Retrain a Pretrained Model

import poutyne

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

First, let’s download a small sample of Canadian addresses from the worldwide-addresses dataset. The dataset tags are already mapped to the deepparse tags. We use the first 1,000 addresses for training and the last 200 for testing.

addresses = download_worldwide_addresses_sample(country_code="ca", num_samples=1200)

Now let’s create a training and test container.

training_container = ListDatasetContainer(addresses[:1000])
test_container = ListDatasetContainer(addresses[1000:])

We will retrain the FastText version of our pretrained model.

address_parser = AddressParser(model_type="fasttext", 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.

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

The checkpoints (ckpt) are saved in the default "./checkpoints" directory, so if you wish to retrain another model (let’s say BPEmb), you need to change the logging_path directory; otherwise, you will get an error when retraining since Poutyne will try to use the last checkpoint.

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

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

address_parser.test(test_container, batch_size=256)

Now let’s retrain the FastText version but with an attention mechanism.

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

Since the previous checkpoints were saved in the default "./checkpoints" directory, we need to use a new one. Otherwise, poutyne will try to reload the previous checkpoints, and our model has changed.

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

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

address_parser.test(test_container, batch_size=256)