Model Description

The Transformer, introduced in the paper Attention Is All You Need, is a powerful sequence-to-sequence modeling architecture capable of producing state-of-the-art neural machine translation (NMT) systems.

Recently, the fairseq team has explored large-scale semi-supervised training of Transformers using back-translated data, further improving translation quality over the original model. More details can be found in this blog post.

Requirements

We require a few additional Python dependencies for preprocessing:

pip install bitarray fastBPE hydra-core omegaconf regex requests sacremoses subword_nmt

English-to-French Translation

To translate from English to French using the model from the paper Scaling Neural Machine Translation:

import torch

# Load an En-Fr Transformer model trained on WMT'14 data :
en2fr = torch.hub.load('pytorch/fairseq', 'transformer.wmt14.en-fr', tokenizer='moses', bpe='subword_nmt')

# Use the GPU (optional):
en2fr.cuda()

# Translate with beam search:
fr = en2fr.translate('Hello world!', beam=5)
assert fr == 'Bonjour à tous !'

# Manually tokenize:
en_toks = en2fr.tokenize('Hello world!')
assert en_toks == 'Hello world !'

# Manually apply BPE:
en_bpe = en2fr.apply_bpe(en_toks)
assert en_bpe == 'H@@ ello world !'

# Manually binarize:
en_bin = en2fr.binarize(en_bpe)
assert en_bin.tolist() == [329, 14044, 682, 812, 2]

# Generate five translations with top-k sampling:
fr_bin = en2fr.generate(en_bin, beam=5, sampling=True, sampling_topk=20)
assert len(fr_bin) == 5

# Convert one of the samples to a string and detokenize
fr_sample = fr_bin[0]['tokens']
fr_bpe = en2fr.string(fr_sample)
fr_toks = en2fr.remove_bpe(fr_bpe)
fr = en2fr.detokenize(fr_toks)
assert fr == en2fr.decode(fr_sample)

English-to-German Translation

Semi-supervised training with back-translation is an effective way of improving translation systems. In the paper Understanding Back-Translation at Scale, we back-translate over 200 million German sentences to use as additional training data. An ensemble of five of these models was the winning submission to the WMT’18 English-German news translation competition.

We can further improved this approach through noisy-channel reranking. More details can be found in this blog post. An ensemble of models trained with this technique was the winning submission to the WMT’19 English-German news translation competition.

To translate from English to German using one of the models from the winning submission:

import torch

# Load an En-De Transformer model trained on WMT'19 data:
en2de = torch.hub.load('pytorch/fairseq', 'transformer.wmt19.en-de.single_model', tokenizer='moses', bpe='fastbpe')

# Access the underlying TransformerModel
assert isinstance(en2de.models[0], torch.nn.Module)

# Translate from En-De
de = en2de.translate('PyTorch Hub is a pre-trained model repository designed to facilitate research reproducibility.')
assert de == 'PyTorch Hub ist ein vorgefertigtes Modell-Repository, das die Reproduzierbarkeit der Forschung erleichtern soll.'

We can also do a round-trip translation to create a paraphrase:

# Round-trip translations between English and German:
en2de = torch.hub.load('pytorch/fairseq', 'transformer.wmt19.en-de.single_model', tokenizer='moses', bpe='fastbpe')
de2en = torch.hub.load('pytorch/fairseq', 'transformer.wmt19.de-en.single_model', tokenizer='moses', bpe='fastbpe')

paraphrase = de2en.translate(en2de.translate('PyTorch Hub is an awesome interface!'))
assert paraphrase == 'PyTorch Hub is a fantastic interface!'

# Compare the results with English-Russian round-trip translation:
en2ru = torch.hub.load('pytorch/fairseq', 'transformer.wmt19.en-ru.single_model', tokenizer='moses', bpe='fastbpe')
ru2en = torch.hub.load('pytorch/fairseq', 'transformer.wmt19.ru-en.single_model', tokenizer='moses', bpe='fastbpe')

paraphrase = ru2en.translate(en2ru.translate('PyTorch Hub is an awesome interface!'))
assert paraphrase == 'PyTorch is a great interface!'

References