Retrain an Address Parser for Single Country Uses

Note

• See the notebook here

• Run in Google Colab

from deepparse import download_worldwide_addresses_sample
from deepparse.dataset_container import ListDatasetContainer
from deepparse.parser import AddressParser
from poutyne import set_seeds

seed = 42
set_seeds(seed)

In this example, we will retrain a pre-trained model to maximize its performance for specific countries (e.g. the UK or Canada).

Retrain a Model

First, to retrain our supervised model, we need parsed address example, as shown in the following figure. Fortunately, we have access to a public dataset of such parsed examples, the worldwide-addresses dataset on the HuggingFace Hub (one configuration per country).

For our example, we will focus on UK addresses since we want to parse addresses only from the UK. So let’s download a sample of the UK (gb) addresses directly from the HuggingFace Hub using the Deepparse download_worldwide_addresses_sample function. It only reads the needed part of the dataset and maps its tags to the Deepparse tags (Suburb and District become Municipality, and the words tagged County or Country are removed from the addresses).

addresses = download_worldwide_addresses_sample(country_code="gb", num_samples=12000)

Now, let’s split our sample into train (10,000 addresses) and test (2,000 addresses) datasets to retrain our parser model.

training_container = ListDatasetContainer(addresses[:10000])
test_container = ListDatasetContainer(addresses[10000:])

We will use the FastText one for our base pre-trained model since it is faster to retrain.

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

But first, let’s see what the performance is before retraining.

address_parser.test(test_container, batch_size=256)

_ = address_parser.retrain(
    training_container,
    train_ratio=0.8,
    epochs=1,
    batch_size=32,
    num_workers=2,
    learning_rate=0.001,
    logging_path="./uk_retrain",
    name_of_the_retrain_parser="UKParser",
)

address_parser.test(test_container, batch_size=256)

To further improve performance, we could train for longer, increase the training dataset size (using a larger num_samples, or the complete gb configuration of the dataset), or rework the Seq2Seq hidden sizes. See the retrain interface documentation for all the training parameters.