tag: text2text-generation
language:
  - vi
  - lo
metrics:
  - bleu

Please use python version 3.10

Direct Use

Load a pre-trained model

Use load_config to load a .yaml config file.

Then use load_model_tokenizer to load a pretrained model and its tokenizers

from config import load_config
from load_model import load_model_tokenizer

config = load_config(file_name='config/config_final.yaml')
model, src_tokenizer, tgt_tokenizer = load_model_tokenizer(config)

Translate lo to vi

Use the translate function in translate.py.

from translate import translate
from config import load_config
from load_model import load_model_tokenizer

config = load_config(file_name='config/config_final.yaml')
model, src_tokenizer, tgt_tokenizer = load_model_tokenizer(config)

text = "   "
translation, attn = translate(
    model, src_tokenizer, tgt_tokenizer, text,
    decode_method='beam-search',
)
print(translation)

Training

Use the train_model function in train.py to train your model.

from train import train_model
from config import load_config

config = load_config(file_name='config/config_final.yaml')
train_model(config)

If you wish to continue training/ fine-tune our model, you should modify the num_epochs in your desired config file, as well as read the following notes (+ is the string concat funtion):

• The code will save and preload models in model_folder
• The code will preload the model with the name: "model_basename + preload + .pt"
• The code will NOT preload a trained model if you set preload as null
• Every epoch, the code will save the model with the name: "model_basename + _ + (current epoch) + .pt"
• train_model will automatically continue training the preloaded model.