Training with spaCy

Natural Language Processing with spaCy

Azadeh Mobasher

Principal Data Scientist

Training steps

Annotate and prepare input data
Disable other pipeline components
Train a model for a few epochs
Evaluate model performance

Disabling other pipeline components

Disable all pipeline components except NER:

other_pipes = [pipe for pipe in nlp.pipe_names if pipe != 'ner']

nlp.disable_pipes(*other_pipes)

Model training procedure

Go over the training set several times; one iteration is called an epoch.
In each epoch, update the weights of the model with a small number.
Optimizers update the model weights.

optimizer = nlp.create_optimizer()

losses = {}
for i in range(epochs):
  random.shuffle(training_data)

  for text, annotation in training_data:
    doc = nlp.make_doc(text)
    example = Example.from_dict(doc, annotation)

    nlp.update([example], sgd = optimizer, losses=losses)

Save and load a trained model

Save a trained NER model:

ner = nlp.get_pipe("ner")
ner.to_disk("<ner model name>")

Load the saved model:

ner = nlp.create_pipe("ner")
ner.from_disk("<ner model name>")
nlp.add_pipe(ner, "<ner model name>")

Model for inference

Use a saved model at inference.

Apply NER model and store tuples of (entity text, entity label):

doc = nlp(text)
entities = [(ent.text, ent.label_) for ent in doc.ents]

Let's practice!

Natural Language Processing with spaCy