Keras ❤️ torchtext

Keras is love
Keras is life
Keras loves torchtext

torchtext is a great library, putting a layer of abstraction over the usually very heavy data component in NLP projects, making the work with complex datasets a pace. Sadly, as torchtext is based and built on PyTorch, using it with Keras is not directly possible.

Keras ❤️ torchtext is to the rescue by providing lightweight wrappers for some Torchtext classes, making them easily work with Keras.

Installation

pip install keras-loves-torchtext

Examples

Wrap a torchtext.data.Iterator with WrapIterator and use it to train a Keras model:

from torchtext.data import Dataset, Field, Iterator
from kltt import WrapIterator

...

fields  = [('text', Field()),
           ('label', Field(sequential=False))]
dataset = Dataset(examples, fields)
iterator = Iterator(dataset, batch_size=32)

# Keras ❤️ torchtext comes to play
data_gen = WrapIterator(iterator, x_fields=['text'], y_fields=['label'])

model.fit_generator(iter(data_gen), steps_per_epoch=len(data_gen))

Easily wrap multiple iterators at once:

from torchtext.data import Dataset, Field, Iterator
from kltt import WrapIterator

...

fields  = [('text', Field()),
           ('label', Field(sequential=False))]
dataset = Dataset(examples, fields)
splits = dataset.split()

iterators = Iterator.splits(splits, batch_size=32)
train, valid, test = WrapIterator.wraps(iterators, x_fields=['text'], y_fields=['label'])
model.fit_generator(iter(train), steps_per_epoch=len(train),
                    validation_data=iter(valid), validation_steps=len(valid))
loss, acc = model.evaluate_generator(iter(test), steps=len(test))

Further and full working examples can be found in the examples folder.

Documentation

Todo

See examples and inline documentation for now.