/LSTM-Additive_attention-TF2.x

Experiments to Build language translators and chat bots using seq-to-seq (LSTM+additive_attention) model.

Primary LanguagePython

seq-to-seq-tensorflow-2.x

There are certainly simplified two steps to start training:

  • Build config.
  • Start trainer.
For demo purpose, we will create english-tamil language translator.

  • Build the tokenizer for the dataset, the structure of the dataset is previewed in the above image.
python build_config.py --csv_path datasets/english_tamil.csv --input_id question --output_id answer 
--data_count 300 --ckpt_dir eng_tam --save_frequency 25 --epochs 400

Arguments short explanation:
-input id   : column name of source values
-output_id  : column name of target values
-data_count : Number of rows to be considered for training
  • A config json file will be auto-generated with training config. we can edit the fields if needed.
{
    "csv_path": "datasets/english_tamil.csv",
    "select": 300,
    "input_node": "question",
    "output_node": "answer",
    "input_vectorizer": "question.json",
    "output_vectorizer": "answer.json",
    "batch_size": 128,
    "epochs": 400,
    "units": 64,
    "embed_size": 32,
    "ckpt_dir": "eng_tam",
    "save_frequency": 25
}
  • Onece the config is generated, execute trainer.py
python trainer.py
  • Testing
inference.ipynb notebook can be used for testing the model