retrain_with_new_seq2seq_params.rst

.. role:: hidden
    :class: hidden-section

Retrain With New Seq2Seq Parameters
***********************************

.. code-block:: python

    import os

    import poutyne

    from deepparse import download_from_public_repository
    from deepparse.dataset_container import PickleDatasetContainer
    from deepparse.parser import AddressParser

First, let's download the train and test data with the new tags, ``"new tags"``, from the public repository.

.. code-block:: python

    saving_dir = "./data"
    file_extension = "p"
    training_dataset_name = "sample_incomplete_data_new_prediction_tags"
    test_dataset_name = "test_sample_data_new_prediction_tags"
    download_from_public_repository(training_dataset_name, saving_dir, file_extension=file_extension)
    download_from_public_repository(test_dataset_name, saving_dir, file_extension=file_extension)

Now, let's create a training and test container.

.. code-block:: python

    training_container = PickleDatasetContainer(os.path.join(saving_dir,
                                                             training_dataset_name + "." + file_extension))
    test_container = PickleDatasetContainer(os.path.join(saving_dir,
                                                         test_dataset_name + "." + file_extension))

We will retrain the ``FastText`` version of our pretrained model.

.. code-block:: python

    model = "fasttext"
    address_parser = AddressParser(model_type=model, device=0)

Now, let's retrain for ``5`` epochs using a batch size of ``8`` since the data is really small for the example.
Let's start with the default learning rate of ``0.01`` and use a learning rate scheduler to lower the learning rate as we progress.

.. code-block:: python

    # Reduce LR by a factor of 10 each epoch
    lr_scheduler = poutyne.StepLR(step_size=1, gamma=0.1)

    # We need a EOS tag in the dictionary. EOS -> End Of Sequence
    tag_dictionary = {"ATag": 0, "AnotherTag": 1, "EOS": 2}

    # The path to save our checkpoints
    logging_path = "./checkpoints"

    # The new seq2seq params settings using smaller hidden size
    # See the documentation for the list of tunable seq2seq parameters
    seq2seq_params = {
        "encoder_hidden_size": 512,
        "decoder_hidden_size": 512
    }

    address_parser.retrain(training_container,
                           train_ratio=0.8,
                           epochs=5,
                           batch_size=8,
                           num_workers=2,
                           callbacks=[lr_scheduler],
                           prediction_tags=tag_dictionary,
                           logging_path=logging_path,
                           seq2seq_params=seq2seq_params)

Now let's test our fine-tuned model using the best checkpoint (default parameter).

.. code-block:: python

    address_parser.test(test_container, batch_size=256)